MODELLING AND CONTROL OF A GAS

PRESSURE REDUCTION SYSTEM

A Thesis Submitted to

THE COUNCIL FOR NATIONAL ACADEMIC AWARDS

in Candidature for the Degree of

DOCTOR OF PHILOSOPHY

by

Ian Fletcher B. Sc.

School of Electrical Engineering

and Applied Physics,

Sunderland Polytechnic,

in collaboration with

The Engineering Research Station,

British Gas plc, Killingworth.

February 1989
BEST COPY

AVAILABLE
TEXT IN ORIGINAL IS
CLOSETO THE EDGEOF
THE PAGE
 11
- -

SUMMARY

In the gas industry there are many cases where substantial

savings could be made if the operational efficiency of the transmission

and distribution network could be improved. The performance of the

current system is influenced by restricted pressure regimes caused by

mechanical stress ratings. This problem is further exaggerated by

redundancies which have to be built into the system because of the poor

accuracy achievable using existing pneumatic hardware. In addition,

recent increases in demand and seasonal variations have compounded

stability and interaction problems.

Modernisation of the transmission system was undertaken to

alleviate these problems using modern microprocessor technology. This

thesis presents an analysis of its application to one particular section

of the system whose role is to supply gas to the regional authorities at

required pressures or flowrates, the Above Ground Installation's (AGI).

The analysis begins by developing equations which characterise the

behaviour of a 'typical' AGI station. The equations are programmed for

digital computer simulation using VAX FORTRAN

77. Results from these

computer simulations are presented and comparisons of experimental and

simulated results show the accuracy of the model to be good.

A linear approximate model is derived using perturbation theory

and used to assess the current system's design and performance using

well documented stability/sensitivity techniques. The model also forms

a basis for the application of linear control design procedures. Two

contrasting design methods are employed. The first using the standard

single-input single-output controller configuration and an automated

version of Ziegler and Nichols' ultimate method based upon phase margin

compensation, whilst the second involves multivariable design using

McFarlane's frequency response based characteristic locus technique and

the step response methods independently proposed by Koivo and Porter.

The resulting closed loop designs are found to be robust and to achieve

their required performance specifications.

 111
- -

ACKNOWLEDGEMENTS

The candidate wishes to thank the Governors and the Rector

of Sunderland Polytechnic for the facilities placed at his disposal

during the course of this project. A similar debt is also recognised

to the Systems Division of the Engineering Research Station, British

Gas plc, Killingworth.

The author is very pleased to acknowledge his obligation to

Mr C. S. Cox and Mr W. J. B. Arden, School of Electrical Engineering

and Applied Physics, Sunderland Polytechnic, and to Mr A. Clarke,

Systems Division of the Engineering Research Station, British Gas plc,

Killingworth, who as supervisors of this project, were a fountain-head

of helpful suggestions, valuable advice and above all stimulus.

Thanks are also extended to Dr I. G. French, Mr A. R. Boucher

and my colleagues of the School of Electrical Engineering and Applied

Physics, Sunderland Polytechnic, as well as Mr E. N. Martin, formerly

I. C. I. Research Associate, Agricultural Division, Billingham, for many

worthwhile discussions.

Finally, but not least, to Miss E. Moss for her patience and

considerable effort in the typing of this thesis.

 iv -
-

DECLARATION

It is hereby declared that while registered as a candidate for

the degree for which submission is made I have not been a

registered candidate for another award of the CNAAor of any

other University.

*LkLQ-r-
,

I. Fletcher

(Candidate)
 -V-

To

Nan, Derek, Christine and Christopher

 - vi -

CONTENTS

Page No.

SUMMARY ii

ACKNOWLEDGEMENTS iii

DECLARATION iv

DEDICATION v
1. INTRODUCTION 1

References 12

2. AND VALIDATION OF A MATHEMATICALMODEL

THE DEVELOPMENT
FOR AN ABOVEGROUNDINSTALLATION 13

2.1 Introduction 14
2.2 AGI Str uctures 16
2.2.1 Regulators 19
2.3 Mathematical Modelling of the AGI System Elements 21
2.3.1 Valve Characterisation 21
2.3.1.1 Steady State Characterisation of
Regulator Flow Profiles 23
2.3.1.2 Evaluation of the Regulator
Coefficients Cv, Cg 24
2.3.1.3 Application of the Universal Gas
Sizing Equation 26
2.3.2 Actuator Dynamics -1 32
2.3.3 Pipeline Modelling 37
2.3.3.1 Pipeline Restrictions 39
2.3.3.2 Pipeline Capacities 43
2.3.3.3 Complete Discretised Pipeline Model 45
2.3.3.4 A Comparison of the Effects of the
Restriction Modelling Techniques
upon the Pipeline Behaviour 45
2.3.3.5 Evaluation of the Minimum Number of
Stages Necessary to Adequately
Reflect the Pipeline Dynamics 50
2.3.3.6 Construction of the Downstream
Dynamics for the Proposed AGI
Simulation 50
2.3.4 Interstage Pipework Modelling 54
2.3.5 System Signal Conditioning 55
2.4 Discuss ion and Validation of the AGI Simulation 56
References 65
Nomenclature 67

3. AN INVESTIGATIONINTO THE EFFECTIVENESS

OF THE PRESENT
AGI CONTROLPHILOSOPHY 70

3.1 Introduction 71
3.2 Ziegler-Nichols Ultimate Method 72
3.3 The Modified Ultimate Method 75
3.4 Phase Margin Compensation 82
 - vii -

Page No.

3.4.1 PI Control with Phase Margin Compensation 82

3.4.2 PID Control with Phase Margin Design 85
3.5 Industrial Applications of the Modified Ultimate
Method 86
3.5.1 Remote Boundary Pressure Control 88
3.5.2 Flow Control Station 93
3.6 A Relay Modification to Provide Noise Rejection 98
3.7 Tuning of the Twin Valve AGI Model 101
3.8 PID Control of the First Stage Cut 102
3.9 Discussion 109
References 112
Nomenclature 114

4. SMALLSIGNAL ANALYSIS- AGI SENSITIVITY AND STABILITY

IMPLICATIONS 115

4.1 Introduction 116

4.2 Construction of a Linear Model 120
4.3 Development of a Linear Model 123
4.3.1 Linearisation of the Pipeline Model 123
4.3.2 Linearisation of the Valve Flow Equation 126
4.3.3 Linear AGI Model 126
4.4 Small Signal Model Validation 128
4.5 Implications on Sensitivity and Stability of Certain
Identified Parameters 133
4.5.1 Sensitivity and Stability of the First Stage
Pressure Cut 134
4.5.2 Sensitivity and Stability of the Second Stage
Pressure Cut 141
4.5.3 AGI Characteristic Equation Development 145
4.5.4 Influence of Controller Design, Actuator
Selection and Interstage Volume on AGI
Stability 148
4.6 Discussion 149
References 158
Nomenclature 159

5. TO THE DESIGNOF AGI CONTROL

A MULTIVARIABLEAPPROACH SYSTEMS 161

5.1 Introduction 162

5.2 Comparison of Frequency Domain Properties of Scalar
and Multivariable Systems 166
5.2.1 Poles and Zeros of Piultivariable Systems 170
5.2.2 Generalised Frequency Response and Spectral
Decomposition 172
5.2.3 Stability Analysis 175
5.2.4 Interaction Analysis 175
5.2.5 System Integrity 177
5.2.6 Robustness and Sensitivity 178
5.3 Frame Alignment Techniques 179
5.3.1 Commutative Control 179
5.3.2 Approximate Commutative Control 181
5.3.3 Characteristic Direction Alignment 181
5.4 Characteristic Locus Design Method 182
5.5 AGI Characteristic Locus Design 184
 - viii -

Page No.

5.6 Implementation of Characteristic Loci Designed

Controllers 191
5.6.1 Validation of the Multivariable PI
Control Scheme 194
5.7 Tuning of a Multivariable PI Controller 199
5.8 Interaction Measurement and Analysis 208
5.8.1 Closed Loop Interaction Measurement 210
5.8.2 Multivariable Input-Output Pairings 211
5.9 Discussion 217
References 219
Nomenclature 221

6. PRACTICALEVALUATION
OF CONTROL
TECHNIQUES
USINGAN
EXPERIMENTALTEST RIG 224

6.1 Introduction 225

6.2 The Experimental Test Rig 225
6.2.1 Test Rig Operation 226
6.3 Static and Dynamic Performance of the Electro-
pneumatic Converter - Valve System on the
Experimental Test Rig 231
6.3.1 Software Positioning of the Control Valves 236
6.4 Comparative Study of the Test Rig Behaviour 238
6.5 Design of SISO Proportional Plus Integral Controllers
Using the Modified Ultimate Method 245
6.6 Digital Multivariable Proportional Plus Integral
Control of the Experimental Test Rig 250
6.6.1 Multivariable P+I Controller Application 253
6.7 Discussion 262
References 266
Nomenclature 267

7. CONCLUSIONS 269

References 275

APPENDICES 276

AI PROGRAMS
COMPUTER 276

AI. 1 Fortran Simulation of the Above Ground Installation

(AGI) 277
AI. 2 Fortran Routine for Linearising the AGI Simulation
at Defined. Operating Points 282
AI. 3 Fortran Routine for Evaluating the AGI's
Characteristic Equation 288
AI. 4 Fortran Routine for Designing PI Controllers Based
Upon Approximate Models 292

All PUBLICATIONS 299

AII. 1 Axial Flow Regulator -A Study 300

AII. 2 Investigations into Remote Pressure Control of Gas
Transmission Pipelines 344
AII. 3 Modelling and Control of a Two-Stage Gas Pressure
Reduction Station 358
 ix -
-

Page No.

AIII ALTERNATIVE
DESIGNMETHODS
FORUSE IN AGI STATION
CONTROL 368
AIII. 1 An Alternative Method of Designing SISO
Proportional Plus Integral Controllers Using
Approximate Models 369
AIII. 2 A Digital Alternative to Control System Design 385
 -1-

CHAPTER1

INTRODUCTION
 L

British Gas plc provide a full range of services from

exploration and production to the selling and maintenance of domestic

appliances, via its twelve regional authorities, making it the world's

largest integrated gas undertaking [1.1].

The growth of the corporation has been dramatic since the

discovery of the North Sea fields and it now supplies over sixteen

million customers providing approximately one quarter of the total

primary energy in Britain.

The production and transmission of natural gas throughout the
United Kingdom is shown in figure 1.1. Briefly, the gas is extracted

from the offshore reserves via production platforms and supplied to the

on-shore reception terminals at the following locations:

i) St. Fercus in Scotland.

ii) Easington on Humberside.

iii) Bacton in East Anglia.

iv) Isle of Grain in the Thames Estuary.

v) Barrow on the West Coast.

At each of these terminals there exists a substantial compressor

installation which is used to boost the pressure of the gas after

refining and then feeds it into the 36"/42" diameter national transmission

pipeline for distribution to the twelve regional supergrids throughout the

country.
St. Fergus is the largest reception terminal comprising 17 Rolls
Royce RB211 controlled compressors. Typical flowrates through St. Fergus

are in the region of 100 Million sft3/hr with the outlet pressure being
typically 75 bar G. As a direct result of these large volumetric flow

rates, there is an appreciable pressure drop due to friction along the

length of the pipeline and therefore there are intermediate compressor

stations, usually comprising two identical RB211 or Maxi Avon compressor

 -3-

to
co

ception

Fig. 1.1 National Gas Transmission System

 -4-

units, installed at 40 mile intervals, which boost the pressure back

up to 75 bar, as shown in figures 1.1,2 (typical inlet pressure to a

station would be in the region of 45 bar G).

In addition to the offshore sources, British Gas plc support

two Liquid Natural Gas (LNG) plants at Dynever Arms in Wales and the

Isle of Grain in the Thames Estuary. These installations extract gas

from the national transmission system pipeline during the summer months

(periods of low demand), and liquefy and store it in high pressure

holder stations. When the demand for gas becomes excessive during the

winter months the LNG is vaporised and exported back into the national

transmission system at 75 bar G to top up the supply (peak shaving).

The Isle of Grain site is by far the larger, comprising six vaporisers

each capable of exporting up to 120 tonnes of LNG/hr.

As stated previously, the National Transmission System (NTS)

supplies gas to the twelve British Gas Regional Authorities via some
3,500 miles of high pressure pipeline. The gas is extracted from the

transmission pipeline through FITSofftakes sometimes referred to as

Above Ground Installations (AGI's); figure 1.2 illustrates the pressure

profile of the national transmission system. There are approximately

200 AGI's throughout the United Kingdom, each one capable of passing up

to 25 Million sft3/hr. These stations are sited at remote locations and

operated in unmannedmode with the controlled parameter being volumetric

flow rate. Each station is monitored from the regional Grid Control

Centre which has the facility to adjust the flow set points for closed

loop control, and open the loop to control directly the position of the

main regulators via telemetry. The outlet pressure is limited between

350 lbf/in2 and 550 lbf/in2 G to ensure continuity of supply and limit

the maximumpressure to a safe value respectively.

 - 5-

V
I-
to >b
c 4-3

1-4
c:
cr-

S-
c
0
o
NC 4-3
to O
Q) "r.
S- 4J
Cl- it
E+- 4-3
O N1 N
U . I-

S-

(A
w
0-S.
-
E

L- 4-)
0 N
NC >
v) O L')
G) "r
i 4)
G (ö 0
E 4J .r-
O v) (A
V N
. I-
E

ä-
r
(ß
C
0
o . r.
'r. CO 4J
+-) c b
n. "r-
a) E
U
()
c4-) a) N

"1

O CD C:) Kt
aý r. 9.0 Ln mNO
S .0
N
N
C)
S..
0

NOO oööö0
dO CO tz Ict N
vr
 -6-

In addition, these NTS offtakes supply the regional Pressure

Reduction Stations (PRS's) which are similar in design to the AGI's

but the controlled variable is outlet pressure. There are approximately

1,500 to 1,600 stations throughout the 12 regions, all monitored remotely

from their respective Grid Control Centres.

Finally, these stations supply the low pressure distribution

terminals of which there are approximately 3,600. The outlet pressure of

these stations is typically 10" WGto 30" WGand represents the final

pressure reduction.
The rising consumption of gas in recent years has not only
increased the average flow rates through the AGI's but has also enlarged

the seasonal variations in demandwith resulting changes in process

dynamics. Inadequacies of the existing pneumatic equipment have been
identified by an instability or hunting causing a severe deterioration
in quality of control as well as maintenance problems because of undue

wear of the first stage regulators.

In a recent study [1.2] the performance of the currently

employed pneumatic control scheme was analysed using a prototype system

consisting of single stream operation with first stage pressure cut and

second stage volumetric with high and low pressure overrides. The main

problems identified were:

a) Sluggish flow control response.

b) Excessive pressure transients when overrides are induced.

c) Poor response and accuracy of pneumatic control/switching

equipment.
d) Poor fault detection and isolation.

Furthermore, problems investigated within AGI stations operating

with the conventional pneumatic controllers have highlighted the

following operational disadvantages:

 -7-

1) The redundancy in the mechanical control hardware to provide

control results in pressure differentials between the set points

of various controllers, causing pressure stacking across the

station. Typically, this can result in the suppression of the

high pressure limit by more than 205 Kn/m2 (30 lbf/in2).

2) Provision of first stage pressure control into a small inter-

stage volume can result in unstable controller interaction. A

temporary solution to this problem has been to inhibit the first

stage pressure control and use the second regulator for the

total pressure cut. This results in high noise and, in some

cases, excessive levels of vibration.

3) Under certain control valve failure modes, an otherwise healthy

station could be totally shut down.

4) Because of capacity and storage requirements of the downstream

system in winter, the pressure set points are sometimes elevated,

their ultimate value being determined by the slamshut settings

and the confidence of the engineer that the pneumatic hardware

will discriminate. However, under pressure transient conditions,

slamshuts can be inadvertently actuated.

5) Because of the inflexibility of pneumatic controllers, capacity

and storage are lost in pipelines that enter restricted pressure

zones.
In 1980, a working party was set up to investigate the application

of the new generation of commercially available microprocessor controllers

to an AGI. The main objective of the new digital control scheme was to

overcome the operational constraints of the existing pneumatically

controlled offtake stations and increase the control flexibility, namely:

 -8-

i) To increase the available storage capacity of existing pipe

systems, by eliminating pressure stacking between regulators

and slamshuts within AGI's allowing normal operation closer to

the pipeline limits.

ii) To increase the flow capacity of existing stations and/or reduce

upstream compressor fuel costs by using multi-stream operation

for both normal and reduced inlet pressures.
iii) To improve regulator failure discrimination and control mode
degradation by using revised control and stream selection

strategies.
iv) To reduce revenue costs by changing the maintenance philosophy

from one based upon frequency to one based on necessity.

v) To reduce operational problems, such as regulator instability,

inadvertent slamshut operation and excess noise emission, to a

minimum.

vi) To provide a system which is 'end user' friendly irrespective

of engineering discipline, while providing the robustness in

terms of integrity, required to prevent operator/microprocessor

malfunction at all times.

The work involved within the scope of this text addresses parts

(i), (v) and (vi) above.

These factors are inter-dependent and, as explained above, are

all expected to benefit from more sophisticated control strategies made

possible by applying microprocessor techniques.

Direct financial savings from (v) and (vi) are difficult to

assess but, to illustrate the potential of (1), we can consider a

particular installation. The Saltwick - Bishop Auckland pipeline could

have its storage capacity increased by 42,500 m3 (1.5 P1ft3) simply by
 -9-

raising the high pressure override by 205 Kn/m2 (30 psi). An estimate

for providing equivalent storage is based on the figure of f25/m3

(50.7/ft3) so the proposed scheme is capable of saving, or at least

deferring, a capital expenditure of 51.05 Million. The cost of

conversion from the existing pneumatic equipment to microprocessor

based control is therefore adequately covered by the savings from this

aspect alone.

Initially, the work was split into three individual modules

which would ultimately lead to the development of a complete station

control scheme. The development modules being:

1) Electronic Stream Selection.

2) Electronic control of the Flow and Pressure Control Regulators.

3) Remote Boundary Pressure Control.

In 1983, a joint collaboration between ERSand Sunderland

Polytechnic was undertaken to address the problems associated within the

second development module. The work presented within this thesis is

dedicated to the development of a control strategy concentrating primarily

upon Above Ground Installations.

Early attempts to address the stability problem encountered

within AGI's commencedin 1970 when an ERSinternal report [1.3]

characterised the dynamics of the first stage pressure cut using a first

order mathematical model to represent the valve/interconnecting volume

network. At the same time this first order relationship was taken by
Brown and Todd [1.4] and used to form the basis of a stability analysis

upon the twin valve system to variations of the interstage volumes

dimensions. However, these techniques were limited in that they only

considered the system's structure, whereas the work presented within

this text is dedicated to the development of a general control strategy

which will solve the stability problems and improve station performance
irrespective of the individual site's topology and/or construction.
 - 10 -

The objectives of the study are thus: -

1) To develop a comprehensive and flexible mathematical model for

an AGI station.

2) To ascertain those design parameters which most influence the

performance of an AGI system, and finally

3) To develop controller policies for these stations that will

enable greater operating efficiency over their entire working

ranges.
The study commencesby first considering the hardware and

design philosophies currently employed within AGI stations. A mathematical

model is developed of a typical pressure reduction station whose

performance is then used as a yardstick against which all further designs

can be assessed.
Chapter 3 presents a controller tuning strategy which uses an
ideal relay to automate the Ziegler-Nichols [1.5] approaches presently

employed. Guidelines are presented for the industrial application of the

technique and its subsequent performance on a variety of Gas systems is
investigated.

A linearised model of the AGI is presented in Chapter 4 using

small signal analysis techniques, identifying the station's multivariable

nature. The resulting simulation is then used to investigate the factors

which most affect the system's performance and to assess its stability

over its complete working range. Multivariable controller designs are

presented in Chapter 5 to offer a formal design solution using frequency

response methods. The resulting structure is then used to investigate a

number of practical solutions to the on site tuning of such a controller,

based upon the results of step response tests.

 - 11 -

Chapter 6 discusses the construction of a pneumatic test rig

with the flexibility to emulate the behaviour of the majority of AGI

stations. The test rig provides a means by which the controller design

techniques developed in Chapters 3 and 5 can be evaluated under

laboratory conditions, thus forming the first stage of a stringent

vetting procedure before applications of the designs to the actual trans-

mission system can be considered. Finally, Chapter 7 is devoted to

conclusions and recommendations for further work.

 - 12 -

References (Chapter 1)

1.1 March, A. D. N. "Controlling the British Gas grid system".

Measurement & Control, Vol. 18, pp 345-351, Nov. 1985.

1.2 Arden, W. J. B. "GASCONTROL -A review of past, present and

future systems". North of England Gas Association, ERS,
Killingworth, 1988.

1.3 Smith, I. "The control stability of series pressure regulation

systems". Gas Council Engineering Research Station Report No.
ERST37, Sept. 1970.

1.4 Brown, G. L., and Todd, A. "Stability analysis of a two stage

pressure reduction station". Sunderland Polytechnic Internal
Report, July 1970.

1.5 Ziegler, J. G., and Nichols, N. B. "Optimum settings for

automatic controllers". Trans. ASME, Vol. 64,1942.
 - 13 -

CHAPTER
2

AND VALIDATIONOF A MATHEMATICAL

THE DEVELOPMENT

MODELFORAN ABOVEGROUND
INSTALLATION
 - 14 -

2.1 Introduction

The development of novel control schemes and the analysis of

system stability is fraught with many difficulties if attempted directly

upon real plant. Amongthe problems that may be encountered are excessive

cost due to plant inactivity whilst the identification tests necessary to

tune the control schemes and/or study the plant's behaviour are carried

out. This is especially true when attempting to derive transfer function

relationships from dynamically slow processes such as the gas transmission

system. Further difficulties arise in keeping the system variables constant

whilst attempting to assess the effect of a specific test on the overall

system behaviour. In particular, the gas systems considered here will

'float' upon the prevailing pressures, whose values differ continuously

with demand. Finally the operation of all tests are limited within a

specific range determined by the plant's safety system, which when

triggered will significantly change the system characteristics and

invalidate the exercise. For example, the high and low pressure overrides

in an Above Ground Installation (AGI). Because of these disadvantages, it

is often desirable to develop alternative schemes initially using a model

of the process.
The word modelling, in an engineering context, has two principal

meanings. The first is associated with physical models where actual hard-

ware is constructed according to appropriate scaling laws such that its

behaviour is predictably related to that of the full scale system. Wind

tunnel testing of aircraft and towing tanks for ship design are examples

of this type of r.iodelling. The second is mathematical modelling in which

the system to be studied is represented by a set of equations developed

by applying the appropriate physical (and chemical) laws to each component.

Solution of this set of equations to various sorts of stimuli

will then represent the behaviour of the system. These equations, together
 - 15 -

with the computational means of solving them, can then be considered as a

simulation of the system allowing systematic investigation of control modes

or, indeed, the effects of varying any parameter within the jurisdiction of
the designer.

In all but the simplest cases, practical considerations will

necessitate compromises being made so that the model, whether physical or

mathematical, will not be capable of reproducing all aspects of full scale

behaviour. Engineering judgement and critical appraisal of the respective
data is therefore required to ensure that the results are trustworthy in

those areas relevant to the investigation.

In particular, for stability and control studies, a mathematical

model must incorporate expressions which describe the dynamic properties

of the plant and this, in itself, can be a formidable task. The dynamic

equations are inherently more complex than the corresponding steady state

relationships (to which they must reduce as a special case) and their
formulation demands a knowledge of, or at least a familiarity with, very

many branches of technology. Moreover, the primary information itself may

be inadequate to define a rigorous model because some of the relationships

are known only as empirical or semi-empirical correlations. Because of

this and the unrelenting factor of time, some approximations are, therefore,
inevitable even at this stage and it is largely a question of judgement
based on personal, or colleagues experiences.

This chapter is concerned primarily with the development of a

mathematical model for an AGI. The model must be capable of implementing

different permutations of the individual system elements within the overall

framework to simulate any of the 200 AGI's that exist in practice. More-

over, it must also permit an investigation into the effects on system

performance of new and enhanced equipment prior to its application. The

potential benefits of such a simulation study are: -

 - 16 -

i) Individual analysis of the elements that comprise the total

system would allow considerable simplification of any subsequent

stability/controller design study.

ii) Information gathered from identification tests upon sections of

the real system could be incorporated for evaluation or updating

of the model.

iii) It will yield. a fuller appreciation of the fundamental aspects

of the system and the task which it has to perform.

2.2 AGI Structures

The AGI station, Figure 2.1 (2.1), comprises an interacting

network of pipes, valves, controllers and associated sensing equipment.

The primary function of these stations is to provide gas, either in

required volumes or at required pressures to the various parts of the

regions. The mode of control currently employed, is based upon single

stream operation with an identical standby stream, the latter assuming

control in the event of a working stream failure. In the case of the

National Transmission System (NTS) offtakes, Figure 2.2a, each region

currently predicts its gas requirements 24 hours in advance and each

stream is thus configured as a first stage pressure reduction followed by

a flow control regulator protected by high and low pressure overrides.

The flow controller set-point is adjusted remotely from the grid control

room at the value requested by the region concerned in an effort to ensure

that the gas produced offshore is used efficiently. In contrast, a

Regional Pressure Reduction Station (PRS), Figure 2.2b, consists of a two

stage pressure reduction providing outlet pressure control only.

Irrespective of the type of station control the downstream

pressure must be maintained between upper and lower limits because of

 - 17 -

ýý
4) vo
0 4J CC
a E
3 vý Co
0N a)
a)
(1) E N
v
"r
LN
L +' 0 I--

NC
> "r
a)
O0 c r0r- a
m
54-3 "ý0)0
tp ii
41
+b Q) C
NV) 0 O
V
N
tio
4)
N
S-
H
Eo
ro4,
asp
ir
4) M
mo N Q1
,- "r Q. N
.4-
c +ý
"r- co O
w 4J
CL.
O
Cr
0 tu
CL
I-

N c)
4)
C cl)
C.i
"r L
OO
a>E
4-3 ns c>
0) i 0
Q) 4-- S.
bý CV
NN C IN
0
"r U
S.- li.
I-

C)
U 4)
"r- 4-,
4-
-
ia
O
 - 18 -

spi SP2

Valve 1 Valve 2

a. Pressure Control

Orifice
plate
b. Volumetric Control

Fig. 2.2: AGI station control structures

 - 19 -

pipeline mechanical properties and operational constraints. These limits

are maintained by providing high and low pressure overrides designed to

assume control of the appropriate regulators when required. In the event

of the high pressure overrides failing to work then ultimately the stream

will be isolated from the transmission system by the actuation of a slam-

shut valve.

Under normal operation, the volumetric/pressure control stations

depend upon proportional plus integral control of the particular process

variables to some set point defined remotely by Grid Control, which, in

the majority of cases, is provided by conventional pneumatic proportional

plus integral (PI) units.

2.2.1 Regulators

The regulating control valves are arguably the most important

item of hardware within the AGI station. As a consequence, various sizes

and types of regulator are employed within the NTS/PRSstations depending

upon the policies, and politics at the time of commissioning, of the

particular authorities as well as the predicted loading requirements of

the future. At present, the majority of installations have employed two

main types of control valve, Fisher's V25 (generally NTS's : typical flow

20-30 Msft3/hr), and Jetstream (usually PRS's : typical flow

requirements

requirements of up to 10 Msft3/hr) regulators.

The V25 regulator (2.2) is available in sizes between 4 to 16

inches diameter, permitting a maximumdifferential pressure of 1000 psi

the regulator. The regulator is a throttling ball valve operated

across
by an air to open pneumatic actuator with built in positioner, Figure 2.3.

Linear movement of the actuator is transmitted through a crank to give

angular rotation of the ball, thus varying the aperture, machined into the
 - 20 -

Cr,
arran

Fig. 2.3: Schematic diagram of a V25 throttling ball valve

 - 21 -

ball, available for gas flow. Its maximum flow capacity is approximately

30% greater than that available through other types of regulator of the

same diameter.
Fisher's Jetstrear regulator (2.3) is available in sizes between

1 and 8 inches diameter, with a maximumdifferential pressure of 1200 psi

allowed across the valve. The regulator is an expansion type diaphragm

motor valve operated by an air to open pneumatic actuator, Figure 2.4.

Any downward movement of the actuator stem is transmitted to the slave

cylinder via hydraulic fluid, applying pressure to the solid rubber inner
bung which expands radially to decrease the annular area within the valve
body and reduce the flow. Because the relationship between the valve stem

displacement and the corresponding change in annular area is unknown and

impractical to measure, no form of positioning is available for this

particular type of regulator.

2.3 Mathematical Modelling of the AGI System Elements

2.3.1 Valve Characterisation

Accurate modelling of the control valves used in the industry

is of great importance because of their critical role in the operation of
the Above Ground Installations.

Because the fluid is a gas, inertial effects in the flowing

stream can be neglected and the valves characterised by combining two

separate factors (2.4) based on the results of static and dynamic tests,
these being: -

i) Valve Flow Profile

The development of a mathematical relationship which can be used

to evaluate the valve's throughput given the stem position and

pressure conditions directly upstream and downstream of the

regulator.
 - 22 -

Rubber
Air in diaphragm

Solid arrangement
rubber
bung

Fig. 2.4: Schematic of a Fisher Jetstream regulator

 - 23 -

ii) Actuator Dynamics

Incorporating all the valve dynamics relating the valve's stem

travel with respect to its applied control input, lumped

together in the form of an actuator model.

The final requirement is that the resulting valve model structure

should be sufficiently flexible to permit its application to any of the

many different types and sizes of control valve that occur within the

syster in practice.

2.3.1.1 Steady-State Characterisation of Regulator Flow Profiles

Various forciulae have been proposed (2.5-2.7) for predicting the

steady state flow rates of gases through the many different types and

sizes of commercially available regulator.

flany are based upon the accepted equation for the flow of liquids

through control valves.

cv 2.1
.....

derived from Bernoulli's equation and the continuity equation and differing

only in the pressure term that is incorporated to describe the gas density

effects, i. e.

Q= 1364 Cv/1 -1.... 2.2

although several of the flow formulae are empirical relationships, developed

to describe data collected from air tests upon particular control valves
(2.7). Most of these formulae are intended for use over a restricted range

of pressure drop conditions and a more general expression to cover the flow

of a compressible medium under isentropic conditions must take account of

the transition in flow regime which occurs at a critical ratio of inlet and

outlet pressures where the fluid passing through a controlling aperture

 - 24 -

reaches sonic velocity. Theoretical treatments of this phenomenonlead

to complex expressions which are computationally cumbersome (2.8).
By introducing sufficient redundancy into the formulation,
Buresh and Schuder (2.9) have proposed a method that is applicable to

virtually all types and sizes of regulator existing today. Utilisation

of this "Universal gas sizing equation" within the proposed valve model

would permit the use of a fixed structure subroutine, whose predicted

output flow would be dependent upon the surrounding system pressure (psi
Absolute) and temperature T°(R) conditions and the various constants

necessary to describe the particular regulator using the equation,

59-64 P1-ß1 ,r/2

Q=T C2 Cg PI sin sft3/hr
.
2910
rads
2.3
.....
where CV and Cg are coefficients which are dependent upon the valve's stem

position; the first indicating the basic flow capacity of the valve for
the critical drop, whilst the latter measures the extent of the pressure

recovery of the valve. The inclusion of constants C2 and G allows for the

various types of gases that-can be passed. Figure 2.5 tabulates the

values. of C2 and G for some of the more commonly encountered types of gas.

2.3.1.2 Evaluation of the Regulator Coefficients Cv, Cg

Since the coefficient Cv is the standard liquid sizing

coefficient, equation 2.1, this could be determined by water tests.

Alternatively the information could be gathered from air tests providing

P2<0,02
i

Only within this range can the incompressible/compressible flow through

the regulator be said to be equal. Here the trigonometrical relationship

sin(X) =X holds accurately and equation 2.3 reduces to equation 2.2, i. e.

 - 25 -

Correction factor for

Gas Specific density specific heat variation
(G) (C2)

Acetylene 0.90 0.98

Air 1.00 1.00

Butane 2.00 0.94

Ethane 1.03 0.96

Helium 0.14 1.04

Hydrogen 0.07 1.00

Methane 0.55 0.98

Natural Gas (Bacton) 0.60 0.98

Nitrogen 0.97 1.00

Oxygen 1.10 1.00

Propane 1.52 0.95

Propylene 1.45 0.91

Fig. 2.5: Gas property dependent-constants for use in the

Universal gas sizing equation [2.2]
 - 26 -

A Pf A ýlý
Q= 59.64 Cv Pl 2.4
.....

The coefficient Cg is dependent upon the critical flow (Qc) through the

regulator, that is, the maximumflow through the valve at a fixed stem

position, over the entire pressure range, when sin(X) = 1, i. e.

Qc = C2 Cg Pi 2.5
.....

Figure 2.6 illustrates the use of equations 2.4 and 2.5 in determining

the coefficients Cv and Ca at a fixed stem displacement of 70% open. The

illustrated characteristic is taken from the manufacturer's details of a

10" V25 throttling ball valve, passing natural gas with a specific gravity

of 0.6 at a constant temperature of 520°R (60°F) with an upstream pressure

of 1000 psiA.
Although the above methodology is presented in terms of imperial

quantities the equation itself will be self scaling if SI unit measure-

ments are incorporated in the above fornulaes, the result being a

multiplication of the Cv and Cg coefficients by a factor of 1.1408E-6,

based upon pressure units of Kn/m2 and volumetric flow rates in m3/sec.

2.3.1.3 Application of the Universal Gas Sizing Equation

to the Fisher V25 Regulator

The first stage in the development of the Universal gas sizing

equation for any regulator is the recording of the device's steady state
flow profile. Figure 2.7 illustrates the flow characteristics of a 10"
V25 throttling ball valve, derived from manufacturer's details. By

applying equations 2.4 and 2.5 to the valve's flow characteristics at

each specific value of valve stem position, in the manner displayed in

Figure 2.6, a pair of expressions relating the Cv and Cg coefficients to

 27
Qnax

100
Incompressible
flow

75

Critical
flow
OOO
50 Measured
000 data
/O 0
00 O
O
O

25
ýdýj

3z,
0
0 0.2 0.4 0.6 0.8 1.0
/67

Fig. 2.6: Fisher V25 70%open characteristic; showing application of Universal Gas
equation to determine Cv and Cg coefficients

0
 -28-

VALVE TRAVEL(%) CL
0 w
cx
°2 O O
r
N
O
N
N
w
cy-
a
J
+J
I-
00
W
cy-
W U
U-
U (0
Cl s
0
Ln W 3
n I-
0
ýO

o,
a,
Lý

tn N
CV G1.

p
Q) Lo

Ln rtsC
iY
o

N
M
4-4-
00
S-
CL

O
N
'F- N
G) d
Ln ro E
ý ro
N 0!

+)
bN
ýö C
v3
++ o
NM

C1

LL-

000°o0000°
ý` f
O 47 %0 -t rn

(°/a Mold 3A1VA

I
 29

the valve's stem position can be determined. Figure 2.8a displays these

relationships graphically for the V25 regulator as percentages of their

maximumvalues, where
Cv (100%) = 2760 ) based upon
' imperial
and C9 (100%) = 56600 measurements

For application within the computer simulation, polynomial

descriptions of the two responses were derived using a least squares

parameter estimation routine, resulting in the following equation format,

with the constants tabulated in Figure 2.8b.

Cx%= a1X%1+ + a3X%3 + a4 X%4 + a5X%5 2.6

a2X%2 .....
To ensure consistency between the valve and its model the stem
travel is saturated at both extremes of operation.
The steady state flow profiles displayed in Figure 2.8a are

representative of all the available sizes of V25 regulator and Jetstream

regulator, allowing the characterisation of these valves using the above

analysis and the individual valves maximumCv and Cg coefficient. Some

of the more popular sizes of these regulators' maximumCv and Cg

coefficients are tabulated in Figure 2.9.

Further validation of this technique is presented in Appendix

II. 1 (2.10) which includes a report upon the modelling of a 4" Axial flow

valve, a device commonly encountered within the regional pressure

reduction stations. Operation of the axial regulator requires the

introduction of gas into a chamber behind the rubber constricting sleeve

at system pressure levels via an appropriate interface unit. Even though

the valve operates in a much different manner to any of the previously

discussed regulators, the Universal gas sizing equation provided a good

representation to the recorded data set. Comparison of the resulting

flow equations for the regulator at two different upstream pressure

ranges indicated a discrepancy which was finally tracked down to the

positioning of the measurement transducers.

 - 30 -
100

3
° 90
U-

80
u..
O

ü 70

60
W
a
50

40

30
CG% CV%
20

10

b 20 30 40 50 60 70 80 90 100
0
PERCENTAGE
OF TOTALTRAVEL

a. Normalised Cv, Cg ti stem displacement relationships

Cx%= a1X%+ a2X%2+ a3X%3+ a4X%" + a5X%S

al a2 a3 av a5

Cv -2.496E-2 1.145E-2 -1.781E-4 1.661E-6 -

Cg -1.040E-1 3.35E-2 -6.121E-4 6.356E-6 -2.484E-8

b. Polynomial fitting of the above coefficients

Fig. 2.8: Evaluation of the Cv, C coefficients for the

modelling of a 10" Fisher V25 regulator using
the Universal gas sizing equation
 - 31 -

Regulator Size Cv (100%) Cg (100%)

(inches)

4 490 10,100

6 1,140 22,300

8 1,680 34,400

10 2,760 56,600

12 3,690 75,600

16 5,820 119,000

a. V25 Throttling ball valve [2.2] with a

line to valve size ratio of 1.5: 1

Regulator Size Cv (100%) Cg (100%)

(inches)

1 14 297

2 51 796

4 86 2,087

6 300 5,371

8 765 13,530

b. Fisher jetstream [2.3) with 60% liners fitted

Fig. 2.9: Maximumtravel flow coefficients for

use in the universal gas sizing
equation; based upon imperial
measurements
 - 32 -

Repetition of this exercise on all the regulators currently

in service would lead to the formation of a valve database, which could
be readily loaded into the Universal gas sizing equation for simulation,

achieving both conformity and flexibility.

2.3.2 Actuator Dynamics

Many different types of actuator are used within the trans-

mission system at present, the most prevalent being the Fisher pneumatic

actuator, consisting of an actuator chamber which when sufficiently

pressurised provides enough force upon a metallic backed flexible rubber

diaphragm to overcome an opposing spring, present to ensure the valve

returns to its starting position, to cause valve stem travel. The

actuator stem travel is ideally proportional to the pressure in the

chamber but may, in practice, be affected also by friction at the valve

gland (pressure seal) and by fluid forces, whether static or dynamic,

acting on the internal components, e. g. valve plug, disc or ball (2.25).

A positioner, i. e. a local feedback loop which compares the

inlet control pressure (or current) and valve stem position, is often

used to reduce these adverse effects.

In general the actuator pressures required to move the valve

stem from one extremity to the other are typically 3-15 psig, these

pressures being provided, in the case of microprocessor control, via an

electropneumatic convertor, a device which regulates its pressure output

using a needle valve connected to a sprung lever with the position of

the lever being determined by an electromagnet, and therefore by the

device's current input. Since the speed of response of the valve/

convertor arrangement is limited by the rate at which the actuator chamber

can be charged and discharged, it is commonpractice to fit booster relays

to the convertors to permit greater flowrates into and out of the chamber
by working from larger pressures.
 - 33 -

The following work details the identification of the dynamics

of a Fisher air to open pneumatic actuator driving a 4" Jetstream

regulator, whose pneumatic input is supplied by a Fairchild T5200

electropneumatic convertor with booster relay, a device frequently

encountered within industry. By monitoring the response of the

actuator's chamber pressure and the valve's stem position to changes in

control voltage, then the dynamics of the actuator/electropneumatic

convertor could be analysed. This was achieved by the application of

appropriately sized pressure and displacement transducers to the valve

in the manner displayed in Figure 2.10, further details of the test

equipment being tabulated in Figure 2.11. All measured inputs and outputs

were conditioned to provide signals in the range of 0- 10 volts, a

constraint imposed by the 8 bit data acquisition unit.

The response of the system to step changes in the control input,

Figure 2.12, illustrates the associated dynamics. Inspection of the

actuator chamber's pressure response, Figure 2.12a, to the square wave

input shows the electropneumatic convertor to be bi-linear, i. e. displaying

different dynamics in different directions. Further investigations

revealed the cause of this behaviour was due to a mismatch in the

convertor's inlet/outlet port sizes. However, this behaviour could be

overlooked as the chamber pressure was not necessary in describing the

actuator's dynamics using the prescribed method. Furthermore, analysis of

the valve's stem displacement to the same control input variations

indicated an almost uniform response in both directions.

A linear model of the valve's stem travel behaviour, Figure

2.12b, using a first order differential equation and standard transient

analysis techniques was obtained for both the charging and discharging

conditions, the resulting system time constants being 0.93 seconds and

1.05 seconds respectively. By averaging the charging/discharging rates a

 34

Air
supply

Fairch
I/P col

Actuator
pressure

Steam di
transduci

Fig. 2.10: Schematic of actuator testing arrangement

upon 4" jetstream regulator
 - i5

C "ý
E
CL
O of
CL
t\
""- LO tD ko
f\ Ul)
V O O O

p
0 U
O O O
ý
v1 N = CO ZA N rr-
M M rn
O Ö OO '.0 0
"ý "O qzr "O I[)
S
cf
44 O(V I CD 1 Or- 1
0
O Q1 C) Q)
" 4- >bm I ß 4) ce
4) 4.. a + to i )4Ä
4-3 a) i«- i"r- S- -r- o
4A a
N GJ II. CJß C. O'3 d
+) CCE CCE CCE
"r- to GJ "r- CO G1 "ý to d 4)
U
S- JmI-- Jcni
o -im

4)
V X
E 0
d a) U U
M
+V r "r 0 p a
N
CL
C O C.
O N 4-3
N O r- c
ý q:r to U
N E
N d

i-/ Q1 Q) QCJ
C Q1 "r- "r
Cl) C GJ VI E N
a E Q G)
YC
"ý- i 14) O UL) 4-3

W a S- r-
M LO
O O
4--
O
N
C
O"
. I-

U
G. i U . I-
O w
P-4 4j I U . I-
of U N U
00 O co T7 ci
CD 0 CD w O.
m N
"ý 54) N 4-3 O to
F- + co S-
O 'C i "r- OE 4)
r- (D u J O r-
r40 OOO S..
t N
LG S- t 40 r- C VI
J(o MO. yl
". - C r- CN to r_ C)
to OO ýp "r i O . r.
1i(3a: NCF-- Na Lt.
 - 36 -

sNN
Li
of cu

Mm

NN

OO

^^
m

1(

NN

r- m

I"1
I N.
ý
"ý.
O
Lri /r1 ON 00 O
rri

"X
1o ý. 0-
 - 37 -

single system time constant of approximately one second was obtained.

Figure 2.12b illustrates the accuracy of the resulting model described

by the following transfer function relationship

XK 2.7
H i+s .....

where K represents the scaling necessary to ensure that the outputs remain

consistent with those of the real system.

The experimental time constant obtained was confirmed through

discussions with site engineers and on site monitoring of valve stroking

times for the range of types and sizes of Fisher regulators employed.

2.3.3 Pipeline Modelling

The dynamics of the system downstream of the second regulator

play a major role in determining the overall stability of the network.

It is therefore necessary to ensure that the model used to represent this

part of the process in the simulation will reproduce the salient features

adequately (2.11-2.14) without incurring an inordinately high cost in

computing time.
Pipelines constitute distributed parameter systems and a

rigorous analysis involves partial differential equations which must be

solved under two-point boundary conditions (2.15-2.20). This is, in

computer terms, a notoriously time consuming procedure. We shall,

therefore, explore the commonly applied approach of converting the

partial differential equations into a set of simultaneous ordinary

differential equations by "finite differencing" (2.21). In effect, this

means discretizing the pipeline into a number of sections (n), see

Figure 2.13a (2.15,2.20 and 2.22). We know that putting n=1, i. e.

treating the pipeline as a single fixed volume, oversimplifies the

situation to a misleading extent (2.15). The first task is, therefore,

 -38-

QIN mm*o L ", 00UT

-I --- L,-I af

P(1) P(2) P(n)

111
171

a. Discretised pipeline structure

b. Restriction modelling schematic

P rý
M1 M2

V
Q1 Q2

R T

c. Capacitance modelling schematic

Fig. 2.13: Pipeline modelling techniques assuming negligible

inertial effects
 - 39 -

to find the minimum value for n which will yield acceptable results.
One major advantage of the approach is that the resulting model is well

suited to the application of standard linear control design techniques.

The nonlinear dynamics of each pipeline element are to be

modelled using a resistance/capacitance equivalent network, since in the

presence of pipeline wall frictional effects and pressure drops inertial

effects are negligible (2.16).

2.3.3.1 Pipeline Restrictions

The pressure/flow relationship of compressible gas propagating

down a pipeline are known to exhibit a high degree of nonlinearity. Even

in the steady state numerous relationships have been proposed which

evaluate the pressure drop corresponding to a given flow magnitude and

direction (2.22). Any such equation which is to be employed to estimate

the discretized pipeline's pressure profile should incorporate the

following: -
(i) Wall frictional forces, which are one of the major

contributions towards the pressure drop within a pipeline (2.17).

(ii) The pressure head that occurs due to the geography of the

pipeline.
One such equation that satisfies the above criteria, and whose

accuracy has been proven in practice (2.23) over medium and high pressure

ranges, is the 'General Panhandle equation' (2.24) indicated below.

Q=K. Lß. , p5] 2.8

Pb L.
avg. cavg. f .....

where
K' = 4.95286

0.031055. G. (h2-hi). Pavg2 2.8a

Ex Tavg
avg. .....

Pb, Tb base pressure and temperature respectively

Zavg average compressibility factor for the gas

 -40-

The correction for elevation changes (Ex) is based upon average

gas density and the pipeline wall friction factor (f) is determined from

a Moody Chart (Figure 2.14) using the following equation which evaluates
the system Reynold's number for natural gas with a specific density (G)

of 0.6
[Q]
Re = 88802.16 2.8b
.....

The friction factor can also be evaluated when under fully

turbulent flow conditions via

[4
-f = log(3.7 x Relative Roughness)]2 2.8c
.....

The above equations provide the basis to evaluate the pipeline's

steady state pressure profile at some specific flow level, under the
following assumptions (2.18): -
(i) The properties of the gas do not vary significantly over a cross

section of the pipe.

(ii) The area of a cross section of pipe is constant.

(iii) The temperature remains constant (2.19), Figure 2.15a, (T=520°R).

(iv) The pipe has a relative roughness of 12,500 (f=0.002872).

(v) The pipeline is horizontal (Ex=O).

Under the above assumptions equation 2.8 reduces to the

following equation which is used to evaluate pipeline initial conditions

for a given flow value

L. Q2
P12 P22
- =
735.581Z D7 2.9
.....

All quantities are in SI units as listed in the nomenclature.

As previously stated, the discretized model has the capacities,

and, therefore, the pressures, concentrated at nodes, requiring a

relationship to permit the calculation of the flow between nodes. A

model of a restriction which incorporates the wall frictional effects was

 - 41 -

cc
LU No-0
w
W CD KZ) Ö C> io
8 w
CID
J= N
Jý UO
W O
J cD
CY O
=x (1) tC
.< (n O Q'
CL, J N
O
cr)
Z

W I
CY G)
W
1I CO O
E
a)

I--
4)
CL
O.

1 . r.
O
IýW
L
O

S-
o

V U
m
4-
i-
O a
O 0
O
x Ln
W . r.
U
L
v-

N
4J
N

rC
w
f-ýt

O
O

ww LL.
CD Co %0 Ln'-
ö00 ö öö r-
ö0 CO 0
ZCY- N 0
c2 C:) O
a
O
VV Ö

LL LL
 - 42 -

Temperature
(°K)

400

300

200

100

10 20 30 40 50 60 flow (m3/sec)

a. Gas temperature ti volumetric flow rate

Density
(Kg/m')

60

50

40

30

20

10

60 flow (m3/sec)

b. Gas density ' volumetric flow rate

Fig. 2.15: Variations of gas properties during axial flow

valve testing over the regulators full working
range
 - 43 -

determined from first principles, Figure 2.13b:

Gas velocity

U=K. (dP)'

where K is a constant depending upon the characteristics of the

particular gas flowing; considering volumetric flow, then: -

Q=A. U = K. A. (dP)ý

= K'. (dP)l 2.10

.....
If Mass flow were being considered, then

A=p. A. U =K P. (dP)l 2.11

. .....

evaluated using the relationship A=p. Q obtained when differentiating

M=p. V for constant specific density. The above equations are applicable

if, and only if, the Mach number of the gas flowing does not exceed 0.2.

By combining the section pressure drop due to wall friction effects

calculated using equation 2.9 with equation 2.10 above, then the pipeline

section constant K' can be determined. This represents a 'non-linear

resistor' which fits directly into the proposed pipeline modelling scheme.
This 'new' equation would have the benefit of reducing the number of

analogue units necessary to simulate each section, therefore achieving

greater computational efficiency (2.22) at the expense of a slight

degradation in steady state accuracy. Further conclusions about the

validity of using this particular restriction model within the proposed

pipeline scheme will be discussed later within this section.

2.3.3.2 Pipeline Capacities

In order to complete the pipeline model, Figure 2.13a, a method

of evaluating the pressures at the nodes is required. This was also

developed from standard thermodynamic theory, Figure 2.13c.

 - 44 -

Consider the ideal gas law

P. V = m. R. T

Differentiation of the above equation with respect to time

results in

= K. M 2.12
.....
Under the following assumptions -
(i) Negligible temperature changes occur.

(ii) Constant enthalpy (no external heat added).

(iii) Constant specific volume, hence constant specific
density, its reciprocal.
Figure 2.15a and b illustrate the variations in temperature

and density respectively, that occurred during the testing of a 4" Axial

valve (2.10). The tests were carried out for flow variations of up to
55 m3/sec (7 Nsft3/hr), at valve differential pressures ranging from
350 Kn/m2 (50 psi) to 2400 Kn/m2 (350 psi). It is evident from these

results that only small changes in system temperature and density have

occurred, providing justification of the above assumptions.

Now
RVT-
K= =p. R.T = constant

In the previous section it was shown that under these

assumptions
A= P"Q

Combining this with the ideal gas law results in (2.13)

a-t =K*. Q 2.13

.....
where
Pb
K* = p. K = p=R _ = constant,

Pb being the base pressure and V the pipeline section volume. Note: Pb

can also be determined as a function of the particular gas flowing.

 - 45 -

2.3.3.3 Complete Discretized Pipeline Model

Each individual section of the pipeline model can now be

replaced by an equivalent nonlinear resistance/capacitance network using

the previously developed equations, 2.10 and 2.13, in the form and

notation, Figure 2.13a, indicated below: -

Q(i) = K. [P(i-1) - P(i)]'

.....
2.14a

PO-I) =I [Q(i-1) - Q(i)]dt + P(i-1)ic 2.14b

.....

The constants C, K and Pic are all functions of the initial

flow (Q), pipeline dimensions (L, D) and the numberof sections considered
(n).

2.3.3.4 A Comparison of the Effects of the Restriction

Modelling Techniques upon Pipeline Behaviour

The effect upon the pipeline behaviour of replacing the generally

accepted Panhandle equation by the nonlinear restriction equation developed

in section 2.3.3.1 required investigation. A valve/pipeline model was

constructed as a means of Bete m-fining the validity of this substitution,

based upon an actual scheme existing within the North Thames region.

This particular arrangement, Figure 2.16, consisted of a 16"

Fisher V25 throttling ball valve, see section 2.3.1, supplying 18 miles of

24" diameter pipeline that connected an A. G. I. station at Winkfield to a

pressure boundary at Ripley. By employing the modelling techniques

presented within this chapter to the Winkfield/Ripley pipe network a

'yardstick' to which all other proposed techniques could be compared was

developed.

In order to obtain a full representation of the pipeline

dynamics within the simulation, the controller arrangement chosen used

remote impulsing, via telemetry lines, of the Ripley boundary pressure.

 - 46 -

"o-0
Eý
cý o
_im

r-
0

0
U
a)
5ý.

C) Q)

r-
U
43

a) d)
E 73

4)
U
N E
U
0

G)
I
E

Co "r
r-- r-
U
0.
. r-
d

d
"r

Cl
+-) S C4
10
0 ý t0 "r
3
4-3
m
N
cr_I

r-
N

LL
 - 47 -

By setting the pipeline outlet pressure as the controlled variable then

the full 16 sections, see section 2.3.3.5, of pipeline model are

incorporated within the control loop.

The first indication of the two restriction modelling equations'

compatibility arose as a consequence of the controller tuning exercise.

This was achieved by applying the standard Ziegler-Nichols ultimate
technique (2.26, chapter 3) to the simulations resulting in the critical

gain (Ku) and period (Pu) values tabulated below.

Panhandle Restriction

Ku 10.904 11.145

PU 1722.0 1705.0

Table 2.1 Comparison of the critical gain and

period of the two restriction
modelling methods presented.

The closeness of the critical values indicate how little the

dynamics of the system have been affected by the substitution. A direct

comparison was obtained by employing proportional plus integral (P. I. )

control of the form detailed in equation 2.15, using the following

Ziegler-Nichols parameters tuned from the Panhandle equation simulation.

GA(S) - Kc TIT] 2.15

.....

where
Kc = 4.907

Ti = 1435.0

Application of this controller to both valve-pipeline

simulations when under a step change in boundary flow (pipeline outlet

flow) of +20% resulted in the responses displayed in Figure 2.17. No

significant differences in the transient behaviour of the two methods is

apparent, allowing us the confidence to utilise the more computationally

efficient nonlinear restriction model in any further studies.

 48 -

Open

56

54

52

50

10 20 30 40 50 60 time (rains)

a. Stem displacement movement - X

Flow rate
Msft3 /hr M3/sec
6.5
50

6.0

45. nt

5.5 le equatio

-essible
mation

5.0. 40
(mins)

b. Valve volumetric throughput -Q

Fig. 2.17: Comparison of pipeline restriction modelling techniques; time

(1 of 2) records illustrating the behaviour of the Winkfield/Ripley
telemetry control scheme to a 20% load flow disturbance
Note: Both techniques give the sane transient response under
these conditions
 - 49 -

Pressure
psi Kn/m2
3590
520

3580

518 1
3570

1 3560
516

3550

514
(mins)

c. Pipeline inlet pressure - P(1)

Pressure
psi Kn/m2
3450
500.

498.

3425
nt
496
le equation

494 essible
oration
3400

492
(mi ns)

d. Pipeline outlet pressure - P(n)

Fig. 2.17: Comparison of pipeline restriction modelling techniques;

(2 of 2) time records illustrating the behaviour of the Winkfield/
Ripley telemetry control scheme to a 20% load flow
disturbance
 - 50 -

2.3.3.5 Evaluation of the Minimum Number of Stages Necessary

to Adequately Reflect the Pipeline's Dynamics

In order to satisfy the transient accuracy of the discretized

pipeline model an infinite number of stages (n) are necessary. However

computationally this is impossible due to the excessive amounts of

programming required, indicating the need to determine the 'optimum'

number of stages which would satisfy the accuracy/computational efficiency

balance.

This was investigated using the Winkfield-Ripley simulation of

the previous section by comparing the system responses, when the system is

experiencing a +20% load flow change, for various values of n. Figure 2.18

illustrates the responses derived from the above test. From these

responses it was found that for 4 and above stages the differences between

the simulations were sufficiently small as to be considered negligible.

Further confirmation of this resulting figure for the 'optimum' number of

stages was found in a similar study (2.16) upon a pipeline model employing

continuity equations for the pressure/flow relationships.

By keeping n to a minimum, without significantly degrading the

pipeline's performance, the order of the characteristic equation developed

in chapter 4 would also be kept to a minimum, ensuring that any stability

study performed upon the A. G. I. simulation would be as uncomplicated as

possible.

2.3.3.6 Construction of the Downstream Dynamics

for the Proposed A. G. I. Simulation

As an example, in using the suggested pipeline modelling

technique consider the 'general system' appearing downstream of the A. G.I.

station, Figure 2.19, consisting of 3 miles of 18" diameter pipeline which

is to be discretized into four sections of equal length. It should be
 - 51 -

% open

56

54

52

50

ne (mi ns )

a. Stem displacement movement -X

Flow Rate
Msft3/hr, m3/sec
6.5.
50

6.0
:oe
/
_. -n=1
ýý'e"'Ool"
45 ---n2
.
5.5
n8
'/_"/ n=16

5.0. 40
time (minn)
b. Valve volumetric throughput -Q

Fig. 2.10: Evaluation of "optimum" number of stages (n) necessary to

(1 of 2) reflect the pipeline dynamics; time records illustrating
the behaviour of the tdinkfield/Ripley telemetry control
scheme to a 20% load flow disturbance
 - 5? -

Pressure
psi Kn/mz
518
3570

3560
516

3550

514
3540
e (mins)

c. Pipeline inlet pressure - P(l)

Pressure

psi Kn/m2
,
500. 3450.
-"-n=1
---n2
498. `ý n=4
-""-
" ýý --- n=8
3425 ýA'n= 16
,
496.

494"
3400 "' `ý --

492.
10 20 30 time (mins)

d. Pipeline outlet pressure - P(n)

Fig. 2.18: Evaluation of "optimum" number of stages (n) necessary to

(2 of 2) reflect the pipeline dynamics; time records illustrating
the behaviour of the Winkfield/Ripley telemetry control
scheme to a 20% load flow disturbance
 - 53 -

a.

a)
c
r-
a)
U
CL

N 4)
I
W
E

T
ai co
I
N
a)
a 0
U
i

r.
"r
E

cý S-
o

N N L
d U CL)
N CD

c
.ý 0
o. .r-
4)
43
N
N

N C3
N X N
d -
r
as
U

aý =
"E
"r cor- C)

C- N( CJ
C-
N U r- 0)
C9

I'ý 4J 0)
w s- LA-

4- 41
c)
r-"r

.I

x ..
.-

N
a
 - 54 -

noted that since the model parameters are all dependent upon the pipe-

line's dimensions the nodes need not necessarily be equally spaced; one

such case is detailed in Appendix 11.2.

The first stage of the mathematical modelling exercise is to

determine the pipeline's steady state pressure profile using equation 2.9.

The resulting pressures for an average steady state volumetric flowrate of

39.329 m3/sec (5 Msft3/hr) and a pipeline inlet pressure of 3447.5 Kn/m2
(500 psi) were

P(1) = 3429.0246 Kn/m2

P(2) = 3410.4492

P(3) = 3391.7720

P(4) = 3372.9914 "

These values formed the initial conditions of the integrators

that model the pipeline's capacity, equation 2.14b, and also provide the

means of evaluating the restriction constants (K) of equation 2.14a, thus

K(1) = 9.1499

K(2) = 9.1252

K(3) = 9.1003

K(4) = 9.0752

Now since each section is of equal volume, then the section capacity

constant (C) remains constant for all stages and is calculated to be

C=1.9551

The above pipeline parameters provide one steady state starting point

from which any simulation could be performed.

2.3.4 Interstage Pipework Modelling

The evaluation of the interstage pressure (Pu) within the twin

valve system, Figure 2.19, affects the operation of both the first and
 - 55 -

second stage regulators. By employing the techniques proposed in the

pipeline modelling section to characterise the interconnecting pipework,

the following conclusion was reached: -
The resulting pressure drop across the small section of pipe-

work is found to be sufficiently small, as to be considered negligible,

thus permitting a single volume representation of the interstage pipework

system, implementing equation 2.13.

Previous studies involved with the modelling of small elements

of pipeline have utilised this conclusion, i. e. Brown and Todd (2.27)

combined a small volume model of this type with a valve's flow character-

istics, modelled as a variable restriction, to form a first order lag

relationship with a variable time constant. Using the resulting model to

assess the system's stability with respect to the volumes dimensions, a

problem which is currently of great interest to the gas industry, in the

guise of A. C. I. stations' stability, and whose solution is pursued in

Chapter 4.

2.3.5 System Signal Conditioning

The scaling of the sensor/interface signals into the correct

format for application to the final modelling scheme is considered below.

Since the sensing lines are constructed from small bore pipe, the sensing

volumes are very small. This in conjunction with the fact that the flow

rates within the sensing lines are small, and therefore the pressure

drops within the lines are small, results in a resistor/capacitance

approximation of the sense line dynamics whose bandwidth is sufficiently

large, with respect to those of the process, that the use of a unity gain

model is permissible.
 - 56 -

Because the controller error signal lies within the range

± Ps*, where Ps* is the transducer input range, and the output signal

within 0- 100%valve stem displacement, corresponding to 4- 20 mA

controller output, then offset compensation is necessary. By combining

all the signal levels, the interface characteristic displayed in Figure

2.20 was derived. Note the inclusion of the signals saturation levels

that occur at the extremes of operation. In the case of the A. C. I.

simulation pressure transducers ranged from 0- 6895 Kn/m2 (0 - 1000 psi)

were implemented.

2.4 Discussion and Validation of the A. G. I. Simulation

The mathematical models of the process elements developed in

the previous sections were placed within a fixed structure framework of

the form detailed in Figure 2.21, from which the multivariable nature of
the process is clearly evident. Since current policy dictates the use

of a single input - single output (SISO) control scheme then the

construction of a high performance/stable system will only be possible

if any of the following situations exist: -
(i) The open loop plant matrix is diagonally dominant.

(ii) Any interaction that is present is beneficial to the overall

system's behaviour.
(iii) Very high controller gains can be applied without adversely

affecting the system's stability/noise sensitivity.

In order to quantify statement (ii), an analysis of the system
by various process interaction measurement techniques is presented in

chapter 5, where statements (i) and (iii) are also expanded.

One consequence of the application of SISO control to a multi-

variable process is that any interaction that is present would prove

 - 57 -

U)

L.1..

S.-

L
c)
v

L
4)

0
44

v
"r
4)
0
O
4-T
U'

xc CD S..

C U
G3
ü
to
Si-
a,

"r

Z3
O
U

I---
C
C)
N

U..
- 58 -

c
0
.I
4J
N

C:
Gr.

O
4-

O
^'
W

as

fL1
T

"r
r"
r`

cv
c4
cý,

LL.
 - 59 -

unfavourable when the SISO controller tuning exercise is performed.

This is considered in greater detail in chapter 3, where the resulting

controller parameters, table 2.2, were utilised to investigate the

complete simulation's performance.

Kc Ti

GC1 1.53 0.57

GC2 29.15 1.43

Table 2.2 Controller parameters for the A. G. I.

simulation tuned using the auto-tuning
technique proposed in chapter 3

The complete mathematical model of the AGI station described,

is now in the correct format for implementation within analogue/digital

computers, requiring only the means of solving the differential equations.

All digital computations were performed upon a VAX 11/750 VMSVersion 4.4,

and programmed in Fortran 77, a listing of which is presented in Appendix

I. 1, with responses displayed using REGIS graphics. Solution of the

differential equations was achieved using a simple euler integration

subroutine.
Validation of the resulting AGI model was not straightforward

since the simulation was 'general' and not of any one particular situation.
In addition, no real data was currently available in a form suitable for

the necessary confirmatory studies.

The answer to this seeming impasse was advanced by British Gas

themselves. Over the years they had independently developed their own

extensive simulation facility called "BSIHS" (2.28). This in-house

simulation package is based around a conventional dynamic modelling

kernel supplemented by highly sophisticated distributed pipe model

representations which describe one dimensional compressible gas flow

based on the conservation of mass, momentumand entropy equations. The
 - 60 -

accuracy of the package has been proven over a number of years and

consequently would provide an acceptable means of comparison.

The following tests were used as a 'bench mark' to assess the

validity of the complete AGI mathematical representation. The parameter

values of the P+I controllers were selected to be those listed in Table 2.2.

Test (i) A 35 Kn/mz (5 psi) change in the interstage volume pressure

set point (SP1) at a steady state load flow of 71 m3/sec
(9 Msft3/hr), Figure 2.22.

Test (ii) A 35 Kn/m2 (5 psi) change in pipeline inlet pressure set

point (SP2) at a steady state flow of 71 m3/sec (9 Msft3/hr),

Figure 2.23.

Inspection of Figures 2.22 and 2.23 confirms the accuracy of the

proposed mathematical representation. Further evidence to support the

validity of the suggested modelling techniques was obtained through work

that has been carried out to provide controller settings for telemetry

control schemes, (2.22) Appendix 11.2. Single valve/pipeline models were

developed to represent various regional supply systems and tuned using

the autotuning technique of chapter 3, the resulting settings derived

from the simulations requiring only minimal adjustment when they were

implemented on the actual system.

The analysis of any data arising from plant testing that
becomes available in the future will provide more information on the

validity and accuracy of the presented modelling techniques; this

objective is a necessary extension to the work presented. However, all

the above simulations require an appropriate set of controller settings

if the results are to make any sense at all. The following chapter

addresses this problem by developing a closed loop tuning strategy based

upon a technique currently in use, whose output amplitude can be controlled

in order to avoid the violation of any plant safety limits.

 ressure 61
- -
psi Kn/m2
756
5210

754 5200

5190

752

5180

750
50 100 time (secs)
Pressure a. Interstage volume pressure - Pu

psi Kn/m2

500.02

3447.6

500.01

500.00 4 3447.5

(secs
b. Pipeline inlet pressure - PIN
Fig. 2.22: Model validation results for the AGI station; time records
(1 of 2) illustrating the response of the AGI simulations to a 35 Kn/m2
(5 psi) change in interstage volume pressures set point
 -6Z-

% Open

42

40

38

36

34

32
(secs;
c. Valve stem displacements - X1, X2

Fig. 2.22: Model validation results for the AGI station; time
(2 of 2) records illustrating the response of the AGI simulations
to a 35 Kn/m2 (5 psi) change in interstage volume
pressures set point
 Pressure - b3 -

psi Kn/M2

5200
750

5000

700
4800

4600

650
(secs)

a. Interstage volume pressure - Pu

Pressure
psi Kn/m2
3490
506

3480

504

3470

502
3460

1 3450
500
50 100 time (secs
b. Pipeline inlet pressure - PIN

Fig. 2.23: Model validation results for the AGI station; time records
(1 of 2) illustrating the response of the AGt simulations to a 35 Kn/mz
(5 psi) change in pipeline inlet pressures set point
 - 64 -

`,ä Open

50

40

30
50 100 time (secs)

c. Valve stem displacements - X1, X2

Fig. 2.23: Model validation results for the AGI station;

(2 of 2) time records illustrating the response of the
AGI simulations to a 35 Kn/m2 (5 psi) change
in pipeline inlet pressures set point
 - 65 -

References (Chapter 2)

2.1 Arden, W. J. B. "GASCONTROL -A review of past, present and

future systems". North of England Gas Association, ERS
Killingworth, 1988.

2.2 "Fisher throttling ball valves - Design V25". GEC- Elliott

Control Valves Ltd, Airport Works, Rochester, Kent. Bulletin
E-103-GB, 1971.

2.3 "The Jetstream regulator". HEECOInternational Ltd, Caxton

Way, Thetford, Norfolk.

2.4 Moore, R. L. "The use and misuse of pressure regulators".

Instrumentation Technology, March 1969.

2.5 "Recommendedvoluntary standard formulas for sizing control

valves". Fluid Controls Institute, Inc., FCI 62-1, May 1962.

2.6 Brockett, G. F. "Correlation of valve sizing methods". ISA

Conference, Newark, New Jersey, April 1952.

2.7 Turnquist, R. 0. "Comparing gas flow formulas for control

valve sizing". ISA Journal, June 1961.

2.8 Kay, J. 11. "Introduction to fluid mechanics and heat transfer".

Cambridge University Press, 1968.

2.9 Buresh, J. F., Schuder, C. B. "Development of a universal gas

sizing equation for control valves". Fisher Governor Company,
TM-15, Marshalltown, Ioha.

2.10 Fletcher, I. "Axial flow regulator -A study". Sunderland

Polytechnic/ERS Killingworth Internal 1987, Report No. EEC/IF/
1/87.

2.11 Tsai, D. 11., Cassidy, E. C. "Dynamic behaviour of a simple

pneumatic pressure regulator". Trans. ASME- Basic Eng.,
Vol. 83,1961.

2.12 Dustin, M. 0. "Analog computer study of design parameter

effects on the stability of a direct acting gas pressure
regulator". NASA Technical Note TN D6267, March 1971.

2.13 Lee, V1. F. Z., Bonner, J. A., Leonard, R. G. "Dynamic analysis

and simulation of a gas regulator". 1971 Symposiumon flow
control - 3, System Design & Application, Paper No. 3-3-100.

2.14 Tatnell, M. L., Redpath, A. "Dynamic modellind of a gas

pressure regulator for low pressure service". Trans.
Measurements and Control, Vol. 11, April 1978.

2.15 Stoves, D., Thompson, B., Graham, P. "Simulation in the design

of gas pressure control equipment". 1983 International Gas
Research Conference.

2.16 Wylie, E. B., Streeter, V. L., Stover, M. A. "Unsteady state

natural gas calculations in complex pipe systems". IEEE Trans.
in Systems, Man. & Cybernetics, Vol. 13, No. 4, January 1974.
 - 66 -

2.17 Kralik, J., Stiegler, P., Vostry, Z., Zavorka, J. "Modelling

the dynamics of flow in gas pipelines". IEEE Trans. in
Systems, Man. & Cybernetics, Vol. 14, No. 4, July 1984.

2.18 Andersen, T. R., Farsoe, H. F. "Estimation of the pressure

in a natural gas pipeline". ACI 83 Applied Control &
Identification Proceedings, IASTEDSymposium, 1983, Vol. 1.

2.19 Rachford Jr, H. H., Dupont, T. "A fast, high accuracy model
by
for transient flow in gas pipeline systems variational
methods". Society of Petroleum Engineers of AMIE, Paper No.
SPE4005,1972.

2.20 Isermann, R. "Process fault detection based on modelling and

methods -a survey". Automatica, Vol. 20, No. 4,
estimation
1984.

2.21 Lakshminarayanan, P. A., et al. "A finite differencing scheme

for unsteady pipe flows". Int. Journal of Mechanical Sciences,
Vol. 21, No. 9,1979.

2.22 Cox, C. S., Arden, W. J. B., Fletcher, I. "Investigation into

the remote pressure control of gas transmission pipelines".
Int. ASMEConference on Modelling & Simulation, Sorrento, 1986.

2.23 Orange, K. "Flow formulae for gas distribution and trans-

mission". Yorkshire Junior Gas Association, ERS Library,
Report No. P/116.3.

2.24 American Gas Association. "Steady flows in gas pipelines".

605 Third Avenue, New York, July 1985.

2.25 Martin, E. N. "Control valve movement limitations". Journal -

Inst. Measurement & Control, Vol. 8, September 1975.

2.26 Ziegler, J. G., Nichols, N. B. "Optimum settings for automatic

controllers". Trans. ASME, Vol. 64,1942.

2.27 Brown, G. L., Todd, A. "Stability analysis of a two stage

pressure reduction system". Sunderland Polytechnic Internal
Report, 1970.

2.28 Wilkinson, R. B., Thompson, B. "Computer program for a non-

linear model of transient flow in a gas pipe for the simulation
package "BSIMS". ERS Internal Report No. ERS R. 2355.
 -67-

Nomenclature (Chapter 2)
Modelling of a Pressure Reduction System

A Area of pipeline (m)

C Pipeline volume coefficient (m3/Kn/m2)

Cg Critical flow coefficient

Cv Flow capacity coefficient

C2 Specific heat correction factor

dP Valve differential pressure (Kn/m2)

D Diameter of pipeline (m)
Ex Pipeline elevation factor

f Pipeline friction factor

G Gas density factor

Gc(s) Controller transfer function

h Height of pipeline (m)

i Discrete section label

K Actuator scaling constant

K' Flow conversion constant

K* Pressurised volume constant

Kc Proportional gain of controller

Ku Critical gain of process
K(l), K(2)
etc. Pipeline restriction constants
L Length of pipeline (m)

M. Mass of gas (kg)

M Controller output signal

A Ptass flow rate (kg/s)

n Number of pipeline stages

P Pressure (Kn/m2)

Pb Base pressure (Kn/m2)

Pic Initial condition pressure (Kn/m2)
 - 68 -

P1 Valve upstream. pressure (Kn/m2)

P2 Valve downstream pressure (Kn/m2)

P(l)'P(2)
etc. Pipeline section pressures (Kn/m2)

Q Volumetric flow rate (m3/s)

Qc Critical flow rate (m3/s)

R Universal gas sizing constant (Nm/kg/K)

Re Reynold's number-

s Laplace operator

t Time (s)

T Temperature (R)

Tavg Average temperature (R)

Tb Base temperature (R)

-Ti Integral gain of controller (s)

U Gas velocity (m/s)

V Pipeline volume (m3)

X% Normalised valve stem displacement

Zavg Average compressibility factor

p Density of gas (kg/m3)

AGI Simulation

Gcl(s) Upstream controller transfer function

GC2(s) Downstream controller transfer function

F1, Upstream controller output

M2 Downstream controller output

Pin Pipeline inlet pressure (Kn/m2)

Ps Station inlet pressure (Kn/m2)

Ps* Transducer range constant

Pu Interstage volume pressure (Kn/m2)

 - 69 -

Qa 2nd stage pipeline inlet flow (m3/s)

Qb 3rd stage pipeline inlet flow (m3/s)

Qc 4th stage pipeline inlet flow (m3/s)

Qload Pipeline outlet flow (m3/s)

Q1 Upstream valve throughput (m3/s)

Q2 Downstream valve throughput (m3/s)

R1 Upstream controller's signal conditioned output

R2 Downstream controller's signal conditioned output
SP1 Upstream controller set point
SP2 Downstream controller set point

X1 Upstream valve's normalised stem displacement

X2 Downstream valve's normalised stem displacement
 - 70 -

CHAPTER
3

AN INVESTIGATIONINTO THE EFFECTIVENESS

OF

AGI CONTROL
THE PRESENT PHILOSOPHY
 - 71 -

3.1 Introduction

As mentioned earlier, modern developments in instrumentation

hardware and control theory make new operational strategies feasible for

the control of natural gas. Not unexpectedly, however, the first attempts

at an AGI control strategy were based around familiar procedures

particularly the ubiquitous proportional plus integral (PI) structure.

The main reasons for this are their successful use in a wide range of

applications and the fact that ground rules for efficient operation are

well understood and documented, while training is reasonably straight-

forward.

The chosen PI strategy has been successfully configured in a

Negretti Automation IIPC84 industrial process controller. Arden (3.1)

details how this is achieved using a range of standard control blocks

interconnected by software link. A high level interpretative language

SEUZTROL,provides a user friendly system to perform these links and

configure the interface channels. It should be made clear that the

programmable system can provide up to eight control loops and 16

simultaneous sequence programmes per unit. Consequently the MPC84can

also, as well as providing the primary regulatory functions, be employed

as an extensive data logging facility, overseen by alarm condition

monitoring routines configured by the user to enhance/replace present

plant safety mechanisms.

A problem which existed with both the original pneumatic

systems and their microprocessor based replacements is that of just how

much proportional and integral action is required, a process referred to

in the literature as the 'tuning' of the controller. Over the years a

variety of methods have been suggested (3.2,3.3 and 3.4), the British

Gas approach being basically trial and error based on operator experience
due to the significant problems that arise practically. Two methods which
 - 72 -

have retained their popularity are the open loop process reaction curve

method and the closed loop ultimate sensitivity method, both proposed

by Ziegler and Nichols in 1942.

The process reaction method requires one first to measure the

'step response' of the open loop system. For this to be effective the

system must be type 0, that is, contain no open loop pure integrators,

an approach that is impracticable in gas pipeline systems since quiescent

conditions are only possible when inflow matches outflow, and this

situation is destroyed by a change In either the load/supply conditions

or the position of any regulatory valve within the system. It follows

that any proposed tuning strategy must be carried out with the system in

closed loop.
The chapter commenceswith. a comparative study between two

closed loop tuning methods, namely, the Ziegler/Nichols ultimate method

and an automated version of the ultimate method proposed by Astrom and

Hagglund (3.5 and 3.6) and concludes with the application of the latter

to the AGI model developed in chapter 2, after the resulting software had

been exhaustively tested within both laboratory and industrial environ-

ments.

3.2 Ziegler-Nichols Ultimate Ptethod

The method requires the determination of the ultimate gain, Ku.

This is the value of gain (for a controller with only proportional mode

of operation) which causes the closed loop controlled variable to cycle

continuously with fixed amplitude. This 'marginally stable' situation

implies that the Nyquist curve of the open loop frequency response must

pass through the critical -1+j0 point on the Argand diagram.

Ku 3.1
UT' .....
 -73-

The period of this marginally stable oscillation, is called

the ultimate period, Pu. In the original Ziegler-Nichols scheme, Ku and

Pu were determined in the following way: tune out any reset or derivative

action from the controller, leaving only the proportional mode. Maintain

the controller on automatic, i. e. leave the loop closed. With the gain of

the proportional mode set to some low arbitrary value impose an upset on

the process (move the set point for a few seconds then return it to the

original value) and observe the responses. If the output response grows,

reduce the controller gain; if the response damps out increase the controller

gain. Continue in this way until sustained oscillations of constant

amplitude are encountered. Finally, the P, PI and PID controller parameters

can then be obtained by using empirical formulae which rely on Ku and Pu.

The ultimate method empirical results are presented in Figure 3.1.

Someof the constraints that are commonly encountered when

implementing the ultimate method are: -

(i) The process transfer function must be at least third order and

contain no more than one free integrator.

(ii) For a system with long time constants the technique is a very

slow process.
(iii) If the critical gain is exceeded then the system becomes unstable.
(iv) It is difficult to automate the procedure, and perform it in such

a way that the amplitude of oscillation is kept under control.

(v) The magnitude of the oscillations is dependent upon the plant

gain as well as the conditions supported by the plant when the

test is initiated, i. e. the final amplitude of oscillation is not

known before the test.

The inability of the user to control the output amplitude

severely limits the application of the ultimate method in industry due to

the possibility of it damaging expensive plant and/or infringing system

 - 74-

KU G(s)

a. "Ultimate" method

Kc Ti Td

p 0.50 Ku - -
PI 0.45 Ku Pu/1.2
-
0.60 K11 Pu/2.0 Pu/8.0
PIO

b. Controller tuning formulae

Fig. 3.1: Ziegler-Nichols ultimate tuning method;

with associated look up tables for P, PI
& PID controller parameters
 - 75 -

safety considerations. Thus another method which can provide automatic

determination of Ku and PU is proposed.

3.3 The Modified Ultimate Method

Because of the possible dangers indicated in the previous section,

a means of controlling the system's output during the ultimate tuning

exercise is desirable. The modification to the method proposed by Astrom

and Hagglund (3.5 and 3.6) is to replace the variable gain element with an
ideal relay. The purpose of this relay is twofold, firstly the possibility

of the system going unstable is prevented since the output will be forced

to limit cycle (3.7) when:-

G(s) = -- 3.2
.....

where the amplitude dependent function N(A) is a quasilinear complex gain

called a describing function, evaluated on the assumption that the input

to the nonlinearity is a sinusoid of known amplitude. In practice, this

would only be true if the plant's frequency response behaved as a low pass

filter making the affect of all frequencies other than the fundamental

Thus for an ideal relay (3.7): -

negligible.
N(A) =0 3.3
.....

Secondly, it is evident from equation 3.3 and Figure 3.2 that

the amplitude of the relay characteristic (D) can be used to limit the

magnitude of the limit cycle oscillations (A), providing the means of

manipulating the final system output so as to satisfy any plant safety

factors. Furthermore, because the ideal relay's describing function

[N(A)B is purely real then the resulting limit cycle must occur when the

plant's frequency response intersects the negative real axis. Under these

conditions the ultimate period (Pu) may be read directly from the output

response. To ensure that this test is comparable with the ultimate method
 - 76 -

(i) Ultimate method (ii) Relay method

1cba
- R(AT 7

Nyquist plots and associated output time

responses

a. I<1
G(jU-Uu)

b. G(Jl1u)I=1

IG(JWu)I >1
C.

Fig. 3.2: Comparison of Ziegler-Nichols Ultimate method and

the relay tuning technique proposed by Astrom &
Hagglund; time responses illustrating the extra
control achievable over the amplitude of the
resulting output signal
 - 77 -

a means of extrapolating the ultimate gain (Ku) from the closed loop

relay test data is required. Inspection of equations 3.1 and 3.2 reveals: -
Ku = rf(A) 3.4
.....
Thus by determining the factor N(A) for the particular relay employed,

then we can directly find the ultimate gain (Ku). In practice, this

requires a peak detection circuit in order to identify the limit cycle

amplitude (A).
The system in Figure 3.3 illustrates the hardware/software

requirements necessary to build the Ziegler-Nichols auto-tuner. The

inclusion of the plant's operating limits into the tuning scheme will

allow the systen to be programmed to adjust the relay's characteristics

accordingly, thus providing an in built safety mechanism. Not only is

this technique less complex than PID"self-tuners (3.8 and 3.9) it also

requires significantly less programing and storage to operate, thus

lending itself readily to small microprocessor applications.

To illustrate how the system works, consider the application

of a software version of the tuning scheme presented in Figure 3.3 to a

plant with the transfer function: -

1
G(s) = Is 3.5
+ 1)(2s + 1)(3s +-TT .....

Table 3.1 contains the results of the Ziegler-Plichols auto-

tuning method using an ideal relay with an amplitude of 10. For

comparison purposes the theoretical values of the ultimate gain and

period determined from the Routh array, are also displayed.

Ku Pu

Theoretical 10.00 6.28

Experimental 9.690 6.39

Table 3.1 Comparison between the Ziegler-Nichols

ultimate and auto-tuner techniques
- 78-

r,

C-

U
()
4)

td

a)
S
E
0
S-
4-)

c)

(/1

4-3
a)

0
4)

N
N
a)
U
4J
C

4)
U
a
N

a)
i
t0
s
4-
0

a)
tO

ri
M

U-
 - 79 -

The P, PI and PID controller parameters recorded in Table 3.2

were determined from the above results and the Ziegler-Nichols look up

tables (Figure 3.1). Figure 3.4 displays the responses of the above

three controllers to unit changes in set point when applied to the

process described.

Kc Ti Td

p 4.845 - -
PI 4.361 5.325 -
PID 5.814 3.195 0.799

Table 3.2 Controller settings evaluated using the Ziegler-

Nichols look up tables and the results of Table 3.1

It should be noted that the responses are quite oscillatory

in nature, as is generally the case with controllers designed using

this particular tuning technique since the formulae were specifically

designed for disturbance rejection. Various methods, both intuitive

and theoretical, of detuning the Ziegler-Nichols designed controllers

have been implemented throughout the years in an attempt to reduce this

oscillatory behaviour to set point changes. One of the most successful

techniques (3.10) detunes the system based upon the empirical choice of

a maxinun limit, placed upon the closed loop gain of the Ziegler-Nichols

tuned proportional plus integral controller/process. A detuning factor

(F) is varied until the maximumclosed loop gain is as near as possible

to the limit, using a suggested starting value for single input single

output systems of F=2, thus: -

Kc = Kcz/n/F

Ti = Tiz/n. F

where Kcz/n and Tiz/n are the PI controller settings obtained from the

Ziegler-Nichols look up tables. The combination of this particular

 - 80 -

Scaled
units

--- Set point

1.5
-" - Output

1"
11 11

V
0. s.

time
0 10 20 30 (seconds)

a. Proportional control

Scaled
units

1.5

t. o

0.5
1 t"
iI
time
00 10 20 30
..
(seconds)

b. Proportional and integral control

Fig. 3.4: Application of the Ziegler-Nichols technique of tuning

(1 of 2) P, PI and PID controllers via the ultimate method; time
records illustrate the response of the compensated
1/(s+t)(2s+1)(3s+1),
system, G(s) = to a unit step input
 - 81 -

Scaled
units
Set point
1.5
Output

I'

1.0

I
0.5
I

0 time
10 20 30 (seconds)

c. Proportional, integral and derivative control

Fig. 3.4: Application of the Ziegler-Nichols technique of tuning

(2 of 2) P, PI and PID controllers via the ultimate method; time
records illustrate the response of the compensated
G(s) 1/(s+1)(2s+1)(3s+1), to a unit step input
system, =
 - 82 -

method with the modified ultimate method presented above will provide a

tuning method whose application is entirely in the closed loop.

Unfortunately because of its adaptive nature this method can be very time

consuming and in practice the initial F=2 estimate is frequently used.

The following section investigates a closed loop design method which will

determine the controller parameters based upon system gain or phase margin

specifications directly from the ultimate values.

3.4 Phase Margin Compensation

Since the ultimate values only provide information of the

plant's gain and frequency when it intersects the negative real axis,

then phase lead compensation must be added if any specified phase margin

is to be accomplished. It is obvious that any manipulation of the open

loop system's phase response is impossible using proportional control;

indeed only the use of a three term controller will provide phase lead

compensation directly. However, a slight adaption to the previously

discussed tuning method provides us with a means of obtaining the

required phase margin when operating under PI control.

3.4.1 PI Control with Phase Margin Compensation

In order to provide phase advance from a PI controller we must

allow for the inclusion of the integral action in the initial tuning

exercise, as is illustrated in Figure 3.5. By determining the ultimate

gain and period for the plant plus integrator system we will leave the

controller's remaining zero to dictate the system's phase margin (gym).

Therefore the following equations must be satisfied if a specific phase

margin is to be found: -
Om= argil + jW. Ti) .....
3.6
 - 83 -

D
S G(s)
o

a. Block diagram

1 1/Ku

N(A) V,,
u
ým

\
\
Kc(1 + sT1) designed
to establish
/
compensated frequency i
response at this point
when w=w
u

G(s)
contour
b. Nyquist diagram

Fig. 3.5: Proportional plus integral controller design using

phase margin specifications
Note: Inclusion of integrator with plant is necessary
to achieve the specification
 - 84 -

IGC(jW)I-IG(jW)l 3.7
=1.....
Since only limited information about the plant plus integrator

system's behaviour is made available from the tuning exercise, Ku and Wu,

then it must form the basis for the following analysis. By designing the

controller at the ultimate frequency then equation 3.6 provides us with

the means of calculating the integrator's time constant (Ti) for any

specified phase margin, i. e.

Ti =- tan 4'rn
3.8
.....
Pu
_ tan

In order to evaluate the plant's gain alone, at the ultimate

frequency, then we must allow for the gain of the integrator (1/Wu).

Thus
w 3.9
lGUWu)I= .....

The gain of the PI controller's numerator at Wu is: -

i

l ='cll +
ýGc(iWu) ..... 3.10

Consider the triangle (equation 3.6): -

Wu.Ti

11

Then
ICý(j4Ju)I 3.11
sinKC 1 .....

c
 - 85 -

Now by placing equations 3.9 and 3.11 into equation 3.7, and

rearranging, we can evaluate the controller gain Kc: -

Ku. sin(4m)
Kc = 3.12
Wu .....

Therefore equations 3.8 and 3.12 allow the controller parameters

which fulfil the specified phase margin requirements to be calculated,

using information gathered from the modified Ziegler-Nichols auto tuner

shown in Figure 3.5b.

3.4.2 PID Control with Phase Margin Compensation

Consider the open loop transfer function of the process G(s)

and a PID controller. At the ultimate frequency the phase will be -7,

giving a system argument of

arg11 + jWu. Td + -ri1 -n 3.13

.....

Now for any specified phase margin the system phase must be equal to

Om- ii, when the open loop gain is unity, therefore

6Ju.Td + = tan(gm) 3.14

WuIT .....

obviously there is an infinite set of solutions to equation 3.14, however

by using the quarter decay ratio suggested by Ziegler and Nichols (3.2)

Ti = 4Td 3.15
.....
Then a unique solution exists; after some manipulation we find that

taným + (1 + tan2ým)i
Td = 3.16
'u .....

To achieve the phase margin at the ultimate frequency then the open loop

system gain must be unity; implementing equation 3.2 gives

`1
Kc +j Wu. Td +-Jý-] = Ku .....
3.17

Ater placing equation 3.14 into the above and applying standard
 - 86 -

trigonometric theory we arrive at

Ku
3.18
cos 4m .....

Thus by using the arrangement discussed in section 3.3 and

illustrated in Figure 3.2 to determine the ultimate values then equations

3.15,3.16 and 3.18 will provide the necessary controller parameter values

needed to obtain the desired phase margin. Hagglund (3.11) takes this

work one step further by producing a look up table for tuning three term

controllers depending upon phase margin.

Figure 3.6 illustrates the results obtained when the phase

margin compensation methods (Table 3.3) described above were applied to

the plant detailed in section 3.3. A comparison of these responses with

those recorded in Figure 3.4 indicates the possible improvements and

greater flexibility that the new method provides, compared to the Ziegler-

Nichols look up tables.

Kc Ti Td

PI 1.143 3.386 -
PID 6.852 4.911 1.223

Table 3.3 The PI and PID controller parameters evaluated

using a phase margin of 45 degrees

3.5 Industrial Applications of the Modified Ultimate Method

The following section describes the tuning procedure adopted

for two types of gas transmission systems, based on the phase margin

compensation design technique developed above.

The control of pressure in natural gas transmission pipelines

is usually constrained to local pneumatic control, where the sensed

pressure is immediately downstream from the station. Recent increases in

demand have resulted in greater frictional pressure losses within the

 - 87 -

Scaled
units
Set point
1.5
"- Output

/
0.5

0f -, ---ý- time
0 20 30 (seconds)
10
a. Proportional and integral control

Scaled
units

1.5

1.0

0.5

- time
0 10 20 30 (seconds)

b. Proportional, integral and derivative control

Fig. 3.6: Application of the phase margin compensation technique of

tuning PI and PID controllers via the ultimate method; time
records illustrate the response of the compensated system,
G(s) = 1/(s+1)(2s+1)(3s+1), to a unit step input, ým = 450
 - 88 -

pipelines. In particular, the regional medium pressure systems which

employ this method could, in some cases, experience a failure to supply

condition. Elevation of the set point pressures cannot be implemented

to alleviate these problems, because of the risk of exceeding the maximum

design limits of the pipeline.

The first application considers one solution to this problem,

thus enabling 'optimisation' of the storage/flow capacity of the pipeline.

The microprocessor control strategy involves maintaining the pressure at a

remote point in the pipeline while ensuring that the maximumpressure

limits are not exceeded at all times. The second example involves a flow

control loop which is the usual mode of control employed at National

Transmission Offtakes where gas is supplied into the twelve regional

supergrids. This latter strategy is a replacement to the existing pneumatic

incremental control schene and enables improved station efficiency.

Both applications require the tuning of PI based control loops and

clearly demonstrate the usefulness of the modified ultimate method in

avoiding potential damage to pipelines and ancillary equipment which could

result if other well established tuning strategies were employed.

3.5.1 Remote Boundary Pressure Control

These particular types of control systems are becoming increas-

ingly important as the gas industry attempts to optimise the operational

of its network (3.12). Problem areas arise as certain

efficiency

components within the network require downrating due to ageing, stress

and demands upon the system increase.

ratings
As an example consider the Whasset/Barrow medium pressure

distribution pipeline, Figure 3.7 (3.13), which requires the control of

the remote boundary pressure at Ulverston. The only means available for

controlling this particular boundary pressure, in the face of load

- 89 -

3 aý
ouc
Lrl
( r. Q)
mE CD-
0 r, o.
+. º
CD
a) to=c'.j
Lt. > .- (1)
E
a)
U

r-0
i

cc
oa) o
(A (0
S- +-º G,1
G)4- i
> 4-
0 u
N
Q.
'
a

0
e
a)
i

0
i
i

a)
N
N

aý
N
O
a
O
L
a-

cJ
-C
a..i
n
4-
G) O
L
i-ý U
N -I

Q) f0
E
"n
v
-s
lp U
v V)

rn
LL.
 - 90 -

disturbances, is a single regulator situated approximately 20 miles

upstream. Therefore, telemetry links are required to feedback the

necessary information regarding the state of the controlled variable.

The control of this system is further complicated by the inclusion of a

line valve situated between the regulator and the boundary whose maximum

operating pressure is rated at less than that of the pipeline itself.

The proposal was to design a remote boundary control scheme

capable of replacing the existing pneumatic system and minimise capital

expenditure costs through system uprating by satisfying the following

specification:
(i) Pipeline rating (at inlet) of 350 psiG, at the line valve

275 psiG.

(ii) Maximumload flow required at Barrow of 1.3 Msft3/hr with a

possible 50% load rejection.

(iii) Maximuminstantaneous (worst case) flow change at the

Ulverston boundary offtake of ± 0.5 flsft3/hr.

(iv) Operate with the Ulverston set point as high as possible to

linepack storage (ideally

guarantee supply and optimise
250 psiG).

The design of the control scheme required a working model of

the system, since: -

a) The system, as yet, was not operational.

b) Testing at the proposed worst case scenarios was both

impracticable and exceedingly dangerous.

c) Online tuning of a practical system can take up to 6 weeks at

present, due to the long time constants involved and the

repeatability problems that arise with load disturbances.

Thus, a mathematical representation of the system was

constructed using the 6" Jetstream regulator/pipeline subroutines

 - 91 -

developed for the AGI station in Chapter 2. Setting up the model with
its maximumpredicted load flow and selecting a boundary set point to

satisfy point (iv) above resulted in the following static conditions for

the remote boundary control system simulation: -

Supply pressure Ps = 463 psiG

Boundary pressure Pb = 240 psiG

Flow rate Q=1.3 Msft3/hr

at which the PI controller could be tuned using the relay/integrator

arrangement necessary to permit phase margin compensation.

Note: The proportional element of the cascade controller arrangement

detailed in Figure 3.7 is in fact a pneumatic pilot/convertor system

required to provide valve actuation. The proportional gain Kp can be

varied, in practice, by using different restrictor settings within the

pneumatic hardware, but for the purposes of this work was considered to

be set to unity.

Selecting a relay height of 1E-4 then the system limit cycled,

Figure 3.8a, at the largest permissible amplitude without violating the

pipeline pressure limits, resulting in: -

Ku = 4.1806E-3 After making allowance for

Pu = 5460.1 the transducer's range

which when tuned for a 60 degrees phase margin provided the controller

parameters
Kc = 3.15
Ti = 1505

The performance of this PI controller was evaluated over the

working range of the system. Figure 3.8b shows the response of the

process to the worst case scenario, namely an instantaneous load demand

of 0.5 1jfst3/hr (approximately a 40% increase) direct from the boundary

at Ulverston when initially operating at the quiescent levels defined

 - 9'L -

Pressure
(psiG)

40

3501-.. _........,..,.. _.. _.. -.. _... _.. _.. _... _.. _.. _

30

. ýýý '/ \r/ `/

ý' ý"

200 time
(hours)

a. Modified ultimate method tuning

Pressure Inlet pressure (PIN)

(psiG)
Line valve pressure (PLV)
Boundary pressure (Pb)
400 Pressure limits
_,. _

350. _.. ý .... ... ... _.. _.. _... _.. ,..... _

300.

250

200 time
01234 (hours)

b. +0.5 Msft3/hr step in flow at Ulverston

Fig. 3.8: Nhasset/Barrow remote pressure boundary control scheme;

time responses illustrate the tuning of the system and
its performance when subjected to its worst case load
disturbance
 - 93 -

above. Examination of the results clearly shows that under these conditions

the pipeline inlet pressure limit would be infringed. Two solutions were

recommended. The first was simply to lower the set point. The second was

to redesign the controller to satisfy a phase margin greater than 60 degrees.

The latter solution was found to be unacceptable because of the consequent

reduction in the speed of response.

The application of controller settings designed using the modified

ultimate method to the actual system, after construction, has provided both

good performance and practical validation of the modelling techniques

presented in chapter 2. Deferring capital expenditure upon system reinforce-

ment requirements of ageing pipelines, typically costing in excess of

£200,000, whereas the development of the more accurate and flexible micro-

processor telemetry control scheme costs approximately £25,000 -a 90%

saving.

3.5.2 Flow Control Station

The following study describes the application of the relay tuning

strategy within an industrial environment. The process considered was a

flow control system sited at the Bishop Auckland test facility and consisted

of a 4" Fisher Jetstream regulator fed via a Negretti MPC84programmable

controller and electro-pneumatic interface. Normally we would simply replace

the PI control algorithm with the relay-integrator network, however for

system security reasons reprogramming of the Negretti controller wasn't

feasible. For this reason the Neqretti controller was replaced by a BBC

microcomputer, Tecquipment 8-bit interface and matching voltage to current

convertor. The system arrangement, illustrated in Figure 3.9, required the

following inputs in order to operate:

i) Present set point voltage, corresponding to 1-5V signal from the

flow measurement devices differential pressure output.

 - 94 -

v
0
L
Q.
r-0
a,
0
S-
0
U 'C7
41
c
L O
41 vo U
c 0 t
b
LL. O
C.
4-
O

-I--
O
4J
5-
0
r- 4-3
(0 (0 "r
ý, C
"r C U
N N C)

C rty
-aI--
C . ýC
=U

+-) (1) a
C U (cc.
u E I- b
U Cs
a a i S.. N
0 Q "r- G) "r
N co
C' C L
a1 "r
I -"
a
r""r +ý
S1-- ýý dr-
..
Q. "r
tC U
b
04-
4J
.. >0 (A
Y S- C)
(4)
1-4
N
NN
GJ tcS

N
C. ý
N
i N "'
G)
4-3 C "r
D N S-
E
N
0 Ld
O U "r .C
0 ý4-)
S- Q'
u (4-)
S. r0
E
G1 b
L) S_ CJ
co (+J
m 3 co
'O
i-
"r
2N
+ý

of
CV)

C,
1-I-
 - 95 -

ii) Present current to pneumatic interface input, to provide the

integrator's initial conditions at the existing steady state.

iii) The relay amplitude for control of the limit cycle amplitude.
For a relay amplitude (D) of 0.02, a corresponding ±2 mA
(approximate) swing in plant input current about the set point value was

initiated, culminating in the results presented at Figure 3.10 for two

different operating conditions. From these results the maximumand

minimum values of output flow differential pressure (volts) and limit

cycle period (seconds) were recorded. However, the translation of these

values into the system's critical gain and period was not straightforward,

as two differences between the Negretti/BBC system had to be accounted

for. Firstly the signal convertors employed had different gains, thus

naking it necessary to evaluate a pseudo relay gain (D') for conversion

from one system to another, i. e.

0' = D. Ki. Kcon 3.19

.....
where Ki is the integrator gain and

Kcon is the relative convertor gains, determined via: -

BBCconvertor gains
Kcon - 3.19a
egre ti convertor gains """".

For the particular system/programme employe4 in this example, then: -

D' = 53.336D 3.19b
.....
Thus calculation of the Negretti system's critical gain is possible via

the modified version of the ideal relay describing function: -

4ir.
Ku = AD 3.20
.....

Secondly, the Negretti also contained software designed to

convert the flow transducer's differential pressure measurements to actual

flow, via equation 3.21, for insertion into the PI control algorithm.
Thus allowance for this nonlinear gain must also be made when determining

the peak limit cycle amplitude (A) for the Negretti PIPC84from the BBC
 - 96 -

NC
EO
"1" U
N

"w
s. E
O aO
lt:r
C, N 4-
4-. HN
N O
O r- 0
M C., -
S. S-
4-) (
>
11 O
U i">
O
N 3
Cl- O-0
N rW
4- r
"r
r- fei
O" c0 4-)
"-- "r
4) -im vo
r
O CA CA
O
-o r
T0 GU
0- O CD "r
G. . LO U
'C M CV C a-ý
tß+ß
. r-
0E
'
{. "r
r-

a)
Cr 4-J
"r f0
4-'
LN
U
au'0a
4J

NC 0) a)
EO
c +ý
N
"r
"r U
+-3 C)
N 4
u "-,
a
nsaj
r- 4-,
aj(1)

O N
Id- - tO
44

4- 4- "r
O
O '^ N
M
C'0
LC) O S- CA
"r 04-
II 4J UO
(y "r
O U S" O
N "r 0.
C)
N G"r-
G)
cl: 43 N

CD
U,
M
0
0
O
,.', oL C)
Q"
CL m NN "-
 - 97 -

data. Where A is the peak value of equation 3.21

Q= 14.874E - 3(814.5 * 12.5 * dPq)l 3.21

.....
and dPq is the flow measurement device's differential pressure trans-

ducer output in volts, and Q is measured in Msft3/hr.

The modified ultimate test was performed at two different

operating points, see Figure 3.10, namely

i) System set point of 11.8 m3/sec (1.5 Msft3/hr)

ii) System set point of 13.8 m3/sec (1.75 Msft3/hr)

The employment of the above equations upon the test data

allowed the following table to be constructed; an inspection of the

ultimate values at the different operating points clearly indicates that,

as anticipated, the system behaves in a nonlinear fashion, providing

valuable information regarding the sensitivity of the system throughout

its working range.
Amin Amax Pu Ku

SP = 1.5 1.3622 1.7916 9.23 6.34E-3

SP = 1.75 1.3755 2.0687 8.75 3.92E-3

Table 3.4 Ultimate values for the Negretti controller

For comparison purposes the controller settings presently

employed are presented in Table 3.5, along with those determined from

the relay tuning tests designed to give a 45 degree phase margin.

Actual SP =1.5 SP=1.75

Kc 9.90E-3 6.59E-3 3.86E-3

Ti 3.300 2.158 1.939

Table 3.5 Comparison of Negretti PI controller parameters

Because the system is nonlinear (and a different tuning method

had been employed), one would not expect an exact reproduction of the
 - 98 -

British Gas controller parameters. However, the fact that they are of

the same order supports the validity of the test results, a point

confirmed by the site engineer. The tuned values suggested that the

British Gas values might have been designed for a 45 degrees phase

margin at a system set point of around 1.25 to 1.5 P1sft3/hr.

The application of the modified ultimate method within the

Negretti software would have considerably simplified and improved the

performance of the tuning exercise, by avoiding the work necessary to

ensure data compatibility between the two systems. The autotuner

strategy has been shown to decrease the commissioning time of the

station controller from up to 6 weeks to that of a few hours. Another

advantage is that because of its high degree of integrity this method

can be safely reinitiated at any time, remaining relatively unaffected

by the slow load disturbance transients that occur in practice.

3.6 A Relay Modification to Provide Noise Rejection

One problem that can arise when applying the modified ultimate

method to real systems is relay chatter, the indeterminate switching of

the relay being caused by excessive noise when operating close to the

system set point. By replacing the ideal relay by a relay with

hysteresis then a certain amount of noise rejection is possible,
depending upon the hysteresis width (E) chosen. Although the benefit

of easing the effects of noise at the relay switching zone is obtained,

a degradation of the ultimate process values results since the point at
which the limit cycle exists no longer lies upon the negative real axis.
This is demonstrated by the following analysis, Figure 3.11, from Gelb

and Vander Velde (3.14) where the describing function representation

of a relay with hysteresis is given as

Pr(A) = [cos(p) - jsin(i)] 3.22

....
 - 99 -

a. Block diagram
Where
D- Relay magnitude
E- Hysteresis width Im

Re

irE
zu
1
T(Ä)
--- ---

Point determining
1. Frequency of limit cycle
- Wu G(s) contour
2. Amplitude of Limit cycle
- Ku

b. Nyquist diagram

Fig. 3.11: Incorporation of a relay with hysteresis during the

ultimate tuning exercise, providing some noise rejection
and helping to reduce problems associated with relay chatter
 - 100 -

rE)
where sin" TAJ 3.23
` .....

Simple algebra reveals that: -

1 (E] l
7, +3 3.24
(A) ITT 7ý MY .....

It is now clear that the term 1/N(A) is complex; however, the

imaginary term is not amplitude dependent and can therefore be selected

by a suitable choice of relay parameters, see Figure 3.11. By keeping

the relay hysteresis width small with respect to its amplitude then

equation 3.24. can be considered as being purely real, i. e. the imaginary

term is negligible, therefore approximating to the ideal relay character-

istic, thus ensuring that the previously described method of critical

stability evaluation is applicable. If operation within defined plant

safety margins is to be achieved then obviously a maximumpermissible

value of relay amplitude will exist. Under these conditions great care

must be taken in the selection of the hysteresis width if the benefits of

noise rejection are not to be outweighed by the accompanying degradation

in stability point detection, although the results would generally err on
the side of safety by designing a conservative controller. The following

example is included in order to assess this noise rejection/accuracy trade

off. Consider the third order plant previously described in Section 3.4: -

1
G(s) =s+s+s+ (equation 3.5)

Table 3.6 compares the critical values determined in section 3.4,

with a relay amplitude of 10, with those determined for various values of
hysteresis width.

% Deviation
E Ku PU in Ku

0.00 9.690 6.390 0.0

0.01 9.564 6.428 1.3
0.05 9.984 6.580 6.2
0.10 8.587 6.752 11.4
0.50 6.312 7.824 34.8
1.00 4.982 8.817 48.6

Table 3.6 Evaluation of relay hysteresis/gain trade off

 - 101 -

A further example is given in an internal report (3.15), which

gave comparable results to the above for a type 1 system, a system containing

one free integrator.

3.7 Tuning of the Twin Valve A. G. I. Model

Because the Above Ground Installation model presented in chapter 2

was of a general system layout, then no actual controller settings existed

that could be utilised in a model comparison exercise. Therefore the

modified ultimate technique presented within this section was applied to the

twin regulator system. The multivariable nature of the system presented

one problem that had not been encountered previously, namely that each

control loop affected the behaviour of its counterpart. Thus a method of

allowing information from one loop to influence the tuning of the other loop

was necessary. After some experimentation, the most successful method

found was to allow the process to reach the required steady state operating

conditions and then follow the procedure below.

i) Ensure that the second stage regulator does not change its steady

state position.
ii) Now, tune loop 1 as if the process was a single input single

output system (SISO) using the ideal relay and integrator.

iii) Set up the loop 1 controller with the parameter values derived
from step ii).
iv) Tune loop 2 as if it was also a SISO system, again using the

relay/integrator combination.

v) Set up the loop 2 controller with the parameter values derived

from step iv).

vi) If the tuned system's performance proves unsatisfactory then the

procedure can be repeated from step ii), but with the loop 2

controller present.
 - 102 -

The above procedure was applied to the A. G. I. model described

in Chapter 2 when a steady state flow of 31.3 m3/sec (5 Msft3/hr) was

being passed by the station, operating under pressure control with a

pressure cut of 6895/5171/3448 Kn/m2 (1000/750/500 psi). The resulting

critical values (table 3.7) and controller settings tuned for a 45 degree

phase margin (table 2.2) were utilised to obtain the closed loop step

responses presented in Figure 3.12.

Ku Pu

Loop 1 0.085 36.95

Loop 2 94.38 6.158

Table 3.7 Ultimate values for the A. G. I. model

presented in section 2.6

Inspection of the step responses illustrates the interaction

that is present between actuator movement and the interstage volume

pressure using the standard SISO PI control strategy, although the

performance of the downstream control loop is acceptable.

3.8 PID Control of the First Stage Cut

The oscillatory response of the interstage pressure control

loop indicates that the use of the present PI control arrangement fails
to provide'sufficient speed to enable the upstream loop to compensate
for changes in its controlled pressure arising from movements of the
downstream regulator.

The extra phase advance a three term controller (PID)

contributes would ensure that the upstream control loop is dynamically

faster without affecting stability margins.
Inspection of the linear transfer function element G11(s),
derived in section 4.5, relating the output of the loop 1 controller to
 - 103 -

U,
C)c
Eo
.'(. )
+-) C)

in c
ý "r
ON
UO
O (M
ä. c
cO
G) s
EU
"r
"r
"0 (/
>1
U
"r LC)
r
O
aN

L\
Q) C
rL
r'
O LC)
S-
CM
0
U (A

ý-+ O
O.. C)
\C
-
G. +3
1- L
O
NE u1
r-y "r N
NNO
S.
GJ Oa
+ý
43 d
OE
LOO
G1-r-
"0 +J 0
C co >
+.
V1 Q)
C Cý
O 4J tC
"r t 4-J
43 43 N
t0 i.
r" 4- G)
OO +3
EO
"r 0) "r
(A N
iC
º-ý O "r
cý a
< vº c
0) 0
O j_ r
LN +)
+) d) +- U
q- .CC
4-) ri
OO 0- 43

L +3 O
tlý +) f-+
O S- C)
"r}3 N
>N
aý
S.. tö Ot G1
7 1.3 43
.O .
N O00
N
a) Eo000
S.. CD o
L1 YN
I N
r- e
G M
"r
E
O 000 rn
Z NN
LL-
 - 104 -

the interstage volume pressure (Figure 3.13a) reveals that it is of

second order. Since the relay tuning method presented requires the

system to be at least third order then in this particular situation it

would not be applicable in its standard form. However, the inclusion

of an integrator with the transfer function element will permit the

application of the tuning method in the design of the three term

controller, providing a maximum of 180 degrees of possible phase advance.

One further important fact that arises as a consequence of including the

controller's integrator within the tuning loop is that it will now limit

cycle at a gain and frequency when the plant's phase contribution is -90

degrees, therefore ensuring that the ultimate frequency will correspond

to the second order plant's undamped natural frequency (WO).

Consider the standard second order system: -

Wo2
GCS, =2K. 3.25
S+2 os + W02 .....

the solution of its characteristic equation when an integrator is included

in the open loop yields the ultimate values

Mo
Ku
K
3.26
.....
wwu= wo

The proportional, integral and derivative gains being evaluated via the

following equations designed for any desired phase margin within the range

0-90 degrees.

Kc =
Ku Wut + ITT
- Wut . Td 2
(since ,
Td - 2wu 3.27
and .....

Ti = 4Td

Which for a phase margin of 90 degrees reduces to

1K
 - 10 5-

Gain (units) Phase (degrees)

1001- 0

10 -45

-90

0.1
-135

0.01
I 1
4-180
100
Frequency (rads/sec)

a. Plant

Gain (units) Phase (degrees)

100.9-

10 -45

-90

0.1 -135

0.01 `i-180
--
I E-6 I E-4 IE-2 1 100
Frequency (rads/sec)

b. Compensated plant

Fig. 3.13: Open loop frequency responses of the first stage pressure
cut and PID compensator, designed to approximate an
integrator using phase margin compensation techniques
 - 106 -

Ti = 4. Td 3.28
.....

11
Td =-_
u 2WO

Resulting in the open loop compensated system:

K (s+410)2 1
3.29
4W0L + os+ oIs .....

Placement of the zero's of the controller using this method will

cancel the second order dynamics of the process when C=1 exactly,
however, for the majority of c values the maximumphase error is approxi-

mately 33 degrees symmetrically about Wo. This error can be improved by

varying the ratio of Ti with respect to TO either using a priori knowledge

of ý or to best approximate the possible range. Cancellation results in an

open loop transfer function of an integrator, which on closing the loop

becomes a first order lag whose bandwidth is determined by the controller

gain.
The application of this method to the A. G. I. stations first stage

cut from 6895 to 5171 Kn/m2 (1000 to 750 psi) whilst passing 39.3 m3/sec
(5 1.,
tsft3/hr), for a 90 degree phase margin, produced the PID controller

parameters,
Kc = 8.1898

Ti = 0.696
Td = 0.174

resulting in the open loop transfer function displayed in Figure 3.13b.

The PID controller was applied to the system in the derivative on output

only form (3.16) to avoid problems with derivative kick on set point

changes; a derivative action filter was also fitted to limit high

frequency amplification. Figures 3.12 and 3.14 illustrate the response

of the A. G. I. system with the controller parameters displayed in table

3.8 to 34.5 Kn/mz (5 psi) set point changes in both loops simultaneously.
 - 107-

Nominal pressure

psi Kn/m2

10

50
5-

time
0- 0 crnnrie
i
5 10 15 20

Interstage volume pressure

-5' _ _
Pipeline inlet pressure
-50
-10. ----Set point

a. PID compensated
Nominal pressure

psi Kn/m2

10

50
5ý _ r.

time
0. 10 1'5 2b (seconds)

-5 -"- Interstage volume pressure

-50 Pipeline inlet pressure

-10 ---- Set point

b. PID compensated (Kc x 5)

Fig. 3.14: Application of a phase margin compensation designed PID

controller within the first stage cut of the AGI simulation,
with PI control of the second stage; time records illustrate
the response of the station to simultaneous 34.5 Kn/m2
(5 psi) changes in both set points
 - 108 -

In order to quantify the control system's improvement a fiqure of merit

was desirable. Chapter 5 investigates several alternatives to assess

both individual loop performance and closed loop interaction. It

concludes that the method proposed by Suchanti and Fournier (3.17)

provided all the necessary information whilst being relatively simple

to implement. Basically, the technique involves the comparison of

performance indices applied to the k'th output of the multivariable system

when

a) a unit step is applied to the k'th input with only the k'th

loop closed (TAE)

b) a unit step is applied to all inputs with all loops closed

(ISE)

using the equation

T5'Ek- ISEk
Indexk =k 3.20
see.

Note* For this particular application the integral of square errors

(ISE) performance criterion has been used to penalise large

amounts of interaction.

The index defines system interaction as lying in three

categories: -
i) -1< index <0 derogatory interaction

ii) index =0 no interaction

iii) 0< index <1 helpful interaction

Since the index is normalised then some reference to the

magnitude of the integral of square errors criterion is also necessary

if a full appreciation of the control system's performance is to be

attained. Application of this method to the various control schemes

designed resulted in the data tabulated below, which is discussed in

greater detail in the following section.

 - 109 -

PI/PI control I PID/PI PID/PI (Kcl*5)

control

Loop 11 Loop 2I LOOP 11 Loop 21 Loop 1 Loop 2

Kc 1.53 29.15 8.19 29.15 40.95 29.15

Ti 0.57 1.43 0.696 1.43 0.696 1.43

Td - - 0.174 - 0.174 -
TSE 115.1 571.6 59.2 379.5 39.8 278.1

ISE 23050.2 296.8 1198.5 258'.5 85.6 249.9

Interaction 0.724
-0.995 -0.951 0.468 -0.535 0.113
Index

Table 3.8 Performance results of PID Loop 1 controller

3.9 Discussion

The modified Ziegler-Nichols closed loop tuning technique

provides an alternative means of determining the critical process values,

but with tighter control over the system output's behaviour. Results

presented in table 3.1 and Fletcher (3.15) show that the critical values

found in this manner generally lie within 4% of their theoretical

equivalents. It should be emphasised that since the precision of the

technique relies on the ability of the describing function to represent

the various relay characteristics it will only provide sensible results

for those cases when the approximation is valid. Accuracy is yet further

reduced when hysteresis is included for noise rejection purposes. This

trade off was examined in the previous section where table 3.6 compares

the critical values found using various relay characteristics against

those obtained using an ideal relay. From the tabulated results the

following guidelines into the selection of the relay parameters were

revealed.
 - 110 -

i) The D/E ratio should be > 100 (3.15) to best satisfy the

accuracy/noise rejection trade off. Since the hysteresis

width (E) must be set greater than the output's noise amplitude
then a restriction is immediately placed upon the possible

range of relay amplitude (D) values.

ii) Any increase in the relay's amplitude (D) causes a subsequent
increase in the amplitude of the resulting limit cycle (A), by

a factor which is dependent upon the plant gain at the critical

frequency (Wu).

If operation within defined safety margins is required, then

this imposition of limits upon the magnitude of the plant's output could

possibly ensure that no relay characteristic will exist that satisfies

the above guidelines. One possible solution to this problem is to filter

the process noise, with a known postcompensator, whilst running the tuning

method and make an allowance for its gain and phase in the calculations.

However the resulting bandwidth of this system will no longer be purely

dependent upon the plant itself, but also the filter's characteristics,

since this dictates the one frequency at which system information is known.

By including various known phase lead and phase lag compensators within the

relay/process forward path then a method of evaluating the system's

frequency response is obtained (3.18), a feature, which because of the

closed loop nature of the test is of great practical value. Furthermore, a

study into system linearity is also possible by examining the variation in

the critical values that occur when the plant is tuned over its entire

working range and the form of the corresponding responses.

Comparison of Figures 3.4 and 3.6 shows the inadequateness of

using the Ziegler-Nichols look up tables to evaluate the controller

parameters especially when plants contain pure integrators or a large

 - 111 -

spread in time constants (3.15), as well as the possible improvements

available using the phase margin design technique. However the inclusion

of the PI controller's integrator during the tuning exercise limits its

application to type zero systems or type one for PID. Overall the

modified Ziegler-Nichols ultimate method provides the user with a simple,

robust technique that requires no a priori knowledge of the process to

produce a stable control scheme, as long as the process is at least second

order, and contains no more than one pure integrator.

The work presented has shown that fast, non-interactive
behaviour is possible from the twin valve system if the bandwidth of the

upstream control loop can be increased. This was achieved simply by

incorporating PID control within the first stage control loop. The use

of three term control in the first stage cut gave a twenty fold reduction
in closed loop interaction over the normally used twin PI control scheme,
table 3.8. A further reduction of one fourteenth was possible by

increasing the gain of the PID controller by a factor of five, a solution

which, because of the extra phase advance supplied by the three term

controller, was permissible without affecting the dynamic form of the

response. Larger gains could be used to completely suppress the process

interaction providing that controller output saturation was avoided.
Unfortunately this simple solution conflicts with current
British Gas policy where electrical interference and noise pick up on the
field instrumentation have been seen to promote unstable behaviour due to

high frequency amplification. The extra cost of ensuring an electrically

clean signal was deemed prohibitive and the cheaper option of employing

only PI controllers was pursued, the corresponding reduction in speed

being acceptable as it isn't of prime importance in the majority of

applications. Unfortunately this particular case is not one of these.

For these reasons another solution is required to meet the objectives.
 - 112 -

References (Chapter 3)

3.1 Arden, W. J. B. "GASCONTROL -A review of past, present and

future systems". North of England Gas Association, January
1988.

3.2 Ziegler, J. G., Nichols, N. B. "Optimum settings for automatic

controllers". Trans. ASME, Vol. 64,1942.

3.3 Cohen, G. H., Coon, G. A. "Theoretical consideration of

retarded control". Trans. ASME, Vol. 75,1953.

3.4 Rovira, A. A., Murill, P. W., and Smith, C. L. "Tuning

controllers for set point changes". Instruments and Control
Systems, Vol. 42,1969.

3.5 Astrom, K. J. "Ziegler-Nichols auto tuner". Report TFRT-3167,

1982, Lund Institute of Technology.

3.6 Astrom, K. J., Hagglund, T. "Automatic tuning of simple

regulators with specifications on phase and amplitude margins".
Automatica, Vol. 20, No. 5,1984.

3.7 Atherton, A. P. "Limit cycles in relay systems". Electronic

Letters, 18, No. 21,1982.

3.8 Gawthrop, P. J. "Selftuning PI and PID regulators". Proc.

IEE Conference on Applications of Adaptive and Multivariable
Control. Hull, 1982.

3.9 Wittenmark, B., Hagander, P., Gustavsson, I. "STUPID -

implementation of a self tuning PID controller". Report
TFRT-7201,1980, Lund Institute of Technology.

3.10 Luyben, W. L. "Simple method for tuning SISO controllers in

multivariable systems". Ind. Eng. Chem. Proc. Des. Dev.,
Vol. 25, pp 654-660,1986.

3.11 Hagglund, T. "A PID tuner based upon phase margin

specifications". Report TFRT-7224,1981, Lund Institute of
Technology.

3.12 Cox, C. S., Arden, W. J. B., Fletcher, I. "Investigation

into the remote pressure control of gas transmission systems".
Int. ASMEConference on Modelling & Simulation, Sorrento,
1986.

3.13 Fletcher, I., Arden, W. J. B., Cox, C. S. "Remote boundary

control of the Whasset/Barrow pipeline". I1orWest Gas
(Company Confidential).

3.14 Gelb, A., Vander Velde, W. E. "Multiple input describing

functions and nonlinear system design". McGrawHill, 1968.

3.15 Fletcher, I. "Automatic control for simple processes".

Report IF/02/85,1985, Sunderland Polytechnic.

3.16 Astrom, K. J.,, Wittenmark, B. "Computer controlled systems -

theory & Design". Prentice-Hall, 1984.
 - 113 -

3.17 Suchanti, N. C., Fournier, C. D. "A new algorithm for pairing

manipulated and controlled variables". ISA Conference
Instrumentation in Chemical & Petroleum Industries, St Louis,
p. 67,1973.

3.18 Luyben, W. L. "Derivation of transfer functions for highly

nonlinear distillation columns". Ind. Eng. Chem. Res.,
Vol. 26, pp 2490-2495,1987.
 - 114 -

Nomenclature (Chapter 3)

A Amplitude of limit cycle

Amin Minimum level of flow's limit cycle (Msft3/hr)

Amax Maximumlevel of flow's limit cycle (Msft3/hr)

dPq Flow measurement's differential pressure output (volts)

D Relay magnitude

E Relay hysteresis width

F Detuning factor

G(s) Process transfer function

Gc(s) Compensator transfer function

TAE Integral of square errors, sin gle loop

ISE Integral of square errors, all loops

Kc Controller gain
Kcon Controller conversion factor
Ki Integrator gain
Ku Ultimate gain
N(A) Amplitude dependent nonlinear function

Pu Ultimate period (sec)

Q Volumetric flow (Msft3/hr)

s Laplace operator

SP Controller set point

Td Derivative time constant (sec)

Ti Integrator time constant (sec)

Phase margin (degrees)

Damping factor
W Angular frequency (rads/sec)

Wo Undampednatural frequency (rads/sec)

Wu Ultimate frequency (rads/sec)

 - 115 -

4
CHAPTER

ShALL SIGNALANALYSIS

AGI SENSITIVITY ANDSTABILITY IMPLICATIONS

 - 116 -

4.1 Introduction

The equations describing the operation of an AGI, as derived

in chapter 2, are inherently non-linear and non-analytic. Because the

principle of superposition is no longer valid in this type of situation,

transfer functions are inapplicable. The consequence of this implication

is that investigations relating to stability and control system design

with the equations in the above form would require an extensive series of

numerical solutions with no guarantee that general rules for system design

would be realised.

By contrast, in the case of linear systems, there exists a consid-

erable body of work, based on frequency domain and root locus techniques,

leading to powerful methods of control system design synthesis. These can,

however, be applied to a non-linear system by converting its defining

equation into a linear form in which the respective coefficients are

functions of the system's operating point [4.1]. This procedure must, of

course, be validated in each case by ensuring that errors introduced by the

linearisation itself and by excursions of the variables from their quiescent

values in response to probable disturbances remain within acceptable limits.

To illustrate these considerations, we can examine the behaviour

of a single pressure control loop of an AGI with the downstream regulator in

open loop (figure 4.1). Figure 4.2 displays the trajectories followed by

the upstream regulator when the stem position of the downstream valve is

stepped between predetermined levels of 5% of full scale travel. The

responses, displayed on the V25 control valve steady state characteristics,

clearly illustrate the changes in process dynamics that occur with operating

point as well as the relative sensitivity of the first stage pressure cut to
the quiescent volumetric flow rate through the station. In its worst case,
this behaviour manifests itself as a hunting instability between the two

regulators. Figure 4.3 illustrates this using data recorded from experi-
 - 117 -

E
v
4)
N
+-) (1)
E "r Q)
G) C 4- >
i3 I--"
(OO
.J -a: > >

E
ý- ctS
Oa

N
"r C
r- 3
"r- O

+ý E
vA O
5-
W 4-
_r_
4J CS

ý "r
+.) N
Ri "r
C)

NC
NO
> "r
C +ý
"r U
rp
OO
4J 4-)
C
"r
C
00
-r- 4--)

CO 4)
+31 U
N 4)
G
CD a)
C
Ct
co 41
4- 3
0 of
4-' C
OO
to U"r
i 4-)
> C") O co
(C i "r
"r L
"C N f0
N>
UO
"r iC
+) mo
f0 r
E (1)+)
C) C`"r
0 (O N
U4-º 0
V) 0G

I-

LL
NOO
O LO
NIA
Lf)'-
 118 -

N
N
d1
L

fO
4)

NN "ý
Ei to
\G) Q

'L7
G) a)
X-
00
43
4-3
c
C
cu a)
U,
M
C)
Cl)
LO 0
CV U, C
N .C
I- C
Q) 4J
o- I
X CL
N aW C)
}
U, ON O+)
OO
N r- M "r
Xp
CO r O=0
+J U"r-
U
I^1 (U Cl.

C-4m
^y ", 7 r- GJ
(C O +)
Z S"" N
ý 4-)
CY)
C c-
oo
II
+-3 UO
U C.
OC
G) OO
Zr-A

iý i
C) 0
OO a
NNOO
NN (0
ZO

Iý i i 0-. -
a)
U i'
Q: N-, ý
O

i O
U)
co _O
1ýý4.3
V1 OU
UO

ýj -co 4-) 'r7

I- N
Si C= 3

ii . r-

4-
ro V)
N (A
"1"

j !ý
OO +)
lD NZ

I
NZZ0
O "F}3
04-10
"'- r-
>O7
co s a)
4-) (1)

i--i i aC)f
>(
vag

V)
4J
N
S-
tO
C)
CC
"r-
4Z
Ob

Ny
NZ,

0N

10 r
CN,
O
0.

GJ O
0 ä-`v

i
Z
r- O
4-
N ""
N
0

4ýý
--I
L
ö
r-
U') O
t0 L.ý
 - 119 -

N
V
OJ
In N

rL
c) O
E N+-
N
I- Lr

4) C
It, L G)

>+3
co O
rr
0 4-
N
"r- r-
-0 to
NX
-0 (CS
L
OO
U to
G)
LC
O
(1) G

4-) >,

"w"r"
Nr
C"1"
OU
vL N "r- fz
N
i (ö
N
S- y) 4)4)
NN
N yL O C)
N n
GJ Z
L y a)
r- ro >)
E
i ()
am Ö L
G/ Or
"r C
> 3 i
L
4.3 p N
aý >) O
O rn X-
to i
4-3 O
ul 0 4-1
ro
4. ) N 0 00
A 4-J .
4J C 4-)
(J
U
N º-ý y
N 4-)
C RS

b ý v
074)
C

+- O
0 4-
O
t G)
C.
II-
ON
4-3
CN
O G)
"r- +-i

rý -
Lr
+) G)
CA "'
=-
r
r 4-
O

C,
"r
C5 --T LL.
CC U
NNMM

VI yi t
CL C In
vvn
13

N
 - 12u -

mental field trials performed on an axial flow regulator at the Low

Thornley regional test facility, Appendix I. 1. The interaction occurring

between the station inlet pressure regulator, a 4" Fisher V25, and the

axial regulator, whilst attempting a step change in the latter's position.

At the same time, we must also be aware of phenomenapeculiar to

non-linear systems, such as limit cycles, jump resonance and sub-harmonics.

The latter two are not often considered in detail but imply that great care

must be taken in choosing input amplitudes and frequency ranges where, as

in the gas industry, frequency response tests on working installations are

still popular, especially since, for security reasons, these tests are

generally carried out under closed loop conditions.

The remainder of this chapter will, therefore, be devoted to

developing linearised forms of the equations describing the AGI and

establishing the validity of the procedure for further study.

4.2 Construction of a Linear Model

In general, a physical variable, y, is dependent for its value

on n independent physical variables [4.2]. That is,

y= f(xl, x2, ....., xn) 4.1

.....
where f is a function which is generally non-linear. In order to use the

standard linear design techniques on this function, it is desirable to

replace the actual non-linear relationship by a linear model: -

y= Ko + Kl. xl + K2.x2 + Kn. xn 4.2

..... .....
where Ko, K= ....., Kn are constants which best approximate the relation-

ship between the input variables and the non-linear function. Clearly,

this form of linearisation will only be effective over a small region

centred about a specific value of the variables, that is, a quiescent

operating point. By considering incremental changes of the physical

variables relative to the operating point then equation 4.2 can be re-
 - 121 -

written in the standard format: -

Ay = K1.oxl + K2.ox2 + ..... + Kn. oxn .....

4.3

so that the construction of a linear model is reduced to the proper

selection of the coefficients. The mathematical formulation of this

concept is available from the Taylor series expansion of the function y

about some chosen operating point xo

f(xo) of (x a2f (x - xo)2 4.4

y=x-2+.... + - xo) + .....
x=xo x=x°

Thus a linear approximation is obtained if only the first two

terms of the expansion are considered. As the number of variables

increases the technique can simply be extended to accommodate them: -

y= f(x10, x209 x309 ... ) +äi (x1

- xio)

(x2 (X3 4.5

+ - x20) +`3X - x30) +
"""" ".. ""
33X2

These form a hyperplane [4.3], as depicted in Figure 4.4, which

is tangential to the non-linear function over a limited region within n-
dimensional space and upon which the solution of the linearised equation

exists. By removing the starting value from the new linear relationship
[Yo = f(xIo, X2o, X30, ... )7, equation 4.5 can then be rewritten into the

format of equation 4.3: -

äX 4.6
Ay =äi AX1 +äf AX2 + ex3+ ...
23 .....

Determination of the partial derivatives now provides the

coefficients necessary to describe the small signal system behaviour. In

order to select the coefficients correctly the following information is

necessary: -
i) A description of the non-linear phenomenon.
ii) The region of interest and point of operation.

iii) A criterion for judging the linear model.

 - 12[ -

X2

xl

Figure 4.4: Geometrical interpretation of multivariable

linearisation
 - 123 -

This particular technique relies upon linearisation at a single

quiescent condition, and its accuracy is largely dependent upon how
drastically the output changes with respect to its inputs. Should this

prove to be a problem, then an approximation to the non-linear function

over the required operating range would be required. One of the more

popular techniques in achieving this is Least squares identification [4.4]

which selects suitable coefficients by minimisation of a performance

criterion, at the expense of small signal accuracy.

The description of the non-linear phenomenaoccurring in an AGI

has previously been derived theoretically in chapter 2. Following the

above analysis the next step involves linearisation of these non-linear

expressions, enabling suitable coefficients to be determined at any chosen

operating point by the inclusion of the prevailing quiescent conditions

within the resulting linear relationships.

4.3 Development of a Linear AGI Model

4.3.1 Linearisation of the Pipeline Model

Of the two basic equations developed to simulate pipeline

behaviour (equation 2.14) one can be seen to be non-linear

IC
P(n - 1) =s CQ(n - 1) - Q(n)]

Q(n) = K(n) n--n

Application of the small signal analysis technique to this

elemental pipeline equation resulted in the following linearised

equivalent (note the dropping of the o notation from now on): -

Q(n) = P(n - 1) P(n)

aann+ an)

= Kn. CP(n - 1) - P(n)7 .....

4.7
 - 124 -

where
3Q(n) -aQ(n) K(n)
Kn = _ -
aP(n - 1) aP(n) 2VP(n--n

Application of equation 4.7 to all the pipeline restrictions

results in the following linear pipeline equations: -

Pit) =s1C [Q2 - Q(2)] 4.7a

.....

Q(2) = K2[P(1) - P(2)] 4.7b

.....

P(2) = [Q(2) - Q(3)] 4.7c

SI .....

Q(3) = K3[P(2) - P(3)] 4.7d

.....

P(a) = [Q(3) - Q(4)] 4.7e

s-'(7 .....

Q(4) = Ka EP(3) - P(4)] 4.7f

.....

P(4) =c [Q(4) - Qload7 4.7g

.....

By placing equation 4.7g into 4.7f for P(4) and arranging in

terms of Q(4)
(sC). P(3) + Qload
Q(4) = 4.8a
.....

Now placing equation 4.8a into 4.7e for Q(4) and rearranging in

terms of P(3)

[sC/Ka + 11. Q(3) - Qload

P(3) = 4.8b
$$y+ .....

Repeating the procedure with equation 4.7d in terms of Q(3)

[sC(sC/K4 + 2)]. P(2) + Qload

Q(3) = 4.8c
(T-. .....
z
F',
4]

Substituting equation 4.8c into 4.7c for P(2)

[sc2 (
(sC). ++ 1] Q(2) - Qload
" + .
P(2) _ 4.8d
.....
(sue (sC)2. _ 4l
+ I. + + (sC). (3)

Repeating for equation 4.8d into 4.7b for P(2)

 - 125 -

(SC)3 3
(Z2- ýl
+ (sC)2. + + (sC). (3)j P(1) + 01oad
.
R(2) =J
3( ý+ ýý
(sC)2. (sC).
+ +24 +T, Iý ++ +1
K3
2. Zý

4.8e
.....
The linear transfer function relationship between the first stage

pressure P(1) and the pipeline inlet and outlet flows, Q2 and Qload

respectively, may now be determined by placing equation 4.8e into equation

4.7a for Q(2) and rearranging, thus
GP1(S). Q2 - Qload
P(1) = 4.9
P2 S .....

where
[T2 13---K]+sc2[
Gp1(s) (sC)3
_ - + +
K4
. 3 21ý aj\"KyJ

(T3
+ (SO +2K3 +1+1.....
T14 4.9a

and
(
Gp2(s) (SC)4 + (sC)3 I
= K22K3 + K22K4 + K32Kal
2R--ý-K

+ (sC)2 Iý +T+-+ (SC)(4) 4.9b

.....

To complete the linearisation exercise, the relationship

determining the pipeline inlet pressure Pin from equation 4.9 was required,

since: -
[K? ]2
Pin=P(1)+ 4.10
.....

then
Pin = P(1) + Kin-Q2 4.11
.....
where
?
Kin
KiePin
= ý-- 1)2
= KIýJ` 4.11a
.....

The combination of equations 4.9 and 4.11 will provide a full

linear description of the pipeline behaviour.
 - 126 -

4.3.2 Linearisation of the Valve Flow Equation

The flow through the valves is described using the Fisher flow

equation (section 2.3) which has the form: -

CV {1 -] i]
Q= Kf1. Cg.P1. sin[Kf2. - 4.12
. .....

where the flow coefficients CV and Cg (Figure 2.11) are non-linear

polynomial expansions of the valve stem displacement (X). Thus: -

j+ ( 3Q)
QXI P1I + P2 4.13
a1 la- 1 .....

where the partial derivatives are

aQ Kf1. Kf. Cv p
_ LF1i-i
(I' cos(h) + Kf1. Cg sin(o)
a1
4.13a
.....

aQ Kf1. Kf2. Cv ýýpJP1

ý__ _ . cos (e) .....
4.13b

ac Kf1. Kf2. EP1(P1 - P2)] [c__]aC,

= Kfl. P1.1 sin(d) + "
Cg
i'
"- Cvý cos(h) 4.13c
.....

where
Cv Pr ý p21ý
ý= Kfz. ý. I I 4.13d
.....

a=6.644E-6
X3 - 5.343E-4 X2 + 0.0229 X-0.02496 4.13e
.....

aCg
- 12.42E-8 X" + 25.424E-6 X3 - 18.363R-4 X2 + 0.067 X-0.104
3- --
4.13f
.....

4.3.3 Linear AGI Model

The development of a linear model requires the evaluation of the

constants described by the partial derivative equations derived in section

4.3. Because the constants are operating point dependent then it is

 - 127-

necessary to establish some quiescent condition at which the linearisation

is to take place. The system operating conditions selected in chapter 2

for the initial simulations were as follows: -

Q= 39.33 M3/sec (5 P1sft3/hr)

Ps = 6895 kN/m2 (1000 psi)

Pu = 5171 kN/mz (750 psi)

Pin = 3447 kN/m2 (500 psi)

This choice of flow ensures that the valves operate at the most

non-linear part of their characteristics, allowing the following non-linear

pipeline constants to be derived.

P(1) = 3429.0246 Kn/m2 (497.32 psi)

P(2) = 3410.4492 Kn/m2 (494.63 psi)

P(3) = 3391.7720 Kn/m2 (491.92 psi)

P(4) = 3372.9914 Kn/m2 (489.19 psi)

K(1) = 9.1499

K(2) = 9.1252

K(3) = 9.1003

K(4) = 9.0752

C = 1.9951

which when placed into equations 4.8 and 4.11a resulted in the linearised

pipeline restriction constants

Kin = 0.9395

K2 = 1.0586

K3 = 1.0529

K4 = 1.0471

leading to the following linearised pipeline relationship (equation 4.9)

between the first stage pressure and associated inlet, outlet flows

Gpl(s) = 6.4051s3 + 17.2256s2 + 11.126s +1

Gp2(s) = 12.5225s' + 40.4582s3 + 36.310s2 + 7.8204s
 - 128 -

Operation at the above quiescent condition requires the valve

stem position's
X1 = 26.8%

X2 = 29.4%

Placing this information into equation 4.13 then the linear

valve coefficients were found to be

(i) Valve 1:

Kll = a0-= 2.6118

K12 ==DQ' 13.225E-3 4.14

.....

K13 =ä_ -10.027E-3

u
(ii) Valve 2:

K21 =a=2.353

K22 =a= 13.953E-3 4.15

.....

K23 = _ -9.522E-3
a-

Figure 4.5 displays the structure of the resulting linear AGI

simulation, clearly illustrating the multivariable nature of the system.

4.4 Small Signal Model Validation

Before we can proceed to analyse the linearised AGI model, we

must first ensure that it exhibits the same behaviour as the non-linear

version from which it was derived. The basis for comparison was the

model's response to the following disturbance conditions about the above

quiescent conditions: -
i) 5 psi change in SP1 (Figure 4.6)

ii) 5 psi change in SP2 (Figure 4.7)

iii) 1 Msft3/hr change in load flow (Figure 4.8)

- 129 -

Z
ý--ý
a

4)
N
E

G'Y
T

"CL

v
0

to
9

LO

C.
L.
 - 130 -

Pressure
psi Kn/m2

10

60
8

6 40-

4-

1 20
2

oý 0
o time
(seconds)
a. Variation in interstage volume pressure - SPu
Pressure
psi Kn/m2

0.1
0.5

00

-0.5

-0.1
j time
(seconds)
b. Variation in pipline inlet pressure - 6PIN

Fig. 4.6: Linearised model validation results for the AGI station;
time records illustrate the simulations response to a step
change in volume set point

Note: The linear/non-linear simulations behave

almost identically
 - 131 -

Pressure

psi Kn/m2

40 300

20.15C

0a

-20
-150

-40-
- 300
P- time
(seconds)
a. Variation in interstage volume pressure - SPu

Pressure

psi Kn/n2 1 Non-linear

101
-"- Linear
60

--- Set point

40
5- 000, #000
0ý0,
0

20

0J 0' ' time

05 10 (seconds)

b. Variation in pipeline inlet pressure - aPIýj

Fig. 4.7: Linearised model validation results for the AGI station;
time records illustrate the simulations response to a
step change in pipeline inlet pressure set point
 - 132 -

Pressure
psi Kn/m2

5
3'

Q.

-3
-time
(seconds)
a. Variations in interstage volume pressure - Ou
Pressure
psi Kn/m2
Non-linear

-'-Linear
5
---Set point

0-

-51

rýtime
0 6 12 (seconds)

b. Variations in pipeline inlet pressure - 6PIN

Fig. 4.8: Linearised model validation results for the AGI station;
time records illustrate the simulations response to a
25% step change in load flow

Note: The P+I controller settings specifically tuned

for unstable response to allow a comparison of
the two simulations stability margins
 - 133 -

where tests i) and ii) use the controller settings developed for chapter 2

and are designed to examine the small scale behaviour of the linear system,

whilst test iii) incorporates an unstable control scheme to investigate

both large scale accuracy and any changes in system stability during the

linearisation exercise.
By visual inspection of Figures 4.6,7,8 it is evident that the

two model forms respond in the same fashion. Figure 4.8 [test iii)]

clearly illustrates the limitations of this particular linearisation

technique in that a degradation in accuracy results as the amplitudes

increase, although the relevant dynamics appear to be unaffected by the
linearisation as the stability margins remain unchanged. For completeness,

numerical analysis of the validation tests was performed using the

Coefficient of Determination [4.5], providing a figure of merit which
indicates the percentage fit between the linear and non-linear models. The

results presented in table 4.1.

Coefficient of Determination
Pu Pin

Test i 99.9999% 99.9831%

Test ii 97.6956% 99.6671%

Test iii 91.8539% 92.0238%

Table 4.1 Linear/Non-linear Model Comparison Figures

4.5 Implications on Sensitivity and Stability of

Certain Identified Parameters

The form of the response obtained from the AGI station is

naturally a function of its structure plus the inputs and disturbance

effects experienced during its normal operation. However, certain

 - 134-

parameters were thought to be particularly important in realising stable

closed loop behaviour. Often, unfortunately, the choice of these

parameters was not clear and their influence uncertain. To help clarify

this situation it was decided that an analysis of the system characteristic

equation [4.6] using the root locus technique [4.7] was desirable. In this

way, stability boundaries, related to the individual parameters, can be

determined and sensitivity of system stability margins assessed.

Before commencing a study into the complex behaviour of the full

AGI system, two initial studies were carried out in an attempt to identify
the critical features relating to the first and second stage pressure cuts.
The studies included investigating the sensitivity of each individual stage
to operating point by solving the appropriate linearised equations over its
full working range and observing the resulting parameter variations. In

addition stability was assessed by identifying the operating regions where

an incursion into the right half of the s-plane occurred by the roots of

the relevant characteristic equation.

4.5.1 Sensitivity and Stability of the First Stage

Pressure Cut

The main transmission inlet pressure to an AGI is ideally 6895

Kn/m2 (1000 psi). Under these conditions the characteristic equation of

the compensated first stage cut, figure 4.9, in steady state conditions

can be shown to be

s3(C. T1) + s2(C-K13. T1) + s(Ka. Kil. Kc1- K13)

{Ka. Kl]
+ K11. =0 4.16
.....
i

where Ka = 7.252E-3 the sensor/interface/actuator conditioning constant.

Under British Gas policy the first stage cuts operational range

would normally be constrained to lie within the boundaries

 - 135 -

P1

Infinite
- Q2
upstream
volume
Fisher
V25
-+-- 14' 18"
,
pipeline

a. Schematic diagram

Pc

b. Linearised system block diagram

Fig. 4.9: First stage pressure cut system; for analysis of

sensitivity and stability margins
 - 136 -

3500 < Pu < 6200 Kn/m2

(500) (900 psi)

when subjected to the maximumdesigned load flow variations: -

10 <Q< 200 m3/sec

(1) (25 Msft3/hr)

The resulting variation in the numerical values of the linearised

regulators' parameters, K11, K12 and K13, when the system lies within the

defined steady state boundaries are displayed in figure 4.10. The effect

these variations have upon the first stage system stability is shown in

figure 4.11a, where the critical flow rates are displayed for interstage

volume pressures within the previously defined ranges and various values

of upstream pressure that can occur in practice, depending upon the

topography of the particular station.

It is evident from these characteristics that the problem areas

regarding the stage's stability occur at low flows and high differential

pressures across the valve, a point that is more apparent from figure 4.11b,

where a single boundary condition between stable and unstable behaviour is

developed by considering the effects of upstream and downstream regulator

pressure variations simultaneously.

A Root Loci study with respect to the length of the interstage

volume, using equation 4.16, illustrated that the inclusion of integral

action within the controller promoted the possibility of instability upon

an otherwise inherently stable system. Figure 4.12 displays the

trajectories that the roots of the characteristic equation take when

varying the reciprocal of the dimension dependent, interstage volume

constant (C). By inspection it is apparent that the system becomes

unstable under the given controller settings and system conditions when
1/C < 29, that is an interstage volume length of greater than 48 metres
(160 feet approx. ) assuming the pipe diameter remains unchanged. More-
 - 137-

Kii

Pu
6210 step
4 ,
Kn/M2
900 step
,
psi)

0' r
0 50 100 150 200 (m3/sec)
Flow 0

05 10 15 20 25 (Msft3 /hr)

a. Valve parameter 73Q'

-xi
K12
Pu
(psi)
0.2

(900)

0.1
(800)
700
600
500

p 50 100 150 200 (m'/sec)

Flow Q

05 10 15 20 25 (Msft3/hr)

b. Valve parameter
s
Fig. 4.10: Linearised upstream valve parameters variation over
(1 of 2) the AGI stations working range
 - 138 -

0.2
Pu

m2 (psi )

(900)

0.1

(800)
(700)
(600)
(500)
0
0 50 100 150 200 (m3/sec)
Flow Q

05 (Msft3/hr)
10 15 20 25

Valve parameter -au

c.

Fig. 4.10: Linearised upstream valve parameters

(2 of 2) variation over the AGI stations working range
 - 139 -
Q

P1sft3/hr f13/sec

20

Ps

6900 step
,;n/m2
10 1000 step
L A)

01 0
4000 5000 6000 -Kn/m2
(Pu) interstage volume
pressure
500 600 700 800 900 1000 psi
Q
a. Individual stability margins
t1sft3 /hr m3/sec

Kn/mz
Valve differential
pressure dPv
100 200 300 400 500 psi

b. Stability/instability bound

Fig. 4.11: First stage pressure cut systems stability margins over
the full working range of the AGI station, below which
unstable behaviour results
 - 140 -

Imaginary

3
Q Open loi

0 Open l oi 2

Q Nominal

Ri

-1

-2

-3

a. Full root loci

Imaginary

Real

b. Stability assessment

Fig. 4.12: First stage cuts sensitivity; root loci display the
roots of the systems characteristic equation variation
with respect to the reciprocal of the interstage
(1)
volume constant
lc
 - 141 -

over, decreasing the volumes dimensions would increase the stability

margins of the system. However, this would make the interstage volume

pressure extremely sensitive to input and output flowrates, and therefore

to stem position changes of the two regulators, whilst possibly
invalidating the assumption of laminar flow within the pipeline sections

since operation within the vena-contracta regions of the regulators would

occur.

4.5.2 Sensitivity and Stability of the Second

Stage Pressure Cut

Following a similar approach to that presented in the previous

section, the sensitivity and stability of the second stage pressure cut,
figure 4.13, were investigated. This initially requires the definition

of the British Gas design limits for the second stage pressure reduction.

Assuming the interstage volume pressure Pu could be held at its nominal

set point value of 5171 Kn/m2 (750 psi) then the defined working range is
2400 < Pin < 4900 Kn/m2

(350) (700 psi)

10 <Q< 200 m3/sec

(1) (25 Msft3/hr)

Under these conditions regulator sensitivity characterised by

the parameter variations illustrated in figure 4.14 was investigated. By

placing the resulting information into the second stage characteristic

equation
{Kin G (s)
1+ (Gc2(s). A2(s). K21 - K23). +,, =0..... 4.17
p2 s

where
Ka
A2(s)= 1 ST2
 - 142 -

Infinite
upstream Qload
MMOO'
pressure
Fisher
V25
-0--1 rile 18" ý pipeline w to

a. Schematic diagram

b. Linearised system block diagram

Fig. 4.13: Second stage pressure cut system; for analysis of

sensitivity and stability margins
 - 143 -

K21 PIN

2760,4839
step 690 Kn/m2
(400,700 step
100 psi)
6

0. ' .º
0 50 100 150 200 m3/sec
Flow Q

05 10 15 20 25 Msft3/hr
.
a. Valve parameter 3Q
K22

PIN
0.3 m2 (psi )
0 (700)

0.2

0.1 0 (600)

0 (500)
0 (400)

0
0 50 100 150 200 m3/sec
Flow. Q

05 10 15 20 25 P1sft3/hr

b. Valve parameter a42

-
u
Fig4.14: Linearised downstream valve parameters variation
(1; f 2) over the AGI stations working range
 - 144 -

K2
Pin

0.3 (psi)

(700)

0.2

0.1
(600)
(500)
(400)
0
0 50 100 150 200 m3/sec
Flaw Q

05 10 15 20 25 Msft3/hr

SQ2
Valve parameter ýi
c.
n

Fig. 4.14: Linearised downstream valve parameters variation

(2 of 2) over the AGI stations working range
 - 145-

then the system's stability could be assessed. Analysis of the

characteristic equation's root locations within the above operating

boundaries whilst varying the second stage upstream pressure between the

limits defined in section 4.5.1 revealed that, under the controller settings

given, there was no possibility that the roots would migrate into the right

hand plane, and therefore stability was assured.

4.5.3 AGI Characteristic Equation Development

The previous studies, although providing valuable knowledge about

the system behaviour, were "idealised" in that the information gathered

confirmed stability in the steady stage operating region defined without
taking into account the true multivariable nature of the problem. Consid-

eration of the effects of the system interaction during transient changes,

so graphically displayed in section 4.4, requires the formulation of the
full system characteristic equation. The first stage of its development
is the construction of a multivariable transfer function description of
the linearised process of figure 4.5. Using block diagram manipulation

techniques the following transfer function was developed, whose Nyquist

Array is displayed in figure 4.15 when under normal operation conditions.
GII(s) G12(s)
GP(s) _
G21(s) G22(s)

K11.A1(s). Y(s) -K21. A2(s). GP1(s)

K11.A1(s). K22.X(s) K21.A2(s). (sC-K13). X(s)

(SC - K13). Y(s) + K22.Gp2(s)

4.18
.....

where
Ka
A, (s )+
srl
sensor/interface/actuator dynamics
Ka
A2(s) (Ka = 7.252E-3)
+ sr2
 - 146-

Imaginary

Rea1

Fig. 4.15: Nyquist array of the linearised AGI model

operating at a throughput of 40 m3/sec
 - 147-

Y(s) = Gp2(s) - K23. X(s)

X(s) = Kin. Gp2(s) + Gpl(s)

Gpl(s) = 6.4051s3 + 17.2256s2 + 11.126s +1

Gp2(s) = 12.5225s" + 40.4582s3 + 36.310s2 + 7.8204s

The polynomials Gpl(s) and Gp2(s) developed in section 4.3,

represent the pipeline dynamics and have orders that are purely dependent

upon the chosen number of stages during the modelling exercise (see section

2.3.3). The selection of a 4th order pipeline model optimised the trade

off between complexity in the design and analysis stage, and modelling

accuracy.
To complete the open loop system we must add the controllers,

noting the single input single output format.

FGci) 0
Gi(s) 4.19
0 Gc2(s) .....

The open loop transfer function can now be described via the

equation
Qt(s) = GP(s). Gc(s)

P111(s) N12(s)
4.20
.....
N21(s) F422(s)

D(s)

In nultivariable theory the denominator of the transfer function

matrix is referred to as its characteristic polynomial. Like SISO theory

the closed loop system's characteristic equation can be found by equating

its characteristic polynomial to zero, therefore from equation 4.20: -

det[I + Gt(s)] =0 .....

4.21

Combining equations 4.20 and 4.21 to form the characteristic

polynomial description (note the dropping of the Laplace subscript)

D2 + D(N11 + N22) + N11. N22 - N12. N21 =0..... 4.22

 - 148 -

Placing equations 4.18 and 4.19 into the above and manipulating

we get

Z(s). Y(s) + K21.A2(s). GC2(s). Z(s). X(s) + K22.Gp2(s) =0..... 4.23

where Z(s) = sC - K13 + Kii. A1(s). Gc1(s)

4.5.4 Influence of Controller Design, Actuator Selection

and Interstage Volume on AGI Stability

Utilisation of the characteristic equation to determine the

effects certain factors have upon the overall behaviour of the system,

must first involve the identification of the system parameters regarded

as being the most important in terms of its response.

The importance of the controller parameters in determining

system stability margins is obvious, however analysis of their individual

effects upon overall system performance could provide valuable information

regarding controller structure and design. Two parameters that also affect

this situation are the individual actuator time constants; the reason for

their inclusion within this analysis arises through maintenance problems

within the system due to regulator wear/reliability. It is the replacement

of such regulators with new devices that could possibly promote stability

problems. The final parameter considered was the length of the interstage

volume, reflected in the volume constant C, the correct selection of which

is of great interest and concern within the Gas industry.

Parameter Nominal Value

Kcl 1.53
Controller KC2 29.15
design Til 0.57
Tie 1.43

Actuator T1 1.00
dynamics z2 1.00

Interstage C 3.072E-3
volume

Table 4.2 System parameter nominal values for application

to the stability analysis
 - 149 -

Evaluation of equation 4.23 with respect to one of the above

parameters, whilst replacing all others with their nominal values enables
the calculation of the root locus pole and zero locations. Appendix 1.3

contains a listing of a program designed to perform this task and the

resulting loci are displayed in figure 4.16.

Closed loop instability is predicted by the root loci study

(figure 4.16) when the following conditions prevail

T1 > 7.7 secs

T2 > 16.7 secs

Til < 0.071 secs

0.023 < Ti2 < 0.161 secs

C>0.145 secs
However, the results of the above analysis are only applicable
at the operating point at which the system was linearised, and the
controllers were designed. It has already been established that

significant changes occur in the system parameters with process operating

conditions. The effect of these variations upon the overall stability of

the system was evaluated by solving for the roots of the characteristic

equation at all expected system conditions. Figure 4.17 displays the

critical flow rates that bound the stable/unstable regions of AGI system

operation, within the previously defined valve differential pressure

ranges.

4.6 Discussion

The stability of the AGI system over its full working envelope,
figure 4.17, indicates few problem areas as the unstable behaviour only

arises at flow rates that are below the minimum capacity of the station.
However, the stability margins are critically dependent upon the tuning
 - 150 -

Imaginary s

.5

0.5

Real s
a. Kc,

-1

-2
-0.5 0

b. "C2

Fig, %16: AGI station sensitivity under SISO control; s-plane

(1 of ) diagrams display the roots of the system's characteristic
equation with respect to the stipulated variables
 - 151 -
0 Imaginary

,5

.0

'. 0
Real s

C. 1/TIi

-4

d. 1/Tie

Fig. 4.16: AGI station sensitivity under SISO control; -s-plane

(2 of 4) diagrams display the roots of the system's characteristic
equation with respect to the stipulated variables
 - 152 -
.5

Imaginary

.0

,5

.0

0.5

1.0

f. -1.5
-0.5 0 0.5 I
-1.5 -1.0
Real s
e. 1/ý1

.5

.o

.5

f. 1/c2

Fig. 4.16: AGI station sensitivity under SISO control; s-plane

(3 of 4) diagrams display the roots of the system's characteristic
equation with respect to the stipulated variables
 -- 153 -
+1.5

1.0

0.5

0.0

-0.5

-1.0

-1.5
0.5

9" 1/c

Fig. 4.16: AGI station sensitivity under SISO control; s-plane

(4 of 4) diagrams display the roots of the system's characteristic
equation with respect to the stipulated variable
 - 154 -
-D

Qý
I-
rt%
N
S..

N
a)
LYr L
o) O G.
4-
4- MO tC
N CL N "r

CL aa "r
ýv -P Q) C G) ýº 3
N
.4
C. O
fq r-
--N
4-
r-
U
cu

Q "r
U, QL
O lD U
O

-a r
N
i 4ý :3
N
4- C)
LL.
OL

C "r "r
G"r
r3
to

C) f3 N
0
C: "r"
M
Q "I-> U
C fý "r
Lt
+ý 3
N

r- 3
r- O
"r
a)
.. -0
QO
CD QrN
0
0 tu (D
N r-" "r

"r
"r
NC
0

C.7
¢ C'
O
N r- "r
O t'0
4- S.
4)
CD G) 0.
-O
0
I-
9-P. -
O(
QL

+- O
"r 4-
I--
"r- N
CJ
"G
(ß 4)

V) L

U
C)
N 0
n7
E
v

s Ln C) L
i3

N
 - 155 -

of the first stage controller, inappropriate setting of which could lead

to the introduction of stability boundaries within the operating region

of the station.
The reduction in stability margins that occurs with decreasing

flow rates, dynamically illustrated in figure 4.2, arises as a consequence

of two factors. Firstly, the non-linear behaviour of the regulators

flow - pressure relationships, figure 2.7, and secondly, the British Gas

policy for transmission valve sizing, the latter stating that any valve

must be capable of passing the maximumrated station flow whilst maintaining

the minimum possible station differential pressure (typically 10 psi) in

order to protect the pressure of the gas supply to the regions. Inspection

of figure 4.18 illustrates the effect that sizing of the transmission valves
in this way has upon normal system operating conditions when applied to a

typical valve flow characteristic (Fisher V25). Consider operating with a

steady state differential pressure of 250 psi, then only 63% of valve travel

is necessary to move from zero to maximumstation flow. Thus valves selected

under this policy are significantly oversized. The maximumstation flow is

only 30% of the maximumflow possible through the valve when operating within

the system.. This implies using the valve in a more non-linear part of its

characteristic, figure 2.7, where it also has a high sensitivity to stem

movement, i. e. in control terms, its gain is high, illustrating the need to

tune the SISO PI controllers at lower station throughputs in order to 'avoid

instability problems.
The above work has also confirmed that greater stability margins

will arise by making the interstage volume smaller. However, this will also

lead to the promotion of larger fluctuations in interstage volume pressure

whenever regulator movement occurs. In order to provide both good

performance and stability then it is necessary to suppress the interaction

that arises as a consequence of the small volume dimensions. Stabilisation

 - 156 -

O
x
CD
O OV CD OOOn U,
s
4-3

t
cý
c
N
. r.
N
i
O
>
O
d)

"r

b
S
4)
N

r (J
r- C
"r O

r) > "r

i U-o
"r C
r-OU O
G

E
"r
NO
"r C
N

NN
>'C
F-- C
(z O
Q1
CC
O "r
"r 4-'
N tic
NL
"r Q)
a
ýn O

i Q)
4
N
r0 N
C.7 C)
U

C) 000000 C'**- 31.1

CD 0 t/1 "C

CD 0) co LO M 0 (N "r
4J
O
i
i->
T- "r
LC
C] "r

ö E
x Co
0 a)
E rn
0 U-
 - 157 -

of the interstage volume pressure could readily be achieved by increasing

the bandwidth of the first stage cut, thereby making it fast enough to

compensate for the pressure transients at the expense of poor noise

sensitivity and excessive controller action, unless the speed of response

of the actuators could be improved. This would generally lead to stability

problems; however the root loci analysis has established that the system

stability is particularly insensitive to the controller gain parameter Kc1

(figure 4.16a). Therefore, by increasing the gain of the first stage

controller we could ensure fast, non-interactive behaviour from the twin

valve system at the cost of a more oscillatory response.

 - 158 -

References (Chapter 4)

4.1 Gibson, J. E. "Nonlinear Automatic Control". McGraw-Hill

Book Company, 1963.

4.2 Thaler, G. J., and Pastel, M. P. "Analysis and Design of

Nonlinear Feedback Control Systems". McGraw-Hill Book Company,
1962.

4.3 Doebelin, E. 0. "System Modelling and Response - Theoretical

and Experimental Approaches". John Wiley & Sons, 1980.

4.4 Ljung, L., and Soderstrom, T. "Theory and Practice of Recursive

Estimation". MIT Press Series, 1983.

4.5 Young, P. "Recursive Estimation and Time Series Analysis - An

Introduction". Springer-Verlag, 1984.

4.6 Evans, W. R. "Control System Synthesis by Root Locus Method".

Trans. AIEE, Vol. 69, pp 66-69,1950.

4.7 Kuo, B. C. "Automatic Control Systems". Prentice-Hall, 5th

Edition, 1987.

4.8 Sinha, P. K. "Multivariable Control - An Introduction

Electrical Engineering and Electronics; 19". Prentice-Hall, 1984.
 - 159 -

Nomenclature (Chapter 4)
Small Signal Analysis AGI Sensitivity and Stability Implications

A1(s) Upstream actuator/sensor transfer function

A2(s) Downstream actuator/sensor transfer function

C Pipeline volume coefficient (m3/Kn/m2)

Cg Valve critical flow coefficient

Cv Valve flow capacity coefficient

dPv Valve differential pressure (Kn/m2)

D Denominator of open loop system transfer

function matrix

Gc(s) SISO controller transfer function matrix

Gcl(s) Upstream controller transfer function

Gc2(s) Downstream controller transfer function

Gp(s) Plant transfer function matrix

Gpl(s) ) Polynomials relating pipeline flows to

Gp2(s) ) its first stage pressure

GI(s) Open loop system transfer function matrix

Ka Actuator/sensor conditioning constant

Kcl Upstream controller proportional gain

Kc2 Downstream controller proportional gain

Kfl ) Fisher valve

Kf2 equation constants

Kin Linearised pipeline inlet restriction coefficient
Kn, K1 etc. Linearised pipeline restriction coefficients

K(n), K(1) etc. Pipeline restriction coefficients

K11, K12 etc. Linearised upstream valve flow coefficients'

K21, K22 etc. Linearised downstream valve flow coefficients

n Discrete pipeline section label

N11,N12 etc. Numerators of open loop system transfer function

matrix
 - 160 -

Pin Pipeline inlet pressure (Kn/m2)

Ps Station inlet pressure (Kn/m2)

Pu Interstage volume pressure (Kn/m2)

P(n), P(1) etc. Pipeline section pressures (Kn/m2)

P1 Valve upstream pressure (Kn/m2)

P2 Valve downstream pressure (Kn/m2)

Q Valve flow (m3/s)

Qload Pipeline outlet flow (m'/s)

Q1 Upstream valve flow (m3/s)

Q2 Downstream valve flow (m3/s)

Q(n), Q(1) etc. Pipeline section flows (m3/s)

S Laplace operator

Til Upstream controller integral time constant (s)

Tie Downstream controller integral time constant (s)

X Normalised stem displacement

X1 Upstream valves normalised stem displacement

X2 Downstream valves normalised stem displacement

X(s) )
Polynomials relating to plant transfer
Y(s) )
function matrix
Z(s) )

Upstream valves actuator time constant (s)

T1
Downstream valves actuator time constant (s)
T2
 - 16,1 -

5
CHAPTER

APPROACH
A MULTIVARIABLE TO THE DESIGNOF

AGI CONTROL
SYSTEMS
 - 162 -

5.1 Introduction

In a single-input single-output (SISO) system any changes in

the output of a process can normally be directly related to corresponding

changes in its input, via the system's transfer function, or to outside

influences, generally classed as disturbances. In a multivariable system,

however, which by definition contains more than one input and/or output,

there are additional influences brought about by interaction: the

behaviour of one loop affecting that of the other(s). This interaction

is normally such as to prejudice system stability margins. It is therefore

essential when designing controllers for multivariable systems, to make an

allowance for these interactive effects as an integral part of the design

exercise. Moreover, the availability of more than one input and/or output

can lead to fundamental difficulties in the best choice of input/output

pairings for control purposes.

It can be concluded from this that the increased structural

complexity associated with multivariable systems means that the well

documented, SISO control techniques are only applicable to the small range

of multivariable systems where interaction is minimal.

The emergence of multivariable control came into prevalence in

the late 1950's, this sudden and dramatic increase in activity being

directly related to two factors. Firstly, the development of the digital

computer enabling more intricate and extensive studies than were previously

possible and secondly, the upsurgence in areas of space exploration and

aircraft design [5.1].

Feedback solutions to the aerospace control problems, whose

dynamics were well known and understood, were formulated to minimise

problem related performance criteria leading to design methodologies which

are now classed as optimal control theory [5.2,5.3,5.4]. The basis for
 163 -

these theories is the state space description of a system, consisting

of sets of first order ordinary differential equations which collectively

describe its behaviour. Although quite easily implementable for aerospace

problems where the relationships between design requirements and

performance criteria were well defined, the method failed to become popular

in other industries where difficulties in the selection of the quadratic

performance criteria arose. Furthermore, a solution of the optimal control

problem was often restricted by states that were unobservable or

uncontrollable, as shown in figure 5.1. This problem was partially

alleviated by the application of observers [5.5] to recover inaccessible

states using the state space model and information about the other inputs

and outputs.

Meanwhile, advances in the study of the structures of linear

multivariable systems had also taken place leading to the foundation of

many new design techniques including modal control [5.6] and pole-shifting
[5.7]. Both techniques basically involve the use of a feedback controller

to position the poles (eigenvalues of the state matrix) of a system at

desired locations within the s-plane, and were therefore plagued by the

same problems when unobservable and uncontrollable states exist.

In general, all the above time domain based techniques, although

relatively simple and systematic, require accurate system models which

become significantly more complex as system order increases. Furthermore,

they do not facilitate the insertion of classical dynamic compensation or

readily handle the standard forms of engineering performance specifications.

These difficulties provided the motivation for the development of frequency

domain techniques as an alternative to the state space approach. The

immediate aim of the fundamental studies was to seek to generalise and

exploit the tools and insights which have made classical techniques so
 - 164 -

Uncontrollable
and
Unobservable

Fig. 5.1 Kalman description of a system

 - 165-

successful. It should be remembered that the classical feedback theory

relied on scalar feedback principles since it generally concerned systems

with one input and one output (SISO). Multivariable systems (MIMO)

are based on vector feedback theory where the system descriptions are

described by transfer function matrices.

To allow system outputs to be adjusted independently, it is

often desirable that, particularly in the steady state, manipulation of

one input will only affect its corresponding output. This translates to

the natural requirement of a diagonal transfer function matrix. In order

to achieve this behaviour, initial attempts involved the construction of

decoupling controllers which then reduces the multivariable problem to
the design of a number of SISO systems using well established techniques.

The problems associated with this particular technique arise from the

fact that the resulting decoupling controllers were invariably complicated

and essentially inverses of the system. Thus, like the time domain

techniques mentioned above, they require an almost exact model of the

system, producing excessively large high frequency gains, which introduced

noise sensitivity problems, when applied to 'normal' systems.

By replacing the principle of exact decoupling with the more

practically feasible concept of diagonal dominance, the Inverse Nyquist

Array (INA) method [5.8] provided a much more powerful tool in the design

of multivariable systems. A formal description of diagonal dominance is

generally considered to be [5.9] when the magnitude of the diagonal

elements of the open loop compensated system are greater than the sum of
the magnitudes of the off diagonal elements in the same matrix row or

column, row and column dominance respectively. However, problems do

arise in the selection of the cascade compensation necessary to achieve

dominance. Generally, it involves the calculation of a simple gain matrix

designed to diagonalise the system at a specific frequency. It is this
 - 166 -

choice of frequency and the ability of the resulting compensator to

achieve dominance at other frequencies that causes the problem.

In 1973 an alternative strategy to INA was proposed by

P1cFarlane which he called Characteristic Locus Design [5.10]. Here,

the characteristic loci (the contours followed by the eigenvalues of

the system transfer function matrix in the s-plane as frequency is

varied) are dynamically compensated at required frequencies in order to

satisfy the stability, interaction and accuracy criteria on which the

technique is based. This is achieved using various 'tools' to provide

systematic synthesis of the controller.

The following chapter commenceswith a library of the tools

required to both design and assess the performance of a characteristic

locus based approach, resulting in a methodology that is particularly

suited to application within a computer aided design package. Several

methods of assessing interaction and its effects are also presented,

along with guidelines for the appropriate choice of input-output pairings

for multivariable control. The chapter culminates with the application

of the technique to the AGI model developed in Chapter 2 and discusses

the practical aspects of implementing such a controller.

5.2 Comparison of Frequency Domain Properties

of Scalar and Multivariable Systems

Naturally, it would be desirable if multivariable systems could

be controlled using well documented SISO control techniques. To illustrate

possible problems that may arise using this approach, consider the scalar
feedback control of a plant with two inputs and outputs described by the
transfer function matrix [5.11].

9li(s) 912(s)
Gis) = 5.1
.....
921(s) 922(s)
 - 167 -

where the elements of the matrix represent the transfer functions between

specified inputs and outputs. The notation, gmn(s), refers to the

relationship between the n'th system input and its m'th output.

Now, the transfer function relating the second output to its

associated input with the top feedback loop closed and the bottom loop

open, shown in figure 5.2, is determined to be

(s) .
u2 _ [922(s) - 912(s). K1. [1 + K1. g11(s)] 921(s)] .....
5.2

If the top feedback loop gain is high, as is generally the case for set

point following, then equation 5.2 becomes,

(s) 912(gl(S)J
(s)
= -
u2
L922
det G(s)
_ 9 5.3
1- .....

where det G(s) is the determinant of the transfer function matrix. Clearly,

an extra transmission term, due to the action of the top controller, has

been added to the direct transference g22(s), and it is this term that
introduces considerable difficulties in control terms.

Repeating the above analysis with the bottom loop closed and the
top loop open, the transfer function relating first input and output can,

for sufficiently large K2, be shown to be

det G(s)
g22 s)

Thus implying that acute multivariable difficulties will arise if det G(s)

has any zero's in the right half of the s-plane, since their presence will
introduce non-minimum phase relationships into the loop transfer functions.

Consequently, the application of scalar techniques for the control of

multivariable systems is generally only possible when interaction affects

are minimal.
It should be noted at this point, that the following work will

only consider the case where the system is square, since if

168 -

U
O
I-
0
d1
.wt
E +)
d)
NC
T fC
N

N
00
.
f0 I-
"r- U

03 CL
>O
"r- 0

C.
EO
tö
Os
4-) 41

'O L
G) +)
"r "r
r- 3
C-
aN
co E
C)
Yt
UU
(0 N
.0
'C7 G)
G) s
G) +-)
N
iO
( 41
r- (CS C
(U S- C)
U C.
Vn V)

LF r G.
Or- O
"r O
V1 I-
"N
rC E
AiO

C "r O
<0 r2

N
L)

cý
"r
LL.
 - 169 -

ym(s) = Gmn(s) . un(s) .....

5.4

and (i) n>m: We have an excess of controls, by selecting the most

appropriate input-output pairings then the others can

then be eliminated from the design.

(ii) m>n: There are not enough inputs to provide unique control.

Therefore we must either ignore or combine some outputs,

or modify the control structure.

Given the definition of a state space model as

x=A. x + B. u
5.5
y=C. .....
x+D. u
then the relationship between the transfer function description of the
system and the state space with no input feedforward terms, i. e. D=0,

can be shown to be
G(s) = C. (sI - A)-1. B

C.adj(sI - A). B
_ 5.6
et s- es...

If only input-output information is known of a system then it is possible

that it may contain internal behaviour which is either undesirable or of

concern. In addition, such internal behaviour may be uncontrollable by

any combination of inputs and/or unobservable in that knowledge of all

outputs is insufficient to deduce the internal state.

Thus a complex multivariable system may be considered to consist

of four partitioned sub-systems, as shown in figure 5.1. Obviously,. the

transfer function description represents only the controllable and

observable portions of the overall system. Controllability and observ-

ability can be assessed by comparing the dimension of their respective

spaces with that of the system state space. Therefore controllability and

observability are assured if the matrices 4c and ýo respectively, are of

full rank (Cayley-Hamilton theorem), where for a m'th order system:
 - 170 -

eo = [B A. B A2. B
....
Am-1.BI
m1
eo = [CT CT.AT .... CT(AT)- ]
.....
5.7

5.2.1 Poles and Zeros of Multivariable Systems

The denominator of the transfer function matrix contains the

poles of the multivariable system and is termed the characteristic poly-

nomial. Expansion of the state space form of the transfer function

described in equation 5.6 using the method detailed in reference [5.12],

reveals
det(sI - A) = det(sI - A)

Thus the poles of the transfer function are also the eigenvalues (A) of

the system matrix. It should be noted, however, that the poles of G(s)

form only a subset of the eigenvalues because of the possibility of pole-

zero cancellations. Extraction of a commondenominator provides

G(s) = 5.8
0(s) .....

where N(s) is the polynomial matrix of the system

and D(s) is its characteristic polynomial

The system characteristic equation is formed by equating the

characteristic polynomial to zero; the roots of this equation, its eigen-

values, can be used to study stability (see chapter 4).

Decoupling Zeros

In the scalar case the system zeros determine the residues and
hence the proportion of the system modes which will be observed in the

output. The modes themselves are determined by the manner in which the

system is driven. Cancellations can occur by particular ways in which

 -1'1-

the output is formulated or by arranging input paths to bring about

equal and opposite effects at the output (for example disturbance

feedforward). The zeros which appear in the transfer function and cause

these effects are referred to as input-output decoupling zeros.

Similar, but more complicated, effects are observed in multi-

variable systems. By suitable choice of the B and C matrices of the

state space system, zeros can be placed at will in the complex plane.

Different arrangements for coupling into the system and for extracting

information from it leads to the different patterns of system zeros,

which represent the system modes and the external environment. That is,

the decoupling zeros consist of the uncontrollable and unobservable

modes which do not appear in the transfer function matrix G(s).

Element Zeros

The zeros of the scalar systems that constitute the transfer

function matrix G(s) are termed element zeros and play no special role
in multivariable theory.

Transmission Zeros

Consider the closed loop system of figure 5.3 with H=I.

It can be derived that

y(s) = [I + G(s). K]-1. G(s). K. r(s) 5.9

.....

where K is a constant precompensator matrix. Then the system's

characteristic polynomial is given by det[I + G(s). K]. By allowing K

to increase without limit then:

det[I + G(s). K] = det[G(s). K] = det G(s) det K .....

5.10
 172

Thus, with increasing loop gains, the closed loop poles approach

the roots of det G(s), that is, the open loop zeros. For the transfer

function matrix the zeros are determined from det N(s) = 0. The finite

zeros of det N(s) are called the transmission zeros of the system.
Eigenvalues of the system which are not present in D(s) have been cancelled

by an appropriate set of decoupling transmission zeros.

5.2.2 Generalised Frequency Response and Spectral Decomposition

The following work attempts to extend the concept of scalar

frequency response theory into the multivariable domain. Before proceeding,

the standard multivariable transfer function notation of McFarlane, in

relation to the block diagram of figure 5.3, is presented as it will form

the basis for the following work.

r(s) = m*1 vector of reference input transforms

e(s) = m*1 vector of error transforms

y(s) = m*1 vector of output transforms

u(s) = n*1 vector of plant input transforms 5.11

.....
K(s) = n*m matrix of controller transfer functions

G(s) = m*n matrix of plant transfer functions

H(s) = m*mmatrix of feedback/transducer transfer functions

Again noting that in this presentation only square systems are

required and therefore n=m. Other terms used to express various

quantities within the feedback scheme are:

Q(s) = G(s). K(s) Open loop transference matrix

T(s) = Q(s). H(s) Return ratio matrix

5.12
F(s) =I+ T(s) Return difference .....
matrix
R(s) = F(s). 'Q(s) Closed loop transference matrix
I Unit matrix
 - 173 -

Fig. 5.3 Block diagram representation of a general

multivariable system
 - 174 -

Note that in the study of matrix manipulators the order of the individual

matrices is of great importance since matrix operations are generally not

commutative. The above matrices form generalisations of the equivalent

scalar functions defined by Bode.

The principle of spectral decomposition is of major importance

in the design of multivariable control systems using frequency response

techniques. The spectral decomposition of the complex square open loop

transference matrix Q(s), of order m, is defined by

Q(s). wi(s) = xi(s). wi(s) for i=1,2 M 5.13
... .....
where the xi(s) and w1(s) are the respective eigenvalues (or characteristic
transfer functions) and eigenvectors (or characteristic directions) of Q.
Note, that the frequency response of these eigenvalues are termed the

characteristic loci of the system. Thus, for the full system

Q(s) = W(s). A(s). V(s) 5.14
.....
where
W(s) = [wl(s) w2(s) wm(s)] the eigenvector matrix
.....
A(s) = ql(s) 0 0
.....
0 q2(s) 0
.....
the eigenvalue
.

matrix

00..... qm(s)

V(s) = W(s)-1 the reciprocal eigenvector matrix

In assuming that the characteristic transfer functions of Q(s) are distinct,
then the set of column vectors W(s) and row vectors V(s) will form the

basis of an m-dimensional complex space. The matrices W(s) and V(s) are

usually referred to as the eigenvector frame and dual eigenvector frame.

Using this characteristic decomposition it can readily be shown

[5.1] that, for unity feedback, H(s) = I, the closed loop transference
 - 175 -

matrix R(s) may be written as

rI qi(s)
R(s) = W(s). diag i} V(s) 5.15
l1 + gi(s)1 . .....

Thus, the relationship between the open loop and closed loop characteristic

transfer functions is exactly the same in both the scalar and multivariable

cases. Note also, that both characteristic transfer functions have the

same eigenvector frame W(s).

The principal of spectral decomposition also provides a

generalisation of the Nyquist stability criteria in the multivariable case.

5.2.3 Stability Analysis

It has already been established that a multivariable system is

open or closed loop stable if the corresponding characteristic polynomial

has no singularities in the right half of the s-plane (stability analysis

of chapter 4). Therefore the generalised Nyquist stability criterion may

be defined as:

A multivariable system is closed loop stable iff, the net sum of

counterclockwise encirclements of the -1, jO point by the characteristic

loci of the return-ratio matrix T(s) is equal to the number of open loop

unstable poles. For the case of open loop stability the net sum of

encirclement must be zero.

The above holds for all systems that have no open loop

uncontrollable and/or unobservable modes existing in the right half plane,

since they cause discontinuities in the D-contour map.

5.2.4 Interaction Analysis

Inspection of the closed loop transference matrix, equation

5.12, shows that in the unity feedback case system interaction can be
 - 176 -

supressed by simply making the characteristic loci of the open loop

transference matrix Q(s) moduli as large as possible implying:
R(s) >I

However use of this particular method at medium and high

frequencies violates the system's stability and noise sensitivity

requirements. Furthermore, for most practical systems the elements of

the transfer function matrix, and hence of the characteristic loci, will

all tend towards zero as frequency increases. Let us consider the open

loop transference matrix over the frequency ranges where the use of high

gain is impractical, here using spectral decomposition

Q(s) = W(s). A(s). W(s)-I 5.16

.....
where A(s) is the diagonal matrix of eigenvalues
W(s), W(s)'1 are the corresponding eigenvector frames

By aligning the characteristic loci of Q(s) over this frequency

range using eigenstructure compensators then

A(s) = a(s). I

The eigenvalues x(s) can now be considered as a scalar over the

frequency range of interest, thus

Q(s) = a(s). I. 4J(s). W(s)'1

= x(s). I
Q(s) = A(S) 5.17
.....
Hence the open loop system transfer function is the same as its

eigenvalues over the frequency range of interest.

A method of evaluating the accuracy of this high frequency

alignment is provided from spectral decomposition theory, where it is

shown that the characteristic directions of the open and closed loop

transference matrices are the same. Since for minimal closed loop inter-

action the matrix R(s) must be diagonally dominant, then either the

characteristic gains are almost equal or, that one of the characteristic
 - 177-

directions of R(s), and therefore Q(s) must be a standard basis vector

e1, where
E= Ce1,e2,.. ej,.. em] =1..... 5.18

Therefore, a convenient way of expressing how good the eigen-

vector alignment has been over this frequency range is possible using the

misalignment angle [5.131

Wi(s), ei
cos ci(s) =sýý..... 5.19
Iw
-i
where small angles imply good alignment.

5.2.5 System Integrity

The eigenvalues of a system will not give a good description of

its behaviour unless the corresponding eigenvectors are an orthogonal set.
This is known as the problem of skewness and will cause loss of information

regarding the system's gain and phases when the system's eigenvectors lie

in similar directions, possibly invalidating the design specifications.

However by using the vector set at right angles to the eigenvectors, and

noting the gain and phase changes of this set as it passes through the

system a full picture is gained which can then be used to assess a system's

design integrity. This vector set is called the Singular Value Decompositions

(SVD) of the system.

The maximumand minimum boundaries of the singular values as they

vary with frequency form the system's Principal gains (d) and phases. The

boundaries illustrate possible gain/phase variations within the system.

Considering the principal gains then

6.t(S) < < arn(s) 5.20

ü(5) ,....

if the characteristic loci and principal gains of the open loop transference

matrix deviate little from each other then we have a high integrity system,
 - 178 -

large deviations showing problem areas. This concept is therefore

extremely useful in the study of the performance of the system to satisfy

the specific design criteria [5.14].

5.2.6 Robustness and Sensitivity

Robustness is generally defined as the ability of a control

system to remain closed loop stable whilst being subjected to the effects

of plant parameter variations, modelling errors or neglected dynamics.

This can be assessed using the classical approach of M-circles and
Gerschgorin rings to define the regions of uncertainty that arise from

these variations. However by formulating the plant variation matrix

AG(s) then the robustness of the system to these changes can be evaluated

via its relative stability matrix [5.15]. That is, the transfer function

relating the disturbance input AG(s) and its effect upon the output of the

closed loop system, the principal gains of this matrix identifying the

uncertainty bounds of the system and hence illustrating the effect of plant

variation on system stability.

By determining the principal gains of the transfer functions

relating any disturbance input to the process output the system sensitivity

bounds can be calculated, thus permitting a full sensitivity analysis of

any multivariable process. It should also be noted how similar the scalar

and multivariable requirements are. For example, consider the case of

parameter sensitivity. In scalar systems with unity feedback good

sensitivity to small changes in plant parameters is achieved when the

modulus of the closed loop characteristic equation is large over all

frequencies of interest. In the multivariable case this is interpreted as

SME(I + G(s). K(s))-1] <ý..... 5.21

where ý is a large positive function.

 - 179 -

Obviously, as in the scalar case, there is. considerable

conflict between individual requirements. The solution however turns

out to be the same; try to make sm[Q(s)] large at low frequencies and,
dl[Q(s)] small at high frequencies. Figure 5.4 illustrates a general

performance specification which highlights system sensitivity to output

disturbances, sensor noise and set point tracking requirements.

5.3 Frame Alignment Techniques

Since we now have the techniques of analysing the multivariable

system and its associated characteristic locus, it becomes apparent that

the next step is to investigate ways of manipulating the system eigen-

structure and characteristic directions. The following section presents

an overview of the techniques to be used in the final design method.

5.3.1 Commutative Control

If we could gain access to the system's eigenvalues directly

then it would be possible to modify the system's eigenstructure using

standard SISO frequency response techniques. Consider the plant matrix

G(s) with the eigenstructure

G(s) = W(s). A(s). V(s) 5.22

.....
where

A(s) is diagonal and contains the plant eigenvalues

W(s) is the matrix of plant eigenvectors

V(s) is the inverse of IJ(s)

Designing the precompensator K(s) to have the same eigenvector

framework as the plant and the eigenvalue matrix Ak(s), then

K(s) = W(s). Ak(s)"V(s) 5.23
.....
resulting in an open loop transference matrix of
Q(s) = W(s). A(s). nk(s). V(s) 5.24
.....
 - 180 -

Gain

Lower bound for

YLEQ(s)]
L

Upper bound for

Yr[Q(s)]

W1 W2 frequency

Generally

L >_1 High gain at low frequencies required for set point following,
disturbance rejection, and robustness with respect to parameter
changes.
<1 Low gain at high frequencies required for sensor noise rejection
and robustness with respect to unmodelled parameters.
The crossover region determining stability and robustness against gain
and phase changes.

Fig. 5.4 General multivariable performance specification; based upon

system sensitivity to output disturbances, sensor noise and
set point tracking requirements.
Note: The similarity between the definitions in the scalar
and multivariable cases.
 - 181 -

This technique is known as Commutative Control and aligns the

eigenvector frames of the plant and precompensator. This ensures that

the open loop system's eigenstructure is totally dependent upon the

product of the diagonal elements of the matrices of eigenvalues A(s) and

Ak(s), before finally transforming the eigenvector framework back to

that of the original plant.

This concept of directly manipulating the characteristic loci

using SISO techniques is very appealing. However in practice the

formulation of the eigenvector frame W(s) is usually irrational and non-

realisable, even when it is rational its complexity is great.

5.3.2 Approximate Commutative Control

As was stated above, the problem with Commutative control is

the formulation of the frequency dependent eigenvector frames. One

solution is to evaluate a constant matrix that best approximates the

eigenvectors at one particular frequency, say wx. Here

K(s) = Y. Ak(s). Y-1 5.25

.....
where y is approximately equal to W(jwx).

This technique is called Approximate Commutative Control and

behaves in a similar manner to the Commutative controller at and about

the chosen frequency whilst always remaining physically realisable.

Problems can arise when the characteristic directions of the

transfer function matrix are badly aligned with the standard basis

vectors at the required frequencies. In these cases each loci would

have to be modified separately making it difficult to design a non-

interactive system.

5.3.3 Characteristic Direction Alignment

The previous problem of bad eigenvector frame alignment

affecting Approximate Commutative controllers application could be

 - 182 -

avoided if the characteristic directions of the system were aligned.

This can be achieved using the precompensator

K(s) = G-'(s) 5.26

.....
which yields the open loop transference matrix
Q(s) = G(s). G-1(s)

Q(s) =1..... 5.27

Such a controller not only aligns the characteristic directions

but also eliminates all interaction by aligning them to the standard basis

vectors. Because normal systems tend towards zero gain as the frequency

increases, the inverse precompensator will require excessive gain and

phase advance at high frequencies making it physically unrealisable.

Since we are only interested in eigenvector alignment, and not

phase compensation, we can approximate the inverse compensator at one

particular frequency, wX. Thus

K(s) = G"1(jwx) 5.28

.....
results in a real constant precompensator matrix which will achieve the
desired alignment.

5.4 Characteristic Locus Design Method

Utilisation of the tools described in the previous section will

permit us to modify the characteristic loci of any system to some pre-

determined specification, thus dictating the final system's closed loop

response. Such a method is called the Characteristic Locus Design Method.

The original method of McFarlane and Kouvaratakis [5.10] aims to

satisfy the following requirements

(i) Closed loop stability as defined in section 5.2.3.
(ii) Good steady state accuracy and minimum interaction at low
frequencies.
 - 183 -

(iii) Reduced high frequency interaction by alignment of the

characteristic directions with the standard basis vectors.

The above is achieved using the following systematic

approach:

At High Frequencies

Improve the misalignment angles between the characteristic

directions of the plant and the standard basis vectors using a high
frequency precompensator (Kh) to approximate the eigenvector frame.

The resulting real constant matrix would be calculated at the desired

system's bandwidth frequency.

At Low Frequencies

Satisfy Stability criteria and align low frequency gains using

an Approximate Commutative controller (Ks,) at some frequency that is

sufficiently low as to not affect the high frequency compensation already

provided. This objective is helped by the inclusion of integral action

within the low frequency precompensator reducing its affect upon high
frequencies whilst also providing the necessary gain at low frequencies

to satisfy (ii) above.

The result is the final precompensator

K(s) =S Kt + Kh 5.29
.....

where
K(s) = Kjz at Low Frequencies

Kh at High Frequencies

the constant gain a being adjusted to provide the best separation of the

two compensators. Should difficulties with stability margins or bad mid

 - 184 -

range loci alignment exist then a further Approximate Commutative

controller (Km) should be included at intermediate frequencies to provide

the necessary dynamics.

As a result of the analysis presented in section 5.2.4' the

design approach could be extended to not only meet the specification of

figure 5.4, but also align all the eigenvalues upon the desired open loop
frequency response of the system, thus ensuring the form of the closed

loop system's response whilst minimising interaction at all frequencies.

5.5 A. G.I. Characteristic Locus Design

The initial design was applied using the system equations

derived to form the basis of the AGI simulation. The first step in the
design technique is to establish the design specifications. After

consultation with British Gas staff the final design was chosen to

satisfy:
(i) A bandwidth in the range of 0.5 }1 Hz to achieve a fast

response without infringing upon valve/actuator rate limits.

(ii) A damping factor of 0.707 -1-1 to produce a fast response with

minimal overshoot.
Figure 5.5 presents a graphical interpretation of these require-

ments. The next step is usually to check open loop stability: however

since this had already been confirmed (see chapter 4) the design could be

initiated by assessing closed loop stability using the Generalised Nyquist

stability criterion presented in section 5.2.3, providing no encirclements

of the -1, jO co-ordinate occurred. The design exercise required the

following three stages of compensation.

(i) High Frequency Compensation

The first stage of the design is to align the eigenvector frames

at a frequency equal to or slightly above the desired system bandwidth,

 - 135 -

9io(w) rads/sec
IE

-vu
. \\

0
loglo(w) rads/sec
IE` IE-2 " ý"ý IE2

wo,

1\1,

J
-180 Phase (degrees)

-"- Characteristic loci of the uncompensated AGI station

, Boundaries of the performance specification translated

into the open loop

Fig. 5.5 Comparison of the characteristic loci of the uncompensated

AGI simulation and the necessary open loop specifications
to achieve the required closed loop performance.
 - 186 -

thus avoiding any future problems associated with the inclusion of

Approximate Commutative controllers. Figure 5.6 shows the effect upon
the characteristic loci of the system when the following precompensator

was applied.
Kh = G"1(jlO) 5.30
.....

(ii) Intermediate Frequency Compensation

Since 10hwas chosen as the approximate bandwidth of the system

then great care must be taken not to violate the alignment already achieved

at this frequency, therefore any further compensation to be applied at

frequencies lower than wh must approach the unity gain matrix at this
frequency, i. e.

lm(s) I 5.31
.....
1s1 Wh
Compensation at an intermediate frequency is normally only

necessary if the low/high frequency compensators are sufficiently spread

as to affect alignment and/or stability margins. Since in this particular

case we are also attempting to shape the loci into some stipulated form

further dynamic compensation is necessary to achieve the required alignment.

Inspection of figure 5.6 showed that dynamic compensation was

needed in two distinct regions.

1. The high frequency alignment beyond 111h

required the inclusion of

a simple lag to loop 2 which satisfied equation 5.16 to ensure

that both loci have the same roll off rates.

2. Phase advance was necessary to shape the loop 2 loci into the

desired form at the intermediate frequencies.

This was achieved using an Approximately Commutative controller

Km(s) at a frequency wm of 0.1 rads/sec, sufficiently low as to not affect

 187 -

Log gain

requency
rads/sec)

Phase

Frequency
(rads/sec)

Fig. 5.6 Characteristic loci of the open loop transference

matrix with high frequency alignment compensation
 - 188 -

Kh and the eigenvalues

1 0
nm(s) _ 5.32
.....
0

producing the characteristic loci of figure 5.7.

(iii) Low Frequency Compensation

In order to achieve good steady state accuracy and interaction

suppression at low frequencies the inclusion of integral action is

necessary. By employing it in the form of PI controllers further shaping

of the loci would be permitted by suitable positioning of the remaining

zero. Again the dynamic compensation would have to be arranged in such a

way as to avoid the disruption of the previously designed compensators,

i. e.

KR(s)I I..... 5.33

I5I; i ým

Thus an Approximate Commutative controller was employed at a

frequency wt of 0.001 rads/sec with the eigenvalues

s+1.0 0
s
AL(s) = 5.34
0s+0.85 .....
s

Because the high frequency alignment of the eigenvector frame-

work had been chosen to take place at a frequency higher than the desired
bandwidth, allowing attenuation affects to reduce the effect of any

alignment problems greater than this frequency, slight retuning was

necessary. The application of a diagonal gain compensator K to the now

diagonal system (figure 5.8) had the desired affect, producing a system

whose eigenvalues, and hence SISO open loop transfer functions were
 - 189

Log Gain

requency
rads/sec)

.\

Pulse

Frequency
(rads/sec)
IE-3 1 IE3

Fig. 5.7 Characteristic loci of the open loop transference

matrix with high and intermediate frequency compensation
 - 190 -

Log Gain

Frequency
(rads/sec)

Phase

Frequency
(rads/sec)

Fig. 5.8 Characteristic loci of the open loop transference

matrix with full frequency compensation
Note: The alignment of the system's eigenvectors
to satisfy a performance specification of
the form illustrated in Fitt. 5.5.
 - 191 -

approximately:
3.1
Xi (s) 5.35
s. Es 0.16 + .....

obtained via the open loop forward path transfer function

Q(s) G(s). Kh. [y Am(s). y ý. [YQ. Ap (S). K 5.36

= . YQ-1]. .....
The corresponding closed loop step responses displayed in figure

5.9, clearly satisfy the design specification. The integrity of the design

exercise was assessed using the principal gains of the system described in

section 5.2.5. Figure 5.10 shows the system to have high integrity to

process gain and phase changes as there are no major discrepancies between

the eigenvalues and principal gains of the system.

5.6 Implementation of Characteristic Loci

Designed Controllers

The prescribed techniques obvious advantages are its simplicity

and reliance upon commonfrequency response manipulation techniques, its

one disadvantage being the practical implementation of the resulting

design. As is shown by equation 5.37 below, the controller tends generally

to be of large order, in this particular case third order.

Ncli(s) Nc12(s)
Gc(s) =
LN21cs)
Nc22(s) .....
5.37

cs
where
Ncll(s) = 1.4606s3 + 34.022s2 + 39.763s + 3.1924

Nc _ -0.9315s3 + 1701.8s2 + 1461.6s + 16.301

12(S)
Nc21(s) _ -2.9564s3 - 13.633s2 - 11.865s - 5.6127

Nc22(s) = 1.8856s3 + 1909.2s2 + 1646.9s + 23.738

Dc(s) = S3 + 10.2s2 + 2. Os

Since a prerequisite of the AGI design exercise was that the

final form of the controller be easily understood and implementable by
 192

Nominal pressure
psi Kn/m2,

6%

4I_.. Interstage volume pressure

_
0.5- I
Pipeline inlet pressure
.-.

21
Time
0"
seconds)
a. SP1 only

Nominal pressure
psi 1 Kn/m21

0.5.

0 Time
. ý, / 24 (seconds)

b. SP2 only

Fig. 5.9 Evaluation of the characteristic loci designated

compensators performance; time records illustrate
the closed loop system's response to 1 psi changes
in set points, at the designed station throughput
(Q =5 Pisft3/hr)
 - 193 -

Log Gain

nd phases

Frequency
(rads/sec)

Pulse

Frequency
(rads/sec)

Fig. 5.10 Evaluation of the integrity of the characteristic loci

compensated system
Note: High integrity is illustrated by small deviations
between the system's eigenvalues and principal gains.
 - 194 -

the plant operator a procedure was required to simplify the results.

These requirements could be met by minimising the controller to a multi-

variable P+I format. The following method outlines a solution which

would maintain all the design specifications already attained. Should

the dynamics of the controller prove too complex for this simple

modelling procedure then a more complicated controller would be required.

(i) Obtain the real matrix necessary to maintain diagonal dominance

at the bandwidth frequency by frame alignment.

(ii) Best approximate the individual elements of the controllers

transfer function matrix frequency responses by careful

consideration of the PI controllers zero's location, given the

gains dictated from stage (i).

Using a frame alignment controller at the bandwidth frequency G-1(j3.1),

the previous controller was minimised using steps (i) and (ii) above to
2.11(s + 1.0) 177.1(s + 1.0)'
s s
GC(s)* _ 5.38
.....
-2.472(s + 1.3) 198.3(s + 1.0)
s s

Figure 5.11(a) and (b) compares the original multivariable

compensated system and its minimised equivalents closed loop Nyquist

arrays. The corresponding time responses to 5 psi set point changes of

both process outputs simultaneously, as presented in figure 5.11(c) and

(d), show their similarity and hence the success of the proposed

minimisation technique. Figure 5.12 shows the practical implementation

the form of the controller (equation 5.38).

of minimised

5.6.1 Validation of the Multivariable PI Control Scheme

Because the design nethod relies upon the system behaving in a

linear fashion, any nonlinearities that are present can cause problems.
 - 195 -

paa1

maginary

-i. 1.

-0.5

a. Full multivariable control

Real

Imaginary

-1.0
0.5

-0.5 1 0.5

-0.5
-1.0

b. Minimised PI control

Fig. 5.11 Effects of ninimisation upon system performance;

(1 Of 2) closed loop Nyquist arrays of the original system
and its minimised equivalent.
 - 196 -

Nominal pressure
psi Kn/m2

ý. ý
1.0 "ý. _
6

4 lume pressure
0.5
t pressure

00 Time
Y (seconds)

c. Full multivariable control

Nominal pressure
psi Kn/m2
1.0

4
0.5 I

2I

0" 0 Time
(seconds)

d. Minimised PI control

Fig5.11 Effects of minimisation upon system performance;

(2; f ) time records illustrate the response of the closed
loop systems to simultaneous 1 psi changes in set
points whilst passing 5 Msft3/hr.
- 197-

a)
r-
.n
i
>
I
I
E
N
t
4-+

O
C
O

rC
4)
C
a)
E
a)
r-
a.
G
"r

r0
U
"r
4)
U

S
a
(D
Q) E
G)
.C
4-3
U
4- N
0
UO
"r S-
(0 C
E0
G1 U

N a.

rn
z
a U-
 - 198 -

Analysis of the AGI system has identified two cases which could cause

some concern. Firstly, the nonlinear relationships describing regulator

operation. As a consequence of this behaviour the system has been shown

to be operating point dependent (see chapter 4) and can therefore be

categorised by the function

f(n, s) =j f(s) 5.39

1 .....

where

n is the number of linear operating regions

s depicts frequency dependence

A problem of this type is normally solved by designing

controllers at the different operating conditions and applying some

form of gain scheduling to the final controller to account for these

variations.
Secondly, the presence of actuator saturation, although

presenting no real problem except a loss of linearity in the loop

concerned, would go unnoticed by the other loops which, expecting linear

behaviour, may overcompensate the system. Avoidance of this phenomenon

is possible by

a) Providing more information about the system's condition

using sensors.
b) Avoiding rapid changes in actuation by slugging the

response and/or designing the controller with actuation

limitations/sensitivity in mind, the latter being

employed in section 5.5 as part of the closed loop

specifications.

Since the operation of an AGI station is dependent'upon

seasonal loadings then the behaviour of the multivariable controller

described in equation 5.38, designed for a flow of 5 P1sft3/hr, was

 - 199 -

investigated throughout the stations working range. The working ranges

of these stations are normally not limited by the regulator capacities,

which are normally vastly oversized, but by the flow metering. Typically

AGI's operate with orifice plates which are rated between 15 and 25 Msft3/hr,
the minimum station flow being determined by the flow measurement devices

accuracy and reliability and is generally accepted to be approximately 7%

of maximumrated flow.
Application of the multivariable controller to the nonlinear

system at these extremes of operation provided the responses displayed in

figures 5.13 and 5.14 to 5 psi changes in regulator set points. Inspection

of these results indicates that there is no significant degradation in

controller performance over the working range, and therefore no need for a

more complex adaptive control scheme to compensate for the plant nonlinear-

ities.

5.7 Tuning of a Multivariable PI Controller

The extension of the classical frequency response design

techniques already used and well understood in the gas industry, into the

multivariable domain provides a solution which has been shown to satisfy

the performance requirements. The fact that this has also been achieved

using a minimised form of the characteristic locus design, the multivariable

PI controller, concludes that the system dynamics at the frequencies

determining stability must be relatively uncomplicated (see section 5.4).

Based upon this knowledge, the following section presents a quick and

effective alternative method of tuning the PI controllers using time response

techniques.

Providing that the multi-input multi-output (MIMO) system is open

loop stable then Davison [5.15] showed that any m*m system described by

the transfer function matrix G(s) and unit step response matrix y(t), t>0,
 - 200 -

Nominal pressure
psi Kn/m2
,. ý
6
A^
4u

I
4
--- Set point
I-
20 "- Interstage volume pressure
2" Pipeline inlet pressure

0 Time
o. 23 (seconds)

a. SP1 only

Nominal pressure
I
psil Kn/m2

0 123
-. -Time
(seconds)

b. SP2 only

Fig. 5.13 Evaluation of the multivariable PI control schemes

(1 of 2) performance over the AGI stations full working range;
time records illustrate the system's step response to
5 psi set point changes whilst passing minimum station
flow (Q =2 Ftsft3/hr).
 - 201 -

Nominal pressure
psi j Kn/m2j

64

4
--- Set point
21 Interstage volume pressure
...
2 Pipeline inlet pressure

0 Time
0J (seconds)
23

c. SP1 and SP2

Fig5.13 Evaluation of the multivariable PI control schemes

(2; f ) performance over the AGI stations full working range;
time records illustrate the system's step response to
5 psi set point changes whilst passing minimum station
flow (Q =2 Msft3/hr).
 - -202

Nominal pressure
psi Kn/m2

6
40

`F
/" Set point
---
/
20 -"- Interstage volume pressure
2I -"- Pipeline inlet pressure

0J 0 Time
.
(seconds)

a.. SP1 only

Pipeline pressure
psi Kn/rn2

6
40

4-

20

00 Time
123 (seconds)

b. SP2 only

Fig. 5.14 Evaluation of the multivariable PI control schemes

(1 of 2) performance over the AGI stations full working range;
time records illustrate the system's step response to
5 psi set point changes whilst passing naximim
station flow (Q = 20 flsft3/hr).
 - 203 -

Nominal pressure

psi I:n/n2
6
40

4.20
--- Set point
/"
-"- Interstage volume pressure
2/ Pipeline inlet pressure

0- Time
0 (seconds)

c. SP1 and SP2

Fig. 5.14 Evaluation of the multivariable PI control schemes

(2 of 2) performance over the AGI stations full working range;
time records illustrate the system's step response to
5 psi set point changes whilst passing maximumstation
flow (Q = 20 P.
1sft2/hr).
 - 204 -

steady state performance can be described by the data

G(0) = Y(03)

where G(O) is the transfer function matrix when s -,. 0, and

y(am) is the steady state gains of the time response matrix.

If G(O) is nonsingular then a stable feedback scheme can be

successfully tuned on-line using the unity negative output feedback

controller

Gc(s) = G(0)-1 5.40

.....

where c is a tuning constant selected by on-line tuning.

This work was extended by Astrom [5.16], Penttinen & Koivo

[5.17], Porter [5.18] and Owens & Chotai [5.19] into the more dynamically

flexible proportional plus integral controller form

CS7
Gc(S) = [P] + 5.41
0....

All of the above techniques basically involve the determination

of the initial slopes and steady state gains of the system's open loop

step responses, therefore avoiding any of the practical problems that may

be encountered when attempting multivariable identification and estimation.

From these tests, the proportional and integral gain matrices, P and I

respectively, can be-formulated via the equation

1
lullul-

5.42
.....
py(-)Icul-! -,. l
[I] =

where
[ý] is a matrix of initial gradients of the step responses

[y(am)] is a matrix of steady state values of the step responses

[u] is a matrix of applied step inputs

Iij are diagonal tuning matrices

 - 205 -

Because this particular method relies upon the assumption that the

individual plant matrix elements can be approximated using first order

models then the tuning matrices are included to compensate for any

situation when this is not true. It is in the degree of flexibility

offered by these tuning matrices, see section 6.7, that forms the

essential difference between the above methods. The following work

considers the application of one of these techniques to the linearised

system upon which the characteristic locus design was performed, where

the open loop step response input sequence

i o
[uJ_
o
produced the following step response data
0* 574 -0.512
Cy] =
0.0084 0.0060

1.889 0

2.768 1.792

By placing this data into equations 5.41 and 5.42 with unitary

tuning matrices then the following multivariable controller results

0.77s + 0.53
66.11
s
Gi(s) = 5.43
0.82 74. lls .....
-1.08s - + 0.56
ss

providing the closed loop behaviour portrayed in figure 5.15, whose

responses show good interaction suppression. Further improvement in

performance is possible by manipulating the tuning matrices. Figure 5.16

displays the new closed loop behaviour when both loops speed of response

has been increased, loop 2 by a greater factor to alleviate the problem

of steady state offset, by selection of the tuning matrices settings

2 0
Tj ý
0 5
 - 206 -
Nominal pressure

psi Kn/n2

1-

' --- Set point

Interstaae volume pressure
0.5.
i
1 -"-
Pipeline inlet pressure

Time
5 10 (seconds)

a. SP1 only
Nominal pressure

psi Kn/m2
1.0 --
6

4
0.5

Time
00 'seconds)

b. SP2 only

Fick. 5.15 Performance of the step response tuned multivariable

PI control scheme, unweighted; time records illustrate
the response of the AGI simulation to 1 psi changes
in set point at the designed station throughput
(Q =5 Msft3 /hr)
 - 207 -
Noninal pressure

psi Kn/m'-

Set point
j---
4!
-"- Interstage volume pressure
0.5-
Pipeline inlet pressure

2j

I
Time
o- 1d (seconds)

a. SP1 only

Nominal pressure
1--
psý'ý Kn/M2

0.5.

o" v' Time

---"-"- 5 ---- 10 (seconds)

b. SP2 only

Fig. 5.16 Performance of the step response tuned multivariable

PI control scheme, weighted; time records illustrate
the response of the AGI simulation to 1 psi changes in
set point at the designed station throughput
(0 =5t. 1sft3/hr).
 - 208 -

The procedure of determining the 'optimum' tuning matrices

can be almost as demanding as that of tuning the controllers. Problems

arise when the use of the first order models, upon which the tuning

techniques are based, fail to provide the accuracy required for design

purposes. Owens & Chotai [5.19, Appendix III. 1] describe an off-line

method of evaluating these tuning matrices based upon approximate model

predictions and show the problems that can arise when the design
techniques are applied to processes that exhibit non-minimum phase
behaviour.

5.8 Interaction Measurement and Analysis

The difficulties involved in designing and operating multi-

variable processes are directly attributable to system interaction.

These behavioural changes introduced into one control loop due to

disturbances occurring in any/all of the other loops are mostly

destabilising, although in some cases they can be beneficial to the

closed loop performance of the entire process. The thrust of this

section is aimed at the development of a multivariable performance

criteria based upon both speed of response and minimal interaction.

The study commenceswith a review of current analysis techniques.

In general, a 'good' interaction index should be capable of

providing all the information necessary to,

i) Quantify the effect all loops have on all other loops and

individual loop performance.

ii) Determine whether the interaction degrades/improves closed

loop performance.

iii) Assist with the choice of manipulated/controlled variable

pairings for best closed loop performance.

 - 209 -

The first method proposed [5.20] simply involved the measurement

of the system's steady state gains, by placing the resulting values into

the extended interaction quotient (shown here for a2*2 system) to form

the Relative Gain Array (RGA).

Diagonal P11 P22 GI, G22

- -- G11G22 G12G21
elements -
5.44
.....
Off-diagonal G12G21
elements
P12
12
'2 P21 _ 11 22 - 12 zl

where Pij is the relationship between the j'th manipulated and i'th

controlled variables thus giving some insight into how these variables

could be paired for improved performance. However since this interaction

index bases all of its information upon the system's steady state gains

processes which contain one off-diagonal element of steady state value

zero will be predicted to be diagonally dominant by the interaction

quotient irrespective of the magnitude of its opposite off-diagonal

element. This loss of information could arise from one loop being

decoupled from another or, as is the case with our particular example,

one element containing pure derivative action. A solution to this latter

problem was provided by the Relative Dynamic Gain Array (RDGA) [5.11]

which extended the technique by looking at the system's gains over a

specified range of frequencies. However, it still fails to work if one

loop is decoupled from another.

The Average Dynamic Gain Array (ADGA) [5.22] relies upon open

loop step responses, these being assessed by a suitable performance

such as Integral of Square Errors (ISE), before being placed

criterion
into the interaction quotient, thus providing information about the

system's response as well as a figure of merit.

Great care must be taken if the interaction quotient is to be

on its own since there will be no indication of its failure until

used
the closed loop system is operated. The RDGAand ADGAtechniques provide
 - 210 -

sufficient information to avoid this problem by inspection of the

individual elements behaviour. The above techniques however fail to

provide a direct measure of true closed loop interaction [5.23].

Suchanti and Fournier [5.24] presented a technique based

loosely upon the ADGAwhich compares the step responses of the system

with all/individual loops closed, via the interaction coefficient.

Solving all the above problems with very little computation unlike the

state space technique proposed by Davison and Plan [5.25] which requires
the solution of a set of diophantine equations, restricting its use to

systems having stable matrices throughout the analysis.

5.8.1 Closed Loop Interaction Measurement

From the previous section the method which provided the simplest

means of satisfying the defined objectives was that presented by Suchanti

and Fournier [5.24]. Briefly, this method utilises the interaction

coefficient described by
TEk - IEk
Ik= -- 5.45
.....

where
< Ik <0 defines bad interaction
-1
Ik =0 defines no interaction

0< Ik <1 defines good interaction

Trk is the integral of error of the k'th output for a

unit step in the k'th input with only the k'th

loop closed

IEk is the integral of error of the k'th output for a

unit step in all inputs with all loops closed

It should be noted that for small scale linearised systems or

signals with their bias levels removed Integral of Absolute Errors (IAE)
 211

is recommended[5.26], or if we wish to weight large amounts of inter-

action more heavily then Integral of Square Errors (ISE) [5.27] has

been found to be more suitable.

5.8.2 multivariable Input-Output Pairings

As previously mentioned, the interaction index can also be

used to determine the input-output pairings for best closed loop

performance. The PI controller settings for the normal input-output

pairings have already been determined using the relay tuning method

presented in chapter 3. This process was repeated for the valve 1

input (U1) - pipeline inlet pressure (Pin), valve 2 input (U2) -

interstage volume pressure (Pu) pairings and provided the settings

presented in table 5.1 below.

Normal Pairings Reverse Pairings

Loop 1 Loop 2 Loop 1 Loop 2

(U1-Pu) (U2-Pin) (U1-Pin) (U2-Pu)

Kc 1.53 29.15 133.34 -3.06

Ti 0.57 1.43 0.53 0.37

Table 5.1 SISO controller settings for input-output

pairings analysis

Completion of the necessary step responses (figures 5.17

and 5.15) to 1 psi changes in set point provided the performance figures

presented in table 5.2, using the integral of square errors criterion.

 - 212

Nominal press ure

psi A Kn/M21

5-
25-

A Time
0- 0
(seconds)

-25- pressure
-51 I'
assure
v
-50

-1o
a. Both loops simultaneously
Nominal pressure

psi Kn/m2

2-

10.

1
5j :

Tine
o0 (seconds

-5
-1 b. Upstream loop only (Ui - Pu)

Fig. 5.17 Manipulated and process variable pairings performance study;

(1 Of 2) time records illustrate the response of the AGI simulation
to 1 psi set point variations when subject to the standard
controller format
 213

Nominal pressure
psi Kn/n2
20
10

p Time
10 20 30 ý" 40' 50 (seconds)
.

-10 /" -" -Interstage volume pressure

-20
Pipeline inlet pressure
-200-

`
-40 "/
"'ý
-300
c. Downstream loop only (U2 - Pin)

Fig. 5.17 Manipulated and process variable pairings performance

(2 of 2) study; time records illustrate the response of the
AGI simulation to 1 psi set point variations when
subject to the standard controller format
 21

Nominal pressure

psi Kn/m2

30 200

15 100

0 Time
!I 10 20 30 40 50 (seconds)

-" - Interstacie volume pressure

-15 -100
Pipeline inlet pressure

-30 - -2
a. Both loops simultaneously

Nominal pressure
psi Kn/m2

100

4
50-

T--v- Time
0 10 20 30 40 50 (seconds)

b. Downstream loop only (U1 - Pin)

Fig. 5.18 Manipulated and process variable pairings performance

study; time records illustrate the response of the
AGI simulation to 1 psi set point variations when
subject to the reversed controller format
 - 215 -

Nominal pressure
psi Kn/m2

S. Se -. -. _.
1. S. S. S" -

Time
10 20 30 40 50 (seconds)

_.. __Interstage volume pressure

_1 pipeline inlet pressure

c. Upstream loop only (U2 - Pu)

Fig5.18 Manipulated and process variable pairings performance

(2; f ) study; time records illustrate the response of the
AGI simulation to 1 psi set point variations when
subject to the reversed controller format
 - 216 -

Normal Pairings Reverse Pairings

Loop 1 Loop 2 Loop 1 Loop 2

TS'E 4.603 21.18 4.661 3.364

ISE 924.2 11.92 4.674 6953

Interaction 0.776
Index -0.995 -0.003 -0.999

Table 5.2 Input-output pairing analysis results

As expected, both methods of pairing the manipulated and

process variables (figures 5.17b, c and 5.18b, c) provide satisfactory

regulation of their respective pressures when only single loops are

closed. However, both of the pairings fail to adequately suppress the

interaction when the full closed loop scheme is employed (figures 5.17a

and 5.18a), resulting in poor control of the interstage volume pressure

(Pu). The standard control scheme provides positive interaction in the

of pipeline inlet pressure (Pin), but inspection of the relevant

control

performance indices indicates that it still fails to perform as well as

the reverse pairings method overall. Analysis of the numerical results

presented in table 5.2 illustrates this interaction (large ISE's and poor

indices) and the relative performance of the two schemes.

Inspection of the full closed loop step responses show the

standard control scheme to have greater stability than that of the reverse

pairings method. This is due to the inclusion of the destabilising

interstage volume pressure control loop within the downstream pressure

loop, a fact that is not apparent from the interaction indices alone.

The following section details the application of the proposed

interaction measurement technique to the SISO and MIMOcontrol schemes

developed in chapters 3 and 5 respectively.

 - 217 -

5.9 Discussion

The work presented has developed an intuitive approach to the

problem of designing multivariable systems to closed loop specifications,

requiring only a knowledge of classical SISO frequency response theory.

By applying this method to a linearised AGI simulation a controller was

developed that achieved all the design specifications, at the expense of

considerable complexity of structure. A method of replacing this

controller with a simplified equivalent multivariable proportional plus

integral controller which only slightly degrades the original performance,
but has the bonus of a more familiar structure for the plant operators

benefit, has also been developed.

The resulting linear control when applied to the full nonlinear

system throughout its working range exhibited more than acceptable

performance, negating any necessity to include a more complex gain

scheduling strategy. A direct comparison with the SISO control scheme

normally employed is illustrated in table 5.3, clearly showing the multi-

variable (MIPM10)
controllers improved interaction suppression and speed of

response, even over the PID control scheme developed in section 3.8.

SISO PI/PI Control Plultivariable PI Control

Loop 1 Loop 2 Loop 1 Loop 2

ISt 126.6 303.6 63.8 51.2

ISE 23050.2 296.8 58.2 51.3

Interaction 0.023 0.097

Index -0.995 -0.003

Table 5.3 Comparison of SISO and MIMOProportional plus Integral

Controller Closed Loop Performances at Q=5 Msft3/hr
 - 218 -

Because the design technique relies upon the provision of a

linear transfer function of the system, and at present multivariable

identification and estimation is still fraught with many difficulties

then some other practical means of evaluating the multivariable

proportional plus integral controllers parameters is necessary. To this

end an autotuning technique which required nothing more complex than a

set of open loop step responses was presented and analysed, providing

excellent results when applied to the linear AGI simulation. A solution

which is easily implementable in both continuous time [5.17] and discrete

time [5.16,5.18] forms, the latter being of greater importance practically.

A means of quantifying closed loop interaction has also been

discussed, and developed using the ISE performance index to compare the

various devised control schemes' ability to maintain their set point values.
The technique also provides information regarding the best input-output

pairings to use in the control scheme. Not unexpectedly, the results

supported the scheme presently employed, although great care must be taken

when applying this particular technique since the numerical information it

provides has been shown to be insufficient on its own to base this choice

upon, requiring an expansion of the performance interaction index to

incorporate an assessment of the system's stability margin, possibly by

analysing the closed loop step response decay rates.

The above results are all based upon the theoretical AGI

simulation developed in chapter 2. As an alternative, chapter 6 presents

the results of an investigation into the application of the multivariable

PI tuning technique to a real system. This takes the form of an experi-

mental test rig which was also constructed to simulate the behaviour of

an AGI station.
 - 1179
-

References (Chapter 5)

5.1 Patel, R. V., and Munro, N. "Multivariable system theory

and design". Pergamon Press, International Series on Systems
and Control, Vol. 4.

5.2 Pontryagin, L. S., at al. "The mathematical theory of optimal

processes". New York, Interscience, 1963.

5.3 Bellman, R. "On the application of the theory of dynamic

programming to the study of control systems". Proc. Symposium
on Nonlinear Circuit Analysis, New York, Polytechnic Institute
of Brooklyn Press, pp 199-213,1956.

5.4 Kalman, R. E. "The theory of optimal control and the calculus

of variations". Mathematical optimisation techniques, Univ.
of California Press, 1963.

5.5 Luenberger, D. G. "Observers for multivariable systems".

IEEE Trans. Automatic Control, Vol. AC-11, pp 190-197,1966.

5.6 Rosenbrock, H. H. "Distinctive problems of process control".

Chemical Eng. Prog., Vol. 58, pp 43-50,1962.

5.7 Wonham,W. M. "On pole-assignment in multi-input controllable

linear systems", IEEE Trans. Automatic Control, Vol. AC-12,
pp 660-665,1967.
5.8 Rosenbrock, H. H. "Inverse Nyquist array design method".
IEE Control Series 9, Chapter 5, Peter Peregrinus Ltd.

5.9 Rosenbrock, H. H. "Computer aided control system design".

Academic Press, London, 1974.

5.10 McFarlane, A. G. J., and Belletrutti, J. J. "Characteristic

locus design method". Automatica, Vol. 9, pp 575-588,1973.

5.11 McFarlane, A. G. J. "Multivariable feedback theory notes".

UPIIST.

5.12 Sinha, P. K. "flultivariable control - an introduction".

Marcel Dekker Inc., Electrical Engineering and Electronics/19.

5.13 McFarlane, A. G. J., and Kouvaritakis, B. "A design technique

for linear multivariable feedback systems". Int. J. Control,
Vol. 25, No. 6, pp 837-874,1977.

5.14 Postlethwaite, I. "Robustness in multivariable control system

design". IEE control series 20, Chapter 3, Peter Peregrinus
Ltd.

5.15 Davison, E. J. "Multivariable tuning regulators: the feed-

forward and robust control of a general servomechanisms
problem". IEEE Trans. Auto. Control, Vol. AC-21, pp 35-47,
1976.

5.16 Astrom, K. J. "A robust sampled regulator for stable systems

with monotone step responses". Automatica, Vol. 16, p 313,
1980.
 220
1-

5.17 Penttinen, J., and Koivo, H. N.

"Multivariable tuning regulators for unknown systems".
Automatica, Vol. 16, pp 393-398,1980.

5.18 Jones, A. H., and Porter, B. "Design of adaptive digital

set point tracking PID controllers incorporating recursive
step-response matrix identifiers for multivariable plants".
IEEE Trans. Auto. Control, Vol. AC-32, No. 5, pp 459-462,
1987.

5.19 Owens, D. H., and Chotai, A. "High performance controllers

for unknown multivariable systems". Automatica, Vol. 18,
No. 5, pp 583-587,1982.

5.20 Bristol, E. H. "On a new measure of interaction for

multivariable process control". IEEE Trans., Vol. II,
pp 133-134,1966.
5.21 Witcher, M. F., and McAvoy, T. F. "Interacting control
systems: steady state and dynamic measurement of interaction".
ISA Trans., Vol. 16, No. 3, pp 35-41,1977.

5.22 Gagnepain, J. P., and Seborg, D. E. "An analysis of process

interactions with applications to multiloop control system
design". Ind. Eng. Chem. Process Des. Dev., Vol. 21, pp 5-11,
1982.

5.23 Jensen, N. "Commentson McAvoy.(1983)". Ind. Eng. Chem.

Proc. Des. Dev., Jan. 1985.

5.24 Suchanti, N. C., and Fournier, C. D. "A new algorithm for

pairing manipulated and controlled variables". ISA Conference
Instrumentation in Chemical and Petroleum Industries, St Louis,
p 67,1973.

5.25 Davison, E. J., and Man, F. T. "Interaction index for

multivariable control systems". Proc. IEE, Vol. 117, pp 459-
462,1970.

5.26 Ogata, K. "Modern control engineering". Prentice-Hall, 1970.

5.27 Shinskey, F. G. "Process control systems". McGraw-Hill Book

Company, 1979.
 - 221 -

Nomenclature (Chapter 5)
A Multivariable Approach to the Design of
AGI Control Systems

adj[. ] Adjoint of a matrix

A State space matrix

B State space matrix

C State space matrix

det[. ] Determinate of a matrix

D State space matrix

D(s) Characteristic polynomial of plant

Dc(s) AGI compensators characteristic polynomial

e(s) Vector of error transforms

ei Standard basis vector

E Standard basis matrix

f(s) Frequency dependent function

F(s) Return difference matrix

gla(s),.. Plant transfer function elements

G(O) Steady state gains of plant transfer function matrix
G(s) Plant transfer function matrix

Gc(s) AGI precompensator transfer function matrix

H(s) Feedback/transducer transfer function matrix
I Identity matrix

[I] Integrator gain constant matrix

Ik Interaction coefficient of the k'th loop

IAE Integral of absolute error performance criterion

IEk Integral of error of k'th loop with all loops closed

IEk Integral of error of k'th loop with k'th loop closed

ISE Integral of square error performance criterion

K, Kh Constant precompensator matrix

K(s) Precompensator transfer function matrix

 ZZZ

N(s) Polynomial matrix

Ncll(s),.. AGI precompensator transfer function matrix elements

[P] Proportional gain constant matrix

in Pipeline inlet pressure (kN/m2)

Pu Interstage volume pressure (kN/m2)

P11900 Interaction quotient elements

qi(s) Eigenvalues of open loop transference matrix

Q Station volumetric flowrate (M3/sec)

Q(s) Open loop transference matrix

r(s) Vector of reference input transform

R(s) Closed loop transference matrix

s Laplace operator

T(s) Return ratio matrix

u Plant input

[u] Matrix of step response input magnitudes

u(s) Vector of plant input transforms

vi(s) Dual eigenvector frame matrix element

V(s) Dual eigenvector frame matrix

Vi(s) Eigenvector frame matrix element

W(s) Eigenvector frame matrix

X Vector of the states of a system

Vector of the derivatives of the states

y Plant output

Initial slopes of step response matrix

y(s) Vector of output transforms

y(am) Steady state step response gain matrix

Proportional/Integral action decoupling constant

61(s) Minimum principal gain

öm(s) Maximumprincipal gain

Integrator tuning constant

C
 -223 -

I Integral gain tuning matrix

ýc Controllability matrix

ýi Misalignment angle

ýo Observability matrix

a(s) Eigenvalue of a matrix

A(s),.. Eigenvalue matrices

Y Constant matrix approximation to the eigenvector frame

n Proportional gain tuning matrix

w Angular frequency (rads/sec)

11.1! Norm of a vector

 - 224 -

CHAPTER
6

OF CONTROL
PRACTICALEVALUATION TECHNIQUES

USINGAN EXPERIMENTAL
TEST RIG
 - 225 -

6.1 Introduction

It is well known that the analysis of nonlinear systems is

fundamentally difficult for all but the simplest of special cases.
Features which identify the behaviour of these systems include limit

cycles, chaos, catastrophes and switching and sliding. Any information

that can be gathered about the operation of such a system prior to any

analysis is of great advantage.

One approach, often favoured by engineers, for obtaining such

supplementary information about the system is by the construction of an

experimental test rig; the test rig providing the hardware which can then

be used to determine the numerical values of the constants and other

parameters needed for system identification. In addition, the test rig

provides a means of verifying the various modelling exercises and also

allows the results of any analysis to be evaluated using real data. In

the case of the gas industry this provides the first stage of a stringent

vetting procedure which must be undertaken before application to the

actual transmission system can be considered.

The following chapter considers the construction of an experi-

mental test rig whose structure is sufficiently flexible to represent both

Above Ground Installations (AGI) and Pressure Reduction Stations (PRS),

utilising as many of their actual components as is practical. The structure

of the test rig is described and experimental work is presented to support

the mathematical simulations developed within chapters 2 and 4.

6.2 The Experimental Test Rig

The basic configuration of an AGI station is largely similar to

that of a PRS system, the major difference being the variable which is

controlled by the downstream regulator and the operating conditions

 - 226 -

encountered (system pressures and flows) because of the pipework and

the location in the transmission system. In the experimental test rig

these system pressures are scaled between 0-100 psiG, giving flowrates in

the range of 0-100 litres/min in pipework of quarter inch diameter.

For ease of operation the test rig was constructed in two

distinct sections, such that the inlet pressures can be controlled using

separate inlet regulators and isolated using solenoids. The main section

consists of the working stream whilst the second section provides the
instrumentation supplies for the electro-pneumatic convertors required to

drive the control valves. The test rig layout is presented in figure 6.1.

Basically, the main working stream consists of two series control valves

feeding a downstream volume via a pipeline, designed to simulate down-

stream conditions.
The test rig is fitted with a variety of transducers, shown

schematically in figure 6.2, which monitor supply pressure (P5), valve

interstage pressure (Pu), pipeline inlet pressure (Pin), downstream

loading pressure (Pd), valve actuator chamber pressures (Pactl, Pact2),

valve stem displacements

(X1, X2) and mainstream volumetric flow (dPq).

The output signal from each transducer is passed through a signal

conditioning unit, designed and constructed using operational amplifiers,

an output in the range of 0-10 volts which can then be recorded

providing
8
using an channel, 12 bit data acquisition system interfaced with a BBC

microcomputer. Further details with regard to the specification of the

test are tabulated in figure 6.3.

rig components

6.2.1 Test Rig Operation

Although the basic construction of the test rig is fixed there

is sufficient flexibility in the design to mimic the operation of the

of AGI systems. This flexibility is brought about by the ability

majority
 - 227 -

xi
Main
regulator Pact 1 Pact 2
and filter
Dal1
tube
F-M

Pressure Control Pd
transducer valve
0

Compressor Receiver bottle

I/P converter
Outlet
flow
Actuation

11
Stream Instrumentation estrictor
inputs £
m isolation regulator P11 M2
0 solenoids

S2

Ps Pu Pin
dPq

Fig. 6.1 Schematic diagram of the AGI experimental test rig;

giving the approximate locations of the various
measurement transducers.
 228
N
V
'p "r
iE
(C ft}

O (L) 'C3
+3 E
r0 OE
r- (0
OO G)
C) > S.
a) +3
OC
'v 4-" 3
rts4- 0
o =-v
J 12 --
0
U
Q- (A "C
3 U, "r-
04--)
ý
N Q) \
0
0 OO
0
N EO
CY)O +)
"r UN
4- E
ä to "rto
"w
"C OÜ
"r "r
5.. s. Q)
O
O {-) >
Kr CA

i C.
C 4)"r
IQ
d fo 4. -
O +- 0-0
L7 CC
r c) (
N E
X "r "r N

N OO S-
M-rd-1
C X +) N
V d"r E
V) "r
tö
O -v
p
U')
ca r-ý 0
C7 Q. r-
r
04-) d
4-J ý
"r
'v E
G. to = aý c
Cif
c cý"ý i-) "r
O aJ i to N
t1') E 0- 4F- i-' CJ
OC i-
O
Xý C. N EE
VI CL r0 ON
C7O i i. +3
a+ {/) .)
ýý c: =C
U r0 V) N
ro "r (ý E
d 0 'II N (1)
E
r- U
"r-NG N
"1-' S.. (CS
CL co O C1
0 EE
00 O G1
N
U"re-
(I) ad
N
NN
N
O
N
rý

LL.
ON 001
O

O
4)
i
OV
N Q1
N +-"i
O> fC N
L r- r- a
a0.
EE a vno
c,.) (n L r-
 - 229 -

Qý "r
C "r U1
"r r U)
ö ä E

+>
r N- cl
O tC " U)
C ý- lo O N
O
U O O O

0 0 U U
O O 0 0
N- N- N- N- LL) N 00 CV 00
N t0 - ti7=\ "_\ "\
C ". O " -O O. O OO
O O C C )
" ý
qcr
^
Iti'
N I +1 11 L
4)
t0 N
a
"r O N O 4)
4- C71 CY) 0)
U +-) 4.) ( +. +3 r0 i-) +) ( 4-) +-) (
C) "r"L70 "rTT= "r'a = -0 = 43
S. 'r. c t- i "
ý N (o 3: " 3 3 cC 3 nJ
" 4) "O C. G) - C. O Z7 C. O- Cl. E
CCE CC CCE CC
"r (0 (L) "r f; G1 "r (o Q) "r fTs (L) i
Jm I- I- m I- M F- N
-i . -J -J
N CL
GJ X
I-- N
a)
r U U Q U
"r p p E C3

C CL O
O C. CV N
N Lr) LO t
(V 4-3

O C
G. G') G1 0
X "r" "r i
L. J Q) (A N (d
G)
C C) C. a E
P--4 Pr- G-
CD r_ cz LO C)
Q -14
i IC O In O

y I I I I Z
C
O CD O O O
U

'oa) V M
N
N G.
ifo
r- GJ C
M i i
"r
(t d
QJ "r" V) w .-N N
o SCC O X
(TJ W CL 4-) 41

Vf (0 fö d
Vf
X '4-
C --
b O

Y N

O
O
co

U
L
S-
O LL- i
N y \O U
U) cu CD L)
S- d
S- N
p" C- N \N
(1) N Q- C
114
"r co m M
L7 0. ý-
> C7
(1) N CQ O U
p O r" 4. )
co . -" (C r- C
co
i S- "r i "r G)
X C) X a) X 4. ) O C7 E
J ä LL
: O= oö u
,N -0 S-10 (0
N C) N > 1 r-
0 r_ o r- U 4- a C: N
C f0 C to 4- (a C
ä
NF- NF" Oc F- cm
 - 230 -

to vary: -
i) Mainstream flow conditions.

ii) System pressure profiles.

iii) Pipeline dynamics.

iv) Valve characteristics.

In order to set up a particular simulation the following

procedure is followed. Firstly establish the required supply pressure

by the appropriate manipulation of the working stream's inlet regulator.
The supply pressure is held constant during each test by the use of an

upstream receiver (buffer volume) to represent the large transmission

storage present on the actual system.

Load flows between 0- 100%of the maximumpermissible value

can be set up using a restrictor which adjusts the bleed rate of the

downstream accumulator. By also manipulating the series regulators'

stem positions, the pressure cuts across each regulator can be controlled,

thereby establishing any required pressure profile for the system.

Variable pipeline dynamics are possible by replacing the

flexible link between the flow measurement device and the downstream

volume with one of any desired length. No instrumentation is present

within this flexible link, however, inlet and outlet flows and pressures

can be established by following the steps outlined above.

Finally because the regulator stations utilise various types

and sizes of control valve, then this feature is also required to be

accommodatedby the rig. By using the miniature air operated valve (M

valve) produced by Platon Flowbits Ltd (the M valve is essentially a

scaled down version of the diaphragm operated valve commonly encountered

within the gas industry) the individual regulator characteristics can be

adapted by simply changing the trim to one of the many types available.
 - 231 -

Figure 6.4 shows the manufacturer's data for this particular type of

valve. The 3- 15 psig supply required to close/open the valve is

supplied via a Fairchild T5200 electropneumatic (I/P) convertor,

illustrated in figure 6.5, which in turn requires a4- 20 mA (0 - 10

volt via a voltage to current convertor constructed from operational

amplifiers) and 20 psig (60 psig for the convertor with volume booster

fitted) air supply.

6.3 Static and Dynamic Performance of the Electropneumatic

Convertor-Valve System on the Experimental Test Rig

Prior to any analysis being performed upon the test rig it is

important to establish that the behaviour of an individual element

reflects that of the component which-it is designed to represent. The

following work investigates the linearity and static/dynamic performance

of the valve/electropneumatic convertor interface, to establish just how

similar the behaviour is when compared to that of the AGI regulator. The

relationship between the voltage input and the actuator stem displacement

is of particular importance.

Inspection of the system step responses and steady state

positioning of the standard valve, figures 6.6 and 6.7, indicate the

presence of at least two nonlinearities:

i) The dynamic behaviour of the system, figure 6.6, is dependent

upon whether the actuator chamber is charging or discharging.

This bilinear response can be attributed to a mismatch in the

sizes of the inlet/outlet ports of the electropneumatic

convertor and the effect of the valve spring opposing charging

forces and supporting discharging forces.

ii) The valve's stem travel shows significant hysteresis, figures

6.7. A phenomenonthat mainly arises because of actuator

 - 232 -

Return
spring

Standard Specification
ctuator
iaphragm
Body: Wrought 316 stainless steel
Air Motor Housing: Die cast aluminium,
epoxy coated finish
Diaphragm: Nitrile
Connections: NPT screwed
Stem/Plug/Seat: 316 stainless steel
MVO: Closes with lack of air
MVC: Opens with lack of air ve
m
Working limits: 350 b$r (wwithgut shock
at 20 C); 200 C
Characteristic: Linear

Connections Trim Cv max Kv max

A 3.10 2.60
1/2" NPT C 1.25 1.10
E 0.50 0.43

F 0.32 0.27
0.11 Illustration
H 0.13
1/4" NPT 1 0.05 0.04 approximately
N 0.006 0.005 quarter full size

Fig. 6.4 Manufacturer's data for the Platon miniature air operated
control valve ('m' valve)
 - 233 -

Nozzle Coil PC Board Zero Adjust

Standard Specifications

Supply pressure: 20 psig (± 2)

Output pressure: 3-15 psig
Supply pressure effect: 0.3% change for 1 psig change
within working limits
Temperature coefficient: < 1% of span/50°F
Hysteresis and Repeatability: Within 0.1% full scale
Maximumair capacity: 0.16 SCFM
Shock & Vibration effect: Negligible up to 2G between
5-200 Hz

Fig. 6.5 Manufacturer's data for the Fairchild T5200

I/P converter
 - 234 -

Full scale

10

Time
p5 10 15 ZO (seconds

Fig. 6.6 Performance of the standard pneumatic control valve;

time records illustrate the response of the regulator
to a 60% step change about its mean setting.
 - 235 -

% Output (X)

1001

(i)
8 i)

1-9 V change in P1
2-8 VM
4 3-7 V" P1
4-6 V" P1

% Input
20 40 60 80 90 (N)
0
a. Fixed step (0.1 V)
% Output (x)

1001 0.025 V steps

_. __0.10 V"
80 V If
---0.40

60

40

20

0-1-- 1 op % Input
0 20 40 60 80 100 (Pi)

b. Fixed magnitudes (3-7 V)

.

Fig. 6.7 Performance of the standard pneumatic control

valve; steady state behaviour illustrating
significant hysteresis.
 236

diaphragm distortion with movement, and further compounded

by stem frictional force nonlinearities.

By increasing the charge/discharge rates, and hence the

forces, into the actuator chamber then the bilinear effect can be

reduced with the added benefit of a much faster response (figure 6.8).

This is achieved in practice by fitting a volume booster relay to the

electropneumatic convertor. However, actuator hysteresis is still

present, see figure 6.8b, requiring a positioner to be fitted to

completely linearise the valve's stem response, a solution normally

employed by British Gas plc.

6.3.1 Software Positioning of the Control Valves

The application of a software positioner to the control valve

can help eliminate the nonlinear behaviour previously described, hence

reducing the problems that might otherwise arise in later investigations.

In order to achieve zero steady state positional error a proportional

plus integral controller is designed using the modified ultimate method

presented in chapter 3. A closed loop network is formed by precompensating

the valve/electro-pneumatic convertor system with an ideal relay and

integrator and employing unity feedback. The data presented in Table 6.1

shows the peak amplitude (A) and period (Pu) of the resulting limit cycles
for various values of relay amplitude (D)
of the valve's stem position,

about a quiescent value of 5 volts.

Relay Limit Cycle Limit Cycle

Amplitude (D) Amplitude (A) Period (P )
(volts) (volts) (secondsl

1 0.25 1.25
2 0.51 0.95
3 0.91 0.95

Table 6.1 Resulting limit cycles of the valve's stem position when
subjected to a modified ultimate test for the tuning of
PI controllers
 - 237 -

Full scale

100

80

60 ment (X)

40

20

Time
0 seconds)

a. Dynamic response
% Output (X)

10

% Input
(M)

b. Static response (0.4 V steps)

Fig. 6.8 Performance of the standard pneumatic control

valve with volume booster; note the charging/
discharging time constant similarity.
 - 238 -

Using the average values of limit cycle amplitude and period

D=4.07 A volts

Pu = 1.07 seconds

then the following ultimate gain and frequency values were calculated.
4D
Ku = =5.1E 6.1
irT .....
2ir

Tru-= 5.87 6.2

.....

Selecting a phase margin 4m of 60 degrees to achieve a response

with minimal overshoot then: -

Ku
Kc = sin(ým) = 0.76 .....
6.3
wu

Ti = tan(cm) = 0.29 6.4

u .....

Application of this controller design to the valve/electro-

pneumatic convertor system resulted'in the static and dynamic responses

illustrated in figure 6.9, displaying good set point following, negligible
hysteresis, as well as retaining the linear speed of response permitted

by the volume booster.

6.4 Comparative Study of the Test Rig Behaviour

Once the experimental system had been constructed a series of

steady state and dynamic responses of the test rig were investigated to

establish the degree of some conformity between the mathematical and

experimental AGI models, in order to provide some practical justification

for the mathematical model developed in chapter 2.

First, the static behaviour of the test rig was analysed by

varying one regulator's position whilst maintaining the second regulator

at a specified bias level (in this case 4 volts).

 - 239 -
% Full scale

100

80

Went (X)
60

40

20

Time
0 seconds)

a. Dynamic response
output (X)

100

80

60

40

20

°%Input
20 40 60 80 100 (M)
0
b. Static response (0.4 V steps)

Fig. 6.9 Performance of the positioning control scheme

when applied to the standard pneumatic control
valve with volume booster.
 - 240 -

Starting with zero pressure in the downstream system and with

the downstream regulator approximately half open the upstream valve was

slowly opened. Both the interstage volume pressure (Pu) and the station's

downstream pressure (Pin) increased with increasing flow from atmospheric

pressure, until the system flow saturated at a level dictated by the

downstream regulators operating point, see figure 6.10a. With the

upstream valve fully open then the inter-stage volume pressure settled

out at approximately the inlet pressure (P5), with some small pressure

dropped across the first stage valve, the final pressure drop across the
downstream regulator being much bigger as its restriction is greater.

Confirming the behaviour seen with the AGI simulation developed in

chapter 2.
Repeating the exercise with the upstream regulator approximately

half open and the downstream valve closed the interstage volume pressure

settles at the system inlet pressure, and the downstream pressure at

atmospheric pressure. As the downstream valve opens the downstream

pressure increases and the interstage volume pressure decreases, the

latter phenomenonaccounting for the negative sign seen in the multi-

variable model's transfer function element G12(s). Figure 6.10b captures

the above results.

Second, in order to identify the dynamic behaviour of the test

rig, open loop step responses were performed about the previously specified

bias levels. Data capture was performed every two seconds. The form of

the results, recorded in figure 6.11, is very similar to those obtained

from the AGI station simulation presented in figure 6.12, the only

significant discrepancy being that of the response between movement in the

downstream control valve and its effect upon the interstage volume pressure,

element G12(s). In the mathematical model, G12(s) has pure derivative

 - 241 -
Full scale

100

80

60

-PS
40-

/I
201
I/
%/
Input
25 50 75 100 (M)

a. Upstream regulator only

Full scale

100

80-

60-

40- /

/
20-

% Input
0 (M)
0 25 50 75 100
b. Downstream regulator only

Fig. 6.10 Steady state behaviour of the AGI experimental

test rig; records illustrate the response of
the system to the position of an individual
regulator, whilst maintaining the second
regulator at 40% open.
 - 242 -

% Variation of
full scale
1.5

1.0

0.5

0
-0.5

-1.0

-1.5 Time
(minutes)

% Variation of a. Pu
full scale

1.5

1.0

0.5

0
-0.5

-1.0

-1.5 Time
0 2468 (minutes)

b. Pin

Fig. 6.11 Dynamic behaviour of the AGI experimental test rig;

(1 of 2) time records illustrate the open loop response to
± 5% changes in the input to the upstream regulator
only, both valves starting from a mean level of 40%
open.
 - 243 -

% Variation of
full scale
11
1.5
1.0

0.5

-0.5

-1.0

-1.5 } Time
0 2468 (minutes)

% Variation of a. Pu
full scale

1.5

1.0

V. 5

-0.5

-1.0

Time
-1.5 'minutes)

b. Pin

Fig6.11 Dynamic behaviour of the AGI experimental test rig;

(2; f ) time records illustrating its open loop response to
± 5% changes in the input to the downstream
regulator only, both valves starting from a mean
level of 40% open.
 - 244 -

Pu Fin

0 time

-1

-2

a. Upstream regulator only

Pu Pin

1
'i
0ý
lo
time
1i
-1 16-
-1
J

-2 i -2

b. Downstream regulator only

Fig. 6.12 Dynamic behaviour of the AGI simulation; time records

illustrating its open loop response to ±1 step changes
in the input to a single regulator, whilst maintaining
the second regulator's input constant. Period of step
changes was 8 hours.
 245

action, signified by zero steady state offset in the appropriate open

loop step responses, whilst the test rig response displays phase advance
like behaviour.

The test rig's process transfer function was determined by

identifying continuous time models of the open loop step responses using

available software [6.1]. Figure 6.13 illustrates the accuracy of the

models derived using the method of Recursive Least Squares (RLS). The

resulting transfer functions of the multivariable system are tabulated

in figure 6.14. Figure 6.15 displays the corresponding Nyquist array,

providing graphical confirmation of the similarity between the behaviour

of the test rig and that of the AGI simulation.

6.5 Design of SISO Proportional Plus Integral Controllers

Using the Modified Ultimate Method

In chapter 3, a closed loop tuning strategy is developed for

the present AGI station SISO control policy, in order to provide a basis

for comparing all future designs. The strategy involved the two stage

application of the modified ultimate method using a relay-integrator

combination to force the system to limit cycle. Repetition of this

exercise upon the experimental test rig using an ideal relay with unity

amplitude produced limit cycles with the following amplitudes and

periods:

Limit cycle information Ultimate values

Amplitude Period Gain Frequency
(A) (Pu) (Ku) (wu)

Upstream 1.41 13.6 0.902 0.463

control loop

Downstream 1.31 132.2 0.974 0.0475

control loop

Table 6.2 Results of the modified ultimate closed loop

tuning technique applied to the test rig
 Z46

Variation of
full scale

_21 " Time

0 1234 (minutes)

% Variation of a. G11(Putih11)
full scale

-2 Time
0 1234 (minutes)

b. G21 (PintiP11)

Fig. 6.13 Identification of the AGI experimental test rig;

(1 of 2) time records illustrate the accuracy of fit obtained
using continuous time models of the individual
elements of the system's transfer matrix.
 247 -

Variation of
full scale

-1

-2 Time
0 123 4 (minutes)

Variation of C. G12 (Pu-. M2)

full scale

0ý

-1

-2 Time
'minutes)

d. G22 (Pin''M2)

Fig. 6.13 Identification of the AGI experimental test rig; time

(2 of 2) records illustrate the accuracy of fit obtained using
continuous time models of the individual elements of
the system's transfer matrix.
 - 248 -

Pu [Gil(s) G12(s) h11

G21(s) G22(s) M2-

n.

Each individual element of the transfer matrix G(s) was

modelled in the continuous time domain using the structure:

Gij(Njs =
2s2 ++s
ls+1
Resulting in the data

Multivariable Transfer Function Element

G11(s) G12(s) G21(s) G22(s)

N1 6.102 -5.600 - -
+-)aý No 0.226 -0.056 0.1709 0.2167
L D2 64.00 - - -
D1 41.60 40.00 40.00 40.00

Fig. 6.14 Results of the continuous time modelling

exercise performed upon the AGI experimental
test rig, using the method of recursive
least squares.
 - 24 9 -
Imaginary

Real

a. AGI simulation

Imaginary

a 0.1-

-0.2 -0.1 0.1

Real

-o -0.1

0.1 0.1
0.1 0.2 1 0r? Or2
-

i 1
-o.
-o.
-0.
-0.21
b. AGI test rig

Fig. 6.15 Comparison of the AGI model's transfer function matrix.

Note: The similarity between the general forms of the

individual elements of the Nyquist arrays.
 - 250 -

Based upon a design specification of 60 degrees phase margin,

the required SISO controller parameters were found to be

Proportional lIntegral time

gain (Kc) ; constant (Ti)

Upstream 1 687 3.741

control loop I'
Downstream 17.746 36.439
control loop

Table 6.3 SISO PI controller parameters designed

using the modified ultimate method for
a phase margin of 60 degrees

The effectiveness of the SISO controllers in regulating the

system's pressures is shown in figure 6.16. Because of the presence of

off-diagonal elements in the system's transfer function matrix (figure

6.15) significant interaction is present when the test rig is subjected

to set point variations in the closed loop. As expected from the

simulation results of chapter 2, the feedback control scheme quickly

damps out these oscillations providing good set point following behaviour.

6.6 Digital Multivariable Proportional Plus Integral

Control of the Experimental Test Rig

It has already been established that the interaction existing

within pressure reduction stations is significant, requiring the need of

a multivariable controller in order to achieve satisfactory closed loop

performance.
Chapter 5 illustrated a frequency response technique, the

characteristic locus method, to design controllers for multivariable

systems. This solution is practically very demanding and time consuming

since it requires the measurement of the system's frequency response.

 - 251 -

% Variation of
full scale

10

Time
0 (minutes)

-10 in
a. SP1 only
Variation of
full scale

10

Time
0 minutes)

-10
b. SP2 only

Fig. 6.16 Performance of the modified ultimate design

method's SISO PI control scheme; time records
illustrate the response of the AGI experimental
test rig simulation to ± 5% changes in set point.
 - 252 -

However, the final structure of the solution, a multivariable

proportional plus integral controller, is conceptually very appealing

since it is in a recognisable form to both plant engineers and operators.

Work presented by Koivo [6.2,3], and also investigated in chapter 5,

proposed a simple means of tuning the necessary proportional and integral

gain matrices based upon information gathered from the open loop step

responses of the system, section 5.7. Weighted versions of the process's

inverse derivative slopes matrix [P] and inverse steady state gains

matrix [I], providing the aforementioned proportional and integral gain

matrices respectively. Porter [6.4-6] uses exactly the same approach,

but in the discrete domain, by utilising information embodied within

discrete time models of the process to form the controller matrices, a

solution which is in the desired form for implementation within a modern

microprocessor based commercial digital controller.

Consider the pulse transfer function G(z-1)

bl. z-1 +b z-2 + +b z-Q
)+ . ... .
Giz 6.5
al. z-, + atz- + ... + aP. z- .....

where

q is the order of the numerator polynomial

p is the order of the denominator polynomial

The steady state gain g(O) of the discrete time process can be

evaluated via
b, + bap
a, + a2"... + 6.6
.....

Whilst the initial slopes h(T) of the response are represented by

h(T) = bi 6.7
.....
Performing this analysis on all the elements that constitute

the multivariable discrete time plant the matrices G(O) and H(T) can be

formulated and employed in the following equation to provide discrete

 - 253 -

time multivariable proportional plus integral control.

uk = [P]. ek + [I]. zk ,....

6.8

Integral action is derived using Euler's method [6.7], to

yield
Zk+1 = zk + T. ek 6.9
.....
The controller parameter matrices are determined from the open

loop step responses and the design equations

[P] = H(T)-'. n 6.10
.....
[I] = G(0)-1"i

where n, j are diagonal tuning matrices.

6.6.1 Multivariable P+I Controller Application

In order to proceed with the design method the structure and

elements of the pulse transfer function (equation 6.5) must first be

determined. This is achieved using the micro-CAPTAIN package [6.8] which

employs the method of Instrumental Variables (IV) to identify discrete

time models from experimental data sets obtained from the test rig. A

comparison of the resulting models and the open loop step responses of

figure 6.11 is presented in figure 6.17, and their corresponding pulse

transfer functions in figure 6.18. By placing this information into

equations 6.6 and 6.7 the following controller parameter matrices were

evaluated.
0014086 -0.13424 -1 II1 0
[P] =
0.00910 0 n2
[0.00133
10 22577 0
C11 =
-0.08171--1 11
0.17088 0.21673 0

In order to develop some means of 'tuning' the controller

the effects of varying the tuning parameters

using the matrices provided,
 - 254 -

% Variation of
full scale

-2 Time
minutes)
a. G11 (PutiMi)
% Variation of
full scale

Time
-2 ;minutes)
b. G21 (Pin'tin11)

Fig. 6.17 Identification of the AGI experimental test rig; time

(1 of 2) records illustrate the accuracy of fit obtained using
discrete time models of the individual elements of
the system's transfer matrix.
 - 255 -

Variation of
full scale

-2 Time
(minutes)

% Variation of C. G12 (PuiM2)

full scale

2
I

-1

-2 Time
0 minutes)
d. G22 (Pin'412)

Fig. 6.17 Identification of the AGI experimental test rig; time

(2 of 2) records illustrate the accuracy of fit obtained using
discrete time models of the individual elements of the
system's transfer matrix.
 - 256 -

Pu Gil (Z) ßi2 (Z) M1

Pig G21(Z) G22(z) M2

Each individual element of the transfer matrix G(z) was

modelled in the discrete time domain using the structure;

biä'Z + b2ä-Z
Gij(z) _ a1z 2

Resulting in the data:

Multivariable Transfer Function Element

G11(z) Gl2(z) G21(z) G22(Z)

bl 0.1409 -0.1343 0.0013 0.0091

ILI b2 -0.1328 0.1276 0.0099 0.0075

al 0.9530 0.9122 0.4966 0.4957

10
a2 0.0114 0.0047 0.4374 0.4277

Fig. 6.18 Results of the discrete time modelling

exercise performed upon the AGI experimental
test rig, using the method of instrumental
variables.
 - 257 -

nl, r
n2 and were investigated. The resulting controller's effect on

the closed loop performance is recorded for the following load conditions:

(1) Set point 1-4±0.5 volts

Set point 2-4 volts
(2) Set point 1-4 volts
Set point 2-4±0.5 volts
(3) Set point 1-4±0.5 volts
Set point 2-4±0.25 volts
Initial results using unity tuning matrices (HI = n2 == 1)

produced oscillatory results with large amounts of interaction, indicating

that the use of first order model approximations of the system transfer

function elements was unsuccessful. By keeping the magnitudes of the

tuning parameters equal (n1 = n2 = 1) it was found that decreasing the

magnitude would reduce the overshoot and the interaction present within

the system at the cost of speed of response. Figures 6.19 and 6.20 illustrate

this effect, displaying the responses of the closed loop system with the

tuning parameters set at 0.15 and 0.05 respectively.

To keep the interaction to the amounts displayed in figure 6.20,

the proportional gain tuning parameters were fixed to 0.05, and the integral

tuning parameter 2 was varied. Because the main source of

gain matrix
interaction was due to changes in the set point of the downstream valve

affecting the interstage volume pressure, it was decided to tune the

controller to best improve this response. Increasing the value of I

provided a faster loop 2 response, large increases producing significant

amounts of overshoot. Figure 6.21 shows the response of the system with

0.055, the best compromise found.

=
Finally the tuning parameter nl was decreased to provide faster

loop 1 dynamics in the closed loop, a value of 0.01 providing the fastest

before overshoot became significant, figure 6.22.

response
 - 25 8-

% Variation of
full scale

10

0 Time
minutes)

-10 411

a. SP1 only
% Variation of
full scale

10

Time
0 minutes)

-10
b. SP2 only

Fig. 6.19 Performance of the multivariable PI control scheme

designed using the step response tuning technique
and the tuning constants n=E=0.15; time records
illustrate the response of the AGI experimental
test rig simulation to ± 5% changes in set point.
 - 259 -

Variation of
full scale

10

Time
0 ninutes)

-10 111

a. SP1 only
Variation of
full scale

10

Time
0 ninutes)

-10
b. SP2 only

Fig. 6.20 Performance of the multivariable PI control scheme

designed using the step response tuning technique
and the tuning constants It =E=0.05; time records
illustrate the response of the AGI experimental
test rig simulation to ± 5% changes in set point.
 - 260 -

Variation of
full scale

10

0 Time
(mi nutes)

-10 Pin

% Variation of a. SP1 only

full scale

10

Time
0 minutes)

-10
b. SP2 only

Fig. 6.21 Performance of the multivariable PI control scheme

designed using the step response tuning technique
and the tuning constants n=0.05, E=0.055; time
records illustrate the response of the AGI
experimental test rig simulation to ± 5% changes
in set point.
 - 261 -

Variation of
full scale

10

Time
0 (minutes)

-10 111

a. SP1 only
% Variation of
full scale

10

Time
0 minutes)

-10
b. SP2 only

Fig. 6.22 Performance of the multivariable PI control scheme

designed using the step response tuning technique
and the tuning constants nl = 0.01, u2 = 0.05,
E=0.055; time records illustrate the response of
the AGI experimental test rig simulation to ± 5%
hanges in set point.
 262
- -

Thus the final controller gain matrices were determined to be

0.0623 4.5961
[P] _
-0.0091 4.8228

0.1896 0.0715
[I] _
-0.1495 0.1974

The application of the above controller to the test rig,

figure 6.22, has resulted in good set point following behaviour, the
final closed loop system having a time, constant of approximately 12

seconds (6 sample intervals). Using the index developed in chapter 5

the interaction between the downstream regulator and the interstage

volume pressure has been shown to be reduced to approximately one tenth

of that achieved using the previously discussed SISO control scheme.

6.7 Discussion

A comparison of the responses obtained from the experimental

test rig with those obtained from the AGI simulation show, quantitively,

that the two systems exhibit much the same behaviour. Indeed, it was

discovered during the controller tuning exercises, that the same

difficulties were apparent when tuning controllers for the test rig as

had been encountered when dealing with the AGI simulation. This,

therefore, provides practical evidence of the ability of the mathematical

model to reflect the important features of the physical system.

By accurately positioning the control valves, not only was the
linearity of the system improved, but it also allowed the use of the

regulators, using the appropriate formulas developed from their flow

profiles, to meter system flows given suitable upstream and downstream

pressure measurements, a technique that could be employed in remote

sites to give approximate flow measurements at low cost, and be capable

of providing inferential control during faults in telemetry systems.

 - 263 -

Using the modified ultimate technique (chapter 3) to design

a SISO control scheme for the multivariable process, produced a closed

loop system whose response showed good set point following behaviour,

but considerable interaction. This situation was improved using the

multivariable proportional plus integral controller design technique

of Koivo and Porter, both being special cases of the multivariable

frequency response technique presented in chapter 5 (the method of

characteristic locus design). This involved the diagonalisation of the

open loop transference matrix at high and low frequencies using the

multivariable proportional plus integral controller designed by:

[P] = EG(jwn)]-'
0
CI] = Gý,iwý )"
0I
where

wg is any low frequency where steady state gains exist

'Oh is any high frequency, normally selected as the desired

bandwidth of the system

The methods of Koivo and Porter also utilise the steady state

gains of the system in the design of the integrator gain matrix [I],

however in the design of the proportional gain matrix [P] they both

differ from the above technique in that they use the more simply evaluated

initial slopes of the open loop step responses, i. e.

Koivo:
-1 ro

n
[G(O)] -1
10
[I] = "
0 ý2
 - 264 -

Porter:
[HT]' -ý [A(T)i 0
[P] = .
01 n2

G(am)
=
0 n2'
-1
10
[Iý G(0)
0 12
However, since the methods are based on abstract measures of

experimental step response data care must be taken if significant inter-

action is to be avoided and stability maintained. A prime example would

be the tendency to evaluate a negative initial slope, hence a negative

proportional gain, when dealing with non-minimum phase systems. One

solution to these problems is to analyse the controller design using the

approximate models analysis presented in appendix III. 1, which quantifies

the accuracy of the first order modelling assumptions in terms of

controller design constraints and predicted closed loop performance bounds.

The selection of the tuning matrices discussed above is extremely

important in the design of the resulting control scheme, indeed in Porter's

adaptive controller the setting of the tuning matrices is a prerequisite to

the design exercise. It would be possible in this particular situation to

set the matrices to detune the response and ensure stability and then vary

them during the adaptive stage using Luyben's method [6.9]. This involves

varying a tuning factor F in each loop to ensure that the maximumover-

the closed loop response does not exceed some prespecified value,
shoot of
i. e.
-i IIi/Fl 0

0 II2/F2

/F
"0 1
12/F2
 - 265 -

Further solutions could arise using similar pattern recognition

techniques [6.10,11]. Here, intuitive retuning decisions are made based

upon the closed loop response of the system to various disturbance inputs.

By studying the effect the tuning matrices have upon the system, possibly

incorporating some form of expert system [6.12], appropriate retuning

could take place to optimise system performance on-line.

 - 26 6-

References (Chapter 6)

6.1 Young, P. C. "Recursive estimation and time series analysis".

Springer-Verlag Publishers, Berlin, 1984.
6.2 Penttinen, J., and Koivo, H. N. "Multivariable tuning
regulators for unknown systems". Automatica, Vol. 16,
pp 393-398,1980.

6.3 Koivo, H. M., and Pohjolainen, S. "Tuning of multivariable

PI-controllers for unknown systems with input delay".
Automatica, Vol. 21, No. 1, pp 81-91,1985.

6.4 Porter, B. "Design of error actuated controllers for unknown

multivariable plants". Electronic Letters, Vol. 17, pp 106-
107,1981.

6.5 Porter, B. "Design of set point tracking and disturbance

rejection controllers for unknown multivariable plants".
Proc. IEE Workshop on Self-tuning and Adaptive Control,
Oxford, 1981.

6.6 Porter, B. "Design of tunable set point tracking controllers

for multivariable plants". Int. J. Control, Vol. 35, pp 1107-
1115,1982.

6.7 Astrom, K. J., and Wittenmark, B. "Computer controlled

systems - theory and design". Prentice-Hall, 1984.

6.8 Young, P. C., and Benner, S. "micro-CAPTAIN handbook".

Version 1, Dept. Environ. Sci., Univ. of Lancaster, Report
TR50,1987.

6.9 Luyben, W. L. "Simple method for tuning SISO controllers in

multivariable systems". Ind. Eng. Chem. Process Des. Dev.,
Vol. 25, pp 654-660,1986.

6.10 Bristol, E. H. "Pattern recognition: an alternative to

parameter identification in adaptive control". Automatica,
Vol. 13, pp 197-202,1977.

6.11 Kraus, T. W., and Myron, T. J. "Self tuning PID controller

uses pattern recognition approach". Control Engineering,
pp 106-111, June 1984.

6.12 Porter, B., Jones, A. H., and McKeown, C. B. "Real time

expert tuners for PI controllers". IEE Proc., Vol. 134,
Pt. D, No. 4, pp 260-263,1987.
 - 267 -

Nomenclature (Chapter 6)
Practical Evaluation of Control Techniques
Using an Experimental Test Rig

Experimental rig variables:

dPq Volumetric flowrate differential pressure (psi)

Pactl Upstream valve's actuator chamber pressure (psi)

Pact2 Downstream valve's actuator chamber pressure (psi)

Pd Downstream loading pressure (psi)

Pin Pipeline inlet pressure (psi)

Ps Supply pressure (psi)

Pu Interstage volume pressure (psi)

X1 Upstream valve's % stem position

XZ Downstream valve's % stem position

Discrete time systems:

ak k'th denominator polynomial constant

A(z-1) Plant denominator polynomial

bk k'th numerator polynomial constant

B(z'1) Plant numerator polynomial

ek Error at the k'th sample

g(O) Plant steady state gain matrix element

G(O) Plant steady state gain matrix

G(z"1) Plant transfer function matrix

G(o) Plant step response initial slope matrix

h(T) Plant step response initial slope matrix element

H(T) Plant step response initial slope matrix

[I] Integ rator gain matrix

P Order of plant denominator polynomial

[P] Propo rtional gain matrix

q Order of plant numerator polynomial

 - 268 -

T Sampling time (sec)

uk Plant input at the k'th sample

z'1 Backwards shift operator

zk Integrator output at the k'th sample

I Integrator gain tuning matrix

IT Proportional gain tuning matrix

Continuous time systems:

A Peak amplitude of limit cycle

D Peak amplitude of ideal relay
G(s) Plant transfer function

KC Proportional gain

Ku Ultimate gain

Pu Ultimate period (sec)

S Laplace operator

Ti Integral time constant (sec)

Om Phase margin (degrees)

Ultimate frequency (rads/sec)

wu
 - 269 -

CHAPTER7

CONCLUSIONS
 - 270 -

The material contained within this thesis develops a line of

work which commencedin the mid seventies when research was initiated

at British Gas Engineering into system reinforcement requirements to

fulfil the increasing demand for gas. As time progressed, this growth
in demand has slowed down significantly, resulting in a shift of emphasis
from construction towards one of efficient operation [7.1].

The delivery of gas from the national transmission to regional

systems and ultimately on to the customer, requires a number of control

functions to be performed. The prime objectives when performing these

functions are: -

i) system safety, integrity and security of supply,

ii) minimum operating costs, and finally,
iii) maximum flow capacity and storage utilisation.

Because of the various operational problems and deficiencies

associated with the existing pneumatic control schemes at PATS

offtakes

and regional PRS's, there arose the need to review and upgrade, wherever

possible, the control philosophy, applicable to the method of flow and

pressure control of gas to the regions.

Modern control technology coupled with major parallel advances

in microprocessor technology provided the catalyst for the development of

schemes to overcome the current operational problems and thereby greatly

improve the efficiency of control [7.2] and it is this task that has been

considered here.
The first half of the thesis presented an analysis into the

performance and stability aspects of an AGI station using the currently

favoured proportional plus integral control philosophy. Performance data

obtained from this system is then used as a yardstick against which -

further design methodologies can be assessed. This analysis begins with

the development of a detailed mathematical model of a 'typical' AGI station

 - 271 -

which contains sufficient structural flexibility to enable it to simulate

any of the 200+ stations that exist in practice.

Validation of the mathematical model was performed using data

provided by the Engineering Research Station (ERS) at Killingworth, and a

specially designed experimental test rig. Results obtained from digital

computer simulations, of the model, are presented and a comparison of these

results and those obtained from ERS and the AGI test rig shows that the

simulations are accurate for the set of test signals considered. A further

indication of the accuracy of the model is provided from the study of

remote boundary pressure control systems [7.3,7.4]. Here, models developed

using the same mathematical relationships employed in the construction of

the AGI simulation have been used to design novel control schemes; the

resulting controller parameters being directly applicable without

modification to the actual systems.

Chapter 3 presented an automated closed loop tuning methodology

designed to establish SISO controller parameters whilst allowing for the

multivariable nature of the system. Based upon the relay tuning technique

proposed by Astrom [7.5,7.6] the method forces each pressure control loop

of the system to limit cycle at a magnitude that can be controlled directly

by a suitable choice of relay amplitude, thereby avoiding any problems that

may occur when using trial and error tuning or Ziegler-Nichols ultimate

method [7.7] as is currently the case. The process of adapting the relay's

amplitude can be automated leaving only the selection of appropriate

starting values before tuning can commence. Basic guidelines for this choice

have been found to be:

i) Slow processes - Small relay amplitudes to avoid controller

output saturation

ii) Fast processes - Large relay amplitudes

Results from the industrial application of this methodology and
its application to a variety of simulations are presented, showing it to

be an extremely useful method for the design of SISO controllers.

 - 272 -

An investigation of the effects on system performance of

variations of critical system design parameters is also presented. The

results of this study indicate that, due to the restricted size of the
interstage pipework, it is the design of the upstream pressure control

loop that most influences system performance, stability problems

occurring mainly at low station flowrates where the gains of the regulators
have been shown to be high, making it the ideal condition at which to tune
the controller to ensure stability over the station's working range. This,

and the inability of the P+I controllers to provide sufficient stability

margins to achieve the speed necessary for efficient non-interactive

operation, have introduced considerable difficulties when commissioning AGI

stations.
One simple solution to this problem was to utilise the extra

phase advance afforded to us by the use of a PID controller in the regulation

of the interstage volume pressure. A comparison of the relative performances

of this scheme and the standard controller configuration, table 3.8, clearly

illustrates the benefits of providing sufficient speed in the upstream

control loop to suppress system interaction. Unfortunately, experience to

date, concerning the use of derivative action within control schemes has

meant that such controllers are looked upon unfavourably within the gas

industry.

The second half of the thesis presented an analysis into the use

of multivariable controllers to achieve high speed operation and large

stability margins simultaneously. Following a formal investigation using

the characteristic locus frequency response design technique a complex

controller structure was identified which achieved this objective.

Subsequent investigations, however, showed that this complex structure could

be reduced to a multivariable P+I controller format. Using this result

several practical alternatives to the tuning of such a controller have also

been considered based upon step response testing. Results are presented in

chapters 5 and 6 for applications on both the AGI simulation, and the
 - 273 -

experimental test rig, the multivariable designs providing both the high

speed necessary for interaction suppression and the large stability margins

required for a minimal overshoot response and acceptable performance over

the station's complete working range. The main problem with these

alternatives is the number of design parameters necessary to ensure sensible

operation. These problems arise when the first order step response models,

upon which the techniques are based, are no longer sufficient to adequately

reflect the behaviour of the system. Appendix III. 1 presents an off-line

method which imposes constraints upon the controller design to compensate

for the modelling inaccuracies. The technique is easily extendable to the

multivariable case where it could be used to predict the boundaries of the

possible closed loop responses of the multivariable PI controlled system,

thus aiding the procedure of selecting the required tuning matrices for

'optimum' performance.

However, despite the numerous theoretical studies, very few

multivariable controllers have been reported in use in industry. Someof

the reasons for the lack of commercial applications have included complexity,

excessive engineering manpower requirements, lack of robustness and

integrity, and the operator non-acceptance. It is hoped that the multi-

variable design methodology presented here is sufficiently simple in

structure to overcome these problems and be of great practical benefit to

the gas industry.

The controller studies considered in the work to date have

concentrated on those stations whose downstream control variable is

pressure, the assumption here being that because the systems are so very

similar, any lessons learnt in the study of one system would apply directly

to the other. Furthermore, it is also felt that the methodology developed

in both the analysis and the controller design for this system could be

applied quite generally to a large range of industrial problems, which need

 - 274 -

not necessarily be related to the process industries. The correctness,

or otherwise of these assumptions must, of course, be tested. This,

therefore, forms one of the recommendations for further work. It is also

felt that the use of a single tuning parameter in the digital pole place-

ment proportional-integral-plus (PIP) design strategy presented in

Appendix 111.2 provides one solution to overcome the problem of operator

non-acceptance. At present, current research [7.8,7.9] has shown the

technique to produce excellent results when applied to a wide range of

both linear and nonlinear SISO applications. Initial investigations of

the technique upon the AGI experimental test rigs SISO control scheme

[7.10] have confirmed the results of chapter 3, that a PID controller is

required in the upstream pressure control loop if the regulators response

is to be fast enough to respond to changes in the interstage volumes

pressure, whilst identifying that a PI controller is sufficient to control

the downstream pressure. These results have provided sufficient evidence

to support the recommendation to extend the method into the multivariable

domain [7.11,7.12]. In addition, it is also recommended that consideration

be given to the introduction of enhanced microprocessors capable of

operating with object orientated languages. The use of these programming

techniques should allow the construction of hierarchal plant models to

facilitate rapid diagnosis, fault detection (and isolation), as well as

providing all the existing (and envisaged) control functions. Ultimately,

this will lead to the development of a comprehensive on-line control/

condition monitoring environment which will use both past and present plant

data, possibly using artificial intelligence techniques, to permit safe and

efficient operation of the stations.

 -L/5 -

References (Chapter 7)
Conclusions

7.1 Arden, W. J. B., and Hawkins, R. W. "System software

specification and reliability analysis for the national
offtakes (AGI) electronic control project". ERS Killingworth
Internal Report, 1983.

7.2 Arden, W. J. B. "GASCONTROL -A review of past, present and

future systems". North of England Gas Association,
Killingworth, 1988.

7.3 Arden, W. J. B., and Fletcher, I. "Whasset/Kendal remote

boundary pressure simulation study". North West Gas Research
Report, 1988.

7.4 Arden, W. J. B., and Fletcher, I. "Mill-hill/Cricklewood

remote boundary pressure simulation study". North Thames Gas
Research Report, 1988.

7.5 Astrom, K. J. "Ziegler-Nichols auto tuner". Lund Institute

of Technology, Report TFRT-3167,1982.
7.6 Astrom, K. J., and Hagglund, T. "Automatic tuning of simple
regulators with specifications on phase and amplitude margins".
Automatica, Vol. 20, No. 5,1984.

7.7 Ziegler, J. G., and Nichols, N. B. "Optimum settings for

automatic controllers". Trans. ASME, Vol. 64,1942.

7.8 Young, P. C., Behzadi, M., Wang, C. L., and Chotai, A.

"Direct digital and adaptive control by input-output state
variable feedback pole assignment". Int. J. of Control, Vol.
46, No. 6, pp 1867-1881,1987.

7.9 Boucher, A. R., Cox, C. S., and Young, P. C. "Direct digital

control using a personal computer -a new algorithm". Proc.
DTI Colloquium "Personal computers in industrial control",
Warren Springs Laboratory, Stevenage, England, 1988.

7.10 Fletcher, I., and Boucher, A. R. "Proportional-Integral-Plus

(PIP) control of a gas pressure reduction station".
Sunderland Polytechnic Internal Report, IF/01/89.

7.11 Wang, C. L., and Young, P. C. "Direct digital pole assignment

control system design for multivariable systems based on
input-output state variable feedback". Control 88, oxford,
pp 212-217,1988.

7.12 Boucher, A. R., and Fletcher, I. "Multivariable Proportional-

Integral-Plus (PIP) Control - An initial investigation".
Sunderland Polytechnic Internal Report, ARB/06/88.
 - 276 -

APPENDIXI

PROGRAMS
COMPUTER
 L//

PROGRAM
1

SIMULATIONOF THE ABOVEGROUND

FORTRAN

INSTALLATION(AGI)
 278

C****** NONLINEAR AGI SIMULATION PROGRAMME

c****** USING
C****** SISO/MIMO PI CONTROL
C******
DOUBLE PRECISION QFI, QFIN, QFA, QFB, QFLOAD, P, Z, RDDOT1, RDOT1, R,
+ ERROR19ERROR2, AVAL19AVAL2, RDDOT2, RDOT2, QDIFF, PU, PS, QFC
REAL KC1, TI1, SPI, TIME, T, PIN, MCON1, THETA1, W01, ZETA1, TAU1,
+ KC2, TI2, SP2, MCON2, THETA2, W02, ZETA2, TAU2, LVOL, CDOT1, CDOT2
INTEGER VLV1, VLV2, AN1, AN2, SPOINT, MFLOW
CHARACTER*3 ANSI, ANS2, FILE, TYP
DIMENSION P(4), Z(4), R(2)
5 FORMAT(A)

c****** SIMULATION INPUT ROUTINE

c******
PRINT *, 'ENTER UPSTREAM VALVE SET POINT (was 5171.25 kN/m2)'
ACCEPT *, SP1
PRINT *, 'ENTER DOWNSTREAM VALVE SET POINT (was 3447.5 kN/m2)'
ACCEPT *, SP2
PRINT *, 'ENTER STATION LOAD FLOW (was 39.329 m3/sec)'
ACCEPT *, QFLOAD
PRINT *, ' '
7 PRINT *, 'IS SISO/MIMO PI CONTROL REQUIRED? (S/M)'
ACCEPT 5, TYP
IF(TYP. NE. 'M'. AND. TYP. NE. 'S')THEN GOTO 7
PRINT *, ' '
PRINT *, 'ENTER T, TPRINT FOR INTEGRATION'
ACCEPT *, T, TPRINT
PRINT *, ' '
PRINT *, 'IS DATA FILE REQUIRED? '
ACCEPT 5, FILE
IF(FILE. EQ. 'Y')THEN
(JPEN(UNIT=15, F1LL='OATAFILE. DAT', STATUS='NEW')
ENDIF
c******
c****** INITIAL CONDITIONS (Q=5 Msft3/hr)

TAU1=1
TAU2=1
TIME=0.0
N=0
PS=6895.0
PU=5171.25
PIN=3447.5
PD=3447.71
P(1)=3429.024641
P(2)=3410.449235
P(3)=3391.772059
P(4)=3372.991458
IF(TYP. EQ. 'M')THEN

CHARACTERISTIC LOCUS TUNED MIMO PI CONTROLLERS

c***c******
 - 279 -

AVAL2=-16.413
AVAL1=-130.436
C11A=2.211
C11B=1.
C12A=177.1
C128=1.
C21A=-2.472
C21B=1.3
C22A=198.3
C22B=1.
ELSE
C******
C****** RELAY TUNED SISO PI CONTROLLERS
C******
AVAL1=-1192.5016
AVAL2=-139.3319
C11A=1.5276
C11B=1.7539
C12A=0.
C12B=0.
C21A=0.
C21B=0.
C22A=29.151
C22B=0.69814
ENDI F
R(1)=-3195.146972656
R(2)=-2835.608154297
QFIN=39.329
RDOT1=0.0
RDOT2=0.0
C******
C****** MAIN SIMULATION LOOP
C******
C****** UPSTREAM CONTROLLER
C******
10 ERROR1=SP1-PU
ERROR2=SP2-PIN
CON11=C11A*(ERROR1+AVAL1*C11B)
CON12=C12A*(ERROR2+AVAL2*C12B)
MCON1=CON11+CON12
C******
C****** UPSTREAM VALVE
C******
RDOT1=(MCON1-R(1))/TAU1
X1=R(1)
C******
C****** STEM POSITION SATURATION
C******
X1=(X1+PS)/(2*PS)
IF(X1. LT. 0.0)X1=0.0
IF(X1. GT. 1.0)X1=1.0
X1=X1*100.
C******
C****** FISHER VALVE FLOW EQUATION (V25)
C******
 280

CV1=1.661E-6*X1**4. -1.781E-4*X1**3. +0.01145*X1**2. -0.02496*X1

CV1=CV1/100.
IF(CV1. LE. 0.0)CV1=0.0
IF(CV1. GT. 1.0)CV1=1.0
CG1=-2.484E-8*X1**5. +6.356E-6*X1**4. -6.121E-4*X1**3.
CG1=CG1+0.0335*X1**2-0.104*X1
CG1=CG1/100.
IF(CG1. LE. 0.0)CG1=0.0000001
IF(CG1. GT. 1.0)CG1=1.0
THETA1=3.57531E-2*CV1/CG1*SQRT(ABS(PS-PU))*DSIGN(1. OD0, PS-PU)
IF(THETAI. GE. 1.5707)THETA1=1.5707
QFI=563.27*CG1*SIN(THETAI)
C******
C****** INTERSTAGE VOLUME
C******
QDIFF=(QFI-QFIN)/3.072E-3
C******
C****** DOWNSTREAM CONTROLLER
C******
CON21=C21A*(ERROR1+AVAL1*C21B)
C0N22=C22A*(ERROR2+AVAL2*C22B)
MCON2=CON21+CON22

C****** DOWNSTREAM VALVE

C******
RDOT2=(MCON2-R(2))/TAU2
X2=R(2)
C******
C****** STEM POSITION SATURATION
C******
X2=(X2+PS)/(2*PS)
IF(X2. LT. 0.0)X2=0.0
IF(X2. GT. 1.0)X2=1.0
X2=X2*100.
C******
C****** FISHER VALVE FLOW EQUATION (V25)
C******
CV2=1.661E-6*X2**4. -1.781E-4*X2**3. +0.01145*X2**2. -0.02496*X2
CV2=CV2/100.
IF(CV2. LE. 0.0)CV2=0.0
IF(CV2. GT. 1.0)CV2=1.0
CG2=-2.484E-8*X2**5. +6.356E-6*X2**4. -6.121E-4*X2**3.
CG2=CG2+0.0335*X2**2-0.104*X2
CG2=CG2/100.
IF(CG2. LE. 0.0)CG2=0.0000001
IF(CG2. GT. 1.0)C02=1.0
THETA2-2.9688*CV2/CG2*SQRT(ABS((PU-PIN)/PU))*DSIGN(1. ODO, PU-PIN)
IF(THETA2. GE. 1.5707)THETA2=1.5707
QFIN=81.693E-3*PU*CG2*SIN(THETA2)
C****** --
C****** PIPELINE SIMULATION
C******
QFA=9.1249*SQRT(ABS(P(1)-P(2)))*DSIGN(1. ODO, P(1)-P(2))
QFB=9.0984*SQRT(ABS(P(2)-P(3)))*DSIGN(1. OD09P(2)-P(3))
QFC=9.0759*SQRT(ABS(P(3)-P(4)))*DSIGN(1. ODO, P(3)-P(4))
Z(1)=(QFIN-QFA)/1.9551
Z(2)=(QFA-QFB)/1.9551
Z(3)=(QFB-QFC)/1.9551
Z(4)=(QFC-QFLOAD)/1.9551
PIN=P(1)+(QFIN**2. )*O. 01195
 281 -

C******
C****** DATA OUTPUT ROUTINE
C******
IF(TIME. GE. TPRINT*N)THEN
PRINT 25, TIME, PU, PIN, XI, X2
25 FORMAT(5(1X, F13.7))
30 FORMAT(1X, F10.4,4(1X, F9.4), 3(1X, F10.3))
IF(FILE. EQ. 'Y')THEN
C******
C****** PRESSURES (psi)
C******
PGU=PU/6.895
PGIN=PIN/6.895
PG1=P(1)/6.895
PG2=P(2)/6.895
PG3=P(3)/6.895
WRITE(15,30)TIME, PGU,, PGIN, XI, X2
ENDIF
N=N+1
END IF
C******
C****** INTEGRATION ROUTINES
C******
CALL INTEG(ERROR1, AVAL1, T)
CALL INTEG(RDOT19R(1), T)
CALL INTEG(QDIFF, PU, T)
CALL INTEG(ERROR29AVAL29T)
CALL INTEG(RDOT2, R(2), T)
CALL INTEG(Z(1), P(1), T)
CALL INTEG(Z(2), P(2), T)
CALL INTEG(Z(3), P(3), T)
CALL INTEG(Z(4), P(4), T)
TIME=TIME+T
C******
C****** COMPLETE SIMULATION LOOP
C****** GOTO 10
STOP
END

SUBROUTINE INTEG(YDOT, Y, CINT)

C******
C****** INTEGRATION SUBROUTINE

DOUBLE PRECISION YDOTjY K1, K2, K3, K4

,
K1=CINT*YDOT
K2=CINT*(YDOT+K1/2.0)
K3=CINT*(YDOT+K2/2.0)
K4=CINT*(YDOT+K3)
Y=Y+K1/6+K2/3+K3/3+K4/6
RETURN
END
 PROGRAM
2

FORTRAN
ROUTINEFORLINEARISINGTHE

AGI SIMULATIONAT DEFINEDOPERATING

POINTS
 - 283 -

c******
c******
C*****k
PROGRAMME TO LINEARISE THE AGI SIMULATION
AT DEFINED OPERATING POINTS
Gýcýcýcýk*9c
REAL N1l, N12, N21, N22, KF1, KF2, Kll, K129K13, K21, K22, K23, KIN,
+ G11, G12, G21, G22, K1, K2, K3, K4
DIMENSION Gi(10), G2(10), X(10), Y(10), D(10), N11(lO), N12(10),
N21(lO), N22(10), VOL(10), ACT(l0), DUM(10), DUM1(10)
c******
c****** OPERATING POINT DEFINITION
c****** PRINT *, 'Enter Initial Flow (Msft3/hr)'
ACCEPT *, QSS
PRINT *, 'Enter Upstream, Interstage & Downstream Pressures (Psi)'
ACCEPT *, PS, PU, PIN
ce 1 krr d. 11
rr%a, 1, -,

PRINT *, 'Linearised A. G. I. Equations are : -'

PS=PS*6.895
PU=PU*6.895
PIN=PIN*6.895
QSS=QSS*(39.329/5. )
c******

C****** CALCULATION INITIALISATION

c*** TAU=1.
CVOL=3.072E-3
KF1=81.693E-3
KF2=2.9688
c****** POLYNOMIAL INITIALISATION
c******
c******
CALL ZPOL(G1)
CALL ZPOL(G2)
CALL ZPOL(X)
CALL ZPOL(Y)
CALL ZPOL(D)
CALL ZPOL(N11)
CALL ZPOL(N12)
CALL ZPOL(N21)
CALL ZPOL(N22)
CALL ZPOL(VOL)
CALL ZPOL(ACT)
CALL ZPOL(DUM)
CALL ZPOL(DUM1)
C******
EVALUATION OF QUIESCENT VALUES OF
C****** VALVE STEM DISPLACEMENT
C******
C****** UPSTREAM VALVE
C******
DO 10 X1=1,100,0.05
CV1=1.661E-6*X1**4. -1.781E-4*X1**3. +0.01145*X1**2. -0.02496*X1
CV1=CVl/100.
IF(CV1. LE. 0.0)CV1=0.0
IF(CV1. GT. 1.0)CV1=1.0
CG1=-2.484E-8*X1**5. +6.356E-6*X1**4. -6.121E-4*X1**3.
CG1=CG1+0.0335*X1**2-0.104*X1
CG1=CG1/100.
 - 284 -

IF(CG1. LE. 0.0)CG1=0.0000001

IF(CG1. GT. 1.0)CG1=1.0
THETA1=KF2*CV1/CG1*SQRT(1. -PU/PS)
IF(THETAI. GE. 1.5707)THETA1=1.5707
QFI=KF1*PS*CG1*SIN(THETAI)
IF(QFI. GT. QSS)GOTO 20
10 CONTINUE
C******
C****** DOWNSTREAM VALVE
C******
20 DO 30 X2=1,100,0.05
CV2=1.661E-6*X2**4. -1.781E-4*X2**3. +0.01145*X2**2. -0.02496*X2
CV2=CV2/100.
IF(CV2. LE. O. O)CV2=0.0
IF(CV2. GT. 1.0)CV2=1.0
CG2=-2.484E-8*X2**5. +6.356E-6*X2**4.
-6.121E-4*X2**3.
CG2=CG2+0.0335*X2**2-0.104*X2
CG2=CG2/100.
IF(CG2. LE. 0.0)CG2=0.0000001
IF(CG2. GT. 1.0)CG2=1.0
THETA2=KF2*CV2/CG2*SQRT(1. -PIN/PU)
IF(THETA2. GE. 1.5707)THETA2=1.5707
QFI=KF1*PU*CG2*SIN(THETA2)
IF(QFI. GT. QSS)GOTO 40
30 CONTINUE
C******
C****** VALVE EQUATION LINEARISATION
C****** UPSTREAM VALVE
C******
C******
40 PH1=THETAI
DCV1=6.644E-6*X1**3-5.343E-4*X1**2+0.0229*X1-0.02496
DCV1=DCV1/100.
DCG1=-12.42E-8*X1**4+25.424E-6*X1**3-18.363E-4*X1**2
DCG1=DCG1+0.067*X1-0.104
DCGI=DCG1/100.
K11=KF1*KF2*SQRT(PS*(PS-PU))/CG1*(CG1*DCV1-CV1*DCG1)*COS(PH1)
K11=K11+KF1*PS*DCG1*SIN(PH1)
K12=KF1*KF2*CV1*PU*COS(PH1)/(2. *SQRT(PS*(PS-PU)))
K12=K12+KF1*CG1*SIN(PH1)
K13=-KF1*KF2*CV1*SQRT(PS)*COS(PH1)/(2. *SQRT(PS-PU))
C******
C****** DOWNSTREAM VALVE

PH2=THETA2
DCV2=6.644E-6*X2**3-5.343E-4*X2**2+0.0229*X2-0.02496
DCV2=DCV2/100.
DCG2=-12.42E-8*X2**4+25.424E-6*X2**3-18.363E-4*X2**2
DCG2=DCG2+0.067*X2-0.104
DCG2=DCG2/100.
K21=KF1*KF2*SQRT(PU*(PU-PIN))/CG2*(CG2*DCV2-CV2*DCG2)*COS(PH2)
K21=K21+KF1*PU*DCG2*SIN(PH2)
K22=KF1*KF2*CV2*PIN*COS(PH2)/(2. *SQRT(PU*(PU-PIN)))
K22=K22+KF1*CG2*SIN(PH2)
K23=-KF1*KF2*CV2*SQRT(PU)*COS(PH2), (2. *SQRT(PU-PIN))
 - 285 -

C******
C****** ACTUATOR & VOLUME POLYNOMIALS
c******
ACT(1)=1.
ACT(2)=TAU
VOL(1)=-K13
VOL(2)=CVOL
C******
C**** PIPELINE INITIAL CONDITIONS
C******
PIPD=(18. /12. )*0.3048
PIPL=0.75*5280*0.3048
PIPC=3.141592654*(PIPD**2)*PIPL/(4. *14.7*6.895)
PIPK=PIPL*(QSS**2)/(735.5812*(PIPD**5))
P1=SQRT(PIN**2-PIPK)
P2=SQRT(P1**2-PIPK)
P3=SQRT(P2**2-PIPK)
P4=SQRT(P3**2-PIPK)
c******
C****** PIPELINE POLYNOMIAL
C******
K1=QSS/SQRT(PIN-P1)
K2=QSS/SQRT(P1-P2)
K3=QSS/SQRT(P2-P3)
K4=QSS/SQRT(P3-P4)
KIN=2*QSS/(K1**2)
K2=K2/(2. *SQRT(P1-P2))
K3=K3/(2. *SQRT(P2-P3))
K4=K4/(2. *SQRT(P3-P4))
G1(1)=1.
G1(2)=PIPC*(3. /K2+2. /K3+1. /K4)
G1(3)=PIPC**2. *(2. /(K2*K3)+2. /(K2*K4)+1. /(K3*K4))
G1(4)=PIPC**3. /(K2*K3*K4)
G2(1)=0.
G2(2)=PIPC*4.
G2(3)=PIPC**2. *(3. /K2+4. /K3+3. /K4)
G2(4)=PIPC**3. *(2. /(K2*K3)+2. /(K2*K4)+2. /(K3*K4))
G2(5)=PIPC**4. /(K2*K3*K4)
C******
Cýic-kýcýcýcýr FORMULATION OF AG I SIMULATION'S
C****** MULTIVARIABLE TRANSFER FUNCTION
C******
AX=50. *K11/6895.
AY=50. *K21/6895.
AZ=K22*AX

C****** X POLYNOMIAL
C******
CALL POLK(G2, DUM, KIN)
CALL POLP(DUM, GI, X)
CALL ZPOL(DUM)
C******
C****** Y POLYNOMIAL
C******
CALL POLK(X, DUM, -K23)
CALL POLP(DUM, G2, Y)
CALL ZPOL(DUM)
 -LÖ6-

C******
C****** NUMERATOR POLYNOMIAL'S

CALL POLK(Y, N11, AX)

CALL POLK(G2, N12, -AY)
CALL POLK(X, N21, AZ)
CALL POLK(VOL, DUM, AY)
CALL POLX(DUM, X, N22)

C****** DENOMINATOR POLYNOMIAL

C******
CALL ZPOL(DUM)
CALL POLK(G2, DUM, K22)
CALL POLX(VOL, Y, D)
CALL POLP(DUM, D, DUM1)
CALL ZPOL(D)
CALL POLX(DUM19ACT, D)
c****** TRANSFER FUNCTION OUTPUT
C *****
C******
PRINT *, ' '
PRINT *, ' N11 N12 N21 N22 D'
PRINT *, ' '
DO 50 I=1,7
PRINT 100, ' s**'9I-1, ' ', N11(I), N12(I)9N21(I), N22(I), D(I)
50 CONTINUE
100 FORMAT(1X, A, I1, A, 5(1X, F9.6))
END

SUBROUTINE POLX(A, B, C)
C***
C****** POLYNOMIAL MULTIPLICATION
C******
DIMENSION A(10), B(10), C(10)
DO 10 I=1,10
C(I)=0.
10 CONTINUE
DO 20 I=1,6
DO 20 J=1,5
C(I+J-1)=C(I+J-1)+A(I)*B(J)
20 CONTINUE
RETURN
END

SUBROUTINE POLK(A, B, CON)

c****** POLYNOMIAL MULTIPLICATION BY A CONSTANT
C******
C******
DIMENSION A(10), 8(10)
DO 20'I=1,10
B(I)=A(I)*CON
20 CONTINUE
RETURN
END
 287

SUBROUTINE POLP(A, B, C)
C******
C****** POLYNOMIAL ADDITION
C******
DIMENSION A(10), B(10), C(10)
DO 10 I=1,10
C(I)=0.
10 CONTINUE
DO 20 I=1,10
C(I)=A(I)+B(I)
20 CONTINUE
RETURN
END

SUBROUTINE ZPOL(A)
C******
C*****k POLYNOMIAL ZEROING
C******
DIMENSION A(10)
DO 10 I=1,10
A(I)=0.
10 CONTINUE
RETURN
END
 - 288 -

3
PROGRAM

ROUTINEFOREVALUATING
FORTRAN THE

AGI'S CHARACTERISTIC
EQUATION
 - 289 -

c******
G****** PROGRAMME TO DETERMINE THE AGI SIMULATION'S
C****** CHARACTERISTIC EQUATION
C******
CHARACTERk4 PARA, CHAR
REAL D(20), X(20), Y(20), G(20), NUM(20), T(20), R(20),
+ P(20), KC1, KC2
C******
Ck***** INPUT OF PARAMETER'S NOMINAL VALUES
C******
PRINT *, 'ENTER KC1, KC2, T1, T2, TI1, TI2, C'
ACCEPT *, KC1, KC2, T1, T2, TI1, TI2, C
R(1)=KC1
R(2)=KC2
R(3)=1/TS
R(4)=1/T2
R(5)=1/TI1
R(6)=l/T12
R(7)=1/C
C******
C****** CALCULATE CHARACTERISTIC EQUATION'S
C****** ASSOCIATED WITH EACH PARAMETER
C******
N=0
1 FORMAT(A)
CALL POLYPLUS(T, L, R)
DO 5 N1,7
IF(N. EQ. 1)PARA='KC1'
IF(N. EQ. 2)PARA='KC2'
IF(N. EQ. 3)PARA='RT1'
IF(N. EQ. 4)PARA='RT2'
IF(N. EQ. 5)PARA='RTI1'
IF(N. EQ. 6)PARA='RTI2'
IF(N. EQ. 7)PARA='RC'
DO 10 I=1,7
P(I)=R(I)
10 CONTINUE
P(N)=0.
CALL POLYPLUS(D, L, P)
DO 20 I=1,10
NUM(I)=(T(I)-D(I))/R(N)
20 CONTINUE
PRINT *, ' '
PRINT *, 'FOR ', PARA, ' THEN'
DO 30 I=1, L
PRINT NUM=', NUM(I), ' DENOM=', D(I)
30 CONTINUE
PRINT *, 'HIT RETURN TO CONTINUE'
ACCEPT 1, CHAR
5 CONTINUE
STOP
END
 290

SUBROUTINE POLYPLUS(Z, L, P)
C******
C****** CHARACTERISTIC EQUATION CALCULATION SUBROUTINE

REAL P(20), X(20), Y(20), G(20), Z(20), K1(20), K2(20), K3(20),

+ DUM1(20), DUM2(20), DUM3(20)
INTEGER N, M, L
C******
C****** AGI SIMULATION INFORMATION
C****** ( Linearised at Q=5 Msft3/hr )
C******
C****** Y POLYNOMIAL
Y(5)=12.6346
Y(4)=40.8813
Y(3)=36.7985
Y(2)=7.9963
Y(1)=9.522E-3
C****** X POLYNOMIAL
X(5)=11.77115
X(4)=44.4358
X(3)=51.3635
X(2)=18.4738
X(1)=1.
C****** G11 POLYNOMIAL
G(5)=12.5225
G(4)=40.4582
G(3)=36.3095
G(2)=7.8204
G(1)=0.0
C****** K1 POLYNOMIAL
K1(6)=1.
K1(5)=P(3)+P(4)+10.0274E-3*P(7)
K1(4)=P(3)*(P(4)+18.938E-3*P(1)*P(7))+10.0274E-3*P(7)*(P(3)+P(4))
K1(3)=P(4)*(10.0274E-3+18.938E-3*P(1))+18.938E-3*P(1)*P(5)
K1(3)=P(3)*P(7)*K1(3)
K1(2)=18.938E-3*P(3)*P(4)*P(7)*P(1)*P(5)
K1(1)=0.
C****** K2 POLYNOMIAL
K2(6)=0.
K2(5)=P(4)
K2(4)=P(4)*(P(6)+10.0274E-3*P(7)+P(3))
K2(3)=10.0274E-3*(P(7)*(P(6)+P(3)))+P(3)*(P(6)+18.938E-3*P(1)*P(7)
K2(3)=P(4)*K2(3)
K2(2)=P(3)*P(4)*P(7)*(18.938E-3*P(1)*(P(5)+P(6))+10.0274E-3*P(6))
K2(1)=P(3)*P(4)*P(7)*18.938E-3*P(1)*P(5)*P(6)
DO 10 I=1,6
K2(I)=17.063E-3*P(2)*K2(I)
10 CONTINUE
C****** K3 POLYNOMIAL
K3(6)-0.
K3(5)=13.9532E-3*P(7)
K3(4)=13.9532E-3*P(7)*(P(3)+P(4))
K3(3)=13.9532E-3*P(7)*P(3)*P(4)
K3(2)=0.
K3(1)=0.
 291

C****** FORMULATION OF C. E
N=5
M=6
CALL POLYX(Y, N, KI, M, DUM1, L)
CALL POLYX(X, N, K2, M, DUM2, L)
CALL POLYX(G, N, K3, M, DUM3, L)
DO 20 I=1,10
Z(I)=DUM1(I)+DUM2(I)+DUM3(I)
20 CONTINUE
RETURN
END

SUBROUTINE POLYX(V, N, W, M, U, L)
C******
C****** POLYNOMIAL MULTIPLICATION SUBROUTINE
C******
INTEGER N, M, L
REAL V(20), W(20), U(20)
L=N+M-1
DO 10 I=L, 1, -1
U(I)=0.
DO 20 J=N, 1, -1
IF(N. GT. L)GOTO 20
U(I)=U(I)+V(J)*W(I-J+1)
IF((I-J+2). GT. M)GOTO 10
20 CONTINUE
10 CONTINUE
RETURN
END
 - 292 -

4
PROGRAM

ROUTINEFORDESIGNINGPI CONTROLLERS
FORTRAN

BASEDUPONAPPROXIMATE
MODELS
 - 293 -

C******
Ckk*kkk PI CONTROL USING APPROXIMATE MODELS
C****k* BASED UPON STEP RESPONSE DATA
c******
C******
DIMENSION T(1000), Y(1000), YMOD(1000), E(1000), SUMN(1000)
REAL K, NINIF, KI, K2
CHARACTER*20 ANS, NAME
OPEN(UNIT=2, FILE='STEP. DAT', STATUS='NEW')
5 FORMAT(A)
C******
C****** STEP RESPONSE DATA ENTRY ROUTINE
C****** PRINT *, ' ********* c******ýrýc***********************************'
PRINT *, '***** STEP RESPONSE DATA ENTRY *****'
PRINT *, ' *****************ýt*************ýký1c*******, kýl ***********'
PRINT
PRINT *, 'IS STEP RESPONSE DATA IN FILE - Y/N'
ACCEPT 5, ANS
IF(ANS. EQ. 'Y')THEN
PRINT *, 'ENTER FULL FILE NAME -'
ACCEPT 5, NAME
OPEN(UNIT=1, FILE=NAME, STATUS='OLD')
N=1
10 READ(1, *, END=15)T(N), Y(N)
PRINT *'N, T(N), Y(N)
N=N+1
GOTO 10
15 N=N-1
ELSE
PRINT *, 'ENTER NUMBER OF STEP RESPONSE READINGS'
ACCEPT *, N
PRINT *, 'ENTER DATA FROM t=0 TO t=tmax'
PRINT *, ' '
DO 18 I=19N
PRINT *, 'ENTER TIME, Y(t) FOR DATA POINT ', I
ACCEPT T(I), Y(I)
18 CONTINUE
ENDIF
C******
C,k,ýnk*** FIRST ORDER MODELLING OF STEP RESPONSE

K=(Y(N-1)+Y(N))/2.
SLOPE1=0.
NSLOPE=O
NTAU=O
DO 20 I=29N
IF(Y(I). GE. O. 632*K. AND. NTAU. EQ. O)NTAU=I
SLOPEI=(Y(I)-Y(I-1))/(T(I)-T(I-1))
IF(SLOPEI. GE. SLOPE)THEN
SLOPE=SLOPE1
NSLOPE=I
ENDIF
20 CONTINUE
THETA«(SLOPE*T(NSLOPE)-Y(NSLOPE))/SLOPE
TSLOPE-(Y(NTAU)-Y(NTAU-1))/(T(NTAU)-T(NTAU-1))
THT=(0.632*K+TSLOPE*T(NTAU)-Y(NTAU))/TSLOPE
TAU=THT-THETA
 294

PRINT *, '*****ýcýc********************************************'
PRINT *, '*
PRINT *, '* 1st ORDER MODELLING OF STEP RESPONSE GIVES
PRINT *, '* *'
PRINT 30, '* ', K, ' * EXP(-S *', THETA, ' ) *'
PRINT *, '* ----------------------------
PRINT 30, '* (1+S *', TAU, ' ) *'
PRINT
PRINT *, '*********************************************** *****'
30 FORMAT(1X, A, 2(F8.3, A))
C******
C****** CALCULATION OF MODELS STEP RESPONSE
C******
DO 40 I=19N
IF(T(I). LE. THETA)THEN
YMOD(I)=0.
ELSE
YMOD(I)=K*(1-EXP((-T(I)+THETA)/TAU))
ENDIF
E(I)=Y(I)-YMOD(I)
WRITE(2, *)T(I), Y(I), YMOD(I), E(I)
40 CONTINUE
CLOSE(UNIT=2)
C******
C****** CALCULATION OF THE NORM OF THE MODELLING ERROR
C******
PEAK=1
L=1
SUMN(1)=0.
DO 50 I=2, N
IF(PEAK. EQ. 1. AND. E(I-1). LT. E(I))THEN
L=L+1
PEAK=-1
SUMN(L)=E(I-1)
ELSE IF(PEAK. EQ. -1. AND. E(I-1). GT. E(I))THEN
L=L+1
PEAK=1
SUMN(L)=E(I-1)
ENDIF
50 CONTINUE
NINIF=O
DO 60 I=2, L
NINIF=NINIF+ABS(SUMN(I)-SUMN(I-1))
60 CONTINUE
NINIF=NINIF+ABS(E(N))
C******
C****** OUTPUT OF CONTROLLER DESIGN LIMITATIONS
C*k**** BASED UPON MODELLING ERROR
C******
PRINT
PRINT *, ' *******ýk** +c***************** ºc************************'
PRINT
IF(NINIF/K. LT. 1. )THEN
PRINT '* INTEGRAL ACTION CAN BE APPLIED IN THE FORM - *'
PRINT *, '* *1
PRINT *, '* K1(1 + K2/(K1*S)) *1
ELSE
PRINT *, '* NO INTEGRAL ACTION CAN BE APPLIED THUS
PRINT *' '* PROPORTIONAL CONTROL ONLY WITH THE BOUNDS - *'
PRINT *, '* *1
ENDIF
 - 295 -

PRINT *, '* *'

PRINT 30, '* Ki < `, 1/NINIF '
,
PRINT
PRINT *, '* NOTE !! The higher the limit of K1 the
PRINT *, '* better the modelling exercise.
PRINT *, '*
PRINT *, ' ýkýcýcýcylnkýºcýnk7lryicýcýcyicýrýkýi
***ýtý1rýkýlkýk**-k***ý4ýcýrnk*** k* kýlnkýcýcýlnk7+cýcýc***
PRINT *, 'TO CONTINUE PRESS RETURN -'
ACCEPT 5, ANS
C******
C****** DESIGN OF PI CONTROLLER TO SATISFY DESIGN SPECIFICATIONS
C****** ON PHASE MARGIN AND BANDWIDTH
C******
100 PRINT *, ' '
PRINT *, k kýkýkýc,
kýºcýcýkýcýcýkýkýkýkýkýkýcyl
' ytcýcýtcýcýrýcýkýlcýcýkýcýcýrýkýlc9cýkýkýkýkytcýc** *** kýcýkýc***'
PRINT *, '***** ENTER SPECIFICATIONS FOR *****'
PRINT *, '***** P. I. CONTROLLER DESIGN *****'
PRINT *, ' *********ýIcýicýcil *****ýkýk, kýº********7+ **********************'
PRINT
PRINT
PRINT *, 'ENTER DESIRED SYSTEM BANDWIDTH (rads/sec) -'
ACCEPT *, WD
PRINT *, 'ENTER DESIRED PHASE MARGIN (degrees) -'
ACCEPT *, PHI
PHI=PHI*3.14159/180
ACON=1. /WD**2.
BCON=(1. +(WD*TAU)**2. )/K*k2.
CCON=TAN(ATAN(WD*TAU)+PHI-3.14159/2. +WD*THETA)/WD
K2=SQRT(BCON/(CCON**2. +ACON))
K1=CCON*K2
PRINT
PRINT *, '**********************, k****************, k*************'
PRINT
PRINT *, '* THE CONTROLLER WHICH SATISFIES THE SPEC GIVEN IS *'
PRINT
PRINT 30, '* ', K1, ' +', K2, ' /S
PRINT
PRINT 30, '* WHERE K1 < ', 1/NINIF, '
PRINT *, '* *1
PRINT *, ' ******* ***ýcýkýnk*******ýkýký ************************ýcýc*'
PRINT
PRINT *, 'IS THIS ACCEPTABLE Y/N -'
ACCEPT 5, ANS
C******
C****** CALCULATE THE MODELS CLOSED LOOP RESPONSE AND
C****** SYSTEMS EXPECTED ERROR BOUNDS
C******
IF(ANS. EQ. 'N' )GOTO 100
PRINT *$'IS CLOSED LOOP BOUNDS REQUIRED -'
ACCEPT 5, ANS
PRINT *, '
PRINT *, ' ýcý ****ýnkytýkýcýc****, ýcýn4ýc****ýc********ýký+cýkýcýk***'
PRINT *, '*** OPEN LOOP ***'
PRINT *, '*** STEP RESPONSE MODEL FITTING ***'
PRINT *, '*** INFORMATION IS RECORDED IN ***'
PRINT *, '*** ***'
PRINT *, '*** STEP. DAT ***'
PRINT *, '*** ***'
PRINT *, '*** ARRANGED AS INDICATED BELOW ***'
PRINT *, '*** ***'
PRINT *, '*** T, Y, YMOD, E ***'
PRINT *, '*** ***'
PRINT *' ýcýcýcýnkýcyYýcýcýtýcýcýcýcýcýIcýkýc7lc,
týcýIr***71ri1cýcýkýc7+
****ýkýkýc***'
 - 296 -

IF(ANS. EQ. 'Y')THEN

CALL OWBAND(T, Y, YMOD, E, N, K19K2, K, THETA, TAU, JFLAG)
ENDIF
IF(JFLAG. EQ. 1)GOTO 100
END

SUBROUTINE OWBAND(T, Y, YMOD, E, N, K1, K2, K, THETA, TAU, JFLAG)

C******
C****** ERROR BOUNDS CALCULATION SUBROUTINE
C******
REAL K, K1, K2, NORMW
CHARACTER*20 NAME
DIMENSION T(1000), E(1000), Y(1000), Y1(1000), YMOD(1000),
+ EPS(1000), WA(1000), VA(1000), ZA(1000), ZETA(1000), ASTEP(1000),
+ NORM4(1000), YUP(1000), YDOWN(1000), BSTEP(1000), STEP(1000)
1 FORMAT(A)
JFLAG=O
OPEN(UNIT=15, FILE='DATAFILE. DAT', STATUS='NEW')
DO 5 I=1, N
STEP(I)=1.
5 CONTINUE
C******
C****** CALCULATE MODELS CLOSED LOOP RESPONSE
C******
CALL BOUNDS(T, STEP, ASTEP, YMOD, BSTEP, N, K1, K2, K, THETA, TAU)

C****** CALCULATE MODELS CLOSED LOOP RESPONSE TO E(t)

C******
CALL BOUNDS(T, E, WA, ZA, VA, N, K1, K2, K, THETA, TAU)
NORMW(1)=ABS(WA(1))
DO 10 I=29N
NORMW(I)=NORMW(I-1)+ABS(WA(I)-WA(I-1))
10 CONTINUE
IF(NORMW(N). GE. 1. )THEN
PRINT *, ' N(Wa(t)) >1 RE-DESIGN K(S)
JFLAG=1
GOTO 100
ENDIF
C******
C****** CALCULATE THE FILTERED MODELS
C****** CLOSED LOOP RESPONSE TO E(t)
C******
CALL BOUNDS(T, VA, ZETA, ZA, E, N, K1, K2, K, THETA, TAU)
ZETAMAX=O
DO 20 I=19N
ZINC=ABS(ZETA(I))
IF(ZINC. GT. ZETAMAX)ZETAMAX=ZINC
20 CONTINUE
C******
C****** CALCULATE MODELS CLOSED LOOP BOUNDS
C******
DO 30 I=19N
Y1(I)=YMOD(I)+ZETA(I)
30 CONTINUE
DO 40 I=19N
EPS(I)=(NORMW(I)/(1. -NORMW(I)))*ZETAMAX
40 CONTINUE
 Lv/

DO 50 I=19N
YUP(I)=Y1(I)+EPS(I)
YDOWN(I)=Y1(I)-EPS(I)
50 CONTINUE
C****** STORE CLOSED
C****** LOOP RESPONSE DATA
C******
OPEN(UNIT=2, FILE='DATAFILE. DAT'9STATUS='NEW')
DO 70 I=19N
WRITE(2,60)T(I), YMOD(I), YUP(I), YDOWN(I), WA(I), ZETA(I), EPS(I)
60 FORMAT(7(1X, F8.5))
70 CONTINUE
CLOSE(UNIT=2)
PRINT *0
PRINT *, '*** CLOSED LOOP ***'
PRINT *, '*** BOUNDS INFORMATION IS RECORDED IN ***'
PRINT
PRINT *, '*** DATAFILE. DAT ***'
PRINT *, '*** ***'
PRINT *, '*** ARRANGED AS INDICATED BELOW ***'
PRINT *, '*** ***'
PRINT *, '*** T, YMOD, YUP, YDOWN, WA, ZETA, EPS ***'
PRINT *, '*** ***'
PRINT *, '71cß+c************************** c***********'
100 RETURN
END

SUBROUTINE BOUNDS(T, X, W, Z, V, N, K1, K2, K, THETA, TAU)

C******
C****** CLOSED LOOP SYSTEM SIMULATION SUBROUTINE
C******
DOUBLE PRECISION ERROR, ERRINT, ZDOT, ZOUT
REAL K, K1, K2
DIMENSION T(1000), X(1000), W(1000), Z(1000), V(1000)
ERRINT=O.
ZOUT=O.
TINT=(T(N)-T(N-1))/10.
INC=INT(1.5+THETA/(T(N)-T(N-1)))
DO 20 I=11N
Z(I)=ZOUT
ERROR=X(I)-Z(I)
V(I)=ERROR
W(I)=K1*ERROR+K2*ERRINT
IF(T(I). LE. THETA)THEN
ZDOT=O.
ELSE
ZDOT=(K*W(I-INC)-Z(I))/TAU
END IF
DO 10 J=1,10
CALL INTEG(ERROR, ERRINT, TINT)
CALL INTEG(ZDOT, ZOUT, TINT)
10 CONTINUE
20 CONTINUE
RETURN
END
 298

SUBROUTINE INTEG(YDOT, Y, CINT)

C****** INTEGRATION SUBROUTINE
C******
C****** DOUBLE PRECISION YDOT. Y, Di, D2, D3, D4
D1=CINT*YDOT
D2=CINT*(YDOT+D1/2.0)
D3=CINT*(YDOT+D2/2.0)
D4=CINT*(YDOT+D3)
Y=Y+D1/6+D2/3+D3/3+D4/6
RETURN
END
 299

APPENDIXII

PUBLICATIONS
APPENDIX NOT COPIED

ON INSTRUCTION FROM

UNIVERSITY
 - 368 -

APPENDIXIII

ALTERNATIVEDESIGNMETHODS

FORUSE IN AGI STATIONCONTROL

 J0 !I

AIII. 1

OF DESIGNINGSISO
METHOD
AN ALTERNATIVE

PLUS INTEGRALCONTROLLERS
PROPORTIONAL

USINGAPPROXIMATE
MODELS
 3/u

Previously the design of single-input single-output controllers

has been undertaken using an automated version of Ziegler-Nichols

ultimate method [1] which utilises limit cycle theory to prevent the

propagation of unstable responses throughout the system, see chapter 3.

Unfortunately, one of the drawbacks with this particular method is that

the system cannot limit cycle, using an ideal relay, if the order of

the process is less than third or second if an integrator and ideal relay

is employed in the design of a proportional plus integral controller,

based upon phase margin specification. Traditionally, in these

circumstances, simple design techniques are implemented which are based

upon the use of approximate models to the plant's open loop step response,

comonly referred to as the process, reaction curve. The transfer function

of such a model is typically

-+SSr
ý(S) - (1)
.....

where
K is the plant's gain

e is the plant's time delay

and i is the plant's first order lag time constant

Many authors [1-4] have proposed methods of tuning conventional

controllers based upon the process reaction curve. The more commonly

employed methods [2,3] provide look up tables to select the appropriate

controller parameters. However, very conservative controller settings

have to be implemented if significant error exists between the

would

step responses of the process and its approximate first order model with

time delay; in these cases stability cannot be assured.

Using classical theory these modelling uncertainties can be

allowed for by ensuring adequate gain and phase margins in the frequency

FF
 - 371 -

domain, however, it would be advantageous if the effect of these

modelling errors could be assessed in a more precise way, especially

when applied to time domain analysis.

One possible method which embodies these latter requirements

was proposed by Owensand Chotai's [5,61. The method allows a confident

assessment of the stability and transient performance of the resulting

closed loop design and is based upon the accuracy of the approximate

model to the process. This is achieved by extending the frequency domain

technique of evaluating confidence bands, like the Gershgorin rings used

in multivariable design, about the approximate model's response in the

time domain, the magnitude of the confidence bands being directly

related to the degree of time domain model uncertainty allowing the

designer the option to:

i) Proceed with a more accurate model and reduce the uncertainty,

at the expense of greater complexity of design.

ii) Utilise an approximate model and simplify the design process

with the consequent conservatism necessary within the controller

to ensure stability and performance.

The strength of this method is its flexibility, the user

the complexity/accuracy trade off in both the step response

controlling

and controller design stages. By utilising the various

modelling

proofs presented in references 5,6 upon which the techniques

mathematical

based, the following design methodology was proposed and implemented

are

in software, a listing of which is presented in Appendix 1.4.

Step 1- Data Acquisition

Measure and record the plant's response to a unit step input,

Y(t). It is necessary to remove as much noise as possible from the

without affecting its dynamics.

signal
 - 372 -

Step 2- Approximate Modelling

Choose an approximate model of the process step response;

the software itself fits the process reaction curve with the simple
first order model with time delay described by equation 1.

Hence, use the model to generate the approximate process

step response Ya(t) and therefore the modelling error signal:

E(t) = Y(t) - Ya(t) (2)
.....

Step 3- Open Loop Error Bounds Evaluation

Evaluate the Norm of the error signal NT(E), which is delined

as the total variation of a vector as deduced from the graphical

inspection of its time variation (figure 1a), until some time T using

the formula:
Krnax
NT(E) = IE(O)I +ý IE(tk) - E(tk-1)
K=1

+ E(T) - E(tKmax) (3)

.....

Kmax is the largest integer value such that K. tkmax < T; the

infinity Norm of the signal N. (E) can be calculated by ensuring that

at the time T the step response has attained steady state conditions.
Use of this equation can cause considerable problems if the

signal to noise ratio is small-since the peaks of the noise will be

interpreted as those of the error signal. However, peak detection

analysis should overcome this problem. Moreover, the use of the

infinity norm provides a conservative estimate of the modelling error

as it sums the effects of the error over all time, therefore providing

constant weighting in the frequency domain, as shown in figure lb.

Indeed, use of this information to predict the closed loop error bounds
 373

Modelling error E(t)

E(t2

E(0), E(t.
time

- E(t2)I

E(tl

a. Norm E(t) evaluation

Magnit

t-requency
b. Open loop confidence bounds

Fig. 1 Techniques involved with the use of approximate models

in the design of continuous time PI controllers
 - 374 -

will fail to provide convergence to zero when applied to type 1 open

loop systems. Should more accurate error bounds be required then

various user identified filters must be applied to weight the infinity

norm and the infimum of all of their outputs used [o(s), figure 1b],

that is a bound formed at each individual frequency from the minimum

output of the entire filter set.

Step 4- Controller Design

The maximumpermissible controller gains are dependent upon

the radius of the confidence bands IGa(jw)1 ' . Noo(E), thus the magnitude

of the error Norm is used to impose limits upon the controller settings
to guarantee stability in the closed loop. In the design software

proportional plus integral controllers K(s)

K(s) = K1 + K2/s (4)

,..,.
designed using the linits:

Integral action allowed (K2 x 0) 'ff

iGa(0)1 1. NL(E) <1 (where for a first order

model Ga(O) = K, the
plant gain)

If this is not the case then the model contains insufficient accuracy to

provide stability predictions of the implemented scheme as the radius of

the confidence bounds will encircle the stability point.

maximumgain range permitted (K1 > 0)

Kl N«(E) <1
.
based upon the predicted critical gain of the process developed from the

model and the confidence bounds.

Since the approximate transfer function is in the form of a

first order lag with time delay the proportional and integral gain

can be evaluated by balancing the gain and phase requirements

constants
 - 375 -

of the approximate plant-PI controller system in order to achieve any

specified phase margin and bandwidth. By comparing these settings with

the limits established above then an interactive design procedure is

established. Should the controller fail to satisfy the desired

specifications then a more complex model would be necessary before

proceeding. Should this also fail to provide the required performance

then a re-assessment of the design specifications must be made, or

ultimately, a different controller format chosen.

Step 5- Assess Closed Loop Stability

Calculate the effect the controller K(s) has upon the real

system's closed loop stability by evaluating the response Wa(t) of the

approximate plant's controller output in the closed loop to the open

loop modelling error signal E(t) described in equation 2,

Wa(t) = (1 + K(s)-Ga(s))-'K(s) E(t)

.
The controller K(s) is said to stabilise G(s) if:

i) The system G(s). K(s) is both controllable and observable.

ii) r(N., (Wa)) <1

where r(M) is the spectral radius of the matrix M, that is, the maximum

magnitude of the absolute value of the matrix's eigenvalues.

Step 6- Evaluate Closed Loop Bounds

For analysis purposes the effect of the approximate modelling

exercise and associated controller design upon the closed loop response

of the system is assessed. This is achieved by evaluating bounds e(t)

about the closed loop response of the approximating feedback system

Yc(t) to a unit step input from zero initial conditions, within which

the real plant's closed loop response will lie. The bounds of the
 -3/b-

plant's closed loop response are

Yc(t) - e(t) < Y(t) < Yc(t) + e(t) (5)
.....
where
Nt(Wa)
e(t) = 1-ta Max 1Y1(t) Y0(t)I
-
O<t<T

Y1(t) = YC(t) + C(t)

c(t) is the response of the feedback system

K(s). (1 - H)
+ s. as

to the open loop error signal E(t). H is a filter used to construct

the closed loop error bounds allowing a trade off between mathematical

complexity of the solution and how conservative are the bounds. A

choice of H=0 yields Y0(t) =0 and simplifies the calculation of t(t)

to C(t) = Wa(t) but gives extremely conservative results, whereas the

design software chooses H for accuracy purposes, the 'best' filter

being
Ga(s). K(s)
H=+ass

This yields Y0(t) = Yc(t) and therefore

Nt(Wa)
e(t) = flax IC(t)I (6)
1- r1t(wa) o<t<T .....
The technique outlined above was applied to two examples, the

final design being based upon a specification of 90 degrees phase margin,

thereby cancelling an ideal first order system's pole with the PI

controllers zero, producing an open loop system that consists of an

integrator and a gain, leaving the choice of gain to dictate the

loop system bandwidth.

closed
 - 377 -

Example 1

Consider the unit step response of the first order system:

G1(s)
S(50) +U
and the corresponding approximate first order model as displayed in

figure 2(1). Using the infinity norm to produce the confidence bounds

the following controller is designed to satisfy the above specification

using the software detailed in Appendix 1.4, i. e.

0.105
K1(s) = 4.786 +
s

Figure 2(2) illustrates the expected closed loop performance

of this design when applied to the actual plant and its model. Not

unexpectedly, the corresponding confidence bounds are extremely small

since the first order system model obtained fitted the actual system

almost exactly.

Example 2

Now, consider the second order system:

(s(35) + 1)
G2(s) =
s+s+

Application of the design software to this system produced

the results displayed in figure 3, and the controller

0.102
K2(s) = 0.477 +
s

The closed loop performance of this system, and its corres-

confidence bounds, clearly illustrate the problems that can

ponding
approximate models are used for controller design purposes,
occur when
the confidence bounds predicting that the actual system will be

than the speed of response for which it was designed.

markedly slower
 378

Piagnitude

1.0
0.5

0 Time
seconds)

-0.5 system
Model

-"- Error

First order modelling of the open loop response of system Gi(s) to a

unit step input providing

G1(5)* -+0.951
1 s(45.518)

permitting the application of integral action within the controller

format
K1 +S

providing
K1 < 22.277

Note: The limit upon the magnitude of K1 is directly relative to

the accuracy of the modelling exercise, reducing K1 when a
more conservative controller is necessary.

Fig. 2 Design of a PI controller for Example 1 using the method

(1 of 2) of approximate models.
 siy

Based upon the controller specification of 900 phase margin

at a frequency of 0.1 rads/sec then the following PI controller results

0.105
4.786 +
s

producing the following predicted closed loop response to a unit set

point change

Magnitude

0 ' Time
0 20 40 60 80 (seconds)

Note: The modelling exercise was so accurate that the system's

response under the PI control scheme will be expected to
be the same as that of the model, i. e., very small
confidence bounds exist.

Fig. 2 Design of a PI controller for Example 1 using the method

(2 of 2) of approximate models.
 - 380 -

Magnitude

0.5

0 Time
seconds)

-0.5 system

--- Model

-"-- Error

First order modelling of the open loop response of system CT2(s) to a

unit step input providing:

G2( 5) *_0.982
1 +s (4.MT
permitting the application of integral action within the controller

foniat
Kl +

providing K1 < 2.761

Note: The limit upon the magnitude of K1 is directly relative to

the accuracy of the modelling exercise, reducing K1 when a
more conservative controller is necessary.

Fig. 3 Design of a PI controller for Example 2 using the method

(1 of 2) of approximate models.
 ßd1

Based upon the controller specification of 900 phase margin

at a frequency of 0.1 rad/sec then the following PI controller results:

0.102
0.477 +
s

producing the following predicted closed loop response to a unit set

point change

Magnitude

1 -----------------
0_-- -------------
".

:ý Model

--- Confidence bounds

Time
oI0 20 40 60 80 (seconds)

The system's response under the PI control scheme will

be expected to lie within the confidence bounds
illustrated above.

Fig. 3 Design of a PI controller for Example 2 using the method

(2 of 2) of approximate models.
 - 382 -

The use of approximate models in the design of real time

control systems is very appealing since it generally simplifies the

design procedure; using the method presented within this section the

robustness of the resulting feedback scheme designed using these

approximate models can be assessed. Furthermore the robustness

analysis techniques are applicable to any design method that uses

models of real time processes. The software produced, Appendix 1.4,

uses first order models with time delay and proportional plus integral

controllers in the simplified design of an appropriate feedback control

scheme based upon phase margin and bandwidth specifications. Should

the modelling errors be significant then a conservative design would be

produced that would guarantee closed loop stability; if its performance

is unsatisfactory then a more complex model would be necessary in order

to meet the design specifications, providing they are achievable with

the present controller, and reduce the closed loop confidence bounds.

Because this technique is easily extendable to the multi-

variable case, it could be used to aid in the selection of the tuning

matrices required for operation of the multivariable PI controller

design techniques presented in chapters 5 and 6. Problems arise with

these methods when the first order system models, upon which they are

based, are no longer sufficient to accurately reflect the process

dynamics. By using the above methodology to establish constraints upon

the tuning parameters and predict their subsequent effect upon the

system's closed loop performance, some insight into the design of the

multivariable PI control schemes would be obtained.

 - 383 -

References (AIII. 1)

1 Ziegler, J. G., and Nichols, N. B. "Optimum Settings for

Automatic Controllers". Trans. ASME, Vol. 64, No. 11,
pp 759-768,1942.
2 Cohen, G. H., and Coon, G. A. "Theoretical Consideration
of Retarded Control". Trans. ASME, Vol. 75, pp 827-834,1953.

3 Smith, C. L., and Murill, P. U. "A More Precise Method of

Tuning Controllers". ISA Journal, May 1966.

4 Lopez, A. M., Miller, J. A., Smith, C. L., and Murill, P. W.

"Tuning Relationships Based Upon Integral Performance Criteria".
Instrument Technology, Nov. 1967.

5 Owens, D. H., and Chotai, A. "Stability and Performance

Deterioration Due to Modelling Errors". Nonlinear System
Design - IEE Control Series 25,1984.

6 Owens, D. H., and Chotai, A. "Approximate Models in

Multivariable Process Control". IEE Proc., Vol. 133, No. 1,
1986.
 - 384 -

Nomenclature

e(t) Closed loop step response modelling uncertainty bounds

0
E(t) Open loop step responses's modelling errors

Ga(O) Steady state gain of plant model

G(s) Plant transfer function

H Error bounds conservativeness filter

K Plant gain

K(s) Proportional plus integral transfer function

K1 Proportional gain constant

K2 Integral gain constant

1«. ) Norm of a time response

r(. ) Spectral radius of a matrix

s Laplace operator

Wa(t) Closed loop step response of plant model to E(t)

Y(t) Open loop step response of plant

Ya(t) Open loop step response of plant model

Yc(t) Closed loop step response of plant model

Yo(t) Filtered models output compensation factor

n(s) Infimum of filtered plant outputs

e Plant time delay (sec)

Plant time constant (sec)

T

r. (t) Closed loop step response of filtered model to E(t)

 JOD

AIII. 2

TO CONTROL
A DIGITAL ALTERNATIVE SYSTEMDESIGN
 386

AIII. 2 A DIGITAL ALTERNATIVETO CONTROLSYSTEMDESIGN

All control design procedures presuppose that a mathematical

description of the process to be controlled is available. The design

methods to be discussed in this paper are those based on a 'state-space

plant description', but as vie shall see later the corresponding 'pulse

transfer function model' will also play an important part in highlighting

some of the features of the new algorithm.

General System Descriptions

Most linear, discrete-time, single-input, single-output (SISO),

systems can be represented by a set of first-order difference equations

of the form
Xk+1 = AXk + buk
(1)

yk = SIXk . ....

where Xk = X(kT) is an n-dimensional vector; uk and yk are

scalars; A, b and c' are of appropriate dimension; and, T is the sampling

time.
In pulse transfer function form, yk and uk are related by

Yk=
c'(zI - A)-'b
Uk

blz-1 + b2z-2 + ..... bqz-q

(2)
- .....
1+ alz-1 + a2z-2 + apz-P

where bq * 0, ap x0 and z'1 is the backward shift operator.

If necessary some of the initial bj, j = 1,2, 9 can take zero values
...
to accommodate the presence of a pure time delay.
 - 387 -

System Identification

In order to proceed with any design method the elements of

A, b and c' (in equation 1) or bi, b2 bq and a2, a2 ap (in

... ...
equation 2) must be known. This, of course, in general may not be the

case. One solution to this apparent impasse is to determine a

mathematical description by carrying out an experiment. This is called

system identification. A perusal of the scientific literature quickly

reveals that there are many sophisticated methods available, but all

essentially aim to provide the same kind of information, namely: (i) the

structure of the model, as defined by the choice of p and q (in equation

2) plus an estimate of the system time delay, and (ii) the values of the

individual parameters bl, b2

...
bq and al, a2, ... ap together with some

additional statistical data inferring confidence (or otherwise) in the

parameter estimates. The system identification phase is an important

basic step in the design strategy to be explained. In this study use is

made of the highly interactive micro-CAPTAIN package [1]. The CAPTAIN

(Computer Aided Program for Time-Series Analysis and Identification of

Noisy Systems) package is based on the instrumental variable (IV) approach

to system identification [2,3,4]. This method has the unique advantage

that, in its basic form, it can yield consistent estimates of the

parameters. The basic and refined IV methods for discrete-time systems

have been described in great detail by Young and Jakeman [5].

State-Space Design Methods

Control system design based on pole assignment has been a

popular topic since the early 1960's. It is well known that, in theory,

arbitrary pole assignment can be obtained by linear state variable

feedback, provided the states are available for measurement and the system
 - 388 -

is controllable (see chapter 5). The need either to measure all of the

state variables or to generate surrogate state variables has tended to

discourage its use in practice. The following approach, a joint

collaboration between Sunderland Polytechnic and Lancaster University,

overcomes this problem by defining the necessary state variables in

terms of the measured input and output plant variables. An immediate

advantage over other plant descriptions is that the state variables

are readily available and consequently there is no need to resort to the

use of an observer.

The NMSSPlant Description

The SISO system described by equation 1 is well known. Here we

will consider a special case of this system where the state description

is non-minimal and given by

4= CYk'Yk-1' Yk-p+1' uk-1'uk-2' uk-q+1'mkJT

0.. ...
(3)
.....
where rnk is the "integral of error state" defined by the

equation
mk = mk-1 + {Ykd - Yk} i4)
.....

and ykd is the "desired output". Note that Xk consists of the

present value of the system output as well as past values of both the

plant input and output, mk is included, as mentioned previously, to

guarantee 'type 1' servo-mechanism performance. The non-minimality of

this state space model is obvious: the state has been extended from its

minimum dimension of p to a dimension of p+q. Using equation 2 together

with equations 3 and 4, it is easy to show that the general plant

description of equation 1 can be represented by the following NPISSequations.

 389 -
-

Xk = FXk-1 + quk-1 + dykd

yk = h'Xk i5)
0....

Proportional-Integral-Plus (PIP) Control

The SVF control law can now be written as

uk =- v-'Xk (6)
.....
where

V' = Cfo, fI, f2 ... fp-1

29992 ... gq-1, - k1]
has p+q elements! (7)
.....
The general PIP control algorithm can be developed in two ways: -

* by state space analysis

* by straightforward polynomial algebra based on a block

diagram model form

Let's consider here the second approach. If we define

U(z) _ c[uk], Y(z) _ ý[yk] and M(z) _ F[mk] then in ý transform terms the

control law can be written as

U(z) (fo + flz-1 + fez-2 + ... )Y(z)
(giz-1 + g2z-2 + ... )U(z) + k1M(z) ..... (8)
-

which can be re-expressed as

(1 + g1z-1 + g2z-2 + )U(z)
...
_- f0Y(z) (flz-1 + f2z-2 + ... )Y(z) + k1P1(z)
-
(9)

0 ....

where
1 [Yd(z) - Y(z)] (10)
M(z) = (1 .....
Z-1)
Using equations 9 and 10 we can now construct the block diagram

of figure 1, where it will be seen that, in addition to the proportional

action and the standard, delay free integrator arising from the feedback
 - 390 -

of mk, the controller is characterised by feedback and forward path,

discrete-time filters which are the consequence of the SVF terms from
the remaining state variables.

Computation of the Control Gains

Having justified the form of the basic PIP block diagram of"

figure 1, and suggested a possible implementation we still have the

problem of evaluating the necessary state variable gains. Using block

diagram algebra it is a straightforward matter to obtain the coefficients

of the closed-loop characteristic equation of figure 1 which can then be

equated to those of the desired closed-loop polynomial. Young [6] has

shown that this equivalence can be expressed by the following set of

linear equations
V=Y0 (11)
L.
. 900

where E is a matrix of dimension (p + q) x (p + q) of the

form

p columns q-1 columns

bi 1 bi

b2-b1 b1 al-1 1 b2

b3-b2 b2-b1 bi a2-al al-1 1 b3

b3-b2 b2-b1 .......

a2-a1
a1-i
....... ..
bq-bq+1' b3-b2 ap-ap_1 a2-al bq
.......
-bq bq-bq-1 -ap ap-ap-1 . ..... .
.......
-bq bq-bq_1 -ap ap-ap-1

-bq -ap
(12)
.....
which depends only on the coefficients of the plant model (i. e. equation 2)

while V is the SVF control gain vector defined in equation 7 and ß is the
 - 391 -

Where P Proportional control (fo)

I Integral control (KI)

F Feedback filter (f1z-1 + f2z-2 + +f Z-P+I)

... P-I.
G Pre-filter (1 + g1. z-1 + + (1q_i. z-q+l)
,..
p, q are the system model's denominator, numerator
orders respectively.

Figure 1 PIP Controller block diagram structure

 - 392 -

vector
B' = Cßi, B2, ... ßp+q] (13)
.....
with Bi = di - (ai - ai_1); ao = 1; ai =0 for i>p

where di are the coefficients of the desired characteristic

polynomial:
D(z'1) =1+ dlz-1 + d2Z-2 + ... + dp+qz-(P+q) (14)
.....
The complexity of choice available using equation 14 is reduced

to that of a single parameter (d), if the following multiple pole

specification is selected

D(z-1) = (1 - dz-1)p+q (15)

.....
where the parameter d lies within the range 0d1. This choice of

design will result in a zero overshoot response whose speed and control

action weighting is d dependent, deadbeat control being possible when

d=0, coinciding with the most excessive control actions. Studies to

date [7] have shown the multiple pole closed loop specifications to

produce smoother control actions than if the desired response was

specified in terms of a single pole.

Exam le

Consider a position control system whose "best" pulse transfer

function structure was identified as

b3Z-3 (16)
+ alz-1 .....

Effectively a first order system with a time delay of two sample periods.

Applying instrumental variables to the data displayed in figure 2,

collected at a sampling rate of 0.1 seconds, identified the model parameters

as
ai =-0.235736
(17)
.....
b3 = 0.625264
 - 393 -

Output (x)

- Time
0246 (seconds)

Error

Time
(seconds)
0246

Figure 2 Positioning control scheme identification; time records

compare the responses of the actual system and its
associated discrete time model to a 30% step change in
input.
 - 394 -

Selection of the multiple pole closed loop tuning coefficient

d=0.3 resulted in the desired characteristic equation
D(z-1) =1-1.2z-1 + 0.54z'Z - 0.108z'3 + 0.0081z'4 (18)
.....
Placing the data in equations 17 and 18 into the design

equations resulted in the controller parameters

KI = 0.3840

fo = 0.1184
(19)
.....
gl = 0.0357

g2 = 0.3484
Application of the PIP controller to the system resulted in

the performance illustrated in figure 3. To date, the design technique

has been shown to handle, without modification, a wide range of processes,

providing a powerful tool which requires only a limited knowledge from the

user in order to design a suitable control scheme.

Currently research into the development of a multivariable

version of the technique has been initiated [8,9]. Initial investigations

have produced encouraging results, however, problems with the multivariable

estimator have, as yet, failed to consistently provide the accurate plant

information necessary to design the pole placement solution.
 - 395 -
Full scale

" Time
(seconds)

Figure 3 Performance of the PIP position control scheme;

time records illustrate the response of the system
to 60% step changes.
 - 396 -

References

1. Young, P. C., and Benner, S.

Micro-CAPTAIN handbook.
Version 1, Dept Environmental Science, Univ. of Lancaster,
Report TR50,1987.

2. Young, P. C.
An instrumental variable method for real time identification
of a noisy process.
Automatica, Vol. 6, pp 271-287,1970.

3. Young, P. C.
Recursive estimation and time-series analysis.
Springer-Verlag, Berlin, 1984.

4. Soderstrom, T., and Stoica, P.

The instrumental variable approach to system identification.
Springer-Verlag lecture notes series, Berlin, 1983.

5. Young, P. C., and Jakeman, A. J.

Refined instrumental variable methods of recursive time-series
analysis.
Pts I, II and II, Int. J. of Control, Vol. 29,1979-80.

6. Young, P. C., Behzadi, M., Wang, C. L., and Chotai, A.

Direct digital and adaptive control by input-output state
variable feedback pole assignment.
Int. J. of Control, Vol. 46, No. 6, pp 1867-1881,1987.

7. Boucher, A. R., Cox, C. S., and Young, P. C.

Direct digital control using a personal computer -a new algorithm.
Proc. DTI Colloquium "Personal computers in industrial control",
Warren Springs Laboratory, Stevenage, England.

8. Wang, C. L., and Young, P. C.

Direct digital pole assignment control system design for
multivariable systems based on input-output state variable feedback.
Control 88, Oxford, pp 212-217,1988.

9. Boucher, A. R., and Fletcher, I.

Multivariable proportional-integral-plus (PIP) control - an
initial investigation.
Sunderland Polytechnic internal report, 1988.