IEEE TRANSACTIONS ON POWER APPARATUS AND SYSTEMS, VOL. PAS-89, NO.

3, MARCH 1970 345

State Estimation in Power Systems ?-

Part I: Theory and Feasibility
ROBERT E. LARSON, MEMBER, IEEE, WILLIAM F. TINNEY, SENIOR MEMBER, IEEE O
AND JOHN PESCHON, SENIOR MEMBER, IEEE (tS 0al

Abstract-State estimation is a digital processing scheme which 2) detailed specification and design of an i9orrnPQ P 1
provides a real-time data base for many of the central control and capable of supplying the required system data at the least cost in
dispatch functions in a power system. The estimator processes the sensing, telemetry, and processing hardware. MAR 17 j9
imperfect information available and produces the best possible
estimate of the true state of the system. The basic theory and
computational requirements of static state estimation are presented,
and their impact on the evolution of the data-acquisition, data-
CONCEPT OF STATE ESTIMATION IN
POTEO CAL LIE
processing, and control subsystems are discussed. The feasibility General Discussion of State Estimation LOUVIEI
of this technique is demonstrated on network examples. The state estimator is a data-processing scheme that computes
the state of a system from the following three sources of in-
INTRODUCTION formation:
THE purpose of this paper and a companion paper [1] is to 1) measurements of system variables
present a methodology for processing measured data from 2) the mathematical model of the system and its instrumenta-
electric power networks in order to obtain the best possible tion
estimates of the system variables. The theory is discussed in 3) prior knowledge of various system inputs and outputs,
this paper, and [1] summarizes extensive computational experi- referred to as pseudomeasurements.
ence gained with these procedures on a large, realistic network
and discusses potential applications. The output of the state estimator approaches the true state
The emphasis of the research was on developing efficient of the system. Discrepancies between the true and estimated
computational procedures capable of processing thousands of states are due to the following factors:
measurements per second on present-day computers. The
feasibility of the method is demonstrated on relatively small 1) noise in instruments and telemetry channels
network examples. 2) incomplete instrumentation, in the sense that many
Information-processing algorithms that are mathematically variables are not measured at all
optimum as well as meaningful in engineering terms were 3) delayed measurements that reflect a prior system state
developed mostly after 1960 [2], [3]. These algorithms have 4) erroneous pseudomeasurements
since been used with great success in the defense and aerospace 5) inaccurate network parameters.
industry. To our knowledge, these advanced information- There are numerous approaches to the design of a state
processing techniques have not yet been used in the power estimator. To compare the performance of various estimator
industries. Schweppe first suggested the development of power- designs, it is customary to define a scalar "cost" function that
system state estimators. He has developed static state-estimator increases with the estimation error. An estimator that minimizes
algorithms for relatively small networks [4], [5] and outlined the this cost function is optimal in the sense that it generates the
computational difficulties associated with their implementation best estimate of the state based on the measurements and
[6]. pseudomeasurements received [7]. The principal papers and
The results presented here and in [1] are based on a joint books on optimal estimation theory are listed by Lee [2].
research effort by Systems Control, Inc., and the Bonneville The optimal estimator not only generates a state estimate
Power Administration [7], [8]. The work was undertaken to but also the covariance or "error" associated with it. The covari-
evaluate the potential impact of state-estimation procedures on ance matrix predicts the magnitude of the estimation error and
the evolution of the data-acquisition, data-processing, and hence provides a measure of confidence for the estimated state.
control subsystems of large interconnected power systems. The predicted and the actual estimation errors should be of the
Interest was focused on the following specific areas: same order of magnitude. Significant differences indicate that a
1) new and powerful techniques for system control, monitoring, major change, such as a fault, has occurred in the system or in
display, and alarm that might be made possible by a real-time the instrumentation, and an alarm signal can be given. Further-
state estimator; more, the most likely of these major changes (or faults) can be
determined.
Paper 69 TP 638-PWR, recommended and approved by the Power
System Engineering Committee of the IEEE Power Group for Application to Power Systems
presentation at the IEEE Summer Power Meeting, Dallas, Tex.,
June 22-27, 1969. Manuscript submitted February 20, 1969; made Centralized automatic control of power system dispatch
available for printing April 14, 1969.
R. E. Larson and J. Peschon are with Systems Control, Inc., requires knowledge of the steady-state performance of the
Palo Alto, Calif. (formerly Wolf Management Services of Palo system. This performance must be inferred from system measure-
Alto, Calif.)
W. F. Tinney is with the Bonneville Power Administration, ments communicated to the control center. If enough measure-
Portland, Ore. ments could be obtained continuously, accurately, and reliably,

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY DELHI. Downloaded on November 01,2024 at 16:04:16 UTC from IEEE Xplore. Restrictions apply.
 346 IEEE TRANSACTIONS ON POWER APPARATUS AND SYSTEMS, MARCH 1970

NOISELESS MEASUREMENTS
AND
PSEUDOMEASUREMENTS STATE ESTIMATE x

NOISY STATE COVARIANCE PX

MEASUREMENTS AND MATHEMATICAL MODEL
ZE + UREMENTS
OF NETWORK AND MEASUREMENT
INSTRUMENTATION SYSTEM ESTIMATE z
COVARIANCE OF
MEASUREMENT ESTIMATE Pz
RESIDUAL z -Z
MEASUREMENT
AND TELEMETRY NOISE;
COVARIANCE R

Fig. 1. Inputs and outputs of state estimator.

this would provide all the information needed for control. How- z processed measurements and pseudomeasurements;
ever, a perfect data-acquisition system is often technically and if the redundancy provided by the power-flow model is
economically infeasible so that the control center must depend sufficient, a filtering effect occurs and the error of z
on measurements that are incomplete, delayed, inaccurate, and is less than that of z
unreliable. The function of the state estimator is to process z-z "residual" or difference between actual and esti-
whatever information is available and to produce the best pos- mated measurement; the residual is used to detect
sible estimate of the true state of the system. system anomalies
The operation of a power system depends upon many different R noise covariance matrix; the diagonal elements Rij
control systems (e.g., turbine governors, voltage regulators, equal the mean noise power (or rms error) of the
boiler controls, load-frequency control, and economic dispatch- component wi; the mean noise power may be defined
ing). State-estimation techniques involving the same funda- as the time average
mental concepts but with different equations can be applied to
estimate the variables and parameters of each of these control I bT
lim - wi2(t) dt
systems. In this study, state estimation is limited to the serving T -> o
00

of the control functions associated with the system operating Px state covariance matrix; the diagonal element (Px)ii
under steady-state or slowly varying conditions. The corres- predicts the mean noise power between the true state
ponding mathematical model is represented by the well-known component xi and the estimated state component ti
network equations; these are given in (15). The variables of Pz covariance matrix of the estimated measurement z;
concern are: the diagonal element (Pz)ii predicts the mean noise
1) the injections I and K of active and reactive power at the power between the noiseless measurement component
nodes' zi* and the estimated measurement component zi.
2) the voltage magnitudes and phase angles V and 0 at the The output of the state estimator constitutes a data base
nodes from which many of the various control, display, monitoring,
3) the flows T and U of active and reactive power in the alarm, logging, and interrogation functions contemplated for
branches. the advanced dispatching centers are fed. This data base is
The state variables are taken here to be V and 0. Once they are optimal in the sense that all available information is so processed
as to provide the most accurate estimate of the system state and
known, all the other electrical variables of the network can be
computed directly. The measurements and pseudomeasurements the system measurements.
could consist of any or all of the listed variables but, in practice,
are predominantly of the type I, K, and T with the possibility of Redundancy, Pseudomeasurements, and Neighboring Systems
some additional measurements of V, U, and perhaps 0. A system of N nodes has 2N 1 state variables appearing in
-

the 2N power flow equations,2 namely, N 1 angles 0 and N

-

The State Estimator as a Data Base for Control voltages V. These 2N 1 state variables can be computed
-

from the 2N -1 primary measured variables, namely, N 1 -

Fig. 1 illustrates the inputs and outputs of the state estimator, active injections I and N reactive injections K.
where In addition, there are numerous secondary measured variables,
z raw noisy measurements and pseudomeasurements such as active and reactive line power flows T and U. These
comprising real and reactive injections I and K, real secondary variables can all be obtained directly from the state
and reactive power flows T and U, voltage magnitudes variables.
and phase angles V and 0 When only 2N 1 variables are measured, the state estimator
-

z* noiseless measurements and pseudomeasurements as reduces to a real-time power-flow solution, except for the covari-
would be obtained from a perfect sensor ance computation. When more than 2N 1 measurements are
-

w additive sensor and telemetry noise made, the redundancy allows filtering (or smoothing) of measure-
x state estimate consisting of voltage magnitude and ment noise, that is, reduction in estimation error and anomaly
phase angle, V and 0, at each node detection. These additional measurements may comprise the

'The term "injection" refers to the difference between local 2 The voltage phase angle at the slack node is set to zero; it is
generation and local consumption. not a state variable.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY DELHI. Downloaded on November 01,2024 at 16:04:16 UTC from IEEE Xplore. Restrictions apply.
 LARSON et al.: STATE ESTIMATION IN POWER SYSTEMS, I 347

secondary variables T and U. If fewer than 2N - 1 measure- TABLE I

ments are made, a mathematical solution remains possible but LINEAR ESTIMATION EQUATIONS
can be shown to be meaningless from an engineering point of
view under most circumstances [12]. Number of
Measurements* Estimator Equations
Since the number of measurements actually telemetered to the
dispatching center will probably be less than 2N - 1 in large x = (A'R-A)-'A)'R-lz
power systems, redundancy is introduced in the form of pseudo-
m >n Px (A 'R-'A )-l
=
z = A(A'R-1A)-1A'R-lz = Ax
measurements, namely, P-= A(A'R-1A)-lA' = APxA'
1) known voltage and other regulator settings x = A-lz
m = n PX = (A'R-'A)-l
2) known generation and interchange schedules z =z
3) short-term load forecasts broken down by nodes.
The problem of information redundancy is further complicated x A'(AA')-lz
m<n Px - A'(AA')-1R(AA')-1A
by the fact that few power systems are operated in isolation. z= z
They are interconnected with neighboring systems that influence
them but for which measurements are frequently not available Px A E { (x - ) (x -
in real time. A method of taking account of the effects of neigh- any m Pz A E{(z - z)(z-z
R = E{ww')
boring systems has been devised [8]. Yet another complication
can be caused by the hierarchical structuring of the information * Symbols n and m denote the number of state variables and the
and control systems. A successful solution of this problem is number of measurements (including pseudomeasurements), respec-
possible by means of network reduction and the design of "local" tively.
estimators [8].
where Ui is the rms value (standard deviation) of noise on the
THEORY OF LEAST SQUARES STATE ESTIMATION ith measurement. Numerical values for these a can be calculated
This section summarizes least squares estimation theory as from knowledge of the accuracy of the instrumentation and
applied to a power network [7] under assumed static conditions. telemetry systems. These values characterize the accuracy of
Modifications for updating the estimate for a continuously the individual measurements and are fundamental in computing
changing system under normal operating conditions are covered an optimum estimate.
in [1]. The primary output of an optimum estimator is a value of the
state variables called the optimum state estimate and denoted as
Definition of Terms £. This estimate is optimum in the sense that it minimizes some
Control and estimation theory is heavily dependent on the criterion, based on knowledge of the equations relating the
concept of state variables. For any given system the state measured variables to the state variables, the statistics of the
variables are defined as a nonredundant set of variables that measurement noise, and a given set of measurements z. In least
completely characterize the operation of the system. In power squares estimation theory, this criterion is a quadratic function
systems the state variables are usually the voltage magnitudes of the residuals. These quantities are defined as the differences
and phase angles at all of the nodes. The set of state variables is between the measurements actually made and the measurements
denoted as x, the number of state variables as n. that would have been made if the estimate were exact and there
The quantities that can be measured are taken to be known were no noise. The quadratic criterion weights the residuals
functions of the state variables. These quantities, called the according to the accuracy of the corresponding measurements. A
measured variables, can be the active and reactive power flows lower penalty is assessed for a given residual in a poor measure-
at the nodes, active and reactive power flows in the branches, ment than is assessed in an accurate measurement.
and voltage magnitudes and phase angles at the nodes (the An important by-product of the optimum estimator is the
state variables themselves). The set of measured variables is estimate covariance matrix P,. This n X n matrix, defined as
denoted as z, the number of these variables as m.
Each measured variable is corrupted by noise. The set of E{(x - I)(x -x)'} = Px (4)
measurement noises is denoted as w. The statistics of the measure-
ment noise are assumed to be known. Specifically, each measure- is a measure of the accuracy of the estimate. This quantity is
ment noise is assumed to have zero mean. (If the noise contains important in a number of applications.
a bias term, this can be accounted for [7 ].) Mathematically, The optimum estimator can also be used to generate an im-
proved estimate of the measured variables themselves. This
E{w} = 0 (1) estimate is denoted as z. The covariance of this estimate, denoted
as PZ, is also obtained.
where the letter E denotes expectation. The spread of the
measurement noise about its mean is also assumed to be known.
This can be expressed as Linear Estimation Theory
E{ww'} = R (2) If the measured variables are linear functions of the state
variables, then the optimum estimation equations can be solved
where the prime denotes transposition, and R is an m X m exactly. For this case the measurement equations are written as
matrix called the noise covariance matrix. If the measurements
are all independent [7], then R is a diagonal matrix with diagonal z = Ax+ w (5)
terms
where A is the n X m measurement matrix, and the vectors z,
Rii = qi2 (3) x, and w are as previously defined. The criterion for optimum

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY DELHI. Downloaded on November 01,2024 at 16:04:16 UTC from IEEE Xplore. Restrictions apply.
 348 IEEE TRANSACTIONS ON POWER APPARATUS AND SYSTEMS, MARCH 1970

Az
COVARIANCE R

A = xo +Al
'z=zzxo+At
+AZ~

ESTIMATES AND COVARIANCES

A A
K,Z, Pxt Pz

Fig. 2. Flow chart for nonlinear estimator.

estimation, denoted as J, becomes where F is the n X n Jacobian matrix defined by

J = (z - Ax)'R-'(z - Ax). (6) Ofi i= 1,2<, (12)
.
Fij= oj, * n, j= 1,2,* ,n
The optimum estimate x is found by minimizing the scalar J. dxi
There are three separate cases to be considered, depending on
the relative values of m and n. The optimum estimation equations and where Az and Ax are small changes in z and x, respectively.
for the three cases (m > n, m = n, m < n) are summarized in Substitution of (9)-(11) into (8) yields
Table I. J = (Az - FAx)'R-(Az - FAx). (13)
Nonlinear Estimation Theory This is precisely the same quantity to be minimized as for the
If the measured variables are nonlinear functions of the problem defined by
state variables, a commonly used approach is to linearize the
measurement equations about a nominal value and then to Az = FAx + w. (14)
use the results of Table I. If there is an a priori estimate of Thus the linear estimation equations of Table I can be applied
the state variables that is sufficiently close to the actual values, directly to estimate the state deviations Ax based on the measure-
then this approach is quite satisfactory. If the a priori estimate is ment deviations Az = z - f(x0). The estimated deviations
not sufficiently close, then an iterative approach in which the Ax can be added back to x° to obtain a new state estimate
measurement equations are relinearized about the newly ob- xt. The flow of information in this procedure is shown in Fig. 2.
tained estimates can be applied. This approach can be justified The accuracy of these equations has been studied extensively
heuristically as follows. If the nonlinear estimate equations are from both a theoretical and an empirical viewpoint [7 ].
written For a power system, the basic ac network equations are
z-= f(x) + w (7) written in the customary admittance matrix form [7], [9]
where f is the nonlinear measurement function, then the least N
squares estimation criterion becomes Ii = EI VVjYij cos (oi -
j=1
j- ij)
J = [z -f(x)]'R-1[z -f(x)]. (8) N
Ki = E ViVjYij sin (oi - j- ij),
Now, if xO is a nominal value of the state variables and if j=1
defined as
zo = f(x0) (9) where
then expansion of both f(x) and z to first order yields Ii real power injection at node i
z = zO + Az (10) Ki reactive power injection at node i
Yij magnitude of admittance from node i to node j
f(x) = f(xO) + FAx (11) 1ij phase angle of admittance from node i to nodej.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY DELHI. Downloaded on November 01,2024 at 16:04:16 UTC from IEEE Xplore. Restrictions apply.
 LARSON et al.: STATE ESTIMATION IN POWER SYSTEMS, I 349

TABLE II
MATRIX F AND VECTORS AX AND AZ
FOR MAXIMUM NUMBER OF MEASUREMENTS IN (14)

Ail
N
AIIN
AK,
N
AKN
ATTj
:N -1 2M
F =
Az =
UA5
aN
2M
.IN
Lfnlj A02
N-1
AON
AV,
N
AVN

The corresponding real and reactive branch flow from node i imposed by these considerations on the size of problems that
to node j can be written as can be treated. The cases m > n, m = n, and m < n are treated
separately.
Tij = ViVjYij cos (0 - Oj - ij) -V2y COS'iJ cos (16) 1) Case where m > n: The basic estimation equations are
Uij = ViVjYij sin (0i - Oj - iij) + V,2Y,j sin Vij. (17) -= (A'R-1A) -1'A'R-'z (18)
The measurements can include N real injections, N reactive
injections, 2M real branch flows, 2M reactive branch flows, Px = (A'R-'A)-l. (19)
N - 1 voltage phase angles, and N voltage magnitudes. Because These computations require basic arithmetic operations on
the injections and branch flows are nonlinear functions of the the order of mnn. Using one of the faster present-day computers,
state variables, it is necessary to linearize the measurement the computations for n = 100, m > n, would require only a few
equations in order to apply the estimation equations. seconds; for n = 1000, m > n, the computations would require
If all of these measurements are made, the basic F matrix and several minutes, which is too long for most practical on-line
Ax and Az vectors of (14) are as in Table II. The Ax vector
applications. The high-speed storage requirements are on the
contains A0i and AVi, while Az contains Ali, AKi, ATij, AVij, order of mn locations [7]. If n = 100, m > n, the number of
AGi, and AVi. locations is on the order of 10', which is large but manageable.
The linearization is performed about a nominal set of state However, if n = 1000, the number is 106, which is clearly in-
variables fi', i = 2, *,N and Vi°, i = 1,2,* N. The feasible for any readily available system.
elements of the submatrices a, b, c, d, e, f, and g are the appropri- Thus for present computers the limit of computational feasi-
ate Jacobian terms of (15)-(17). bility by this approach is reached for a value of n somewhere
If fewer than (4111 + 4N - 1) measurements (the maximum) between 100 and 1000. Furthermore, unless precautions are
are made, then the matrix F and vector Az have the same struc- taken, severe numerical difficulties may be encountered at
ture, except that the unmeasured elements of the Az vector and the values of n well below 100, and straightforward application of
corresponding rows of A are deleted. The vector Ax is not (18) and (19) may lead to meaningless results. For example, a
changed. problem arose in connection with a nonlinear estimation problem
COMPUTATIONAL CONSIDERATIONS having 51 state variables [7]. The problem centered on the phe-
nomenon that the matrix (A'R-'A) became nearly singular so
Straightforward Implementation of Estimation Equation that inversion was, for all practical purposes, infeasible. This
The computational requirements of directly implementing the difficulty occurred because some of the measurements were
equations in Table I are easily derived and relatively well known. assumed to be orders of magnitude more accurate than others.
However, the numerical difficulties, particularly in the case Since most of the linearized measurement equations are propor-
where the measurement equations must be linearized, are not tional to differences between state variables, the situation arose
generally understood or appreciated, nor are the limitations where two state variables were not determined precisely, but

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY DELHI. Downloaded on November 01,2024 at 16:04:16 UTC from IEEE Xplore. Restrictions apply.
 350 IEEE TRANSACTIONS ON POWER APPARATUS AND SYSTEMS, MARCH 1970

their difference was known accurately. This problem was elim- groups, the n basic measurements used in the power-flow solu-
inated by placing a lower bound on the accuracy of all measure- tion and (m - n) redundant measurements. It can be shown
ments. that the optimum estimate is the sum of the optimum estimate
A number of other difficulties also arose in this problem. based on the n basic measurements and a correction term based
However, this typical example is enough to show that a straight- on the (m - n) redundant measurements. The optimum estimate
forward implementation of textbook equations is not sufficient based on the n basic measurements can be obtained using the
for large practical systems. power-flow program, exactly as before. However, in order to
2) Case where m = n: The equations for m = n are compute the correction term, it is necessary to process matrices
of order n X n and (m - n) X (m - n). If n and (m - n) are
x= A-lz (20) not too large, this approach is practical.
P., (A'R-1A)-1. (21) 3) Sequential Processing of the (m - n) Redundant Measure-
ments: This technique is the same as in 2), except that the
The computational difficulties here are much the same as in the (m - n) redundant ineasurements are processed one at a time.
case m > n. Computation time is proportional to n3, while As a result, the inversion of the (m - n) X (m - n) matrix is
storage requirements are on the order of n2. Thus the limit of replaced by (m - n) scalar divisions. If (m - n) is large, this
computational feasibility is again somewhere between n = 100 saving can substantially increase the desirability of the ap-
andn = 1000. proach.
The numerical difficulties are not as severe in this case. First, 4) Sequential Processing Using Only the Diagonal Elements of
since A is a square nonsingular n X n matrix, Px can be computed the State Covariance Matrix: The approach used in 3) still
as requires storage and manipulation of an n X n matrix. An
Px = A-'R(A')-1. (22) approximate approach that eliminates this difficulty is the use of
sequential processing,which carries only the diagonal elements of
Since (A'R-'A) does not have to be computed or inverted, many the covariance matrix in all calculations. As a result, the computa-
of the problems encountered for m > n are either lessened tion time grows as nt rather than n3, and the storage grows as n
significantly or eliminated completely. Second, the computation rather than n 2. It is then possible to handle values of n on the
of x requires only the inversion of A; again, this is considerably order of a few thousand even with a modest computer. This
easier than the corresponding calculation for m > n. approach, unlike the previous methods, does not produce an
3) Case where m < n: Although the optimum least-squares optimum estimate. However, it can be shown [7] that the
estimate for m < n exists in principle, there are fundamental difference is very slight. Furthermore, many of the numerical
objections to its validity in a practical system. There are many difficulties present when a full covariance matrix is used are
estimates x for which the criterion eliminated, and in practice this nonoptimal estimator may
J = (z - At)'R-l(z - At) (23) actually perform better than the optimal estimator [7 ].
This last computational procedure forms the basis for the on-
is exactly zero. The optimum estimate is arbitrarily chosen to line implementation of the efficient state-estimator algorithm,
be that estimate from this set that minimizes x' I. There is no which is discussed in [1].
physical significance to this choice in power systems, and the
expected result, that the estimate will be very poor, was con- COMPUTATIONAL RESULTS FOR A STATIC STATE ESTIMATOR
firmed in a number of experiments [7]. In order to demonstrate the performance of a power-system
state estimator, a computer simulation of this process was
Improved Computational Procedures carried out. The structure of this simulation is discussed, and
Only the practical cases, m > n and m = n, will be considered numerical results are given for an 8-node network.
in this summary of some of the techniques developed [7].
1) Use of the Power-Flow Solution for the Case m = n: When Structure of the Simulation
m = n and z consists of the n-nodal injection measurements, the The general structure of the simulation of the static state
problem is equivalent to the ac power-flow solution. If z is the estimator is shown in Fig. 3. The data input consists of the
vector of measured real and reactive injections, £ is the solution actual values of the system state variables, a selection of the
vector of complex node voltages; if z is the vector of real in- measurements to be made, and statistics of the associated
jections only, x is the solution vector of node voltage phase measurement noise. The actual state is specified directly, while
angles only. In linear estimation theory the optimum estimate x the measurements to be made are selected from a list of allowed
is given by (20). This is in agreement with the power-flow measurements. The standard deviation of the noise on each
solution where A-1 is the Jacobian matrix evaluated at the measurement is specified, and the measurement noise covariance
solution point (A is a nonsingular square matrix). matrix is constructed from this information.
If the power flow is solved by Newton's method [9] with The actual values of the state variables were used to generate
optimally ordered triangular factorization [10], A-I is available noiseless values of the selected measurements using the measure-
in triangular factored form. Systems with n = 100 require at ment equations (Fig. 3). Noise was added to each measurement
most a few seconds for solution, and the storage requirements using a random noise generator; the noise sample for each
are modest because the factored A-1 is sparse [10]. The solution measurement was scaled according to its standard deviation.
of (20) for n = 1000 takes less than a second on a fast computer The optimum estimator (linear or nonlinear) operated directly
when A-' is in factored form. on the noisy measurements z and the measurement noise covari-
2) Partitioning of m Measurements for m > n into n Basic ance matrix R. The output of the estimator was the optimum
Measurements and (m - n) Redundant Measurements: For the state estimate x, the associated estimate covariance matrix
case m > n the power-flow solution cannot be used directly. P., the estimate of the measured quantities z, and the associated
However, it is possible to split the m measurements into two measurement covariance matrix P,.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY DELHI. Downloaded on November 01,2024 at 16:04:16 UTC from IEEE Xplore. Restrictions apply.
 LARSON et al.: STATE ESTIMATION IN POWER SYSTEMS, I 351

Fig. 3. Structure of the simulation for optimum estimator. Fig. 4. 8-node example used in computational experiments.

All of these quantities were then fed into the estimator per- sets of measurements. These quantities are given as a percent of
formance evaluator along with the actual values of state variables the actual value of 06 which is 30.78 degrees. The first set of
and measurements. It was then possible to evaluate estimator measurements contains only the injections at all nodes except
performance by comparing the errors between estimated and the slack node; there are 7 measurements and 7 state variables.
actual values of state variables and measurements; for state The other sets of measurements contain 8, 20, and 21 measure-
variable xi this error is defined as ext, where ments, respectively. This table indicates the degree to which
redundant measurements improve the accuracy of the estimate.
Exi = xi - xi (24) Case 2) State Estimation with AC Power-Flow Model: The
while for the jth measured quantity fj(x) the error is defined as voltage-phase angles of Case 1) were used while the voltage
Ezj, where magnitudes were selected arbitrarily. The noise model had a
fixed standard deviation for each type of measurement (real
62} = zj- f/(x). (25) injection, reactive branch flow, etc.) [7 ]. Because certain measure-
It was also possible to compare these errors against their ments, particularly reactive injections and branch flows, varied
standard deviations; if Pxii denotes the ith diagonal element of over a wide range, selecting the standard deviation to be a
PX, then the standard deviation of the error in the estimate of percent of the actual value led to numerical difficulties. Choosing
xi is given by axi, where a fixed value for each type of measurement assumes that identical
meters are placed at each measuring station. It would have
/P'= (26) been just as easy to use a different standard deviation for each
while if P,jj denotes the jth diagonal term of Pz, then the stan- individual measurement.
dard deviation of the error in the estimate of f,(x) is given by ozj, The results in these experiments show good agreement be-
where tween the estimation errors and their standard deviations.
Since an ac model is used, both the voltage magnitudes and phase
OzJ = /PzjJ. (27) angles are estimated (except the phase angle at the slack node);
Clearly, if |Exil is considerably larger than axi, then the validity there are a total of 15 state variables. Results for 06 and V6 using
of the estimate is questionable; similarly for |Ezi and oz 15 and 75 measurements are shown in Table IV.
Case 3) Parameter Estimation with AC Model: It is straight-
forward, at least in principle, to modify the estimation equations
Experimental Results for an 8-Node Example to perform parameter identification as well as state estimation
The simulation was used to perform a number of computational [4], [7]. In one experiment the branch admittance Y36, as well
experiments and detailed analyses [7]. An 8-node network is as the 15 state variables, was estimated. Two cases were con-
shown in Fig. 4 (its parameters were taken from [11]). Voltage sidered. In the first case Y36was at its nominal value, while in the
phase angles were chosen to give the injections shown. second case Y36was set to zero, corresponding to an open circuit
Case 1) State Estimation with Linear Power-Flow Model: in the branch between nodes 3 and 6. In both instances the state
When a linear model [11] of the network was assumed, linear variables were estimated with the same accuracy as in the case
estimation theory could be applied directly. The measured where no parameter identification was attempted. Interestingly
quantities were selected from the 8 nodal injections and the 13 enough, the accuracy of the estimate of Y36 was approximately the
branch flows. The standard deviation of each measurement was same in both cases; these estimates are shown in Table V. Thus,
arbitrarily taken to be 2 percent of the actual value. although the parameter estimation accuracy is on the order of 10
Different runs were made using several different sets of mea- percent, there is no significant loss of accuracy when the param-
surements. In all cases the actual estimation errors agreed well eter varies by 100 percent. This type of performance is very
with the standard deviation of the estimates. Typical results desirable in fault location. The number of state variables is
are shown in Table III where the actual estimation error and n = 15, and the number of measurements is m = 75. Estimation
the standard deviation of the estimate 06 are shown for several error and standard deviation are given as actual values.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY DELHI. Downloaded on November 01,2024 at 16:04:16 UTC from IEEE Xplore. Restrictions apply.
 352 IEEE TRANSACTIONS ON POWER APPARATUS AND SYSTEMS, MARCH 1970

TABLE III The state-estimator concept is most useful when the number
ESTIMATION ERROR 6 AND STANDARD DEVIATION 0- FOR 06 of measurements exceeds the number of unknown state variables
IN 8-NODE EXAMPLE WITH LINEAR NETWORK MODEL
USING DIFFERENT NUMBERS OF MEASUREMENTS (case m > n). The resulting redundancy of information allows
filtering (reduction of estimation errors), anomaly detection, and
Estimationc-ft
Error Standardq06Deviation parameter identification. Since the number of centrally available
Number of measurements is usually less than the number of unknown state
Measuremenlts (degrees) (degrees)
7
variables, particularly when the state of neighboring systems is
1.844 1.037
8 0.634 0.757 included in the estimate, it is imperative to include as many
20 -0.068 0.209 pseudomeasurements as are reasonably predictable from load
21 0.006 0.050 forecasts, generation schedules, and regulator settings.
The actual value of 06 is 30.78 degrees. The number of state vari- If the number of measurements and pseudomeasurements
ables is n = 7. equals the number of unknown variables (case m = n), the
estimator reduces to a standard load-flow solution, to which,
however, a covariance computation should be added.
TABLE IV If the number of measurements and pseudomeasurements is
ESTIMATION ERROR e AND STANDARD DEVIATION e- FOR 06 AND V6 less than the number of unknowns (case m < n), a least squares
IN 8-NODE EXAMPLE WITH AC NETWORK MODEL
USING DIFFERENT NUMBERS OF MEASUREMENTS estimate can still be obtained but is of doubtful value.
The fundamental computational routine for the case m > n
Number is an efficient load-flow program which processes n of the m
of Estimation Error Standard Deviation measurements and produces a preliminary estimate. The addi-
Measure- E06 EV6 06 ,V6
ments (degrees) (kV) (degrees) (kV) tional m - n measurements can thereafter be processed in
15 -1.721 0.343 1.197 1.351 several ways; these additional measurements correct the pre-
75 -0.083 0.252 0.292 0.312 liminary estimate of the power-flow solution.
Although several state-estimation algorithms have been made
The actual value Of 06 is 30.78 degrees, and the actual value of to work reliably and efficiently on relatively small example
V6 is 100.8 kV. The number of state variables is n = 15. problems, the difficulties of extending these algorithms to large,
interconnected power systems and real-time operation should not
TABLE V be underrated. An efficient scheme for on-line state estimation
ACTUAL AND ESTIMATED VALUES OF Y36 IN 8-NODE NETWORK which has a potential for real-time operation in practical power
WITH AC MODEL systems (with several hundred to several thousand nodes) is
developed in [1].
Estimation Standard
Actual Estimated Error Deviation REFERENCES
Y36 Y36 EY36 OJY36
[1] R. E. Larson, W. F. Tinney, L. P. Hajdu, and D. S. Piercy,
5419.9 4981.7 -438.2 543.6 "State estimation in power systems, pt. II: implementation and
0 -495.5 -495.5 543.6 applications," this issue, pp. 353-363.
[2] R. C. K. Lee, Optimal Estimation, Identification, and Control.
Cambridge, Mass.: M.I.T. Press, 1964.
[3] R. E. Kalman and R. S. Bucy, "New results in linear filtering
Case 4) Outage Detection and Location with AC Model: In line and prediction theory," Trans. ASME, J. Basic Engrg.,
outage detection the basic approach used was that of computing [4] F. C.D.,Schweppe
ser. vol. 83, pp. 95-108, March 1961.
and J. Wildes, "Power system static-state
the difference between the estimates of the measured quantities estimation, pt. I: exact model," IEEE Trans. Power Apparatus
and the measurements themselves and then comparing the dif- and Systems, vol. PAS-89, pp. 120-125, January 1970.
[5] F. C. Schweppe and D. Rom, "Power system static-state estima-
ferences with the standard deviation of the measurement estimate tion, pt. II: approximate model," IEEE Trans. Power At
error. If the difference greatly exceeds the standard deviation, paratus and Systems, vol. PAS-89, pp. 125-130, January 1970.
then it is likely that an outage has occurred somewhere in the [6] F. C. Schweppe, "Power system static-state estimation, pt.
III: implementation," IEEE Trans. Power Apparatus and
network. In the numerical experiments based on the ac model of Systems, vol. PAS-89, pp. 130-135, January 1970.
the 8-node network, various types of outages were introduced. [7] R. E. Larson and J. Peschon, "Feasibility study of system
state estimation," Wolf Management Services, Palo Alto,
Generally, a line outage (i.e., an open circuit) was readily Calif., Contract 14-03-79378, Bonneville Power Administra-
diagnosed [7]. However, the detailed description of the method tion, Portland, Ore., Interim Rept., October 1968.
and the numerical results obtained are beyond the scope of this [8] R. E. Larson and L. P. Hajdu, "Potential applications and
on-line implementation of power system state estimation,"
paper. Wolf Management Services, Palo Alto, Calif., Contract 14-03-
79378, Bonneville Power Administration, Portland, Ore.,
CONCLUSIONS Final Rept., January 1969.
[9] W. F. Tinney and C. E. Hart, "Power flow solution by New-
The least squares estimator provides the most accurate ton's method," IEEE Trans. Power Apparatus and Systems,
estimate of the system state since it utilizes all of the information [10] vol. PAS-86, pp. 1449-1460, November 1967.
W. F. Tinney and J. W. Walker, "Direct solutions of sparse
available under the form of raw measurements, prior knowledge network equations by optimally ordered triangular factoriza-
of certain system variables (pseudomeasurements), and the net- tion," Proc. IEEE, vol. 55, pp. 1801-1809, November 1967.
work equations. It processes this information in a manner which [11] J. Peschon, J. C. Kaltenbach, P. Henault, M. W. Siddiqee,
and L. P. Hajdu, "Power system planning and reliability,"
is statistically optimum. Stanford Res. Inst., Menlo Park, Calif., Contract 14-03-72910,
Although the straightforward estimation algorithms dis- [12] Project 6619, Final Rept., June 1968.
R. E. Larson, "Limitations of the right pseudoinverse for state
cussed in the literature are generally not suitable for real-time estimation," Wolf Management Services, Palo Alto, Calif.,
power-system state estimation, several rigorous and approximate Contract 14-03-79378, Bonneville Power Administration,
computational approaches have been developed that can serve Portland, Ore., Project U-926, Tech. Memo 13, January 1969.
as a point of departure for a real-time computational procedure
for large networks. For Combined Discussion see pp. 360-363.
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY DELHI. Downloaded on November 01,2024 at 16:04:16 UTC from IEEE Xplore. Restrictions apply.