Energies 12 02713
Energies 12 02713
Energies 12 02713
Article
Designing Intelligent MIMO Nonlinear Controller
Based on Fuzzy Cognitive Map Method for Energy
Reduction of the Buildings
Farinaz Behrooz 1, * , Rubiyah Yusof 1,2 , Norman Mariun 3,4 , Uswah Khairuddin 1,2 and
Zool Hilmi Ismail 1,2
1 Malaysia -Japan International Institute of Technology (MJIIT), Universiti Teknologi Malaysia,
Kuala Lumpur 54100, Malaysia
2 Centre for Artificial Intelligence and Robotics (CAIRO), Universiti Teknologi Malaysia, Kuala
Lumpur 54100, Malaysia
3 Department of Electrical and Electronic Engineering, Faculty of Engineering, University Putra Malaysia,
Serdang 43400, Selangor, Malaysia
4 Centre for Advanced Power and Energy Research, Faculty of Engineering, University Putra Malaysia,
Serdang 43400, Selangor, Malaysia
* Correspondence: [email protected]; Tel.: +60-10-4315144
Received: 27 May 2019; Accepted: 9 July 2019; Published: 16 July 2019
Abstract: Designing a suitable controller for air-conditioning systems to reduce energy consumption
and simultaneously meet the requirements of the system is very challenging. Important factors
such as stability and performance of the designed controllers should be investigated to ensure the
effectiveness of these controllers. In this article, the stability and performance of the fuzzy cognitive
map (FCM) controller are investigated. The FCM method is used to control the direct expansion
air conditioning system (DX A/C). The FCM controller has the ability to do online learning, and
can achieve fast convergence thanks to its simple mathematical computation. The stability analysis
of the controller was conducted using both fuzzy bidirectional associative memories (FBAMs) and
the Lyapunov function. The performances of the controller were tested based on its ability for
reference tracking and disturbance rejection. On the basis of the stability analysis using FBAMS
and Lyapunov functions, the system is globally stable. The controller is able to track the set point
faithfully, maintaining the temperature and humidity at the desired value. In order to simulate the
disturbances, heat and moisture load changed to measure the ability of the controller to reject the
disturbance. The results showed that the proposed controller can track the set point and has a good
ability for disturbance rejection, making it an effective controller to be employed in the DX A/C system
and suitable for a nonlinear robust control system.
Keywords: intelligent control (IC); fuzzy cognitive map (FCM); direct expansion air-conditioning
(DX A/C); nonlinear; coupling effect; MIMO; stability analysis; Lyapunov function; fuzzy bidirectional
associative memories (FBAMs)
1. Introduction
Heating, ventilating, and air-conditioning (HVAC) systems are a major cause of energy
consumption in the building automation systems (BAS). Because of the depletion of energy sources in
the world and the worsening fuel crisis, the need for advanced, improved control systems designs
aimed at less energy usage and better performance are becoming a significant challenge for control
engineers [1]. Because of the critical influence of HVAC systems on energy and power consumption,
it is imperative to be familiar with the operation as well as the structure of HVAC systems [2].
Nowadays, the use of different controllers on HVAC systems is considered as an important issue, with
the aim of increasing the system performance, by reason of their high installation demand in buildings
and their huge energy consumption.
The direct expansion air-conditioning (DX A/C) system is considered as a subgroup of HVAC
systems. The mentioned system has two types of units, window units and split units, which are
frequently employed in small- to medium-sized buildings by reason of having a simple configuration,
low cost maintenance, and higher energy efficiency [3,4].
According to the authors of [2], energy efficacy and indoor thermal conditions are the main
objectives one needs to take into account when designing HVAC or A/C systems. Because of the
complicated features of HVAC and A/C systems, attaining the mathematical model of HVAC and
air conditioning systems is intricate and difficult. Likewise, designing an appropriate controller
becomes a big challenge [5,6]. HVAC system is a multiple-Input and multiple-output (MIMO) system,
sometimes with coupled parameters [7]. Moreover, it is a complex nonlinear system, which makes
deriving the exact mathematical model a challenging task [3]. Consequently, the control system must
be able to deal simultaneously with cross-coupling effects, nonlinearity, and uncertainty of the system.
The high energy consumption of these systems calls for an intelligent control system that can adjust
the parameters according to the systems’ demands, so that it could prevent energy loss and balance
between the energy consumption and thermal comfort. Needless to say, an appropriate controller
could prevent a significant amount of energy loss [8]. However, designing a suitable controller for
the air conditioning system as a nonlinear, complex MIMO system with coupled parameters is still a
challenging task [8]. According to the authors of [9], a small increase in system operating proficiency
causes major energy savings. Therefore, various studies have investigated different control methods
and optimizations in the HVAC areas.
In the case of the DX A/C system, although one can utilize single-input and single-output (SISO)
control approaches to control each parameter separately, by decoupling the effects of control variables,
yet it degrades the performance of the system [10]. For solving the aforementioned problems and
improving the system performance, the control of the DX A/C system can be viewed as a MIMO control
system [4]. However, based on the literature, MIMO control systems have mostly been employed
on the linearized model of the systems. Thus, the control system is locally stable. In other words,
the controller has the ability to stabilize the system around a certain operating point, and stability
of the system is not guaranteed outside of a small vicinity of the operating point [11]. Therefore,
nonlinear control techniques have been introduced into the DX A/C systems [12,13]. Owing to the
complexity and uncertainty of the system, the application of nonlinear control techniques to the
system is limited. Complicated mathematical analysis, stability analysis, and the need for the entire
states are some aspects that hinder the use of nonlinear control methods [13]. Feedback linearization,
Lyapunov stability theory, and adaptive control are some of the approaches that have been applied to
the system [14].
Some other control methods like model predictive control (MPC) are successfully applied in this
field, but it is necessary to identify a suitable model for the system [15] and installation could be
expensive. This means that an accurate dynamic model from the system is required and, considering
the uncertainties, makes the problem very difficult to solve.
On the other hand, conventional controllers like a proportional-integral-derivative (PID) controller,
usually implemented in a feedback control scheme, are the most applied controllers in this area
because of the simplicity of the control law and easy implementation in continuous-time [16–18].
The feedback control scheme depends heavily on feedback sensors, which have some disadvantages
such as cost, reliability, and noise, which can affect the output of the system. Moreover, control
systems using PID controllers usually have overshoot or undershoot, which means they put sudden
pressure on the actuators and reduce the actuators performance over time. This method suffers from
disadvantages such as long testing times and limited performance. Tuning the parameters of the
controller is cumbersome, and choosing the improper gains of PID controllers can cause the system to
Energies 2019, 12, 2713 3 of 32
be unstable [19]. Considering the above mentioned challenges, a simple control method that can deal
with the complexity of the system and be practically feasible is of interest.
Energies 2018, 11, x FOR PEER REVIEW 3 of 34
106
Figure 1. Direct expansion air-conditioning (DX A/C) framework. EEV, electronic expansion valve.
107 Figure 1. Direct expansion air-conditioning (DX A/C) framework. EEV, electronic expansion valve.
The system mathematical model is extracted from the mass and energy balance conservation
108 The system
principals and themathematical
dynamic systemmodelequations
is extracted [3].from
Thethe mass and
working fluidenergy
of thebalance conservation
refrigeration plant is
109 principals and the dynamic system equations [3]. The working fluid of the refrigeration
Refrigerant 22 (R22), which operates on the basis of the vapor compression cycle’s principle [20,21]. plant is
110 Refrigerant
On the basis22of(R22),
the workwhich operates
of [3], on the basis
the evaporator actsofasthe
an vapor compression
air cooling cycle’s principle
coil to simultaneously [20,21].
dehumidify
111 On the basis of the work of [3], the evaporator acts as an air cooling coil to simultaneously
and cool down the passing air through the evaporator. The emerging moisture and temperature content dehumidify
112 and cool
to the airdown
side ofthethepassing air through
DX evaporator are W the evaporator.
2 and The emerging
T2 , respectively, by the no moisture
fresh airand temperature
assumption [22].
113 content to theair
The passing airtemperature
side of the DX evaporator
through are W2 and
the evaporator hasT2declined
, respectively,
to theby T3the no fresh
value alongair
theassumption
dry-cooling
114 [22]. The[3].
section passing
Becauseair of
temperature through
the small area of thethe evaporator
dry-cooling has declined
section, the entire toevaporator
the T3 value along
wall the dry-
temperature
115 cooling
could besection
assumed [3].toBecause
have theofsame
the small area of(Tthe
temperature w ). dry-cooling
Equation (1)section,
is obtained the using
entirethe
evaporator wall
energy balance
116 temperature
principle forcould be assumed
the dry-cooling to have
section [3].the same temperature (Tw). Equation (1) is obtained using the
117 energy balance principle for the dry-cooling
! section [3]. !
dT3 (T2 + T3 )
Cp ρV
dT
h1 = Cp ρ f (T2 − T3 ) + α1 A1 (Tw −(T T ) (1)
3 dt 23
C V p h1 C f T T A (T
p 2 3 1 1 w
2
(1)
dt 2
118 The temperature of evaporator tube wall is shown by Tw. The airflow rate is f. The volume of the
119 dry-cooling section is illustrated by Vh1, and the area of the heat transfer for dry-cooling region is
120 demonstrated by A1. Oppositely, the heat transfer and the sensible heat transfer obtained in the wet-
121 cooling section between the evaporator wall and air. It is worth mentioning that the coupling effect
122 of air cooling and air dehumidification mostly occurs in the wet-cooling section. The supplied air
Energies 2019, 12, 2713 4 of 32
The temperature of evaporator tube wall is shown by Tw . The airflow rate is f. The volume of
the dry-cooling section is illustrated by Vh1 , and the area of the heat transfer for dry-cooling region
is demonstrated by A1 . Oppositely, the heat transfer and the sensible heat transfer obtained in the
wet-cooling section between the evaporator wall and air. It is worth mentioning that the coupling
effect of air cooling and air dehumidification mostly occurs in the wet-cooling section. The supplied
air temperature by evaporator is shown by T1 , the wet-cooling section volume is Vh2 , the wet-cooling
section area in air side is A2 , and the vaporization of water’s latent heat is shown by hfgh . Thus,
the wet-cooling section energy balance is attained as follows:
! ! !
dT1 dW1 (T3 + T1 )
Cp ρVh2 + ρVh2 h f gh = Cp ρ f (T3 − T1 ) + ρ f h f gh (W2 − W1 ) + α2 A2 Tw − . (2)
dt dt 2
The DX evaporator outlet supply air is assumed to be 95% saturated. As a result, the coupled
supplied air humidity and air temperature by evaporator could be extracted using curve fitting and
plotting [3,4] as follows: ! !
dW1 dT1
− (2 × 0.0198T1 + 0.085) = 0. (3)
dt dt
As the division of the refrigerant side and air side is the evaporator wall, and because of the
considerable different thermal inertia of the air and refrigerant, dynamics responding to changes
on the refrigerant side are considerably faster than on the air side. In other words, it is quite time
consuming for a full response on the air side compared with the refrigerant side. Consequently, using
the same mass flow rate assumption for both inlet refrigerant to evaporator and outlet refrigerant from
evaporator, the energy balance equation of the evaporator wall can be considered as follows:
! ! !
dTw
T2 + T3
(T3 + T1 ) (sVcom )
Cp ρV = α 1 A1 − Tw + α2 A2 − Tw − 1
(hr2 − hr1 ). (4)
w dt 2 2
vs ( PPce ) β
The calculation of the rotor compressor swept volume (Vcom ) was done by applying the geometric
parameters of compressor. The parameter s is speed of motor compressor and superheated refrigerant
specific volume is vs . According to the energy conservation principle for the air-conditioned room,
the sensible energy balance equation is as follows:
!
dT2
Cp ρV = Cp ρ f (T1 − T2 ) + Qspl + Qload . (5)
dt
The feedback of air temperature (T2 ) and air humidity (W 2 ) from the air-conditioned room are
considered equal to the thermal condition of the air-conditioned room. The volume of the flow rate is
demonstrated by f. The supply fan heat gain is shown by Qspl . The volume of the air-conditioned room
is V, and sensible load of room is illustrated by Qload . Referring to the temperature, by increasing the
air flow rate, the supply fan heat gain increases as follows:
On the other hand, for humidity, the air-conditioned room’s moisture mass balance is as follows:
!
dW2
ρV = ρ f (W1 − W2 ) + M. (7)
dt
The humidity load generation in the air-conditioned room is M. According to the authors of [3],
Equations (1–5) and (7), which are all first order differential equations, could form the dynamic model
of the DX A/C system. On the basis of the work done by Qi [22], Equations (1–7) are used as a dynamic
mathematical model of the experimental DX A/C system and have been experimentally validated.
Energies 2019, 12, 2713 5 of 32
Hence, these equations can be used as the DX A/C system model to design a multivariable control
system. In order to design the multivariable nonlinear controller on the DX A/C system, the mentioned
differential equations should be re-written in the state-space format. Therefore, the state-space form of
the model can be expressed in the following compact format:
.
X = D−1 ·g1 (X, U ) + D−1 ·g2 (Z). (8)
0
Qload
0
g2 (Z) = ,
0
0
M
Cp ρVh2 ρVh2 h f gh
0 0 0 0
Cp ρV
0 0 0 0 0
0 0 Cp ρVh1 0 0 0
D =
.
0 0 0 (Cp ρV )w 0 0
−(2×0.0198T1 +0.085)
0 0 0 1 0
1000
0 0 0 0 0 ρV
173
Figure 2. Block diagram of the open loop in step response of fan flow rate.
174 Figure 2. Block diagram of the open loop in step response of fan flow rate.
The results of Figure 3 show that the temperature of the air-conditioned room after
Energies 2018, 11, x FOR PEER REVIEW
290 s started to
6 of 34
increase to 24.15 ◦ C and stabilized at 600 s. By reducing the fan flow rate, the effect of the flow rate of
the fan175
on decliningThe results temperature
of Figure 3 show that the temperature of the air-conditioned room after 290 s started
Energies 2018, the
11, x FOR PEER REVIEW of the room decreased. As is clear from Figure 3, 6 ofthe
34 temperature
176 to increase to 24.15 °C and stabilized at 600 s. By reducing the fan flow rate, the effect of the flow rate
of the177
air conditioned room (T 2 temperature of the room decreased. As is clear from Figure 3, the decreasing
) by decreasing the speed of the fan and, consequently,
175 of theThe fan on declining
results of Figurethe3 show that the temperature of the air-conditioned room after 290 s started
◦ C at 290 ◦ Cand,
the flow
176 ratetoof
178 the fan,
temperature
increase of started
the
to 24.15 °C andto
air increase
conditioned
stabilizedroom from
(Ts.2) By
at 600 24
by reducing
decreasing
thethe sflow
fanspeedto rate,
24.15
of thethe
fan and
effect stabilized
flow rate at 600 s and
ofconsequently,
the
179
remained decreasing
177 stable
of thein the
fanthis flow rate
condition.
on declining of
the temperature of the room decreased. As is clear from and
the fan, started to increase from 24 °C at 290 s to 24.15 °C stabilized
Figure 3, the
180
178 at 600 s and remained
temperature of the air stable in this condition.
conditioned room (T2) by decreasing the speed of the fan and, consequently,
179 decreasing the flow rate of the fan, started to increase from 24 °C at 290 s to 24.15 °C and stabilized
180 at 600 s and remained stable in this condition.
181
Figure 3. Open loop redo simulation result of air conditioned room temperature in step response of fan
182 rate. Figure 3. Open loop redo simulation result of air conditioned room temperature in step response of
flow
181
183 fan flow rate.
182 the results
Also, Figure 3. Open loop redo simulation result of air conditioned room temperature in step response of
184 Also, theof Figure
results 4 indicate
of Figure thatthethe
4 indicate that humidity
humidity ratio ofratio
the air of the air room
conditioned conditioned
after 290 room after
183 fan flow rate.
185 s declined
290 s declined from from 0.01130
0.01130 kg/Kgto
kg/Kg to 0.01113
0.01113 kg/Kg
kg/Kgand remained stable after
and remained 600 s. after
stable As the 600
humidity
s. As of the humidity
186
184 the air-conditioned
Also, the results room
of is is affected
Figure less that
4 indicate by varying the fan
the humidity flow rate,
ratio byconditioned
reduction of the fan
afterflow
of the187
air-conditioned room affected less by varying theofmore
the air
fan flow rate, byroomreduction290 of the fan
185 rate, the humidity
s declined from 0.01130of the air-conditioned
kg/Kg to 0.01113 kg/Kgroom andwas affected
remained by 600
stable after compressor speed, and
s. As the humidity of
flow rate,
188 the humidity
186 consequently
of room
the air-conditioned
decreased.
the air-conditioned
room was affected more by compressor speed, and
is affected less by varying the fan flow rate, by reduction of the fan flow
consequently
187 rate, decreased.
the humidity of the air-conditioned room was affected more by compressor speed, and
188 consequently decreased.
189
189
Figure 4. Open loop redo simulation result of air conditioned room humidity ratio in step response of
fan flow rate.
Energies 2018, 11, x FOR PEER REVIEW 7 of 34
204
Figure 6. Open loop redo simulation result of air conditioned room temperature in step response of
204
205 compressor
Figure speed.
6. Open loop redo simulation result of air conditioned room temperature in step response of
206 compressor speed.
205 Figure 6. Open loop between
The comparison redo simulation
resultsresult of air
by [22] andconditioned room
simulation temperature
results of thisinwork
step response of
shows that the
206 compressor
dynamic speed. model used in this research has almost similar behavior to the experimental
mathematical
model and validated simulation model in the work of [22]. Thus, the mathematical model used in this
work represents the experimental model of the DX A/C system.
Energies 2019, 12, 2713 8 of 32
Energies 2018, 11, x FOR PEER REVIEW 8 of 34
207 Figure 7. Open loop redo simulation result of air conditioned room humidity ratio in step response of
compressor speed.
208 Figure 7. Open loop redo simulation result of air conditioned room humidity ratio in step response of
209 compressor
4. Fuzzy speed.
Cognitive Map
210 The
The development of technologies
comparison between increases
results by [22] andcomplex and nonlinear
simulation results of behaviours in the systems.
this work shows that the
211 Therefore, the requirement
dynamic mathematical modelforused
newinmethods for analysing
this research has almostthesimilar
systems is feasible
behavior [23].
to the The fuzzy
experimental
212 cognitive
model and map (FCM) method
validated has been
simulation modelused
in in
thevarious areas
work of such
[22]. as economy,
Thus, virtual reality
the mathematical modelsimulation
used in
213 in
thisplace
workofrepresents
expert systems [24], sociology,
the experimental or knowledge-based
model of the DX A/C system.expert systems with the purpose of
analysing the complex systems [25]. The developed cognitive maps method was first presented and
214 used for modelling
4. Fuzzy Cognitive the
Map causal relationship among the concepts by the authors of [25,26]. Prior to that,
this method was introduced by a political scientist [27], for decision making in the social sciences and
215 The development of technologies increases complex and nonlinear behaviours in the systems.
political areas. The mechanism of this method is based on replacing the actual values of concepts
216 Therefore, the requirement for new methods for analysing the systems is feasible [23]. The fuzzy
with fuzzy values [28–30]. On the basis of a specified scenario, the fuzzy-graph structure of FCM
217 cognitive map (FCM) method has been used in various areas such as economy, virtual reality
illustrates the parameters of the system by nodes or concepts, and the system characteristics, condition,
218 simulation in place of expert systems [24], sociology, or knowledge-based expert systems with the
and effects of parameters on each other are depicted by causal reasoning [25,26,31]. In actual fact,
219 purpose of analysing the complex systems [25]. The developed cognitive maps method was first
the structure of the FCM method works based on the combination of fuzzy logic and neural networks
220 presented and used for modelling the causal relationship among the concepts by the authors of
methods. The robust properties of both methods are evident in the FCM method [28,31].
221 [25,26]. Prior to that, this method was introduced by a political scientist [27], for decision making in
The FCM structure is formed according to the requested outcome from the system [28].
222 the social sciences and political areas. The mechanism of this method is based on replacing the actual
Consequently, the FCM method has the potential to form as a nonlinear MIMO control system
223 values of concepts with fuzzy values [28–30]. On the basis of a specified scenario, the fuzzy-graph
possibly even with some coupled parameters. Choosing the required and effective parameters from the
224 structure of FCM illustrates the parameters of the system by nodes or concepts, and the system
system as nodes and assigning the suitable weights among them leads to the design of a convenient
225 characteristics, condition, and effects of parameters on each other are depicted by causal reasoning
controller for the HVAC or A/C system. This method is able to keep the structure of the system like
226 [25,26,31]. In actual fact, the structure of the FCM method works based on the combination of fuzzy
MIMO, as well as coupled parameters; accordingly, an improvement in the accuracy and sensitivity of
227 logic and neural networks methods. The robust properties of both methods are evident in the FCM
the controller can be obtained. Moreover, application of this control method on the nonlinear dynamic
228 method [28,31].
model of the system makes the results as similar as possible to the real results from the actual system.
229 The FCM structure is formed according to the requested outcome from the system [28].
230 Consequently, the FCM method has the potential to form as a nonlinear MIMO control system
231 possibly even with some coupled parameters. Choosing the required and effective parameters from
232 the system as nodes and assigning the suitable weights among them leads to the design of a
233 convenient controller for the HVAC or A/C system. This method is able to keep the structure of the
234 system like MIMO, as well as coupled parameters; accordingly, an improvement in the accuracy and
235 sensitivity of the controller can be obtained. Moreover, application of this control method on the
236 nonlinear
Energies 2019,dynamic
12, 2713 model of the system makes the results as similar as possible to the real results 9 of 32
237 from the actual system.
238 FCMs consist of concepts or nodes and edges between the nodes. The mathematical structure of
239 FCM FCMscontains consist
a 1 × of concepts
n state or (C)
vector nodes and edges
(it includes thebetween
values of thethe
nodes.
nodes) The mathematical
and an n × n weightstructure
matrix of
240 FCM contains a 1 × n state vector (C) (it includes the values of the nodes) and
(Wij) (it includes the relationship between concepts), and the number of the nodes is shown by n [32]. an n × n weight matrix
241 (Wij )value
The (it includes
of every theconcept
relationship between
depends on the concepts), and the number
related concepts, of the
with their nodes and
weights is shown by n [32].
the concept’s
242 The value of every concept depends on the related concepts, with their weights
previous value. Value of nodes (Ci) at a particular time reflect the activation degree of nodes in the and the concept’s
243 previousThe
system. value.
nodesValue of nodes
values (Ci ) at a by
are attained particular time reflect
transformation the real
of the activation
valuesdegree of nodes
of concepts intointhe
the
244 system. [0,
interval The 1].nodes
Edgesvalues
betweenare the
attained
nodesby(W transformation of the real values of concepts into the interval
ij) show the influence of each node on the other ones. Each
254 Concept2:3:air
Concept supply humidity
conditioned by evaporator
room’s temperature (W 1(T
). 2).
255 air conditioned
Concept3:4:supply
Concept humidityroom’s humidity
by evaporator (W(W
1).2 ).
256 Concept4:5:air
Concept supply fan flowroom’s
conditioned rate (f ).
humidity (W2).
257 compressor
Concept5:6:supply
Concept speed
fan flow rate(S).
(f).
258 In spite of
Concept 6: the advantageous
compressor speedof(S).
FCMs in design to control the complicated and complex nonlinear
259 systems, the most
In spite of theconsiderable
advantageous weak of point
FCMsofinFCMs design is their potential
to control thetocomplicated
converge to and the undesired
complex
260 steady state.
nonlinear Thus,the
systems, themost
initialconsiderable
weights were defined
weak pointbased
of FCMson theis predetermined
their potential to information
converge to ofthe
the
261 system around an operating point. Thus, to allocate proper initial
undesired steady state. Thus, the initial weights were defined based on the predetermined weights for the FCM network,
262 the generalized
information of the predictive controlan
system around (GPC) method
operating wasThus,
point. chosen to find the
to allocate minimum
proper initial required
weights forvalues
the
263 for control
FCM network,signals, which is suitable
the generalized for the
predictive MIMO(GPC)
control structure [20].was
method Thechosen
relationships
to findofthe concepts
minimum 1 to
264 4 were extracted
required values forfrom the signals,
control direct andwhichindirect relationships
is suitable for the MIMO between the concepts
structure in the
[20]. The linearized
relationships
265 model
of of system.
concepts 1 to 4 wereIn other words,
extracted fromthethe matrixes A and
direct and B of relationships
indirect the linearizedbetween
model were applied in
the concepts to
266 the linearized model of system. In other words, the matrixes A and B of the linearized model wereA
describe the concept weights. For the concepts where the relationship exists directly in matrix
267 and B, the
applied mentioned
to describe the values
conceptwere used For
weights. as athe
direct relationship.
concepts where the Forrelationship
the other concepts with zero
exists directly in
268 values,Athe
matrix and indirect
B, the relations
mentioned with the other
values wereconcepts
used as awere
direct considered and For
relationship. employed.
the other For the actuators
concepts with
269 concepts,
zero values, thetherelationships of them
indirect relations withwiththe other
other concepts
concepts are were defined on the
considered and basis of the KFor
employed. matrix
the
270 calculation
actuators for the GPC
concepts, model of the system.
the relationships of themAs theother
with DX A/C systemare
concepts is MIMO,
definedthe onsteady
the basisspace
of model
the K
271 of the calculation
matrix system wasfor employed
the GPC for
modelfinding
of thethe K matrix
system. As the of theDX GPC model is
A/C system [31]. Regarding
MIMO, the steadythe space
use of
272 these values
model of the as weights
system wasfor the FCMfor
employed method,
findingthey
the have
K matrix to beof normalized to the [31].
the GPC model rangeRegarding
of [−1, 1] [34].
the
Energies 2019, 12, 2713 10 of 32
Thus, the linearized dynamic model of the DX A/C system in state-space representation, which is more
appropriate for designing MIMO control, can be written as Equation (9):
·
X = Ax + Bu
, (9)
y = Cx + d
where the output variables y = [ dT2 , dW2 ]T are the dynamic deviations of air-conditioned room
temperature and humidity from their set points, respectively; and A, B, and C are the coefficient
matrices. According to the GPC method, the compact prediction notation is Equation (10) [35].
A B 0 ··· 0 uk|k
2
A AB B ··· uk+1|k
0
x = .. xk +
→k + 1
.. .. .. .. ..
. . . . .
.
An An−1 B An−2 B ··· B uk+n−1|k
| {z } | {z }| {z }
Px Hx u
→
x = Px xk + Hx u . (10)
→k+1 →k
The output predictions follow a similar method, as shown in Equation (11). The Px depends on
past and the Hx depends upon decision variables [35].
CA CB 0 ··· 0 uk|k dk
CA2 CAB CB ··· u
dk
0 k+1|k
y = . xk + + .
.. .. .. .. .. ..
..
.
→k + 1
. . .
.
CAn CAn−1 B CAn−2 B ··· CB uk+n−1|k dk
| {z } | {z }| {z } |{z}
P H u Ld
→
These are defined relative to an expected steady-state. Implicitly, their values assume a best
estimate of the system disturbance, and of course knowledge of the desired target.
∧ ∧ ∧
e = y = Px xk + Hk u . (12)
→ →k + 1 →k
Finally, by minimising the performance index and rewriting in terms of the actual state and input,
one gets Equation (13) [35]:
−1
uk − uSS = −ET1 [HxT QHx + R] HxT QPx [xk − xSS ]. (13)
| {z }
K
Energies 2019, 12, 2713 11 of 32
By minimizing the (uk − uss ), the least control efforts to reach to the defined set points are calculated.
Equations (14)–(16) show the finalized equations for finding the minimization of performance index.
Entire values should normalize into the interval [−1, 1].
Table 1 shows the numerical values and operating point of the DX A/C system based on the works
of [3,4,22,34], which are used for making the matrices A, B, C, and K using the GPC method for weight
matrix values between the concepts [34].
343
344
T2
345
W1
346
347
348 T1
349 W2
350
351
352 f
S
353
354
355 Schematicofofgraphical
Figure9.9.Schematic
Figure graphicalstructure
structureofofthe
theFCM
FCMcontroller.
controller.
356 Thisgraphical
This graphicalstructure
structureofofthe
thecontroller
controllerfollows
followsthe
thesimple
simplemathematical
mathematicalmodel modelofofFCM,
FCM,which
which
357 contains a 1×n state vector (C), which is the values of the n concepts, and an n × n
contains a 1×n state vector (C), which is the values of the n concepts, and an n × n weight matrix, weight matrix, which
358 is the weights (W
which is the weights ) between the concepts. The number of concepts is six [32,34,36]. On
ij (Wij) between the concepts. The number of concepts is six [32,34,36]. On the basis the basis of the
359 ofconstruction of FCMs,
the construction all nodes’
of FCMs, real values
all nodes’ are normalized
real values in the
are normalized inrange [0, 1][0,1]
the range by Equation
by Equation(17)(17)
as a
360 transformation mechanism. Next, Equations (18) and (19) are employed
as a transformation mechanism. Next, Equations (18) and (19) are employed for calculation of for calculation of concepts
361 activation
concepts level (Ci )level
activation for every concept
(Ci) for every [23,32,34,36]. Lastly, Equations
concept [23,32,34,36]. (18) and (19)
Lastly, Equations (18) were continued
and (19) were
362 continued until the values do not change at all, and the control signal values were defuzzifiedto
until the values do not change at all, and the control signal values were defuzzified and applied the
and
363 actuators [34].
applied to the actuators [34].
if Si t aii f Si t < ai
0, 0,
( S(iSit a−i )a/i )(2(
t
− ai ))if, ai Siift aim≤i Si t < mi
t
/(m2i (maii )),
g(gS(i tS)i
) (17)
(17)
0.5+( S(iSi m−i )m/ (2(
0.5
t t
i ) /b(i 2(mbii )), bi i < Si t ≤ bi
− mifi ))m,i iSfi t m
if Si bii f Si t > bi
1, 1, t
C=new (= f ΣC
old
×j ×)W+ij + Ci
old
(18)
i (18)
1
f (x) = 1 (19)
( )= (1 + e−λx
) (19)
(1 + )
The detected values of the ith state at time t, max, min, and mean are shown by St ti , bi , ai , and mi ,
364 The detected values of the ith state at time t, max, min, and mean are shown
respectively. Cnew depicts the activation level of concept i at time t + 1 and Ccold
by Si , bi, ai, and mi,
discloses the activation
inew j
365 respectively. Ci j atdepicts
level of concept time t. the
Theactivation level of is
threshold function concept
showniby at f.time t + 1 and C cold j discloses the
366 activation
FCMlevel of concept
has been widelyj atused
timein
t. The threshold
practice, function
mainly owingistoshown by f.
its capabilities to deal with complex
367 FCM has
nonlinear been widely
systems. On theused
otherinhand,
practice, mainly
crucially bigowing
steadytostate
its capabilities
error is onetoofdeal
the with complex
disadvantages
368 nonlinear
of the FCMsystems. On the
technique. Toother
solvehand, crucially big steady
the aforementioned state one
problem, errorcan
is one of the
deploy disadvantages
learnability of
of FCM.
369 the FCM technique. To solve the aforementioned problem, one can deploy
Updating the weights by applying supervised learning techniques on FCM makes the FCM techniquelearnability of FCM.
370 Updating the weights
more practical by applying
for online real-timesupervised learning
control problems. techniques
Herein, on FCM makes
the following the FCM
supervised technique
δ learning rules
371 more
were practical
chosen: for online real-time control problems. Herein, the following supervised δ learning
372 rules were chosen: Wij (K + 1) = Wij (K) + ∆Wij (K), (20)
( + 1) = ( ) + ∆ ( ), (20)
∆Wij (K) = ηAi (K)C j (K) 1 − C j (K) d j − C j (K) . (21)
∆ ( )= ( ) ( )(1 − ( ))( − ( )). (21)
The expected value of input node j is dj ; the state value of the nodes connected to this input node
373 is Ci ; theexpected
The value
state value of input
of input node
node j isjCisj ;dthe
j; the state value
learning rate of thewhich
is η, nodesitconnected to this
is in the range [0,input node
1]; and the
374 isiteration
Ci; the state value of input node j is C j; the learning rate is η, which it is in the range [0, 1]; and the
step is K. The concepts’ preferred values should be placed in matrix d.
375 iteration step is K. The concepts’ preferred values should be placed in matrix d.
406
407 Figure Transformation
10. 10.
Figure Transformation of theFCM
of the FCMmethod
method to two
to two layers.
layers.
Energies 2019, 12, 2713 14 of 32
For FCM stability analysis, a specific T-norm should exist, which is obtained from the S-T
composition of the triangular matrices from decomposition of the weight matrix. For this matter,
this T-norm should be equal to the original weight matrix. If this S-T composition exists, the FCM
is globally stable. The product of the S-T composition for two matrices L = (lij )m×p and U = (uij )p×n
should be as follows [37]:
L # U = (w'ij ), (24)
where
w'ij = (li1 Tu1j )S(li2 Tu2j )S...S(lip Tupj ). (25)
In other words, the following steps should be done for the stability analysis of FCMs:
1. Factorization of weight matrix to L and U matrices;
2. Getting a product of S-T composition for L and U matrices based on Equation (25);
3. If such an S-T composition exists for the L and U matrices, which is equal to W, then the
assumed FCM structure is globally stable; otherwise, the proposed stability method cannot solve the
stability problem of the assumed FCM method [37].
Following the above processes and the values from Table 1, the weight matrix and factorization
to the matrix L and U for the proposed designed controller are as below. Accordingly, the designed
controller is globally stable.
−0.279 0.9 0.6 0 0.9295 0.94095
0 −0.9 0 0 −0.26662 0.2267
0.63795 0 0 0.45 0.7707 0.85448
W = ,
0.63795 0 0 −0.9 −0.7707 0.99689
0.4238 −0.417 0 −0.7 1 0
0.3471 0.4686 0.84 0.98 0 1
−0.43734 0 0 0 0 0
0 1 0 0 0 0
1 0 0 0 0 0
L = ,
1 0 0 1 0 0
0.66432 0.46333 0 0.73996 0.98796 0
0.54409 −0.52067 1 0 0 0
0.63795 0 0 0.45 0.7707 0.85448
0 −0.9 0 0 −0.26662 0.2267
0 0 0.84 0.73516 −0.55815 0.65312
U = .
0 0 0 −1.35 −1.5414 0.14241
0 0 0 0 1.7735 1.0402
0 0 0 0 0 −1.8057
Proof of stability analysis by the Lyapunov direct method can be written as below for the fuzzy
cognitive map control method. Lyapunov design refers to the synthesis of control laws for the desired
closed-loop stability properties, using the Lyapunov function for the mentioned nonlinear control
system as follows [39,40]:
.
x = f (x; u(x)), (26)
Energies 2019, 12, 2713 15 of 32
where x Rn is the state and u Rm is the control input. As a result, Equation (11) can be written in the
following form [41]:
( P
Ci = f ( C j (x)wij (x)), f or x = 0
Ci (x) P . (27)
Ci (x) = f (Ci (x − 1) + C j (x)wij (x)), f or x , 0
The f(x) is the unipolar sigmoid threshold function as Equation (15), which squashes the values in
the range of 0 to 1. As the fuzzy cognitive map is a nonlinear dynamical system, the system operates
in the form of dynamics. As a result, the system can be evaluated in the way of the energy function
reduction direction, and it reached a stable state. Therefore, a Lyapunov function for fuzzy cognitive
map, which is called the energy function, by considering the f function as the S shaped continuous
function like Equation (15), can be written as follows [41]:
n n n Z fi ( x )
1 XX X
E(t) = − wij fi (x) f j (x) − f −1 (x)dx. (28)
2 0
i=1 j=1 i=1
As for the activation function f(x), the S-type threshold function is used. The energy function is
bounded defined above, consequently, only proving of the dE dx will show the stability of the system.
∂E ∂ fi (x)
n
dE X
= (29)
dt ∂ fi(x) ∂t
i=1
By combining Equations (29) and (30), Equation (29) can be rewritten as follows:
n X n !2
dE X d fi(x) d fi(x)
= − ( wij f j (x) + . (31)
dt dt dt
i=1 j=1
Because of the choice of the S-type threshold function, which is monotonic type as an activation
function, the function is increasing. It is clear that its inverse function f −1 (x) is also a monotonic type
d fi(x)
and increasing function. Therefore, the value dt ≥ 0, so that dE dt ≤ 0. It is proven that the fuzzy
cognitive map must be stable based on Lyapunov’s theory [39,40].
times of 380 and 1325 s, respectively. The air-conditioned room humidity is shown in Figure 12. The set
initial room humidity is 80% or 0.02157 kg/Kg. It is clearly seen that the humidity stabilized at the
desired set points, with settling times of 1200 and 1357 s, respectively. The humidity declines from 80%
to 48% or 0.02157 kg/Kg to 0.00948 kg/Kg at about 1200 s. After changing the set point at 1300 s, the
humidity reaches to 0.009977 kg/Kg or 53.5% at 1357 s and stabilizes. The humidity is still in the tropical
thermal comfort range, which is in the range of 40% to 60%. The supply air temperature is depicted in
Figure 13. The set initial supply air temperature is 11.25 ◦ C. It is clear that the controller stabilized the
supply air temperature at 180 and 1327 s, respectively. After changing the set point, it increases at
1300 s to 15.85 ◦ C and then decreases gradually to 13.66 ◦ C at near 1327 s and is maintained. Figure 14
illustrates the supply humidity. The set initial supply humidity is 0.009 kg/Kg. It is evident that the
controller stabilized the supply humidity at 1200 and 1329 s, respectively. The humidity rises from
0.009 kg/Kg to 0.01728 kg/Kg at around 29 s, becomes stable at 129 s, and then declines to 0.006 kg/Kg
at almost 1200 s, after which it increases a little to 0.006555 kg/Kg and then stabilizes. After set point
changing, the humidity increasingly changed to 0.009481 kg/Kg at around 1329 s, then decreased to
0.008754 kg/Kg and maintained. The fan flow rate is illustrated in Figure 15. The set initial flow rate is
0.34 m3 /s. It is clear that the proposed controller stabilizes the flow rate at 1200 and 39 s, respectively.
It raises from 0.34 m3 /s to 0.382 m3 /s and stabilize at just about 1200 s. After 1200 s, it decreases
rapidly to 0.325 m3 /s and remains stable at 1300 s. As a result of set point changing at 1300 s, it is
sharply increased to 0.8815 m3 /s, then dropped and finally stabilizes at 0.34 m3 /s at around 1339 s and
maintained. The compressor speed is demonstrated in Figure 16. The set initial compressor speed is
3000 rpm. It is obviously seen that the controller is able to stabilize the compressor speed at 1200 and
1334 s, respectively. The speed raises from 3000 rpm to 6359 rpm, and becomes more or less stable
around 6320 rpm. Afterward, at around 1200 s, it declined to 3960 rpm. From 1300 s, the speed moved
Energies
upwards 2018,to
11,6326
x FOR PEERand
rpm, REVIEW
from 1327 s, it leveled off and maintained at 3180 rpm at 1334 s. 17 of 34
485
Figure 11. Temperature of the air-conditioned room in set point changing.
486 Figure 11. Temperature of the air-conditioned room in set point changing.
485
486 Figure 11. Temperature of the air-conditioned room in set point changing.
Energies 2019, 12, 2713 17 of 32
487
Energies 2018, 11, x FOR PEER REVIEW 18 of 34
Figure 12. Humidity of the air-conditioned room in set point changing.
488 Figure 12. Humidity of the air-conditioned room in set point changing.
489
Figure 13. Supply temperature exit from evaporator in set point changing.
490 Figure 13. Supply temperature exit from evaporator in set point changing.
489
Energies 2019, 12, 2713 18 of 32
490 Figure 13. Supply temperature exit from evaporator in set point changing.
491
Energies 2018, 11, x FOR PEER REVIEW 19 of 34
492 Figure 14. Supply
Figure 14. Supply humidity
humidity exit
exit from
from evaporator
evaporator in
in set
set point
point changing.
changing.
493
Figure 15. Supply flow rate by fan in set point changing.
494 Figure 15. Supply flow rate by fan in set point changing.
493
Energies 2019, 12, 2713 19 of 32
494 Figure 15. Supply flow rate by fan in set point changing.
495
496 Figure 16. Compressor
Figure 16. Compressor speed
speed in
in set
set point
point changing.
changing.
518
Figure 18. Humidity of the air-conditioned room in a disturbance rejection test.
519 Figure 18. Humidity of the air-conditioned room in a disturbance rejection test.
518
519 Energies 2019, 12, Figure
2713 18. Humidity of the air-conditioned room in a disturbance rejection test. 21 of 32
522
523 Figure
Figure 20.
20. Supply
Supplyhumidity
humidityexit
exitfrom
fromthe
the evaporator
evaporator in
in aa disturbance
disturbance rejection
rejection test.
test.
522
Energies 2019, 12, 2713 22 of 32
523 Figure 20. Supply humidity exit from the evaporator in a disturbance rejection test.
525 Figure
Figure 21. Flow rate
21. Flow rate of
of the
the fan
fan in
in a
a disturbance
disturbance rejection
rejection test.
test.
526
Figure 22. Speed of the motor compressor in a disturbance rejection test.
527 Figure 22. Speed of the motor compressor in a disturbance rejection test.
528 8. Discussion
529 Because of the fact that the A/C system is MIMO, and in most previous works it is considered a
530 SISO system, artificially decoupling two strongly coupled feedback loops results in poor transient
531 control performance [22]. The FCM construction has the ability to consider the coupled variables, and
Energies 2019, 12, 2713 23 of 32
8. Discussion
Because of the fact that the A/C system is MIMO, and in most previous works it is considered
a SISO system, artificially decoupling two strongly coupled feedback loops results in poor transient
control performance [22]. The FCM construction has the ability to consider the coupled variables, and
accordingly, the accuracy and sensitivity of control can be increased. In addition, the MIMO controllers
have frequently been applied on the linearized model of the system in neighbourhoods of the operating
point for simplicity. In other words, the system can merely work and be stable nearby a specific
operating point. Thus, the controller cannot be implemented in a wide operating range. In contrast,
the MIMO FCM based controller works on a wide operating range. Subject to HVAC and A/C systems,
applying the nonlinear control methods was restricted, because of the difficult procedure to find the
Lyapunov function, using of the entire states of the system, which causes the nonlinear observer
requirement and difficult mathematical analysis. Because of having the complex and uncertain system,
the need for a simple control algorithm and simple mathematics is feasible. As regards to the FCM
structure and its ability for designing the controller on the basis of the system requirements and what
outcomes are expected from the system, not in what way it works, the FCM method could be an
appropriate solution for HVAC and A/C systems. The FCM method is able to consider the mentioned
requirements of the system such as coupled parameters, MIMO, and nonlinear structures, in a single
control structure. Not only the simple graph structure of the FCM method for designing the controller,
but also the simple mathematics of the FCM method, are beneficial for obtaining the required control
signals values with the intention of reaching the desired set points. In other words, the improvements
by applying the FCM controller are avoiding complex and difficult mathematical analysis to design
the controller; working in a wide operating range; making the nonlinear control design simple by
observer elimination; decoupling cancellation and improving the sensitivity and accuracy of the system;
reducing energy usage and improving energy efficiency; cancelling the overshoots and the undershoots
due to the flexible and adaptable structure of the FCM; improving the transient behaviour of the system
due to flexibility, adaptability, and convergence tendency of FCM; and using single control algorithm
to integrate the DX A/C system’s characteristics.
This section compares the results achieved by the designed FCM controller with the LQG controller
by the authors of [22]. In order to compare the FCM controller with the LQG controller, the results
of the both controller on the linear model of the system around 24 ◦ C are compared in this section.
Both controllers were tested on the linear model of the DX A/C system in initial room temperature
of 24 ◦ C and 50% room humidity or 0.0095 kg/Kg humidity. The results of the LQG controller were
adopted from the work of [22]. Figure 23 shows the results of the air-conditioned room temperature by
applying LQG controller on linear model of the system around the operating point, Figure 24 shows
the air-conditioned room humidity, Figure 25 shows the compressor speed, and Figure 26 shows the
supply fan speed.
The air-conditioned room temperature is indicated in Figure 23. It is clearly seen that the
temperature of the air-conditioned room stabilized around 24 ◦ C. When the set point changed at 500 s
to 23.5 ◦ C, the air-conditioned room temperature declined from 24 ◦ C to 23.5 ◦ C. Increasing the supply
fan flow rate and compressor speed, the temperature of the air-conditioned room decreased from 24 ◦ C
in 520 s to 23.5 ◦ C at around 2000 s. The temperature reaches the preferred set point in 25 min with no
error and is maintained at the desired set point.
The air-conditioned room humidity is illustrated in Figure 24. As it is clear from the figure that
the air-conditioned room humidity stabilized around 0.00965 kg/Kg. By changing the set point in
500 s to 23.5 ◦ C, the humidity of the air conditioned room declined from 0.0095 kg/Kg to 0.0093 kg/Kg.
The humidity of the air-conditioned room decreased from 0.0095 kg/Kg to 0.0093 kg/Kg in about 2750 s
with 2.5% error from 50% humidity by varying of the compressor speed and supply fan flow rate.
Then, the air-conditioned room humidity stabilized at 0.0093 kg/Kg.
558 results of the both controller on the linear model of the system around 24 °C are compared in this
559 section. Both controllers were tested on the linear model of the DX A/C system in initial room
560 temperature of 24 °C and 50% room humidity or 0.0095 kg/Kg humidity. The results of the LQG
561 controller were adopted from the work of [22]. Figure 23 shows the results of the air-conditioned
562 room temperature by applying LQG controller on linear model of the system around the operating
Energies
563 12, 271324 shows the air-conditioned room humidity, Figure 25 shows the compressor speed,24 of 32
2019,Figure
point,
564 and Figure 26 shows the supply fan speed.
565
Figure 23. Simulation results of air-conditioned room temperature by applying the Linear Quadratic
566 Figure
Energies 23.xSimulation
2018, 11, results of air-conditioned room temperature by applying the Linear Quadratic
FOR PEER REVIEW
Gaussian (LQG) controller on the linear model of the system around the operating point. 25 of 34
567 Gaussian (LQG) controller on the linear model of the system around the operating point.
568 The air-conditioned room temperature is indicated in Figure 23. It is clearly seen that the
569 temperature of the air-conditioned room stabilized around 24 °C. When the set point changed at 500
570 s to 23.5 °C, the air-conditioned room temperature declined from 24 °C to 23.5 °C. Increasing the
571 supply fan flow rate and compressor speed, the temperature of the air-conditioned room decreased
572 from 24 °C in 520 s to 23.5 °C at around 2000 s. The temperature reaches the preferred set point in 25
573 min with no error and is maintained at the desired set point.
574
575 Figure 24. Simulation results of air-conditioned room humidity by applying the LQG controller on
576 the linear model of the system around the operating point.
577 The air-conditioned room humidity is illustrated in Figure 24. As it is clear from the figure that
578 the air-conditioned room humidity stabilized around 0.00965 kg/Kg. By changing the set point in 500
579 s to 23.5 °C, the humidity of the air conditioned room declined from 0.0095 kg/Kg to 0.0093 kg/Kg.
574
580 The humidity of the air-conditioned room decreased from 0.0095 kg/Kg to 0.0093 kg/Kg in about 2750
575
581 Figure
s with 24. Simulation
Figure
2.5%24.error fromresults
Simulation ofofair-conditioned
results
50% air-conditioned
humidity by varying room
room humidity
humidity
of the bybyapplying
compressorapplying
speed the
theand
LQGLQG controller
controller
supply flowonrate.
fan on the
576
582 linear
Then,themodel
linear
the of the of
model system
air-conditioned around
the system
room the the
around
humidityoperating point.
operating
stabilizedpoint.
at 0.0093 kg/Kg.
577 The air-conditioned room humidity is illustrated in Figure 24. As it is clear from the figure that
578 the air-conditioned room humidity stabilized around 0.00965 kg/Kg. By changing the set point in 500
579 s to 23.5 °C, the humidity of the air conditioned room declined from 0.0095 kg/Kg to 0.0093 kg/Kg.
580
The humidity of the air-conditioned room decreased from 0.0095 kg/Kg to 0.0093 kg/Kg in about 2750
581 s with 2.5% error from 50% humidity by varying of the compressor speed and supply fan flow rate.
582 Then, the air-conditioned room humidity stabilized at 0.0093 kg/Kg.
583
Figure 25. Simulation results of compressor speed by applying the LQG controller on the linear model
584 Figure 25. Simulation results of compressor speed by applying the LQG controller on the linear model
of the system around the operating point.
585 of the system around the operating point.
586 The speed of the compressor is depicted in Figure 25. Decline in the air-conditioned room
587 temperature was obtained by increasing the compressor speed from 4000 rpm to 4001.6 rpm from 500
583
588 s to 1500 s. In order to keep temperature and humidity of the room in preferred set points, the
589
584 compressor
Figure 25.speed maintained
Simulation results ofatcompressor
4001.6 rpm.
speed by applying the LQG controller on the linear model
Energies 2019, 12, 2713 25 of 32
The speed of the compressor is depicted in Figure 25. Decline in the air-conditioned room temperature
was obtained by increasing the compressor speed from 4000 rpm to 4001.6 rpm from 500 s to 1500 s.
In order to keep temperature and humidity of the room in preferred set points, the compressor speed
maintained at 4001.6 rpm.
Figure 26 displays the supply fan speed. The fan speed raised from 2180 rpm at 500 s to 2193 rpm
at 1500 2018,
Energies s and11,maintained.
x FOR PEER REVIEW
While the air-conditioned room temperature achieved the preferred 26 ofvalues,
34
the fan speed maintained to the essential speed for keeping the air conditioned room humidity and
592 temperature
values, the fan speed maintained to the essential speed for keeping the air conditioned room
in chosen set points.
593 humidity and temperature in chosen set points.
594
Figure 26. Simulation results of supply fan by applying the LQG controller on the linear model of the
595 Figure 26. Simulation results of supply fan by applying the LQG controller on the linear model of the
596 system
system around
around the operatingpoint.
the operating point.
597 The
The results
results of the FCM
of the FCMcontroller
controlleron onthe
thelinear
linearmodel
model of of
thethe system
system around
around the the operating
operating pointpoint
598 are
are shown as follows. The temperature of the air-conditioned room is indicated in Figure 27. It is It is
shown as follows. The temperature of the air-conditioned room is indicated in Figure 27.
clearly ◦ C. At 500 s, the set
599 clearly seen thatthe
seen that theair-conditioned
air-conditioned roomroom temperature
temperature stabilized
stabilized around around 24 500
24 °C. At s, the set point
point changed ◦ ◦
600 changed from from
24 °C24to C to 23.5
23.5 C. Therefore,
°C. Therefore, by changing
by changing the point,
the set set point,
the the air-conditioned
air-conditioned room room
601 temperature
temperature reduced
reduced byby varying
varyingthethesupply
supplyfanfanflow
flowrate
rate and
and compressor
compressor speed.
speed. TheThe air-conditioned
air-conditioned
room ◦ C to 23.5 ◦ C at around 898 s. The temperature stabilized in the
602 roomtemperature reduced from
temperature reduced from24 24°C to 23.5 °C at around 898 s. The temperature stabilized in the
603 preferred
preferredset set point
point in 6.63 min
min with
withno noerror.
error.
604 The humidity
The humidity of of the
the air-conditioned
air-conditionedroom roomisisillustrated
illustratedin in
Figure
Figure28.28.
As is
Asclear fromfrom
is clear the figure,
the figure,
605 theair-conditioned
the air-conditioned room room humidity
humidity stabilized
stabilizedaround
around0.0095
0.0095kg/Kg.
kg/Kg.ByBychanging
changing thethe setset
point in in
point 500500 s
606 tos to
23.5 ◦ C,
23.5 °C,thethehumidity
humidity of of the
the air conditioned
conditionedroom roomdeclined
declined from
from0.0095
0.0095kg/Kg
kg/Kg to 0.0093 kg/Kg.
to 0.0093 kg/Kg.
607 By altering the supply fan flow rate and compressor speed, the air-conditioned
By altering the supply fan flow rate and compressor speed, the air-conditioned room humidity reduced room humidity
608 reduced
from from
0.0095 0.0095tokg/Kg
kg/Kg 0.0093to kg/Kg
0.0093 in
kg/Kg
aboutin about
1044 s1044
withs 1.25%
with 1.25%
errorerror
fromfrom50%50% humidity.
humidity. Then,
609 Then, the air-conditioned room humidity stabilized at 0.009257
the air-conditioned room humidity stabilized at 0.009257 kg/Kg in about 1600 s. kg/Kg in about 1600 s.
Figure 29 shows the speed of the compressor. For declining the temperature of the air-conditioned
room, the compressor speed increased from 4000 rpm to 4003 rpm from 500 s to 689 s. By declining
the air-conditioned room temperature to the desired value, the speed of the compressor decreased
and maintained at 4001 rpm at 803 s to keep the air- conditioned room temperature and humidity in
desired set points.
Energies
Energies 2018,12,
2019, 11,2713
x FOR PEER REVIEW 27 of2634of 32
Energies 2018, 11, x FOR PEER REVIEW 27 of 34
610
610
611 Figure
Figure27.27.Simulation
Simulation results of air-conditioned
air-conditionedroom
roomtemperature
temperature
byby applying
applying thethe
FCMFCM controller
controller on on
611 Figure 27. Simulation results of air-conditioned room temperature by applying the FCM controller on
612 the
thelinear
linearmodel
modelofofthe
thesystem
system around the
the operating
operatingpoint.
point.
612 the linear model of the system around the operating point.
613
613
614 Figure
Figure28.28.Simulation
Simulationresults
resultsofofair-conditioned
air-conditionedroom
roomhumidity
humidity by
by applying
applying the
the FCM
FCM controller
controller on
on the
614 Figure
linear 28.
model Simulation
of the results of air-conditioned room humidity by applying the FCM controller on
615 the linear model ofsystem around
the system the the
around operating point.
operating point.
615 the linear model of the system around the operating point.
616 Figure 29 shows the speed of the compressor. For declining the temperature of the air-
617 conditioned room, the compressor speed increased from 4000 rpm to 4003 rpm from 500 s to 689 s.
618 By declining the air-conditioned room temperature to the desired value, the speed of the compressor
619 decreased
Energies 2019, and maintained at 4001 rpm at 803 s to keep the air- conditioned room temperature
12, 2713 and
27 of 32
620 humidity in desired set points.
621 The supply fan speed is shown in Figure 30. The fan speed increased from 2180 rpm at 500 s to
The supply fan speed is shown in Figure 30. The fan speed increased from 2180 rpm at 500 s to 2317
622 2317 rpm at 1056 s and, after reaching the preferred values, it reduced to 2268 rpm at 1779 s and
rpm at 1056 s and, after reaching the preferred values, it reduced to 2268 rpm at 1779 s and maintained.
623 maintained. After reaching the chosen set point for temperature of the room, at 1056 s, the fan speed
After reaching the chosen set point for temperature of the room, at 1056 s, the fan speed started to
624 started to decline until it met the essential speed to keep the temperature and humidity of the air-
decline until it met the essential speed to keep the temperature and humidity of the air-conditioned
625 conditioned room in chosen set points. The fan speed reduced to 2268 rpm at 1779 s and maintained
room in chosen set points. The fan speed reduced to 2268 rpm at 1779 s and maintained this speed.
626 this speed.
627
Figure 29. Simulation results of compressor speed by applying the FCM controller on the linear model
628 Figure 29. Simulation results of compressor speed by applying the FCM controller on the linear model
of2018,
the system around the operating point.
Energies 11, x FOR PEER REVIEW 29 of 34
629 of the system around the operating point.
630
Figure 30. Simulation results of supply fan speed by applying the FCM controller on the linear model
631 Figure 30. Simulation results of supply fan speed by applying the FCM controller on the linear model
of the system around the operating point.
632 of the system around the operating point.
633 The comparison between both controllers shows that the FCM controller managed to reach the
634 desired set point for temperature of the air-conditioned room in 6.63 min with no error, while the
635 LQG one managed to do so in 25 min with no error. In addition, the humidity of the air-conditioned
636 room by the FCM controller reaches the desired set point after 544 s with 1.25% error, but the LQG
Energies 2019, 12, 2713 28 of 32
The comparison between both controllers shows that the FCM controller managed to reach the
desired set point for temperature of the air-conditioned room in 6.63 min with no error, while the LQG
one managed to do so in 25 min with no error. In addition, the humidity of the air-conditioned room
by the FCM controller reaches the desired set point after 544 s with 1.25% error, but the LQG one
reaches the desired set point after 2250 s with 2.5% error. The benefits and shortcomings of both control
methods are discussed as follows. Although the LQG controller has attractive features of thermal
comfort and energy savings, rapid response [15], increased energy savings [15], and multivariable
control [15], the disadvantages of the LQG control method are as follows.
• It was applied on the linear model of the system. Thus, it is not applicable for a wider
operating range.
• The method is inherently complex. As a result, the optimal feedback control’s evaluation and
online implementation turn into a daunting challenge.
• The need to the model of the system. Thus, the appropriate model of the system should be
identified [15].
On the other hand, the FCM controller avoids complex and difficult mathematical analysis to
design the nonlinear controller. It works in a wide operating range and simplifies the designing of
the nonlinear controller by removal of the observer. It improves the sensitivity and accuracy of the
system by keeping the coupled structure of the system. The FCM controller is designed in a single
control scenario for consideration of the system’s characteristics. However, this control method is quite
sensitive to the initial state values and to the type of concepts used in the FCM model, and it can give
different results as they are different from and non-compatible with classic fuzzy.
The comparison of the two techniques’ performances could be assessed by evaluating criteria such
as integral square error (ISE), integral of absolute value of error (IAE), and integral of time-multiplied
absolute value of error (ITAE). For these tests, the two controllers were tested under the same condition.
The LQG controller was tested in initial room temperature of 30 ◦ C and 80% room humidity or
0.02157 kg/Kg humidity, and a desired temperature of 25 ◦ C and 50% humidity for comparison with
the proposed FCM controller. Figure 31 shows the temperature of the air-conditioned room, which
declined from 30 ◦ C to 22.88 ◦ C in 380 s, and then after oscillation around 25 ◦ C, it stabilized at 1200 s
with small error. Figure 32 illustrates the humidity of the air-conditioned room, which reduced from
0.022 kg/Kg to 0.008 kg/Kg, and it was unable to stabilize at certain values, which makes the system
unstable. It is obviously seen that the LQG controller does not have the ability to control the humidity.
The LQG controller is only able to control the condition around the operating point.
Table 2 shows the performance indexes results for both controllers based on the room temperature error.
Table 3 indicates the performance indexes results for both controllers based on the room humidity error.
It is obvious from the results that the FCM controller has the ability to reduce the performance
index. This means that for the real conditions of the room, the FCM controller works better than
the LQG one. In other words, the FCM controller is more suitable and applicable. Also, the LQG
controller was not able to control different conditions from the operating point and shows instability
in the system. In other words, the LQG controller is severely dependant on the operating point, and
changing the operating point makes the system unstable. In conditions where the controller is more
dependent on the operating point, there is a need to design separate controllers for each operating
point in the system. On the other hand, the Kalman filter in the LQG structure is suitable for linear
systems or the linearized model of the nonlinear systems. As a result, working in wider operating
points requires designing controllers for every operating point or using the multimodal controllers, or
applying the extended Kalman filter (EKF) for calculating the operating point of the system in every
second and extracting the required system parameters. By using each option in order to work in
wider operating points, the volume of calculation is increased or designing the controller becomes
complicated. The proposed FCM controller is not dependent on the operating point, and hence there is
not a need to design more controllers at different operating points.
Energies 2018, 11, x FOR PEER REVIEW 30 of 34
663 the LQG controller does not have the ability to control the humidity. The LQG controller is only able to
664 control
Energies the12,
2019, condition
2713 around the operating point. 29 of 32
667
Figure 32. Humidity of the air-conditioned room by the LQG controller.
668 Figure 32. Humidity of the air-conditioned room by the LQG controller.
669 Table 2 shows the performance indexes results for both controllers based on the room
670 temperature error. Table 3 indicates the performance indexes results for both controllers based on the
671 room humidity error.
672 Table 2. Results of the performance indexes for both controllers based on the room temperature error.
Energies 2019, 12, 2713 30 of 32
Table 2. Results of the performance indexes for both controllers based on the room temperature error.
Performance index
Controller
ISE IAE ITAE
LQG 2.308 × 104 4096 1.763 × 106
FCM 2857 1576 8.217 × 105
Table 3. Results of the performance indexes for both controllers based on the room humidity error.
Performance index
Controller
ISE IAE ITAE
LQG 0.0602 7.719 3457
FCM 0.05038 6.699 2674
9. Conclusion
In this technical research, a new intelligent FCM control system was designed on the DX A/C
system. This FCM control system could be easily designed for HVAC systems. The offered FCM
controller is advantageous because of the graph structure of the FCM method, as well as its simple
mathematics. Therefore, the FCM controller avoids the complex mathematical analysis, which
is required to design the controller for a MIMO and nonlinear system with a complex structure.
The proposed FCM controller met the requirements of designing a suitable controller for the DX A/C
system. The requirements for designing the controller are consideration of nonlinearity, coupling effect,
and MIMO structure, which are all considered in the offered FCM control method. The significance of
this research is offering a new intelligent MIMO nonlinear controller for the DX A/C system. According
to the results of the controller, the offered control system was designed easily and applied efficiently.
The results obviously indicate the efficiency of the controller in the reference tracking signal. It is not
successful only in the normal conditions but also in the occurrence of the heat and moisture loads
disturbances works successfully. Designing the controller based on the FCM method does not require
the mathematical model of the system and the suitable model of the system not like MPC method.
It does not involve difficult mathematic analysis to derive the control law opposite to designing the
nonlinear control or adaptive control methods. It has a wide operating range, unlike the MIMO
controller, and it does not have overshoot and undershoot, which is different from the PID control
method. Moreover, the stability of the FCM control method was proven by the Lyapunov function,
as well as the FBAMs method. Therefore, the FCM control method can be designed for different
nonlinear MIMO systems. Furthermore, it can be a suitable solution for designing the controller for
the systems with no mathematical model or complex mathematical models.
References
1. Tachwali, Y.; Refai, H.; Fagan, J. Minimizing HVAC Energy Consumption Using a Wireless Sensor Network.
In Proceedings of the IECON 2007 33rd Annual Conference of the IEEE Industrial Electronics Society, Taipei,
Taiwan, 5–8 November 2007; pp. 439–444.
Energies 2019, 12, 2713 31 of 32
2. Tashtoush, B.; Molhim, M.; Al-Rousan, M. Dynamic model of an HVAC system for control analysis. Energy
2005, 30, 1729–1745. [CrossRef]
3. Qi, Q.; Deng, S.S. Multivariable control-oriented modeling of a direct expansion (DX) air conditioning (A/C)
system. Int. J. Refrig. 2008, 31, 841–849. [CrossRef]
4. Qi, Q.; Deng, S.S. Multivariable control of indoor air temperature and humidity in a direct expansion (DX)
air conditioning (A/C) system. Build. Environ. 2009, 44, 1659–1667. [CrossRef]
5. Wang, J.; An, D.; Lou, C. Application of Fuzzy-PID Controller in Heating Ventilating and Air-Conditioning
System. In Proceedings of the 2006 International Conference on Mechatronics and Automation, Luoyang,
China, 25–28 June 2006; pp. 2217–2222.
6. Lei, J.; Hongli, L.; Cai, W. Model Predictive Control Based on Fuzzy Linearization Technique FOR HVAC
Systems Temperature Control. In Proceedings of the 2006 1ST IEEE Conference on Industrial Electronics and
Applications, Singapore, 24–26 May 2006; pp. 1–5.
7. Matysko, R. Theoretical model of the operation parameters regulated by the MIMO and SISO system in a
cooling chamber. Int. J. Refrig. 2015, 58, 53–57. [CrossRef]
8. Huang, G. Model predictive control of VAV zone thermal systems concerning bi-linearity and gain nonlinearity.
Control Eng. Pract. 2011, 19, 700–710. [CrossRef]
9. Lü, H.; Jia, L.; Kong, S.; Zhang, Z. Predictive functional control based on fuzzy T-S model for HVAC systems
temperature control. J. Control Theory Appl. 2007, 5, 94–98. [CrossRef]
10. Mirinejad, H.; Ghasemian, M.; Torab, H.; Sadati, S.H. Control Techniques in Heating, Ventilating and Air
Conditioning (HVAC) Systems. J. Comput. Sci. 2008, 4, 777–783. [CrossRef]
11. Venkatesh, S.; Sundaram, S. Intelligent Humidity Control for Healthy Home to Wealthy Industry: A Review.
Res. J. Inf. Technol. 2012, 4, 73–84. [CrossRef]
12. Kassai, M. Experimental investigation on the effectiveness of sorption energy recovery wheel in ventilation
system. Exp. Heat Transf. 2017, 31, 106–120. [CrossRef]
13. Gruber, P.; Balemi, S. Overview of non-linear control methods; Technical Report; Swiss society for automatic
control: St. Gallen, Switzerland, 2010.
14. Afram, A.; Janabi-Sharifi, F. Theory and applications of HVAC control systems—A review of model predictive
control (MPC). Build. Environ. 2014, 72, 343–355. [CrossRef]
15. Perera, D.W.U.; Pfeiffer, C.F.; Skeie, N.O. Control of temperature and energy consumption in building—A
review. Int. J. Energ. Environ. 2014, 5, 471–484.
16. Kozák, Š. State-of-the-art in control engineering. J. Electr. Syst. Inf. Technol. 2014, 1, 1–9. [CrossRef]
17. Kassai, M.; Kajtar, L.; Nyers, J. Experimental optimization of energy consumption for DC refrigerator by PID
controller tuning and comparison with On-Off refrigerator. Therm. Sci. 2019, 23, 941–952. [CrossRef]
18. Raut, K.H.; Vaishnav, S.R. A study on Performance of Different PID Tuning Techniques. In Proceedings of
the International Conference on Electrical Engineering and Computer Science, Trivandrum, India, 12 May
2012; pp. 250–254.
19. Yan, Y.; Song, Y.; Wu, S. Control strategies for indoor environment quality and energy efficiency—A review.
Int. J. Low-Carbon Technol. 2013, 10, 305–312.
20. Shiming, D. A dynamic mathematical model of a direct expansion (DX) water-cooled air-conditioning plant.
Build. Environ. 2000, 35, 603–613. [CrossRef]
21. Shiming, D. The application of feedforward control in a direct expansion (DX) air conditioning plant. Build.
Environ. 2002, 37, 35–40.
22. Qi, Q. Multivariable Control of Air Temperature and Humidity in a Space Served by Direct Expansion Air
Conditioning System. Ph.D. Thesis, Polytechnic University, Hong Kong, China, 2009.
23. Papageorgiou, E.; Stylios, C.; Groumpos, P.; Papageorgiou, E. Active Hebbian learning algorithm to train
fuzzy cognitive maps. Int. J. Approx. Reason. 2004, 37, 219–249. [CrossRef]
24. Huerga, A. A Balanced Differential Learning Algorithm in Fuzzy Cognitive Maps. In Proceedings of the
16th International Workshop on Qualitative Reasoning, Barcelona, Spain, 2 June 2002; pp. 10–12.
25. Khor, S.; Khan, M. Scenario Planning Using Fuzzy Cognitive Maps. In Proceedings of the ANZIIS2003 8th
Australian and New Zealand Intelligent Information Systems Conference, Sydney, Australia, 10–12 December
2003; pp. 311–316.
26. Kosko, B. Fuzzy cognitive maps. Int. J. Man. Mach. Stud. 1986, 24, 65–75. [CrossRef]
Energies 2019, 12, 2713 32 of 32
27. Axelord, R. Structure of Decision: The Cognitive Maps Political Elites; Princeton University Press: Princeton, NJ,
USA, 1976.
28. Stylios, C.; Groumpos, P. Fuzzy cognitive maps in modeling supervisory control systems. J. Intell. Fuzzy Syst.
2000, 8, 83–98.
29. Kim, M.C.; Kim, C.O.; Hong, S.R.; Kwon, I.H. Forward–backward analysis of RFID-enabled supply chain
using fuzzy cognitive map and genetic algorithm. Expert Syst. Appl. 2008, 35, 1166–1176. [CrossRef]
30. Bertolini, M. Assessment of human reliability factors: A fuzzy cognitive maps approach. Int. J. Ind. Ergon.
2007, 37, 405–413. [CrossRef]
31. Aguilar, J. A survey about fuzzy cognitive maps papers (Invited paper). Int. J. Comput. Cogn. 2005, 3, 27.
32. Behrooz, F.; Ramli, A.R.; Samsudin, K. A survey on applying different control methods approach in building
automation systems to obtain more energy efficiency. Int. J. Phys. Sci. 2011, 6, 2308–2314.
33. Stylios, C.; Groumpos, P. Modeling Complex Systems Using Fuzzy Cognitive Maps. IEEE Trans. Syst. Man
Cybern. Part A Syst. Hum. 2004, 34, 155–162. [CrossRef]
34. Behrooz, F.; Mariun, N.B.; Marhaban, M.H.; Radzi, M.A.; Ramli, A.R. New Design Approach to MIMO
Nonlinear Controller for Direct Expansion Air Conditioning System in Building Automation System.
In Proceedings of the 2015 IEEE 15th International Conference on Environment and Electrical Engineering
(EEEIC), Rome, Italy, 10–13 June 2015; pp. 1706–1712.
35. Rossiter, A. Generalised Predictive Control, Videos on model predictive control. Creative Commons,
USA: Attribution 2.0 UK: England & Wales Licence, 2014; Vol. chapter 2. Available online: https:
//www.sheffield.ac.uk/acse/people/jar/mpcmaster (accessed on 12 July 2019).
36. Behrooz, F.; Ramli, A.R.; Samsudin, K.; Eliasi, H. Energy saving by applying the fuzzy cognitive map control
in controlling the temperature and humidity of room. Int. J. Phys. Sci. 2017, 12, 13–23.
37. Martchenko, A.S.; Ermolov, I.L.; Groumpos, P.P.; Poduraev, J.V.; Stylios, C.D. Investigating Stability Analysis
Issues for Fuzzy Cognitive Maps. In Proceedings of the 11th Mediterranean Conference on Control and
Automation MED’03, Rhodes, Greece, 18–20 June 2003.
38. Cheng, Q.; Fan, Z.T. The stability problem for fuzzy bidirectional associative memories. Fuzzy Sets. Syst.
2002, 132, 83–90. [CrossRef]
39. Ge, S. Lyapunov Design, in Knowledge Foundations; Eolss Publishers: Oxford, UK, 2004.
40. Cannon, M. Nonlinear systems. Lectures Hilary Term 2016. Available online: https://markcannon.github.io/
teaching/ (accessed on 12 July 2019).
41. Li, H.; Yang, B.; Xie, Y.; Qian, W.A. Multi-relationship Fuzzy Cognitive Map Stability Method. J. Inf. Comput.
Sci. 2013, 10, 5373–5379. [CrossRef]
© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).