IEEE Transactions on Power Systems, Vol. 14, No.

4, November 1999 1219

DYNAMIC STATE ESTIMATOR USING ANN

BASED BUS LOAD PREDICTION
A. K. Sinha J. K. Mondal
Department of Electrical Engineering
Indian Institute of Technology, Kharagpur, INDIA

A b s t r a c t : This paper presents an algorithm for dynamic The EKF theory is developed for linear systems [5]. This
state estimation of power systems. The method uses ANN requires that the nonlinear power system measurement equa-
based bus load prediction for the prediction step in the tions he modelled as linear perturbations (nonlinearity in
DSE. The proposed DSE uses rectangular coordinate for- measurement function is neglected). Under certain circum-
mulation for measurement equations. A second order dy- stances, such as sudden changes in load and/or genera-
namicstate estimator which incorporates the full non-linea- tion, the nonlinearities of the measurement function be-
rities of the measurement function is used for the filter- come quite significant and can no longer be neglected. Un-
ing step. The inclusion of non-linearities makes the pro- der such situation the performance of EKF is considerably
posed state estimator perform better in case of sudden large degraded [6,7].
changes in load/generations. In recent years enormous interest has been generated in
Key words: State estimation, dynamic state estimation, the area of Artificial Neural Network (ANN). One of the
ANN, load prediction, power system state estimation. application areas wkich has generated very wide interest is
INTRODUCTION Short-Term Load Forecasting (STLF) using ANN [8,9]. As
The initial atten;pt to model dynamics of power sys- discussed earlier. a realistic dynamic model of power sys-
tem into the state estimation process was by Debs et al[l]. tems for DSE should be driven by bus loads, this requiies
They proposed a simple state transition model in which short term bus load prediction.
the change in state is represented by the Gaussian noise. In this paper we present an algorithm for dynamic state
Nishya et al. [2] added B trend factor to take into account estimation which incorporates ANN based short term lrad
the snstained componi;w of load changes. A major inade- forecasting to obtain a realistic dynamic model for thn pre-
quacy of this model is h t it considers the trend factor as a diction step. For the filtering step we propose a rectangu-
deterministic input whiie in reality it is random in nature. lar coordinate formulation for the measurement equationc
To overcome the problems of the previously proposed dy- and a second order DSE scheme to include complete non-
namic models Leite De Silva et al. [3] proposed a dynamic linearity of the measurement function. The characi,r:ristic
model based on on-line estimation of system state transi- and performance of ?he proposed DSE algorithm is ascer.
tion. This model has a faster dynamic response and better tained by testing on a number of test systems. T 4 resuhs
state prediction capabi1it.y. However, in order to reduce the showing the salient features st the proposed DSF.! algorithm
computational burden both the state transition matrix and are presented.
covsriance matrix are assumed diagonal. This means that
the state variables are assumed to he independent. Rut PROBLEM FORMULATION
in reality bus voltage variations depend 'on load changes The power system operation is considered quasi-static and
and each load change influences many state variables (bus the state estimation is carried out at time intervals of a few
voltages). minutes (time samples). At each time sample (k) the state
A more realistic approach in dynamic model for DSE was X ( k ) is estimated from observations Z(k) (which is geuer-
proposed by Mallien et al. [4].They proposed a dynamic ally corrupted by noise) using the power system dynamic
model based on bus load prediction. In this method the model:
prediction step is performed on the basis of bus power in- The state transition equation of the model is
,jections instead of the conventional state transition model.
The approach is based on the fact that: i) Loads actually X ( k ) = f ( X ( k - 1),G(k- 1 ) ) + w ( k ) (1)
drive the system dynamics, ii) Loads are more or less inde-
pendent of each others and iii) Dynamics of bus loads can
and the measurement equation is
be identified through simple and accurate load forecasting
models. Z(k) :r +
h ( X ( k ) ) u(k) (2)
PE-063-PWRS-0-10-1998 A paper recommended and approved by
the IEEE Power System Analysis, Computing and Economics where, w and U are the white noise vector with zero mean
Committee of the IEEE Power EngineeringSociety for publication in the and are uncorelated in time as well as, between each other.
IEEE Transactions on Power Systems. Manuscript submitted
I.e.,
February6, 1998; made available for printing November I O , 1998.
-
w N(O:Q); U N ( 0 , R )- (3)
0885-8950/99/$10.00 0 1998 IEEE

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY MADRAS. Downloaded on August 29,2024 at 10:03:38 UTC from IEEE Xplore. Restrictions apply.
 12m

with, variance estimate of the state vector X. Since the state

trajectory is discretised in time steps of a few minutes, a
linear dynamic model for state transition, in general, is
suffcient [4].DSE using EKF consists of two main steps:
i) The prediction step :
Predicted state for the kth time instant is given by,
- +
.%(k) = F ( k ) X ( k 1) G(k - 1) (11)
Among the possible choices for state vector the most conve- where, F(X) is the state transition matrix and G represents
nient and universally used in power system is the complex the vector for any known control action. The prediction
voltages at all the busbars. The complex busbar voltage error covariance matrix is given by,
can be represented in both polar (VLB) and rectangular
+
( e jf) coordinates. In this paper the rectangular coor-
- +
Po(k)= F(k)P+(k l ) F T ( k ) Q ( k )
where, Pt is the estimation error covariance matrix.
(12)
dinate formulation for the state vector is used. Therefore, ii) The filtering step :
the state vector for the power system is given by Filtered state vector at time instant K is given by,
x = [ e l , f i , e z , f z , . . .,~ N I ~ . (5) + -
X ( k ) = X ( k ) K ( k ) [ Z ( k ) h(.%(k))] (13)
where, N = Number of busbars; and without loss where,
of generality the Nth busbar is chosen as reference K ( k ) = Po(k)HT(k)R-'(k) (14)
busbar with f~ = 0. W) = [H(k)PO(k)H*(k) R(k)l + (15)
The measurements used t o estimate the state vector are: R(k) = W-'(k) (16)
i) active power and reactive power bus injections. ii) ac- ~ + ( k ) = [I - ~ ( k ) ~ ( k ) ] ~ ~-( ~k () k[ )r ~ ( k ) p
tive power and reactive power line flows. iii) bus voltage
magnitudes. +K(k)R(k)IP(k) (17)
These measurements (nonlinear function of the state vari- SECOND ORDER DYNAMIC STATE ESTIMATOR
ables) are expressed in the rectangular coordinates as fol-
The fact i s t h a t linearised EKF performance gets degraded
lows:
when prediction error is of the same order as measuremeri,
Real (P) and reactive (Q) power injections at busbar p :
noise has been discussed in reference [7]. In case of sudderi
N
load variations, prediction is no longer accurate and the
PP = C{el,("GPn - f,BP,) +fP(f,GP, -t e P p l l ) I (6) prediction error is increased and hence effect. of higher order
,=1
terms can not be neglected.
N
The relation between the measurements for the power
Qp = CIfp(eqGpq -f q B p 0 - ep(fqGpp + eqBm)l (7) system DSE and the state (e, f) is given by equations (6)
,=1
- (10). It is interesting to note that all these equations
Real (P) and reactive (Q) power flows in line m-n : are quadratic. The Taylor series expansion of these equa-
tions contain terms upto second, order only. Therefore, the
Pmn = em(emGmn- enGmn - fm&n + fnBmn)
inclusion of second order terms in t h e EKF results in a
+fm(emBmn - -t f m c m n - f n G m n ) DSE scheme in which full nonlinearity of the measurement
+Gk,,(ek + f,?J ( 8 ) function is retained.
Qmn = fm(emGmn -enGmn - fmBmn + fnBmn) At time (k), the predicted state %(k) and bhe set of mea-
surement Z ( k ) are available. Considering the least square
-em(emBmn - edmn + fmGmn - f n G m n ) approach for filtering process we minimize the objective
-EL,(& + f): (9) function
Voltage magnitude at pth busbar :,
-
J(X) = [Z h(X)ITR-'[Z - h(X)] t [X - RITP;I(X - (18)

%z = e: + ji (10)
The first term in R.H.S. of (18). accounts for the measure-
ments Z ( k ) weighted by inverse of their error covariance,
where, Ypq= G,, +jB,, = (p,q)th element of nodal ad- whereas, the second term accounts for our faith in the pre-
mittance matrix. dicted state which is weighted by the inverse of the predic-
+
Yd, z G& jB;, = ;(line charging admittance of tion error covariance matrix. The minimum of the objec-
line m-n). tive J ( X ) is obtained when
It is seen from equations (6) - (10) that power system
measurements are related to state (e,f) through itonlinear
(quadratic) equations in rectangular coordinate formula-
tion. In polar coordinate formulatioQ these equations in- OT, -
H T ( X ) R - ' [ Z h ( X ) ] - q ' ( X - 6 ) = 0 (20)
volve trigonometric functions.
Extended Kalman Filter (EKF)
Dvnamic state estimators based on an EK:F algorithm
-
where, H = ah(X) ; t h e Jacobian matrix
ax
+
Y

using the models (1) and (2) provides the linear minimum ~ i i dX = ? , A X ; t h e estimated d a t e vector.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY MADRAS. Downloaded on August 29,2024 at 10:03:38 UTC from IEEE Xplore. Restrictions apply.
 1221

Using the Taylor series we expand h(X) about to obtain Thus, the complete prediction vector S is defined as

h(X) = h(?)+H(?)AX+S (21)

S [Pz, Qz
I
T
P G ,VG ]
T T
(27)

a2h(X) S consists of 2N-1 components, where N is the number of

where, S = (Ax)2
~

ax= (22) +
busbars and N = NL N G . In order to make regular
exchange of variables S ej X possible, the voltage angle
Substituting from (21) in (22), we get of slack busbar is considered as reference angle and is set
to zero. This makes the number of,variables in both S and
HT(%)R-'[Z - h(%) - H(%)AX - S ] - P,'[% + AX 21 = 0 - X equal to 2N-1,
(23) Active and reactive loads Pr,;,Q&i;for each busbar in-
From which, we obtain cluding the load at generator busbars (if any), is predicted
AX = [HT(R)R-'H(X) + P;'l-'tHT(R)R-'[z - h(%) -SI) (24) for next time instant f k + l \ usine the ANN based bus load
\ . I Y

prediction 191.
This equation is derived using full Taylor series. As this Active bus injections (generation - load) for generating
equation contains the term S which is a function of A X 2 . busbars can not be directly predicted for the simple reason
Therefore, solution of AX requires iterations. In an itera-
that generation has t o adapt itself to load variation. A
tive scheme, the incremental state vector AX for (i+l)th simple method for adapting this variation in load to the
iteration can be written as generation is to use generation participant factors. That is,
AXi+, = [HT(%)R-'H(%) + Pr']-'{HT(X)R-'[Z - h(%) -Si]} for the ith busbar, the real power injection a t time instant
(25)
The iterations are continued till 1 AX! - A X , - , I< E , and k+l is given by
the filtered (estimated) state a t time index (k+l) is ob- +
P a ( k 1) = P~i(k) ai + C
A f ~ j ( k 1) (28) +
?ainL-d as
+
X ( k ) = ?(k) A X ) (26)
jeloads

It can be seen from equation (25) that the Jacobian matrix where
H and calculated measurement h(?) are not functions of +
P~i(k 1) = Predicted active power bus injection ai.
A X , so they are required to be calculated once only us- busbar i at time instant k f l .
ing the predicted state vector and during iterations they p~i(k)= Estimated active power bus injection a t busbar I
remain constant. Only the second order terms Si are re- a t time instant k.
quirr:d to be calculated in each iteration. This considerably ai = generation participation factor (0 5 ai 5 1; oi :
redures the computational burden. 1).
and
DYNAMIC LOAD PREDICTION APr,j(k+ 1) = F ' ~ j ( k 1) - &j(k) + (29) ,

The DLP method is based on more realistic and physically The generation participation factors are calculated using
meaningful foundation in comparison to the conventional the economic load dispatch and are kept constant between
stare prediction model. The main arguments which form two successive economic dispatch period.
the basis of DLP method are: The DLP method uses bus power injections (P,Q) as
i) It is loads and generation which actually drive the sys- prediction variables, while for filtering step complex bus
tem dynamics which is of concern in DSE of power voltage (e,f) are more appropriate. The prediction vector
system. S and filtering step state vector X are related through 2 N - i
ii) Bus loads are more or less independent of each other. nonlineared equations
iii) Bus loads follow a reasonably regular pattern and so s = g(X) (30)
are easier to predict with reasonable accuracy.
The change of variables from S ---* X at the end of the
For the purpose of DLP two types of busbars are defined: prediction step is performed using load flow solution [ll].
a) Load busbars -where only loads Pr, and QL are con- Change of X + S, which is required after the filtering step,
nected. Let, NL be the number of these busbars, and is trivial as it requires simply computing equations (6 - 10).
b) Generator busbars -where generators are connected. The covariance matrix for injection prediction error is
Since generator busbars may also have loads in addi- given by
tion to generation so the net power injection is con- P:(k + 1) +
= Cov [S(k 1) - s ( k + I)] [31)
sidered a t these busbars. Also, a t generating busbars
the voltage is controlled, therefore real power injec- = JoPo(k 1)JT + (32)
tion PG and voltage magnitude VG are used as the
prediction state variables a.6 these busbars. Let NG where .Jo = X = X.
(E+iI , and
be the number of these busbars. In a d e r to satisfy Po(k+ 1 ) = (J;')P:(k+ I)(JgT)-l (33)
the active power balance in the system one generating
bnshars is taken as slack busbar, its active power PG ANALYSIS OF ANOMALO-US DATA
can not be specified and therefore, is not. included in (,ne of the objectives of state estimation is to detect. d e n -
the prediction variables. tify and remove or correct the anomalous data froni the

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY MADRAS. Downloaded on August 29,2024 at 10:03:38 UTC from IEEE Xplore. Restrictions apply.
 1222

incoming information, so as, t o maintain the integrity of Te1-tered

data
the data base. One major advantage of DSE is the avail-
ability of the predicted state 2 which can be very useful
in anomaly detection. Pre-filtering scheme for anomaly de-
tection consists of computing the innovation vector
Historiaal
O(k) = Z ( k ) - h ( Z ( k ) ) . (34)
In case, any anomalous data is present, the hypothesis
I X,(k) I< A,, Vrn (where m is the measurement in-
dex, X,(k) = O , ( k ) / p , ( k ) is the normalised innovation
Id
P

Pigure 1: RWI basad DEE schema

for measurement rn and p m ( k ) is the rnth diagonal ele-
Z(k+l) obtained after prediction step provides the com-
ment Of R(k) matrix not The choice Of thresh- plete predictive data base, whereas, X ( k ) obtained after
Old X V m C is based On ’ Under normal Operating
filtering step provides the complete data base for the
condition all the normalised innovation X,(k) will be small
rent time, The proposed DLP based DSE model faithfully
and within threshold value. follows the recursive scheme of EKF, except that instead
Discrimination between Anomalies and Treatment of of the state transition based dynamic model it uses a com-
Anomalous Data bination of ANN based bus load forecasting and load flow
One important property of the innovation vector is that no +
solution to obtain the predicted state x ( k 1). It also in-
“smearing” takes place in innovation process, that is, at corporates non-linearities of the measurement function in
time sample k, X(k) will have abnormal values for only the filtering step.
those measurements which correspond to the erroneous Computational aspect: To reduce the computational
(bad) measurements(see Appendix A for details). burden and speed up the computational process .>fthe pro-
Although, the occurrence of bad data can be detected from posed DSE scheme, following simplification in the algo-
the abnormal values of X .~ . . of mediction.
; ( k ) . the aualitv
I . . ef- rithm is introduced:
fect the value of A, and a poor dynamic model for prediction For computing predicted states from predicted bus
may lead to X value comparable to the magnitude of bad loads fast second order load flow [lo] is used. This load
data. In such cases, medium size bad data is likely to go flow requires computation and factorization of $he Ja-
undetected. A more reliable method for detection of bad cobian matrix only once in the heginning of the first
data is based on ‘skewness’ of the innovation vector. Un- iteration. For time step (k-tl), the estimated state
der normal operating condition, the distribution of d ( k ) is I(k)is used as the starting value for the load flow
symmetrical. In case of sudden load change, though the computation.
normalised innovation will have large values corresponding (ii) Since Jacobian matrix is already available from the
to measurements in the vicinity of the busbar where sud- load flow at the prediction step, so no Lomputation
den load/generation change has occurred, the innovation for Jacobian elements is required for computation of
vector O(k) will maintadn its symmetrical property. But
in presence of bad data, the distribution of O(k) becomes
Po(k I). +
(iii) Computationof predicted state error covariance Po(k+
asymmetrical and its ‘ & w a e t r y inder’ (skewness mea- l), equation(33) , introduces the heaviest compnta-
sure) y(k) = M3(k)/u3(k)will be large. where, M3 is the tional burden in prediction step. Since tho Jacobian
third moment and U is the standard deviation of the nor- matrix is very sparse, sparse matrix techniques are
malised innovation vector at time sample k(see Appendix
B for details). The presence of bad data is determined
+
used to obtain Po(k 1) column by column.
from the value of asymmetry indez. When bad data is TEST SYSTEM SIMULATION
present I y(k) I> Y ~ where ~ ymar~ ,i s a threshold deter- Test Systems : Simulation tests were performed on a num-
mined through simulation. Once presence of bad data is ber of systems, however due to limitation of space here re-
detected, measurements for which X i ( k ) values are greater sults for only IEEE 118 bus test system is presented. The
than the threshold, are considered bad and are eliminated simulations were carried out for 30 time samples.
from the measurement set. Generation of Bus loads: In the absence of real system
In case of sudden load/generation change, a number of data the generation of bus load data was carried out as
measurements surrounding the changed injections will ex- given below:
hibit large values for X i ( k ) . This is mainly because of the
error in prediction of X(k) which has not takeu the sud- (i) 17 sets of hourly load data from the IEEE-24 bus ( 17
den injection change iioto accuunt. Such a situation can be load bus) reliability test systcm data were used.
taken care of by the proposed filtering scheme as it kicor- (ii) Hourly loads for each set were nivmalised with respect
porates the non-l,oearities of the measurement function. to the peak load of that data set
(iii) These 17 load patterns were randomly assigned to var-
PROPOSED DSE SCHEME ious load busbars ( one load pattern fin each busbar)
The overall procedure for t,he proposed DSE scheme is of the test systems.
schematically shown in Figure 1. (iv) Actual hourly load for each busbar of the test systeni

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY MADRAS. Downloaded on August 29,2024 at 10:03:38 UTC from IEEE Xplore. Restrictions apply.
 1223

was obtained by multiplying the normalised hourly the ANN model is more accurate, the anomaly detection
load of the assigned load pattern with the peak load and discrimination will also be better with this dynamic
of the busbar. model. Plots of normalised innovation under normal op-
(v) In the absence of any hourly load data for reactive erating condition for the 118 bus test system is shown in
power, hourly reactive power load for each busbar was Figure 3 which shows that the performance of ANN model
obtained by keeping the power factor a t each busbar is better as compared to the linearized prediction model.
constant a t the nominal value for that busbar as given Anomaly Detection: Prefiltering for anomaly detection is
in the test system data. carried out in DSE using the normalised innovation vec-
Measurement Generation : The selection of measurements tor Am(k). The components of the normalised innovation
used, was based on random choice and redundancy (m/n) vector A, ( k ) for various operating conditions were com-
N 2 was maintained (460 measurements were used). The
puted. For normal operation with no anomaly present, the
time evaluation for the system state was simulated by a value of A , ( k ) is quite small. The values of A , ( k ) for vari-
load curve for each busbar. The power factor for the loads ous operating conditions is presented in Table 1. Based on
were kept constant. Changes in generation to take care of experience with number of simulation tests the threshold
the load variations were obtained using generation partic- values for was chosen equal to 1.5
ipation factors.
For each time sample, a load flow was carried out using Table 1: Anomaly detection test result for different
the load and generation values obtained through load curve operating conditions.
and generation participation factor, to obtain the true state IOperatinn I sample I measure I I I
(Xt) and true value of measurements ( Z t ) . The actual
measurements (2) were obtained by adding a normally dis-
tributed noise (error) with standard d&al,ion ( U ) of 2% for Load Change 0.793
power injections and flow measurements and U = 1%for Bad Data 3.847
voltage measurements, to Z + .
Initialisation : To start the DSE process the initial values
I Bad Data I 21 I 55,116 [ 6,5 1 3.932 [
for the various parameters were specified as follows: In order to be sure about the type of error for the sus-
Initial state f(0) was obtained using static state estimation pected measurements indicated by large value of &, asym-
from the measurements at k = 0 and its error covariance metry index was calculated for the measurement vector.
was specified as & ( O ) =diag(10-6). Table 1 show the results for asymmetry index for some
Bad Data : Bad data condilions were simulated by intro- typical bad data tests conducted. Once presence of bad
ducing error in the measurement set a t 5th time sample in data is confirmed (lr(k)l> ymaz, the suspected measure-
real power at bus 80, a t 11th time sample in PSOP s ~& Q I 1 6 ments are removed from the measurement set and the fil-
a t 21st time sample in P s ~ . tering process proceeds. On the basis of simulation tests
Sudden Load Changes : Sudden load change condition was the threshold value ymae was chosen 1.1.
simulated by changing the load a t bus number 54 a t 11th In case the innovation components are lager but asym-
and 16th time samples. metry index is below threshold ( i.e., bad data hypothesis is
Performanu ivaluation : The performance of the algo- rejected), t,he situation is identified as sudden load change.
rithms were assessed using the following indices. Filtering Step
Prediction step : The index used for assessing the per- The performance of the proposed DSE scheme using
formance or the prediction model is ANN based DLP was evaluated using performance index
J m and plots for normal operating condition as well as
sudden load change condition are shown in Figures 4(a)
& 4(b) and also compared with the DSE scheme with lin-
Filtering step : For assessina the effectiveness of filter, ear prediction model. The comparative plots for J m show
that the performance the ANN based DLP model is better
under normal operating condition because of more accn-
rate prediction of states by this model. It is also seen that
ANN based DSE does not require any time for settling and
RESULTS AND DISCUSSIONS it gives good performance from k = 1 itself. For sudden
Prediction Step load change condition, the performance of the two models
The oerformance of the ANN b a e d dvnamic model is same as both prediction models are unable to predict
was evaluated using the performance index ( ( k ) and was this situation x ith any accuracy and so, the performa,nce
compared with the linear dynamic model (;.l). The plots depends mainly on the measurements. The profile showing
of ( ( k ) lor 24 (hourly) time samples for the 118 bus test the time evolution of the estimated states (a, f ) for busbar
system are piesented in Figure 2. From the figure it can be 54 (where sudden load change is simulated) in each system
clearly seen that the ANN based dynamic model performs is shown in Figure 5(a) & 5(b). It can be seen that the
much better than the linear state transition ;node1 for real- ANN based DLP model tracks the time evolution of t,he
istic load variations. Since, the predicted state obtained by states better than the other mode!.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY MADRAS. Downloaded on August 29,2024 at 10:03:38 UTC from IEEE Xplore. Restrictions apply.
 1224

CO"S
Simulation tests reveal the following characteristics:

(i) ANN based DLP model provides more accurate pre-

diction than the conventional linear state transition
model.
(ii) The proposed ANN based DSE provides better state
estimates than the DSE with conventional prediction
model.
. .. (iii) Computation time required for the proposed ANN based
FIB 2 Performance of prediction model
DSE scheme is almost same as that for the conven-
5.01 tional DSE scheme ( 10.7 sec and 11.2 sec respec-
tively).

REFERENCES
1. Debs, A S and Larson, R.E., "A dynamic estimator
for tracking the state of the power system", IEEE Trans
on Power Apparatus and Systems, vol. PAS-89, pp, 1670-
1678, Sep/Oct 1970.
2. Nishiya, K., Hasegawa, J . and Koike, T . , "Dynamic
state estimation including anomaly detection and identi-
fication for power systems", Proceedings IEE, Pt-C , Vol.
129, No. 5, pp. 192-198, September 1982.
3. Leita da Silva. A.M., Do Coutto Filho, M.B. (:'antera,
J.M.C., "An efficient dynamic state estimation algorithm
including bad data processing", IEEE Trans on Power Sys-
t e m , Vol. PWRS-2, No. 4, pp. 1050-1058, November
1987.
4. Mallieu, D., Rousseaux, P., Van Cntsem, Th. and
Ribbens-Pavella, M . , "A new hierarchical approach for dy-
namic state prediction and filtering in electric power sys-
tems", Proc. IFAC Conf. on Power Systems and Piiwcr
Plant Control, Beijing, China, pp. 372-377, August 1986
5. Anderson, B.D.O. and Moore, J.B., "Optimal Filber-
ing", Prentice Hall Inc., 1979.
6 . Bretas, N.G., "An iterative dynamic state estirnatloii
and bad data processing", International Journal of Elec.,
trical P o u w & Energy Systems Vol. 11, No. 1, pp. 70--74,
January 1989.
7. Mandal, J.K., Sinha, A.K. and Roy, L., "Incorporat-
ing Nonlinearities of Measurement Function in Power Sys-
tem Dynamic State Estimation", IEE Proceedznga - Gmer.
fiunsm. Distrib, Vol. 142, No. 3 , May 1995, pp. 289 -
296.
8. Park, D.C., El-Sharkawi, M A , Marks 11, R.L., Atlas,
L.E. and Damborg, M.J., "Electric load forecasting using
an art.ificial neural network", IEEE Trans on Power Sys-
tems, Vol. 6 , No. 2, pp, 442 .- 448, May 1991.
9. Mandal, J . K. and Sinha, A. K. " Application of Arti-
ficial Neural Network for Short Term Load Forecasting in
Electric Power Systems", IE(I) JoUTnal EL Vol. 77, May
~

1996, pp. 40-43.

10. Rao, M., Roy, L. and Rao, N.D.: "A fast exact secord
order load flow method for large power systems," Pior.
L6:rq.e Energy systems, pp. 35--41, June 1982.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY MADRAS. Downloaded on August 29,2024 at 10:03:38 UTC from IEEE Xplore. Restrictions apply.
 1225

Appendix A:
The innovation vector (19) is the difference between the
incoming measurements and their predicted values. It is
defined in equation(34) as

B(k) = Z(k)- h(2(k)) (37)

Linearizing equation(2) about 2 ( k ) , we get
Z ( k ) = h(*(k)) + H ( k ) [ X ( k )- X ( k ) ]+ ~ ( k ) (38)
Substituting for z(k)from equation(38) into equation(37)
gives
+
d ( k ) = v(k) H ( k ) [ X ( k )- 2 ( k ) l (39)
In the presence of normal noise in measurements, this vec-
tor is a zero Gaussian white noise with covariance

% ( k ) = H ( k ) P o ( k ) H T ( k )f R(k) (40)
From equation(40) it can be seen that there is no corre-
lation between innovation vector ff ( k ) and measurement
noise v(k). Also from cquation(37) and (39) it can be seen
that B ( k ) is a direc,t estimate of incoming measurement
error relying only on predicted state 2 ( k ) and not on es-
timated state X ( k ) as in post filtering scheme. This is
the main reason why only those components of d ( k ) which
correspond to erroneous measurements exhibit large val-
ues. That is, no smearing effect takes place.

Appendix
_ __ B: _
Under normaloperating conditions the distribution ofin-
novation B(k) is symmetrical, even in case of sudden load
change condition, though the normalised innovation will
have large values corresponding to measurements in the
vicinity of husbar where load/generation change h a s oc-
curred, the distribution of B ( k ) will still be symmetrical[2].
However, in the presence of bad data the distribution of
8 i k ) no longer remains symmetrical. Therefore, the skew-
I . ~ ofS distribution of 19(k) can be used to detect thr pres-
ence of had data.

The skewness of a distribution can be measured in terms

of its third moment about mean. If the distribution is sym-
metric, its third moment Ma = E [ ( X - P ) ~ will
] obviously
be zero. Incase, the distribution is asymmetric its third
moment M3 = E [ ( X - P ) ~ will
] not be zero. Therefore.
symmetry, or lack of it, of a distribution can be measured
using the 'coefficient of skewness' ('asymmetry index')

where, U is the standard deviation of the distribution, ai-

viding M3 by u3 makes y dimensionless.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY MADRAS. Downloaded on August 29,2024 at 10:03:38 UTC from IEEE Xplore. Restrictions apply.