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

1, JANUARY 1970

Power System Static-State Estimation,

Part I: Exact Model
FRED C. SCHWEPPE, MEMBER, IEEE, AND J. WILDES, MEMBER, IEEE

Abstract-The static state of an electric power system is defined The static-state estimator is based on classical mathematical
as the vector of the voltage magnitudes and angles at all network techniques such as estimation, detection, probability, statistics,
buses. The statik-state estimator is a data processing algorithm far and filtering. However, an attempt is made to present the ideas
converting redundant meter readings and other available informa- without drawing on an extensive background in these theories.
tion into an estimate of the static-state vector. Discussions center onA more theoretical presentation would have had some advan-
the general nature of the problem, mathematical modeling, an
interative technique for calculating the state estimate, and concepts tages, but the chosen approach hopefully makes the concepts
underlying the detection and identification of modeling errors. more widely accessible. Most of the specialized mathematical
Problems of interconnected systems are considered. Results of some jargon is relegated to background discussions which can be
initial computer simulation tests are discussed. skipped.
This paper is the first of a three-part series on the static-state
estimator. In this paper, the general problem, model, and theory
INTRODUCTION of solution are developed. In [2] approximate models and solutions
A REAL-TIME central control system can be used to improve are developed. In [3] consideration is given to implementation
the security (reliability) of electrical power system bulk gen- problems related to computational speed, dimensionality of the
eration and an EHV transmission system. It is possible to con- state vector, and the fact that a power system is never truly in
sider the operation of such a central control system in two steps: steady state.
1) raw information is processed in real time by a digital com- The static-state estimator results from a combination of two
puter into a more useful form; and 2) control decisions are made big fields, load flow, and statistical estimation theory so no
from the processed information either by the digital computer attempt will be made to provide a complete bibliography. In [4]
or by a human operator. In this paper only certain aspects of the a clear, readily available load-flow reference with an extensive
first or information processing stage are discussed. In particular, bibliography is provided. Basic statistics references are given in
real-time digital-computer programs (algorithms) for converting [1], [6]. However, there are many other good papers and books
the available information (direct meter readings plus other in- which also contain the same related material. Material is con-
formation) into an estimate of the present state of the power tained in [7] that is very similar to the static-state estimator
system are discussed. Only the quasi-steady state or static oper- concept; differences lie primarily in details and emphasis.
ating behavior of the power system is considered. The vector The following notational conventions are used. A bold face
whose elements are the voltage magnitudes and angles at all the letter denotes a vector or matrix. All vectors are column vectors.
generation and load buses is called the state vector of the system. A prime denotes matrix transpose and -1 denotes matrix in-
Because of the quasi-steady-state assumption,- this vector is verse. The letter E denotes the expectation or averaging opera-
called the static-state vector. tion on a random variable or vector. Complex numbers or matrices
The subsequent use of the state estimate as an input to com- are denoted by a tilde; i.e., Y is a real number while i is a com-
puter control programs or to drive displays for the system opera- plex number.
tor is not discussed. However, it should be clear that no control
system (man or computer) can effectively tell a power system NATURE OF PROBLEM
where to go in the future without some knowledge of is present A power system rarely achieves a true steady-state (static)
state. operating point, as loads and generation patterns are continually
Static-state estimation is related to conventional load-flow changing. However, it is often a reasonable approximation to
calculations. However, the state estimator is designed to handle consider a power system to be in steady state over some short
the many uncertainties associated with trying to do an on-line interval of time.
load flow for an actual system using meter readings telemetered The power systems of much of the United States and Canada
in real time to a digital computer. Uncertainties arise because of are interconnected. However, this paper is concerned with the
-meter and communication errors, incomplete metering, errors in state estimation problems associated with the central control
mathematical models, unexpected system changes, etc. These of only some small portion of the overall power system. The term
uncertainties make for large differences between the usual load- own system (OS) refers to that portion of the total power system
flow studies done in the office for system planning and on-line of direct concern. OS is connected to the interconnected systems
estimation done as part of a control system. (IS) by tie lines. OS need not be a single utility; it may be a pool
or region, etc.
Paper 69 C 2-PWR, recommended and approved by the Power Meters (watt, volt, var, etc.) can be placed on generator buses,
System Engineering Committee of the IEEE Power Group for EHV transmission lines, tie lines, and load buses, and the readings
presentation at the IEEE PICA Conference, Denver, Colo., May can be telemetered in real time to the central control system.
18-21, 1969. Manuscript submitted January 20, 1969; made available
for printing August 1, 1969. This research was supported by the However, real-time meter readings from the IS are not assumed
American Electric Power Service Corporation and NSF Fellowships. to be available to OS. Furthermore, for the sake of generality,
The authors are with the Massachusetts Institute of Technology,
Cambridge, Mass. 02139. real-time measurements on all OS loads (stepdown transformer
SCHWEPPE AND WILDES: SYSTEM STATIC-STATE ESTIMATION, I 121

to a lower voltage distribution network) are not assumed to be R eal-Time Meter Measurements
available. Real-time measurements are assumed to be available Many meter measurements are transmitted to the central
only on OS generators, some EHV transmission lines, and all tie control system in real time. Let the vector Zmeter denote the re-
lines. Situations where more real-time metering is available are sultant set of numbers fed into the computer. Let the vector
special cases or require only simple modification. The real-time Zideal denote these numbers assuming no errors are made. Then
meter readings will not be perfect. Small random errors are
always present. Bad data points will occur. Sometimes a meter Zmeter = Zideal + nmeter.
reading will be lost completely. The vector nmeter represents errors which come from many sources
The transmission network is modeled by a complex admit- such as meter inaccuracies, communication errors, unbalanced
tance matrix Y. The vector x is used to denote the voltage magni- phases, and the effects of analog to digital conversion. The error
tudes and angles at all the buses (relative to some reference). vector nmeter is modeled as a zero mean random vector with
The mathematical model based on Y is only an approximation to
the actual system behavior. E{flmeteilnmeter } = Ometery where Ometer is a positive definite, sym-
OS control has to know the present network flows (voltage, metric matrix, the error covariance matrix. Let Zideal, model
power, current, etc.) on OS in order to predict the flows that will denote the vector of measurements that would be obtained if
result if a transmission line or generator is removed (or added) perfect meters were placed on the static-structure model. Since
in the near future (several hours). To estimate the present flows, the static-structure model is not an exact representation of the
it is reasonable first to estimate the state vector x and then use Y actual system, Zideal is not exactly equal to Zideal, model) but this
and Kirchoff's laws to obtain the desired estimated flows. The difference is ignored under the assumption that other error sources
prediction problem is conceptually quite similar. Since the state dominate.
vector x contains the bus voltages, it will change if a line is Let x denote the state vector associated with the model, and
removed. However, it is reasonable to assume that the generator let Xtrue denote the true (but unknown) value of the state of the
real powers and voltage levels and the load real and reactive model (system). Then the real-time meter readings are modeled
powers will not change (except during a brief transient). Thus an by
estimate of the state vector x after the line is removed can be Zmeter = f(Xtrue) + tlmeter (1)
calculated from the present x, and hence the resulting flows
can be predicted. A change in the generation-load pattern can be where f(x) is a nonlinear function of x determined by the ad-
handled in the same way. If predicting many minutes or hours mittance matrix Y and Kirchoff's laws.
into the future is desired, load forecasting and knowledge of
generator control laws must also be included. The problems of Pseudomeasurements
estimating the present flows and predicting future flows are Real-time meter measurements do not contain all the available
conceptually similar, but in practice they are very different. information on the power system. However, nonreal-time metered
To estimate present flows on OS, the state vector x actually types of information are handled and modeled in the same
need contain only the voltages at the OS generators, loads, and manner as real meter measurements. Such pseudomeasurements
tie lines. To predict the effect of changing the transmission are now discussed.
network or generation pattern, the state vector x must include To illustrate the pseudomeasurement concept, consider a load
buses of both IS and OS. bus whose real power demand is not sent to the central control
Since real-time meter measurements are assumed to be available system in real time. Assume it is known from records of past
only from meters on OS, more information must be added in the behavior that the load should be close to some nominal value of
form of assumptions or data transferred in nonreal time from IS real power Pnom. This information is fed to the computer as a
to OS. Similar additional information may also be needed on OS pseudomeasurement z1 whose value is determined by z1 = Pnom.
load buses that are not metered in real time. Let fi(x) denote the real power at the load point that would be
calculated from the static structure model. Then the pseudo-
MODELING measurement z1 is modeled as
Static Structure
The network generator-load geometry and the network im-
Z, = fi (Xtrue) + ?7 (2)
pedance values determine the static structure of a power system. where i)l is a zero mean random variable with variance Oi. The
One approach to modeling the static structure envolves two value of O1 is determined from past records; for example, as the
concentric rings. The inner ring contains OS plus all systems average of the squares of the deviations of the actual real power
with direct interconnections to OS and possibly others which is determined from the nominal value Pnom.
may have a major effect on OS. The model to be used includes Some of the many types of nonreal-time metered information
the static structures of only the systems in the inner ring. Each are now discussed. All of these can be incorporated into the model
inner ring system, except OS, may have tie lines to outer ring as pseudomeasurements with corresponding error variances.
systems. However, in the model all tie lines from any one inner- Certain load powers remain close to constant nominal values
ring system to the outer-ring system are assumed to be tied to a which can be treated as pseudomeasurements. Other loads vary
single bus along with a single hypothetical generator which widely during a day but remain close to some nominal percentage
models the total interchange between the particular inner-ring of the total load (or portion of load). Such percentages can be
system and all of the outer-ring systems. The transmission net- viewed as load-pattern coefficients which can be treated as
work of all the inner-ring systems is modeled by the one admit- pseudomeasurements. Nominal load-power factors can be treated
tance matrix Y. Of course, all network elements, generators, and as pseudomeasurements. If load-pattern coefficients are used, a
loads are not explicitly included. Lower voltage distribution lines total load daily cycle may also be needed which can be modeled
are ignored in calculating Yl, and small generator or load buses as Pr, total(t) = Ppeakb(t), where Ppea. is the nominal daily peak
may be dropped or lumped together. and b(t) (0 < t < 24 hours) is the nominal daily load cycle.
122 IEEE TRANSACTIONS ON POWER APPARATUS AND SYSTEMS, JANUARY 1970

Information on the nominal Ppeak can be treated as a pseudo-

measurement while b(t) can be treated as being exact where
actual deviations are accounted for by making the error variance
associated with P,k sufficiently large. Some of the generators
may be base loaded, and this base load output can be treated as a
pseudomeasurement. Other generators will follow the load in a
manner determined by control laws. Generator pattern coeffi-
cients can be based on the control laws to specify how the genera-
tion is distributed, and these coefficients can be treated as pseudo-
measurements. Nominal real power sales between interconnected
systems may be known and can be treated as pseudomeasure-
ments. Voltage levels at generators and some buses are often
controlled, and the nominal values can be treated as pseudo-
measurements. Fig. 1. Battery with voltage Xtrue.
The actual pseudomeasurements to be used obviously depend
on the type and amount of real meter measurements that are GENERAL THEORY
available. Thus generalizations are difficult. However, the follow-
ing list indicates one reasonable set of pseudomeasurements that Example
is sufficient to yield a complete state estimate when combined An example is used to introduce the state estimation logic.
with the meter measurements assumed to be available. This Consider the simple network of Fig. 1.
list is only an example. For OS pseudomeasurements are used for Let z1 and Z2 denote the ampmeter and voltmeter readings, re-
1) real power load-pattern coefficients and nominal values; spectively. Assume that the values of R1 and R2 are known while
2) nominal power factors at nonvoltage controlled load buses; the battery voltage Xtrue is unknown. The value of the state
and 3) nominal voltages at voltage controlled load buses. For (voltage) Xtrue is to be estimated from the meter readings z1 and
each system in the inner ring, other than OS, pseudomeasure- Z2.
ments are used for 1) real power load-pattern coefficients or Let t1 and fl2 denote the errors in the meters. The measure-
nominal values, 2) nominal power factors at load buses, 3) ments z1 and Z2 are modeled by
voltage magnitudes at generators, 4) real power generation dis-
tribution pattern coefficients or nominal values, 5) power sales Zk =Cfk(Xtrue) + nck, k = 1,2
to other companies in inner ring, 6) power sales to other com-
panies in outer ring, and 7) total daily load cycle. Obviously all fi(x) = (1/HR + R2)Xtrue (4)
the information on the IS must come from some place. It could be f2(x) = (R2/H1 + R2)xtrue.
simply best guesses made by OS engineers. In a more ideal world,
each inner-ring system would furnish OS the nominal values on a Assume that 'il and rp2 are modeled as independent zero mean
daily or weekly basis and instantaneously transmit to OS major random variables with variances 0/c, k = 1,2. In vector form this
unexpected changes as they occur. If best guesses are used, the model is in the class of (3) when 6 is a diagonal matrix with main
associated error covariances are of course larger. diagonal elements 0//.
Let x denote the estimate of x made from z1 and Z2. Then f
Discussion of Model is defined to be the value of x which minimizes
The static structure model is given by the admittance matrix 2
f which includes only the important elements of the inner-ring J (X) = E7> [Zk - fk(x)]2 (5)
systems. All real-time meter measurements and other types k= 1 Ok/k
of information are modeled in the same way in terms of
or in equivalent vector notation
Z = f(Xtrue) + n (3)
J(x) = [z - f(x) ]'- [z - f(x) ]- (6)
where x is the state of the static structure model and f(x) is
determined by Kirchoff's laws and the admittance matrix Y. It follows by simple calculus that
The error n is modeled as a zero mean random vector with
E{nn'} = 6. HX +R- r; 1 Z1 + R2l - Z2 (7)
In the discussions to follow on state estimation, the model
(i.e., Yif(x),6) is assumed to be perfect. Obviously the model where
itself is often in error. Not all network elements are included in
calculating Y. If a transmission line is unexpectedly lost, Y,
= -(R
011= + R2202
and hence f(x), is in error. The actual line admittances are not 2 - ~ ~ 022-'
011-1 + R2
(8)
known perfectly. Modeling errors are also caused by bad data
points resulting from meter, communication, or computer input It also follows that E{x } = xtrueand E { (x - ftre)2} = 2.
failure (or worse, partial failure). Bad data points are viewed as Now consider why the estimate £ was defined as the value of x
an error in the modeling of 6 as they are measurements with which minimizes J(x) of (5). It is reasonable to choose as an
errors much larger than assumed in the modeling of 6. Modeling estimate the value of x which best fits the observations where the
errors are handled by detection and identification. Detection is fit is weighted by the expected accuracy of the measurement.
concerned with spotting the existence of a modeling error while Thus the estimate fits the measurement made by the accurate
identification is concerned with locating and correcting the error. meter better than the measurement made by the inaccurate
SCHWEPPE AND WILDES: SYSTEM STATIC-STATE ESTIMATION, I 123

meter. This reasonability argument for the choice of estimate biology, aerospace trajectories). Actually, the resulting x can be
can be supplemented by mathematical proofs if desired. For shown to have a variety of mathematical optimum properties,
example, it can be shown that t of (7) yields the smallest estimate and the criterion can be formally derived in various ways. If the
error variance E {x-Xtue }I2 for all unbiased estimates. errors are assumed to be Gaussian random variables, x is a
maximum likelihood estimate which has many nice properties,
State Estimation and the Cramer-Rao or information inequality can be used to
In the problem of interest (3), the dimensions of the vectors interpretl [1], [6].
are larger, and f(x) is a nonlinear function of x. Given z, the state
estimate xi is defined to be the value of x which minimizes Detection
The state estimator algorithm of (15) is based on the assump-
J(x) = [z -f(x)]'6-l[z -f(x)]. (9) tion that the model (i.e., k,f(x),6) is perfect. Detection of model-
ing errors is now discussed. Let J denote J(x), i.e., J(x) evaluated
The minimization of J(x) of (9) with respect to x cannot be for x = x.
done in closed form as in the example because f(x) is a nonlinear One approach to detection can be based on the fact that if the
function of x. However, the following is a standard approach. model is correct, J is a random variable whose probability dis-
Let x0 be an initial guess for x. Then a Taylor series expansion tribution can be calculated (at least approximately) if the
gives measurement errors n are assumed to be Gaussian (and small).
f(x) = f(xo) + F(xo) [Ax] + * (10) Thus if a value of I is obtained which lies on the tails of this
probability distribution, it can be assumed that the model is
where F(xo) is the Jacobian matrix whose elements are no good, i.e., an error in Y and/or 6 has been detected.
A different approach to detection does not rely on probability
afk1(x)/0xk2, X = xo
distributions but requires more calculation. To illustrate the
and Ax = x - xo. If the higher order terms of (10) are neglected, method, consider the case of detecting a lost line. Assume that
substitution of (10) into (9) gives there are K transmission lines in the static-structure model but
that one line may actually be out of service. Let ko denote the
J(x) = [Az - F(xo)Ax]'6'-[Az - F(xo)Ax] (11) admittance matrix with all K lines in the model, and let Y5
denote the admittance matrices with the kth line removed from
where Az = z - F(xo). Equation 11 can be rewritten in the form the model, k = 1... K. Thus there are K + 1 possible models
J(x) = Az'[0-1 -0-YF(xo)X(xo)F'(xo)O-L]Az + [Ax - Z (xo) for the observations given by

*F'(xo)O-'Az]'X-1(xo) [Ax - X(xo)F'(xo)O-'Az] (12) z=fk(x)+n, k=0*..K (17)

where Z (xo) = [F'(xo)h-'F(xo) ] -1 and (xo) is assumed to exist. where fk(x) depends on Yk, k = 0 ... K. There are also K + 1
Let Ax' denote the value of Ax that minimizes J(x) of (12). It is possible state estimatesXk, k = 0 ... K depending on which model
obvious from (12) that is used, where 2k is the value of x which minimizes Jk(x) which
is J(x) for f(x) = fk(x). Let A? denote Jk(x) evaluated for x =
Ax = Z(xo)F'(xo)O-&Az (13) Xk. J1k can be viewed as a measure of how good a fit the static
or structure model with Yk [and fk(x) I provides to the observation
z. If it turned out that for some kl, Jkh << Jk, all k # 1ki it would
=x Xa + Z(xo)F'(xo)O'[z-f(xo)]. (14) be reasonable to assume that the observations z were made on a
If xo is close enough to x to justify the dropping of the higher network with line ki removed. Hence a modeling error caused
terms in (10), then (14) actually gives the desired x. In practice by a lost line (whose loss is not modeled) can be detected by
this is often not true so one defines a sequence xn, n = 1, evaluating
..
lk, k = 0 ... K. This concept can be extended to the
given by problem of detecting bad data points, i.e., errors in the modeling
of 6. To illustrate, assume 6 is diagonal. Then a particular
Xn+1 = xn + X(xn)F'(xn)0-¶[z - f(xn)]. (15) measurement can be modeled as a bad data point by making the
corresponding main diagonal of 0-1 equal to zero. This leads to a
The iterations of (15) are continued until J(xn) approaches a set of possible models and corresponding Jk, k = 1 ... which can
minimum. Possible stopping rules are to iterate until IJ(xn) - be used as before.
J(xn+l) or until the magnitude of all components of x+ -Xn The detection logics are presented in terms of reasonability
are less than some predetermined value. Note that it is possible arguments. However, they are actually drawn from the theory
for J(x) to have local minimum and flat spots, and thus x, may of hypothesis testing and can be given a more formal mathe-
converge but not to x. It is also possible for xn to never converge. matical base.
If x0 happens to be the true but unknown value Xtrue, then it
follows from (14) that Identification
E{IlXtrue - I][Xtrue - x] } = -(i). (16) After a modeling error has been detected, the next step is to
identify the error so that it can be corrected. The detection logic
Equation (16) is actually an approximation valid for small errors. of (17) also identifies the lost lines. The problem of identifying
Thus Z (x) can often be viewed as a measure of the accuracy of x. modeling errors caused by erroneous transmission line admit-
The state estimate x is defined as the value of x that minimizes tances requires an extension of the concept, wherein J is viewed
J(x) of (9). This criterion has a long history of providing good as a function of the line admittances and identification is accom-
estimates in a wide variety of problems (surveying, economics, plished by finding the admittance values which minimize J.
 124 IEEE TRANSACTIONS ON POWER APPARATUS AND SYSTEMS, JANUARY 1970

The identification problem can be viewed as a state-estimation INITIAL TESTS

problem, where the state vector has been expanded to include The behavior of the static-state estimator has been investi-
unknown model parameters as well as the bus voltage magnitudes gated by digital-computer simulation. The nature and results of
and angles. The special term, identification, is used to reduce these initial tests are briefly summarized. Many numerical re-
confusion. sults and much more extensive discussions can be found in [5].
The computer program can be viewed as a Monte Carlo
Discussion of Theory simulation. It can be divided into four blocks.
Return first to the simple example of Fig. 1. x is the weighted 1) Given the desired network and bus powers, establish the
sum of the ammeter and voltmeter readings, z1 and z2. The true bus voltages by a conventional load flow.
weights are determined by the error variances 6kk, the resistance 2) Given the desired meter placement, types, and accuracies,
values Rk, and the network structure. If R2 = 1, and if 0 > 022 simulate a set of meter readings by first calculating the readings
(meter 1 noisier than meter 2), then the more accurate meter perfect meters would make and then adding errors obtained fromn
reading Z2 is given the most weight, i.e., has most effect on the a random number generator.
estimate. However, if 1 > R2, the more inaccurate measurement 3) Given the simulated measurements; calculate the best
z1 may be actually given more weight. Thus as far as the estimate estimate of the voltages at the buses using the static-state
is concerned, the weighting due to the 6kk, may be overcome by estimator.
geometry and network effects. This weighting effect carries over 4) Analyze the errors by subtracting the true bus voltages
to the general case. Thus the state estimator algorithm (15) from the estimated bus voltages and repeating blocks 2) and 3)
is a way of combining measurements of many types into a single many times using different random errors so that an average
estimate in such a way that the importance of any one measure- squared error can be obtained.
ment is automatically determined. The need to combine both The initial tests to be discussed here were based on a five-bus
pseudo and actual real-time measurements of many types makes system which is an example of [4, p. 284]. Many trials were
this automatic weighting very important. made with various numbers, types, locations, and accuracies
The iteration sequence (15) depends on the matrix X which for the meters which could measure real power, reactive power,
exists only if the matrix F'f-'F has an inverse. Let K2 denote the or voltage at any bus. A measurement with a large error variance
dimension of the observation z, and let K1 denote the dimension was assumed to be a pseudomeasurement.
of the state x. l; does not exist if K1 > K2. If K1 = K2, and if F-' Performance depended heavily on meter placement, accuracy,
exists, (15) becomes and type. For a good choice of meter placement and types, the
iteration of (15) converged rapidly with an average rate approxi-
Xn+1 = xn + F-¶(xn) [z - f(xnjJ. (18) mately independent of initial starting value or variance of the
measurement error. However, the actual convergence behavior
If z consists of specified generator and load bus watts, vars, could depend on the particular noise values. For a bad choice of
and voltages, this iteration is approximately the Newton- meter placement and type, convergence was sometimes slow or
Raphson method which has been successfully used to obtain nonexistent. Meter placement and type as well as meter accuracy
conventional load flows [4]. If K2 > Ki, some of the measure- effected the accuracy of the final estimate. The effect of adding
ments are redundant. Thus an alternate to using (15) is to throw or removing one meter was sometimes dramatic and sometimes
away redundant observations until K2 = K1 and then do a con- unobservable; it all depended on the situation. Usually varmeters
ventional load-flow calculation as typified by (18). However, this were more effective than voltmeters but a critically located
alternate approach is not recommended for the following reasons. voltmeter could be a big help in some cases. Pseudomeasurements
Throwing away of information reduces the overall accuracy. with large error variances could have major or minor effects,
It is difficult to decide which measurements to ignore; for once again depending on location and type.
example, is a watt meter measurement on an EHV transmission At no time in the initial tests did the iterations appear to
line more or less valuable than knowledge of the nominal voltage converge to a local minimum or hang up on a flat spot. However,
level at a load pointy Equation (18) is not much easier than (15) there is of course no way to be positive this did not happen with-
to implement on a computer. Finally and perhaps most im- out evaluating J(x) for all possible values of x.
portant, the redundant measurements are the quantities which A wide range of meter placement and types was explored in
enable detection and identification. the hopes of developing some rules of thumb for choosing meter
The importance of the use of redundancy for detection and placement and types which are economical and yet provide
identification should be emphasized. Data processing algorithms good estimates. Unfortunately, no such insight that could be
for handling large quantities of telemetered data made on a extrapolated to larger systems was obtained. This is the main
complex changing system must have internal checks to detect reason a larger than five-bus system was not investigated.
modeling errors such as unexpected system changes and bad We are convinced the static-state estimator will work well on any
data points. Without such tests, very little faith can be given to order system provided that good meter placement and types
the estimate. are chosen. However, at our present level of insight, we would
The interpretation of the matrix w as the covariance matrix have to study the exact system to see if a particular meter place-
of the errors in x has important applications. The value of any ment pattern is good. (Of course, if cost is no consideration,
estimate is greatly enihanced if its accuracy is known. I is also complete direct metering of all buses and lines in OS and IS
very useful in initial design and development as x can be calcu- is good.)
lated before the state estimator has been implemented. I can be The simulations did verify the fact that I does indeed
used to study where meters should be placed and what type and approximate the average squared error if the errors are small. In
accuracy of information is actually needed from IS. I is not a fact, most of the studies on meter placement and accuracy were
perfect measure of performance, but it is a very useful design based on calculating only I rather than using the program in a.
tool. complete Monte Carlo mode of operation.
IEEE TRANSACTIONS ON POWER APPARATUS AND SYSTEMS, VOL. PAS-89, NO. 1, JANUARY 1970 125

CONCLUSIONS REFERENCES
The basic concepts underlying a static-state estimator and the
corresponding detection-identification logics have been pre- [1] H. Cramer, Mathematical Methods of Statistics. Princeton, N. J.:
sented. The paper can be viewed as an exercise in modeling Princeton University Press, 1946.
and formulating the power system problem so that classical [2] F. C. Schweppe and D. Rom, "Power system static-state
estimation, pt. II: approximate model," this issue, pp. 125-
estimation and detection theory can be employed. The resulting 130.
equations are reasonable and worked well during simulation. [3] F. C. Schweppe, "Power system static-state estimation, pt.
III: implementation," this issue, pp. 130-135.
Many important aspects of the problem were not discussed. [4] G. W. Stagg and A. H. El-Abiad, Computer Methods in Power
Some of these will be considered in [2] and [3]. Systems Analysis. New York: McGraw-Hill, 1968.
[5] J. Wildes, "A static-state estimator for a power system network,"
ACKNOWLEDGMENT Power Systems Engrg. Group, Dept. of Elec. Engrg., Massa-
chusetts Institute of Technology, Cambridge, Rept. 8, August
The author wishes to thank American Power Service Corpora- 19, 1968.
tion engineers for their help in explaining power system problems. [6] S. Wilks, Mathematical Statistics. New York: Wiley, 1962.
These engineers may not agree with the results, but the paper [7] R. Larson and J. Peschon, "Feasibility study of system state
estimation," Wolf Management Services, Palo Alto, Calif.,
would have been impossible without their help. Interim Rept., 1968.

Power System Static-State Estimation,

Part II: Approximate Model
FRED C. SCHWEPPE, MEMBER, IEEE, AND DOUGLAS B. ROM, STUDENT MEMBER, IEEE

Abstract-The static state of an electric power system is defined In the first paper [3] mathematical models and general state
as the vector of the voltage magnitudes and angles at all network estimation, detection, and identification algorithms were dis-
buses. The static-state estimator is a data-processing algorithm for cussed. In this paper, an approximate mathematical model and
converting redundant meter readings and other available informa- the resulting simplifications in estimation, detection, and identi-
tion into an estimate of the static-state vector. Discussions center on
fication are discussed. In the following paper [4] various imple-
an approximate mathematical model (related to the dc load-flow
model). This model yields noniterative-state estimation equations, mentation problems associated with dimensionality, computer
simplified prediction of effects of network and generation-load speed and storage, and the time-varying nature of actual power
pattern changes on network flow, and simplified detection and identi- systems will be discussed.
fication of modeling errors. Results of some initial computer studies The following notational conventions are used. A bold face
on the real power-voltage angle portion of the approximate model letter denotes a vector or matrix. All vectors are column vectors.
are discussed. A prime denotes matrix transpose, and -1 denotes matrix in-
verse. The letter E denotes the expectation or averaging opera-
tion on a random variable or vector.
INTRODUCTION
T HIS PAPER is the second of a three-part sequence on a APPROXIMATE MODEL
static-state estimator and related detection and identifica- The mathematical model developed in [3] is of the form
tion problems. The static-state estimator is a data processing
algorithm for use by a digital computer to reduce meter measure- z = f(x) + n (1)
ments and other information on an electric power system into where x is the state vector (voltage magnitude and angle at all
an estimate of the system's steady-state or static-state vector. buses), f(x) is the nonlinear function of x determined by network
The static state is a vector composed of the voltage magnitudes admittance matrix Y' which gives the measurements of ideal
and angles at all the buses. meters, n is the measurement (meter) error vector (zero mean,
random vector with error covariance matrix 0), and z is the vector
Paper 69 C 2-PWR, recommended and approved by the Power of measurements fed into computer. Certain approximations of
System Engineering Committee of the IEEE Power Group for
presentation at the IEEE PICA Conference, Denver, Colo., May f(x) yield an approximate mathematical model which is closely
18-21, 1969. Manuscript submitted January 20, 1969; made available related to the dc load-flow concept.
for printing August 1, 1969. This research was supported by the Assume there are Nb buses and Nt transmission lines. Let
American Electric Power Service Corporation and NSF Fellowships. Pk and Qk denote the real and reactive electric power into kth
The authors are with the Massachusetts Institute of Technology,
Cambridge, Mass. 02139. bus, k = 1... Nb from load or generator sources. Let Vk and 8,k