IEEE T r a n s a c t i o n s on Parer Apparatusand

Systems, v o l . PAS-94,

no. 2, MarchiApril1975

BAD DATA ANALYSIS FOR POWER SYSTEM STATE ESTIMATION E. Handschin* F. Schweppe** C.

J. Kohlas*

Fiechter* A

*Brown Boveri Research Center, Baden, Switzerland

**Electric Power System Engineering Dept., Laboratory, EE Mass. Institute of Technology, Cambridge, Mass.

SUMMARY The state estimation problem in electric power systems consists of four basic operations: hypothesize structure; estimate; detect; identify. This paper addresses the last two problems with respect to the bad data and structural error problem. The paper interrelates various detection and identification methods (sum of squared residuals, weighted and normalized residuals, nonquadratic criteria) and presents new results on bad data analysis (probability of detection, effect of bad data).The theoreticalresults are illustrated by means of a 25 bus network. 1. INTRODUCTION State estimation algorithms fit measurements made on the system to a mathematical model in order to provide a reliable data base for other monitoring, security assessment andcontrol functions. These algorithms should be able to handle four types of errors: 1. measurement (random metering and communication errors), 2. parameter (uncertainty in model parameters such as line admittance), 3. bad data (large unexpected meter and communication errors) and 4. structural (errors in the structural form of the model). This paper addresses solutions to the bad data and structure error problems based on residual analysis andnonquadratic estimation criteria. The tradeoffs between these various approaches are compared to helpa system designer decide which combination of techniques will best meet his particular needs. New bad data analysis techniques are presented which enable: 1. evaluation of the bad data spreading effect and hence lead to the concepts of interacting - noninteracting bad data and local redundancy, and 2. calculationof the probability of detecting bad data and of false alarm. 2. PROBLEM DEFINITION The purpose of a static state estimator is summarized in Fig. 1. To be more explicit defme: System state 5: complex bus voltages, n-dimensional. Measurements 3 : telemetered line flows, businjections,bus voltages or pseudomeasurements, mdimensional. Measurement error yz: zero mean random vector. Covariance matrix E: diagonal matrix whose i:th component ui2 gives the variance of the i:th measurement error E [ Y ~ y z T 1ui2 = ~may ; depend on the size of zi. Model parameter p: value of parameters used to defiie: 1) individual line, transformer, etc. models (usually impedances of n-equivalent network) and 2) covariance matrix E. Obtained from system design data. switch and circuit Model structure 2: telemetered or telephoned breakerpositionswhichdetermine network structure and meter location.
Paper T 74 309-1 recommended and approved by the IEEE Power System f the IEEE Power Engineering Society for presentation Engineering Committee o at the IEEE PES Summer Meeting & Energy Resources Conf., Anaheim, Cal., July 1419,1974. Manuscript submitted August 31, 1973; made available for printing April 3 , 1 9 7 4 .

measurements 2 (meter) structuralinfor mation g (breake position etc. )

Reliable estimate of: s t a t e (bus voltages) (Computer

2 ) model

E ( d e s i g nd a t a )
________)

(structure and pararneter values)

Fig. 1. Purpose of static state estimation Hypothesized model h (5): If the values of fectly, thenz, 5 and & are related by

E and 5 are known per(2.1)

2=b@)+2
where

(5) is specified by p and 5 so thatto be more precise

State estimate 2: estimate of the value of state x computed from z, p and 5. Parameter errors Q: zero mean random vector describing mall deviations of assumed values of parameter p from true model values &E=P+!p. Bad data b: unusually large measurement error added to 3 (caused by meter-communication system failures)

z = h(x,2,9+ h.

z =

~ ( x ) + ~ w i t h ~ = ~ + ~ .

If the i:th meter is bad then b = y a with giT = [O,O, . . . ,1,. . .O] and a = size of bad data. Structure error E: errors in system structure5. Detection:test to determinewhether bad dataor structuralerrors are present. Identification: logic to determine which measurement(s) is (are) bad or which part of the structure is wrong. Redundancy I): ratio of number of measurements to number of unknowns (states) = m/n. Local redundancy ek: redundancy for each bus counting only measurements and unknowns at bus k plus at all buses up to two switchyards away. Some of the above definitions are not necessarily standardterminology but itis essential to have a set of selfconsistent terms. Using the above definitions, an expanded description of the state estimator of Fig. 1 is shown in Fig. 2 in terms of four basic operations: 1) hypothesize structure, 2) estimate, 3 ) detect, and 4) identify. Methods to hypothesize structure i.e. to convert E and 1 .in to &(&) are not discussed in this paper. They are conceptually straight-forward network theory but in practice they are nontrivial. The weighted least squares (WLS) approach to estimate is summarized in Appendix A. This paper is primarily concerned with detect and identify. Thispaperconsiders the processing of onlythe observation z made at one time instant. In practice it is desirable to use time series curve fitting on the individual measurements and plausibility checks on the structural data2 and on the magnitude of the individual measure-

329

7
WSURMENTS

INPUT

Parameter

inforpation E 4 structure information a

I
m
I

HYPOTHESIS MODEL

Determine equations for model ICE) and covariance matrix 8.

I
next ment measure-

criterion miniual. becomes

Fig. 3. 25 bus network. T h e numbers of metered switchyards encircled. T h e measurements are given in kV, MW, MVar.

are

Yes

small-enough? .---------

- z-h(g)
I

If not, then there exist 1) bad data g and/or 2) structure errors 2

IDENTIFY iogic to find the location of bad data in 5 and/or structure errors in E.

l l possible measuremunication links are the most expensive items. A ments at a selected switchyard are taken with the constraint that no line is measured at both ends. All generator switchyards were selected plus the minimal number of other switchyardsrequired for a complete state estimation (the minimization was done by eyeball, not rigorously). T h e redundancy r) = 115/49 = 2.3. Bus 3 was chosen for most bad data studies. Two meter confirations will be discussed. Case 1 is the configuration of Fig. 3 and leading to a local redundancy 3 = 2. In case 2, the voltage and injection measurements at bus 3 are omitted, giving a local redundancy 3 = 1.7. The standard deviations ui of the error 4 are set equal to 0.02 p.u. (on a 100 MVA base) for real and reactive power measurements and 0.002 p.u. for voltage magnitude measurements. Many different types of numerical studies were done on this system. Because of space limitationsonlya few of theseresults are explicitly mentioned here. Ref. 22 contains more details.
4. BAD DATA ANALYSIS

modify inputs

Fig. 2. T h e four basic operations for state estimation in electric power systems ments. Such preprocessing is important but is not discussed in this paper to save space. Parameter errors vp are not discussed in this paper. If ~rp is small and modelled as a zero mean random vector, the parameter error problem can be viewed as being equivalent to a problem with-no parameter errors but with a measurement error with a covariance matrix larger than E. Of course it is difficult fo determine this new R; but fortunately WLS estimation has the property that the estimation,detectionandidentification logics tend to be insensitive to the value of E. Thus, provided the parametererrors are not too large, most of the approaches to be discussed still apply. However, the analysis of theperformance (see sect. 5) can be influenced heavily by the value of E. 3. EXAMPLE CASES STUDIED

T h e key to the analysis of bad data is the residual sensitivity matrix of (A9) which was obtained by linearization. With itand relatedequations such as(A6), (A81 and(A1 l), a wide variety of results can be obtained. Nonlinear simulation showed the validity of the linearization. For example, the component of & , corresponding to P3-13 is shown in Table I for different sizes a of bad data. Numerical values of the residual sensitivity matrix 3 show that theeffect of one bad data in is only noticeable in theneighbourhood of the bad data location (the size of theneighbourhood depends on the local redundancy). For example, for bad data in P3-13, wp3-13, p3-13 = 4.3 IO-1 whereas Wp3-13, p5 = 1.1 IO4; so bad data in P3-13 and P5 are noninteracting. The limited spreading of bad data allows the reduction of themultiplenoninteracting bad data problem to several one bad data problems (sec. 7). The multiple interacting bad data problem has to be solved separately (sec. 8).
5. DETECTION

T h e numerical results to follow are based on the 25 bus network shown in Fig. 3. T h e basic meter configuration indicated in Fig. 3 was chosen using a minimal cost criterion which assumes that the com-

Consider J(&)

be expected that bad data or structural

as defined by (Al) with z=Z. Intuitively, it is to errors should cause an unex-

Table I: Effect of bad data in P3-13 on p3-13,$eoretical: (A1 1) with v =%a (vz=O) simulated: (2.2) with -= ~ J + xz + I

Case size Bad data a in !4W

2 ICase theoreticallsimulated Itheoreticallsimulated 22.8 45.6 22.1 44.8

330

35.2 70.5

35.8 71.9

pectedly large value of J Q . This concept is formalized in (B1) in terms of the J(i>test. As indicated in Appendix B, it is possible to compute P,, the false alarm probability and Pd, the detection probability. Fig. 4illustratesthe type of results which can be obtained. For case 1 a single bad data can be detected with 90% confidence provided the bad data is at least 20 MW. If the bad data is not removed, the error in P3-13 is then - 1 1 MW. If this error is too big, the metering configurationhas to be improved or a larger false alarm probability used. T h e probability calculations underlying Fig. 4 are based on the distribution of J(@ in the presence of bad data, (A27). Monte Carlo simulations using the nonlinear simulation verified these formulae. T h e linearized analysis methods for one bad data point apply to the N multiple (N) bad data case by using (2.2) with b = C % a i . i= 1 Instead of using J(&) it is just as reasonable to test the magnitudes of the individual residuals. This leads to the rr-test of (B2)and the rr-test of (B4). Results such as Fig. 4 can be developed using the bounds of (B3). Suchstudies(andthenonlinearsimulations) have shown that for a single bad data, the gq-test is usually superior to the J(i>test for large networks.

6. DECOMPOSITION OF NETWORK

The limited spreading of bad data makes it possible to decompose the network into a contaminated part and a healthy r (sec. 5 ) . If this is done, onpart by inspection of their 3 and line state estimation can continue to update the system state in a tracking modefor thehealthy part even if no identification is accomplished.Identificationtechniques may be used just onthe contaminated part plus the boundary buses to the healthy part.
7. IDENTIFICATION
7.1. Single or Multiple Noninteracting Bad Data

For the ordered residual search approach the weighted (normalized) residuals r of (A13) are put intoa descending order of m a g nitudeandthe measurementzicorresponding to the largest residual Gax is removed first. T h e result of the subsequent WLS esti-

m)

(r)

l.O&

1-

MMW

mation with the reduced measurement vector Zr is then passed through the detection schemes of sec. 5. If bad data is still detected a further estimation is carried out with the second largest residual removed, etc. For one very large bad datapoint(A16)statesthatthe largest normalized residual rEax directly indicates the bad data location. In thegrouped residual search the p largest residuals in I-or 3 are removed simultaneously andthen put back one afterthe otheruntil bad data is detected. It is particularlysuited for identification of multiple noninteracting bad data. Grouped residual search may requirea new Hypothesizemodel (Fig. 2) when the removal of residuals leads to asituation where some of the states cannot be estimated.

10

20 a[MW]30-

Fig. 4. Performance of J(3) test and effect of undetected bad data of size a inP3-13 Pd: probability of detection Pe: false alarm probability Error: error in estimate of P3-13 when bad data is not removed (A1 1).

T h e following statements compare the three detection tests: 1) on-line implemeTtation is simple and fast. 2) m-test requires off-line calculation of Q in (AI 2) while J(i>test and-r-test do not. 3) %-test is more effective than-g-test in all cases. 4) =-test is more effective than J(&>test for a single bad data point (for large networks). 5) J(g)-test is sometimes better than gq-test and vice versa for multiple bad data (interacting or noninteracting) or structure errors.
These statements lead tothe recommendation to implement:either JGFtest and -r-test or J(g>testand 3-test. A detection is indicated if either test of a pair fails; i.e. if either J($)-test or m-test detects trouble. The J(f) - I-- pair is preferable if can be calculated. If sparcitytechniquesandcareful programming are used then as a very rough rule of thumb, time to compute _D is less than or equal to the time to do one WLS estimate. 331

10

12

14

dS/dr

gradient
4

Fig. 5. Various estimation criteria 1. Quadratic 2. Quadratic-straight 3. Multiple segment 4. Quadratic-Square root 5. Quadratic<onstant

An alternate to residual search is the use of nonquadratic error criteria (see Fig. 5 and Appendix C). The main numerical problem with thenonquadratic errorfunction is thepotential difficulty of many local minima. In order to get a good starting value the result of a WLS estimation is used. Table I1 shows the normalized residuals= using a quadraticsquare root error functionfor bad data in P3-13.

7.3. Structural errors

The most reliable approach to identification of structural errors is to use a search logic (based on the same philosophy as in residual search for bad data).This however can be difficult to automate and demands much computer time. Network decomposition is almost mandatory. errors p, and A distinction between bad data b, parameter structure errors 2 was made in Section 2 because they represent different physical phenomena. However, from a mathematical point of view, (and also) may be interpreted as bad data in E. For example: astructuralerror c can arise when a line is assumed to be in service although it is actually disconnected atboth ends.This can be viewed as bad data in the line admittance; i.e. in 2. This kind of generalization tends to obscure theimportant physical differences, is not entirelysatisfactoryfrom an intellectual point of view and has some practical shortcomings. However it provides a useful engineering approach to the structure error identification problem. To be more precise, consider the estimation of both 5 and p from (7.1)

three Table 11: The

largest normalized residuals P3-13 of case 2.

for bad data in

Measurement Q2 [War
44-21 War] 3-13

I[

a=10 MW 2.82 1.52 3.18

20 MW 32.95 3.20

40 MW 2.81
50.

2.89

3.18

In agreement with the results of Fig. 4 bad data of less than 20MW cannot be detected. Larger errors result in a clear identification of the bad data point. The other nonquadratic error functions of Fig. 5 have shown a similar behaviour. Residual search and nonquadratic criteria are actually closely related concepts. Consider the quadratic-constant criteria of Fig. 5 where the break point hl between the quadratic and constant part is lowered after each trial estimation untilthe bad data is identified. Thiscorresponds to ordered residual search. Inthe grouped residual search the break point X1 is increased after each trial estimation.

7.2. Multiple Interacting Bad Data

Multiple interacting bad data points may occur on occasions such as when all the measurements in one switchyard are transmitted over the same channel, when several power flow meters use a common potential transformer, etc. Furthermore, due tothe random nature of bad data they may occur in two adjacent buses. There are many reasonable ways to extend the residual search logics of sec. 7 for the multipleinteracting bad data case. Unfortunatelythebestdirection to follow is not obvious and it appears that there will always be cases where an extensive (almost exhaustive) search will be required. The difficulties of this search will be reduced if thenetwork is decomposed (sec. 6 ) into healthyandcontaminated areas. Nonquadraticerrorcriteria are effective when 1)the local redundancy Ek is large enough and 2) thenumber ofinteracting bad data is nottoo big. Someresults are summarized in Table 111 for case 1 with bad data in P3-13 and P3-14 corresponding to 20 MW each. For WLS the residuals are smeared while for all formsof thenonquadraticcriteria, the bad data location is obvious. In situations such as these thenonquadraticcriteria have adefinite advantage over residual search procedures since only one nonquadratic iterative sequence is requiredandsophisticated search logics do not have to be developed. It is not necessary to decomposethe network (although it canbe done if desired). Nonquadratic criteria are not effective when the local redundancy ek is too small or when the number of interacting bad data points is too large as e.g. when bad data on all line flows at bus 3 are generated by setting measured quantities equal to zero. It is not clear whether some sophisticated residual search would be better.

Bad data identification logics can now be applied directly, where bad data in the measurementq is of interest.

7.4. Discussions
The following discussions are based partly on solid theory, partly on the numerical studies (only a few of which were actually presented in this paper) and partly on personal judgement. They shouldbe viewed as guidelines, not as absolute truths. Some of the options in the choice of the identification logic are summarized in .Fig. 6 in terms of three levels; Level I: Attempt to identify bad data; Level 11: Decompose the network; Level 111: Use more sophisticatedtechniques to handle bad dataand/orstructural errors. The orders of Levels I and I1 can be exchanged. Each level has several options, including the option to do nothing. The minimum computer cost solution is to choose to do nothing at all levels so that theonlyactiontaken after a detection is to warn thesystem operator. A Level 111 capability requires the largest computer capacity (a Level I11 solution involving real-time display-input and an operator controlled search is a promising approach). Fig. 7 provides a comparison between the use of residual search and nonquadratic estimation. If a nonquadratic estimator approach is used the choice of criteria is not straightforward. In general terms, the quadratic-constant curve worksbest (when it works) but is least reliable (most sensitive to iterative solution difficulties) while the quadratic-straight line is the most reliable but least effective. As a compromise one of the middle curves (quadratic-square root or quadratic multiplesegment) is to be preferred. The use of anormalizednonquadratic error criteria (see Appendix C) provides better results but requires the computation of E.

Table 111: Welghted and normalized residuals for five different estimation criteria and interacting bad data (case 1)

Measurements

WLS

straight

rw
3-13
i3-14

rN

rN

21-16 3-1

3.00 4.50 6.48 7.34 3.04 3.51 2.85 5.20 2.49 2.85

5.43 8.16 7.96 9.03 1.55 1.47

2.83 1.68
332

QM u au dlrta t ilc ip esegment rW rN 10.05 15.09 9.57 14.28 9.88 11.17 10.10 11.42 1.83 2.11 2.57 2.97 1.75 2.02 0.53 0.29 0.15 0.27 0.10 0.11

Quadratics q u a r er o o t rW rN

Quadraticflat rw rN 9.75 14.68 9.72 10.98 1.71 2.03 0.25 0.46 0.21 0.40 0.18 0.20

e s t a t e es timation1

Remove b a d d a t a a n d / o r c o r r e c model s t r u c t u r e

Fig. 6. Options in design of identification procedure.

F
One b a d d a t a
N multiple nonin-

Number o f q u a d r a t i c e s t i m a t i o n s w h i c h 9 be needed using a search Weighted re- Normalized resid u a l s rN siduals % detection detection threshold y : large y : small several one several
several N several N

Performanceofone nonquadratic estimator

threshold E x c e l l e n ti fl o c a lr e d u n d a n c y i s highenough

eracting bad data

one t o N depend i n g on l o g i c Satisfactoryif 1) l o c a l redundancy i s high 2) number of b a d d a t a s m a l l

cting bad data

Very l a r g e . A l l possiblecombinations may have t o b e s t u d i e d

Notes: 1. As a very rough rule of thumb: time for one nonquadratic estimate = ( I to 2) time for one WLS estimate. 2. High detection threshold 7 implies only large bad data detected. Fig. 7. Comparison of residual search with nonquadratic estimation for bad data detection. 8. CONCLUSIONS The residual sensitivity matrix W or WN of (A9) or ( A 1 9 is a powerful toolfor analyzing bad datawithout resorting to multiple, nonlinearMonte Carlo simulations. Study ofthismatrixshowsthe limited spreading of bad data, the important distinction between interacting and noninteracting bad data, and that local redundancy ek is more important than redundancy r) in evaluating a meter configuration. Using the W or W N matrix, and an assumption of normal measurement errors, it is possible to compute the false alarm probability and probability of detection (or bounds there on). This provides a design tool for choosing a metering configuration. The choice of which detection tests to implement is relatively straightforward. T h e choice of identification logic involves a tradeoff between computer timeandperformance.Identification by residual search and by nonquadratic criteria are relatedconcepts.Both have advantages and disadvantages. Appendix A - W L S Estimation and Linearized Analysis. This appendix summarizes the mathematical background of WLS estimation and its error analysis as used for state estimation in electric power systems. The WLS estimate &is based on the criterion The rest of the equations of Appendix A and B are based on linearization obtained by assuming & is small and setting &(x)=h(s)+Bx. From (A4) ri = zi hi(5). &satisfies the optimality condition

where H = d&/d&denotes the Jacobian matrix. T h e nonlinear equation (A2) is solved iteratively [Refs. 1-31 using an algorithm such as

where & denotesthe WLS estimate of the k:th iteration and the choice of affects the convergence properties (B=HT&-18 is often used).Complete discussion on thechoice of E, use of Sparcity Programming etc. is beyond the scope of this paper. With (2.2) it follows from (A21 that

HT

E-[h(5)+~-5(i)] 2.

(A4)

Define the state estimation error as

JP[z-l(x)ITR-[z-h(x)]= iil(q) =i 2

(AI)
333

and for v=v, the covariance matrix of the estimation error dx is

E [ &

hT] =

I & .

(A71

Similarly, for the residual vector!

; -

4 2 - i = h_@)

+v-h_@ - w,
1

where (if 3 is normal)thefirst term is x2- distributed, the second term is normal distributed,and thethirdterm is a constant. As K becomes large I(&)in (A21) approachesanormaldistributionwith mean I.(I and variance 0I 2 .

(A81 where Wi is the (i,i):th element of the residual sensitivity matrix W. Thus, the standardized variables 51 and 32 of (A18) become

with the residual sensitivity matrix W given by

4

1-E &E TE-

(A9

and for x=% the residual covariance matrix

E[-

iT] = &=s-_HT = 5 .
=

(A101 Appendix B - Detection Theory

The detection of bad dataorstructural errors can be viewed as an hypothesis testing problemwith two hypothesis Ho and H1 where Homo bad dataorstructural errors present H1 :Ho is nottrue. Let Pe denote the false alarm probability, i s . Pe is the probabilityof rejecting H,, when Ho is actuallytrue. Let Pd denotethe probability of detection; i.e. Pd is theprobability of accepting H1 when H 1 is true.Threemethods of testing these hypothesis will be summarized. 1. The I(&)-test is Accept H , i f 5 1 < y ( o r 52 < y ) (B1) Reject H , (accept H1) I otherwise

Finally, the difference between the true and estimatedmeasurement is givenby

&

h _ c g -h_&

cw,-gv.

(A1 1)

Let the diagonal matrix _D denote

D = diag
Defme the weighted residual and the normalized residual Then the inequality holds k2 -%. For v=yz the covariances are
- T Elrwgw

(A12)

& .as

a V, =

-1: JR-I and J E L

(A141

where the diagonal elements of UN are all equal to one. Defrne the normalized residual sensitivity matrix V J N by

s=flw

(AIS)

Let WN,jk denote the WN element of the j:th column and k:th row. Then from (A10) and (A15)

the

where c1 and 52 are given by (A18) and 7 is the detection threshold level. The choice of 7 determines P,. For example, when yz is normal, K is large, and Ho is true, then 51 (52) is N(0,l) so ~ 1 . 6 corresponds 5 to Pe = 0.05. In a similar fashion it is possible to compute Pd for a given 7 and a using (A23). 2. The 1_m-testis Accept Ho i f I&,kl < y I k = l . . . m (B2)

so that from the positive semidefinite nature of Cr

Reject H , (accept H1) otherwise. When l , is normaland Ho is true, thenfrom (A14), %,k is N(0,1) which provides a basis for choosing thethreshold 7 . For a given 7 (and a),it is awkward to compute exact values ofP,, (and Pd). Howis ever if Pe,k is probability l k , k l > 7 when Ho is trueandPd,k probability l k , k l > 7 when H 1 is true then pe 5 p e l k ; pd 2 Pd,k 3. The rw-test is Accept HO i f I;W,k] < y I k = l . .
(B3)

(B4) Reject H , (accept HI) otherwise If the same value of 7 used in (B2) is used in (B4), the w-test is more sensitive, since due to (A13): k 219. Appendix C - Nonquadratic Criteria For a nonquadratic cost function (Al) is replaced by both become zero mean, unit variance Gaussian; i s . N(0,l). For one bad data in the i:th measurement of z the error 1 of (2.2) is given by

. I l l

v = 3 + :?a. -

(A191

Substituting (A1 9) in (A8) the (normalized) residuals and r~ take the form

where the vector p(r) is a function of the residual r such that the condierror criteria shown in Fig. 5 are obtained. The optimality tion is obtained via chain rule

:=Ez+ weia and

&=WJVZ

%%a.

(A20) where

dJ-dJ -- de df: = dx de dy dx

2( & p )

g g=o

(C2)

J(3)in (A1 7) becomes

334

Rewriting the optimality condition (C2) as

HT E-1 S -

Q = ~

(C4)

Table CI Nonquadratic estimation functions and gradients for i = 1, 2, m

...,

shows the similarity with (A2). The nonlinear equation (C4) is solved iteratively using the modified algorithm (A3)

E E 51

T T - 1

- G ,[,+,-,]=E

QQuadratic

pi = ri 91

= 1

T -1

E Se

(C5)

@Quadratic-straight lri/uilsA:pi = ri gi = 1

Table CI contains thefunctionspi of e and the diagonal elements gi of the diagonal matrix G for the criteria shown in Fig. 5. Values of the breakpoints are chosen by engineering judgement, normally X=2.5. . . .5 for @ , @ and @ . In @ usually X1=2.5. . . 3.5, h2=4.5. . .6.5 and X3=9.5. . .12. An important variation of (Cl) is the normalized nonquadratic cost function which is the same as (Cl), (C2) except that the breakpoints X i are different for each residual or determined by @ of (A12 . Thus, for example, thecondition lri/uil < h is replaced by Iri/ D;~ I< X .

@Multiple segment lri/uilLXl:

p i = r i gi
= 1

JL

Appendix D

- References
-1/2 gl=[ 2 I k

References and some alternate approaches are now discussed. These discussions are of necessity very cryptic but hopefully serve as a guide to the literature. Apologies are made to authors whose work was accidently overlooked or misinterpreted. The use of J(f), -J+J and l u for detection and identification was first discussed in Ref. 1-3. The X2-nature of J(&) is discussed in Ref. 4. Ref. 5 discusses the use of J(i(). The use of nonquadratic criteria for power system bad data detection and identification is introducedin Ref. 6 (which used the quadratic-square root criteria). Ref. 7 considers the use of quadraticconstant criteria with a varying breakpoint which decreases exT h e technique used in ponentially with the number of iterations. corresponds to the quadratic-constant criteria. Ref. 8 apparently Independent of these attempts t o , handle bad data, the theory of 9) has been used in mathematicalstatistics robust estimation(Ref. for parameter location estimation. The quadratic-straight and quadratic multiple segment criteriacomefrom Ref.9. This work is presently extended towards robust regression theory (Ref. lo), using nonquadratic error criteria. A stimulating discussion with Huber and Hampel was helpful in clarifying some of the issues discussed in this paper. A bridge between robust estimation and the power systems bad data problem has been formulated in Refs. 1 1 and 12. An excellent study on the effect of parametererrors p on is found in Ref. 13. The structural error identification approach is taken from Ref. 14. References15and16 discuss bad datadetection and identification based on a deterministic load flow, using the redundant information to identify bad data. One difficulty is the choice of an accurate nonredundant set of input data to the load flow. The combined spatial and time correlation approach of Refs. 17, 18 assumes that a sensor which has failed at time k+l will show a markedly different reading from the previous estimated measurement at time k. For any large difference, an automatic search logic is used to check other differences which are spatially correlated with the measurement in question (spatially correlated is probably equivalent to interacting) to see if the sudden change is caused by a change in the operating condition or bad data. This method appears to have difficulties in detecting multiple interacting bad data, structural errors,anda singlebad data point caused by a meter which drifts slowly over several hours or days. Thehypothesis testing approach of Ref. 19 is based on the assumption that for bad data the measurement error covariance matrix E is multipliedwitha scalar u2 i.e. all variances are multiplied by the same factor 0 2 . This seems to be a mathematical artifice which does not correspond to the physical problem. In practice, however, the
335

I-11

gi = 0

@ Quadratic-square
/r./uilLA:pi = ri
1

root gi = 1

@Quadratic constant gi = 1

logics of Ref. 19 are closely related to the and g q tests, the main difference being in the extra mathematical complexities resulting from the introduction of 02 (see also Ref. 20). The network of Fig. 3 is taken from Ref. 2 1. The present paper is a condensed version of Ref. 22 which cont a i n s many more details. The linearized analysis of Appendix A is straightforward and can be found in many places. Ref. 23 contains discussions on the general approach and also on the hypothesis testing concepts of Appendix B. The X2-nature (and limit distributions) of J(g) (when a = 0) for linear be found in many standard statistical texts (see also models can Ref. 23). Methods to solve (A2) for i (and other approaches) are discussed in many of the above references and elsewhere. Ref. 24 can be viewed as a companion paper to the present paper which discussed the overall static state estimation problem in more detail. Ref. 24 uses the terminology of this paper.

Je)

(22) E. Handschin, F. C. Schweppe, J. Kohlas and A. Fiechter, Bad data analysis in electric power systems, Brown Boveri Research Report 1973. (23)F. C. Schweppe, Uncertain dynamicsystems, Prentice Hall, Englewood Cliffs, N. Y. 1973. (24) F. C. Schweppeand E. Handschin,Static state estimation in electric power systems, to be published in Special issue of IEEE Proceeding devoted to Power Systems and Computers.

Discussion
A. C. Sullivan, G. H. Couch, and J. A. Dembecki (Electricity Commission of New South Wales, Australia): The authors are to be congratulated for their description of some alternative methods for detecting and identifying bad data in power system estimation. W e in the Electricity Commission of New South Wales have been actively assessing the feasibility of applying state estimation techniques in our system and feel qualified to comment onthispaper. Our observations are, in large, of a general nature. We have experienced similar effects using nonquadratic criteria as the authors[ 11. In particular we have found it desirable to perform one or two iterations with weighted least squares (WLS) before introducing, say, the quadratic-square root criterion. This hasthe effect of helping to speed convergence, avoiding spurious minima and allowing more flexible selection of penalty function breakpoints. The authors discussion of network decomposition, and their emphasis of the requirement for local redundancy are most interesting. In fact, the relatively small effect of one measurement error on the estimate of states electrically remote from that error would indicate that state estimation is a task which is ideal for application of decomposition techniques. This is, however, clearly conditional on the redundancy conditions in each locality. In applying state estimation to the N.S.W. system[ 11 it has been found convenient to inhibit the nonquadratic penalty function with respect to selected measurements. For example, if there is only one voltage measurement or where busbar real and reactive power injections are known to be zero, such measurements should not be subject to this form of bad data suppression. In general, care should be exercised when there are interacting measurements having widely different standard errors. The residual search procedures described in the paper, particularly the rn-test, are a result of the fact that the largest residuals from a W L S estimate do not necessarily correspondwith the bad data. The nonquadratic criteria attempt to avoid the necessity of making several estimates to eliminate bad data. It would seem that, provided available computingtime is sufficient, the matrix G , used to implement the modifiedpenalty function, should not simply be a function of the residuals r and measurement errors R, but also of the residual sensitivity matrix W. Both the sensitivity of the rn-test and the performance of thenonquadratic criterion may be then used to greater advantage. Depending on the algorithm chosen for estimation, thestate covariance matrix may be available in triangular factorised form from which the residual covariance may be computed with little additional effort. In practice it would be well to remember, that in the final outcome, it is the system operator who, once alerted, is perhaps the most competent identifier of bad data. With this qualification, the detection phase must be regarded as having prime importance, with precise identification being subordinate. However, any information which can be given to the operator concerning the probable locality of the bad data would be of the greatest assistance.
REFERENCE
[1 ]

REFERENCES

(1) F. C. Schweppe and J. Wildes, Power system static-state estimation - Part I: Exact model, IEEE Trans., PAS-89, pp. 120-125, 1970. (2) F. C. Schweppe h d D. B. Rom, Power system static-state estimation - Part 11: Approximatemodel, IEEE Trans., PAS-89, pp. 125-130, 1970. (3) F. C. Schweppe, Power system static-state estimation - Part 111: Implementation, IEEE Trans., PAS-89, pp. 130-135, 1970. (4) R. D. Masiello andF. C. SChWeDDe. Power svstem tracking state estimation. PSEG Report EE-Dept. M.I.T.-1970, see also IEEE Trans., PAS-90, pp. 1025-1033, 1971. (5) J. F. DODZO. 0. A. Klitin and L. S. VanSlyck. State Calcula. _ tion of Power Systems from Line Flow Measurements, Part II, IEEE Trans., PAS-91, pp. 145-151, 1972. (6) H. M. Merrill and F. C. Schweppe, Bad data suppression in power system static-state estimation, IEEE Trans., PAS-90, pp. 2718-2725, 1971. (7) Handschin. H. Glavitsch and J. Kohlas. Estimation und . , E. Entdeckung Lhlechter Messdaten, EDV in der Elektrizitatswirtschaft, Seminar des Oesterreichischen ProduktiviMtszentrm, Wien, 11. und 12. April, 1973. (8) B. Poretta and R. S. Dhillon, Performance evaluation of state estimation from line flow measurements on Ontario Hydro power system, IEEE Winter Power Meeting, T 73 086-6, 1973. (9) D. F. Andrews et al., Robust estimates of location , Princeton University Press, 1972. (10) P. J. Huber and F. Hampel, Private communication, Swiss Federal Inst. of Technology, Zurich; 1973. (1 1) J. Kohlas, On bad data suppression in estimation, IEEE Trans., AC-17, pp. 827-828, 1972. (12) E. Handschin and J. Kohlas, Statistical data processing for power system operation, Proc. 4th PSCC, Grenoble 1972. (13) T. A. Stuart and C. J. Herget, A sensitivity analysis of weighted least-squares state estimation for power systems. IEEE Winter Power Meeting 1973, T 73 085-8. (14) H. M. Merrill and F. C. Schweppe, On-line system model error correction, IEEE Winter Power Meeting 1973, C 73 106-2. (15) G. W. Stagg, J . F. Dopazo, 0. A. Klitin and L. S. VanSlyck, Techniques for real-time monitoring of power system operation, IEEE Trans., PAS-89, pp. 545-555, 1970. (16) F. Ariatti et al., Methods for electric power system state estimation, Proc. 4th PSCC, Grenoble 1972. (17) A. S. Debs, R. E. Larson and L. S. Hajdu, On-line sequential state estimation, Proc. 4th PSCC, Grenoble 1972. (18) E. Handschin (editor), Real-time control of elecmc power systems, Chap. 2, Elesevier, Amsterdam, 1972. (19) J. F. Dopazo, 0. A. Klitinand A. M. Sasson, State estimationfor power systems: detection and identification of gross measurement errors, Paper X-2, PICA Conference Minneapolis, 1913. (20) F. C. Schweppeand E. Handschin, Discussion of the paper by J. F. Dopazo et al.: State estimation for powersystems: detection and identification .of gross measurement errors, presented at PICA conference 1973, to be published in IEEE Trans., PAS. (21) R. G . Gungor, N. F. Tsang and B. Webb, A technique for optimizing real and reactive power schedules, IEEE Trans., PAS-90,pp. 1781-1790, 1971.

.,

G. H. Couch, A. C. Sullivan, and J. A. Dembecki, A State Estimator Oriented to a Semi-Isolated 5 GW Power System, IEEE 1974 Summer Power Meeting, paper number C 74 346-3.
Manuscript received July 25, 1974.

Hyde M. Merrill (American Electric Power Service Corp., New York, N. Y.):This very informative paper consists of a survey and important extensions of previous work in addition to work reported here for the first time. It is required reading for anyone interested in power system
Manuscript received August 5, 1974.

336

state estimation because the bad data/modelerrorproblems are so critical in this application. One point that I wish the authors had stressed more than they did is their new and remarkable result that for h = Q the largest normalized residual flags the bad data point. This lends additional confidence to residual search methods. In equation (CS), which is part of the descriptionof the nonquadratic criteria, the messy term HTGTE-IGWJ can be replaced with the simpler @ E l l , or with any other gain matrix suitable for WLS estimation. Furthermore, since the right hand side of (C5) involves G and p only as a product 2 = Q, most of the complicatedterms in Table CI vanish by cancelation. For instance, for curve @ , si = ri or si = s i g n ( r i ) / m . similar cancelations occurfortheother nonquadratic criteria. This means thatthe practicaldifferencebetween programming and solving estimators based on WLS and nonquadratic criteria is very minor: WLS: solve B(kk+l-&k) = HTB-i nonquadratic: solve B(gk+l-jk) = _HTB-Is, where 5 is a simple function ofz. When nonquadratic criteria were fvst considered, it was with the hope that an estimator could be designed that would operate directly on raw data to give good answers. WLS estimators must converge to a wrong answer, detectand identify bad data, throw itout, and reestimate. It was hoped that the nonquadratic estimator would weight out bad data in the process of converging and thus not require that it be thrownoutafter a detection and identification process. The example given in Table I11 seems to indicate that this hope may be justified. This example involved two 100 bad data points. The residuals for nonquadratic criteria @ , @ , and @ were all close to 100. In effect, this means that the nonquadratic estimators subtracted 100 (the gross measurement error) from each of the bad measurements in the process of fitting an estimate 2 to the measurements p . Removing the bad measurements altogether would not have given a better &. Is this typical of the results the authors obtained, and if so, do they really feel that the nonquadratic estimators should be used merely to identify bad data to be removed, as in Fig. 6? Or do they feel that with nonquadratic estimators the detection-identification-removal steps may no longer be necessary? The application of nonquadratic criteria given involves replacing the residual I with 5. This is fine for the algorithm of refs. 1 4 , which operates directly on I, but this quantity does not appear explicitly in the AEP line-only algorithm of ref. 5. Dr. R. D. Masiello suggested in aconversationrecently thatthenonquadraticconcept could be applied directly to the AEP algorithm. One possible way of doing this follows (using notation of ref. 5 ) : the WLS AEP algorithm computes asequence go, El, . . . that converges to anestimate of g, the bus voltages. &+I is found by solving (BTDB)Ek+l = BTD(!mk-A&g), (a)

and Computing &+I

(BTDB)Ek = BTD(lrck-!!@g). directly by subtracting (b) from (a) gives (BT!?B@k+l

= BT!?(Xmk-Xck)

(b)

(BTPB)&k+l = gT!?Gkl&-S:k). (C) The equivalent AEP algorithm is: compute &+I from (c). If it is = 0, convergence has occurred. In any event, &+I = Ek + &+I. Finally, - k k = 1,the residue, since
= BTpB21*. (dl (BTpB)AEk+ 1 Applying the nonquadratic concept is now simple: replace I* in (d) withz*, where 5 is calculated as in the present paper. Do the authors feel this exercise is legitimate? Or do they have a better way of applying the nonquadratic criteria to the AEP estimator? Or do they feel it cannot be done?

E. Handschin, F. C. Schweppe, J. Kohlas, and A. Fiechter: We appreciate very much the efforts of the discussers in commenting and expanding on a few points of our paper.This contributions are a welcome addition. The important result mentioned by Dr. M e d l , thatfor =Q the largest normalized residual coincides with the bad data location is based on the condition that only one bad data point is present. For identificationsufficient computer capacity is required for calculating the normalized residuals. T h i s must be emphasized because the computation of the normalized residuals qq involves the calculation of the covariance matrix & of the residuals. As shown in Table CI,thecalculation of the product G p leads to thecancelation of manyterms.Thus, there is no significant difference between programming a WLS and aNonquadraticestimator. Furthermore,the use of nonquadratic estimators yield results close to thoseobtained by a WLS estimator with bad data removed before processing. Indeed, the results in Table I11 indicate the equivalence between a WLS estimator with bad data removed and a nonquadratic estimator. However, there are twopointsto be stressed. First, nonquadratic criteria give good results provided the normalized cost functions are used; i.e., the break points Xi in Fig. 5 are calculated as mentioned at the end of Appendix C. From a computational point of view this approach requires a considerable amount of extra calculation. The results in Table 111 are obtained with normalized nonquadratic error criteria. Second, from a convergence point of view it is advantageous to start with WLS estimation. The nonquadraticcriteriaproposed are nonconvex. Thus,their performance is excellent if local redundancy is high and the number of bad data is mall. Otherwise the nonquadratic approach may lead to unstablesolutions. It has been found that the four stepapproach of Fig. 2 is safer; thus special detection and identification steps are necessary in order to be applicable in a wide variety of cases. The use of a nonquadratic criterion with the line-only method seems to be a useful approach. It has not yet been numerically tested.

eax

where y m k is a function of the actual system measurements and Ek. The test for convergence is applied by computing PEk+ 1 = Ek+ 1 - Ek. If &+I C = Q, convergence has occurred. A theoretically and computationally equivalent restatement of the algorithm will now be given. From equation (1 1) of ref. 5 , BEk = Xck -

Manuscript received November 14, 1974.

337