972

[82]
[83]
[84]
[85]

PROCEEDINGS OF THE IEEE, VOL.

transmission anddistributionplanning,in
Proc.
4th
PSCC
(Grenoble, France, 1972). Paper 1.1/13.
R. A. Evans, A quadratic programmingalgorithmfor
optimal
generation planning, presented at IEEE PES Winter Meet., New
Ywk, 1973. Paper C 73 112-0.
R. R. Y t h , Power system simulation modelbaaed on probability
analysis. IEEE Trans. Pown App. Syst., vol. PAS-91, pp. 62-49,
Jae/Feb. 1972.
-,
Optimalgenerationplanningconsideringuncertainty,
ZEEE Trans. P w n App. Sysf., VOL PAS-91, pp. 70-77, Jan./Feb.
1972.
A. Breton and M. Falgarone, Application dela theorie de la commande optimale au probleme du choix des equipment8 de production a E.D.F, in Proc. 4th PSCC (Grenoble, France, 1972). Paper
1.1/7.

puters inpower system planning, this issue, pp. 872-883.

1871 W.R. Puntel et al., An automatedmethod for long-range planning
of transmieaion networks, in Proc. 8th P I C A C m f (Minneapolis,
Minn, 1973), p,p. 3 M :
(881 L; L. Garver, Transrmssion network estimation using linear pro.
vol. PAS-89, pp. 1688gramming, ZEEE Trans. P m n A p pSyst.,
1697, Sept./Oct. 1970.
[ 8 9 ] Y. P. Dusonchet and A. H. El Abiad, Transmiasion planning using discrete dynamic optimizing, IEEE Trans. Pwer App. Syst.,
vol. PAS-92, pp. 1358-1371, July/Aug. 1973.

[90] G. Cuervo, G. Navas, and R. Querubin, Optimal design of radial

distributionnetworks(inSpanish),presented
Cong., Bogota, Colombia, 1973.

1974

Capacitor Opfimizdion
[PSI S.S.Sachdeva and R. Billinton, Optimal network var planning by
[96[
1971
[98]

1991
(1001

Disbibution Design and O p a a t i a

JULY

[91] R. N. Adams and M. A. Laughton, A dynamic programming network flow procedure for distribution system planning, in Proc. 8th
P I C A Cmf. (Minneapolis, Minn., 1973), pp. 348-354.
1921 K. Matsueda and M. Matsuda, Secondary bus voltage regulation
method a t distribution substation, in Proc. 4 f h PSCC (Grenoble,
France, 1972). Paper 2.1/4.
[93] T. Okada and C.Uenoaono, Mathematical modelsfor optimal
design of power distribution systems for long terms, in Proc. 4th
PSCC (Grenoble, France, 1972). Paper 1.1/3.
[94] J. P. Ludot and M. C. Rubinstein, Methodes pour la planification
acourttermedes
reseaux dedistribution,in Proc.4th PSCC
(Grenoble, France, 1972). Paper 1.1/12.

Transmission Nctcaork Expansion Planning

[MIU. G. Knight, R. R. Booth, S . A. Mallard, and D. M. Lewis, Com-

62, NO. 7,

[loll

a t Panamerican

nonlinearprogramming,
ZEEETrans. Power A H . Sysf., vol.
PAS-92, pp. 1217-1225, July/Aug. 1973.
-,
Optimal network var planning using real and reactive power
decomposition
non-linear
analysis,
in Proc.
8th
P I C A Conf.
(Minneapolis, Minn., 1973), pp. 33%347.
A. Kishore and E. F. Hill, Static optimization of reactive power
use of sensitivityparameters,
IEEETrans.
sources bythe
Power App. Sysf.,vol. PAS-90, pp. 1166-1173, May/June 1971.
R. M. Porter andH. K. Amchin, Var allocation-An optimization
technique,in Proc. 7th P I C A Conf. (Boston,Mass., 197l), pp.
310-318.
J. A. Bubenko and B. M. Waern, A method for the unit commitment of thereactive compensation, in Proc. 7 f h P I C A Conf.
(Boston, Mass., 1971), pp. 3 0 3 3 0 9 .
A. M. Pretelt, Automatic allocationof network capacitors, I E E E
Trans. P o w n A p p Syst.,
.
vol. PAS-90, pp. 54-61, Jan./Feb. 1971.
R. M.Maliszewski, L. L. Garver, and A. J. Wood, L i y a r programming asan aidinplanningkilovarrequirements,
IEEE
Trans. Power App. Syst., vol. PAS-87, pp. 1 9 6 3 1 9 6 8 , Dec. 1968

Static State Estimation in ElectricPowerSystems

FRED C. SCHWEPPE,

MEMBER, IEEE, AND

EDMUND J. HANDSCHIN,

M E ~ E R ,IEEE

Invited Paper

Abstracl-A static state estimator is a collection of digital computer programs which convert telemetered data into a reliable estimate of the traagmission network structure and statebyaccounting
for 1) small randommetering-communicationerrors; 2) uncertahties
in system parameter values; 3) bad data duetotransientsandmetercommunication f*-;
and 4) m ~ r in
s thenetworkstructure
to fadQ
bre*er
information The Overall state
estimationprocessconsists of four steps: 1) hypothesize m
a
t
h
e
mm:
.2). =timate state vector: 3) detect bad b t a and/or
structure errors: and identify bad data andior structure errors. The
Problem is characterizedby high dimensionalityandtheneedfor
real-time
limited -pukr
time and storee
various
methods of solution are discussed and compared.

I. INTRODUCTION
HERE IS NO single structure or list,offunctions for a

modern centralcontrolsystemas each utility or group

of utilitieshasits
own sDecia1 Droblems and needs.
Nevertheless, Fig. 1 shows a typical central control system
in
order
to
indicate
the
role of a static
state
estimator.
The
many action oriented tasks of Fig. 1 work with and from a
central data base. Indeed, the life blood of the control system
is a base of clean pure data defining the system state and
status(voltages,networkconfiguration).This
life blood is
obtainedfromthenourishment
providedby
the measurements gathered from around the system (data acquisition).A
static
state estimator is the digestive system which removes
Manuscript received January 9, 1974.
F.c. M w e p p e is with ~ l ~ c tpower
r i ~ Systems ~ n g i - r i ~hbora- the impurities from the measurements and converts them into
tory, Energy Laboratory, Department of Electrical Engineering, Mass+
a form which the brain (man or computer) of the central conchusetts Instituteof Technology, Cambridge, Mass.02139.
E. J. Handschin is with Brown Boveri Resear& Center,
trol systemcanreadily
use tomakeaction
decisions on
Switzerland. security.
and
quality, economy,
system

973

XEWEPPE AND HANDSCHIN: STATIC STATE ESTIMATION I N P O W E R SYSTEMS

Fig. 2.

Fig. 1. Schematic diagram for typical control center.

There are many ways to

design a static state estimator
and the choice depends on the particular systemsneeds. The
purpose of this paper is to summarize and compare various
possible approaches.
11. BACKGROUND

Purpose of static state estimator.

(dataacquisition)simplycannotprovide
a clean complete
data base and so has to be rejected for most central control
systems. The second approach (data logging, limit checking,
etc.) can provide a partially cleaned up but often still incomplete data base. The third approach (static state estimation)
provides a complete data base that is as clean as possible with
the given measurement system (or, equivalently, static state
estimation requires less metering redundancy to achieve the
same level of cleanliness). The disadvantage of static state
estimation is the increase in digital computer software
sophistication and dataprocessing costs.
The preceding definition follows the conceptsof [ 11- [4]
by viewing a static state estimatoras all the computer codes,
logics, and algorithms needed to clean up the data. The fact
that static state estimation is based onanoverallmathematical model is perhaps its most distinguishing feature.

The static state estimation concept

(as i t is beingdiscussed in this paper) has evolved rapidly from its initial inception [l] to on-line implementations; the first is apparently
111. PROBLEM
FORMULATION
the Tokke system installation in Norway[43]with the larger
A . Definitions and Notation
AEP installation[ 141 following soon after. Many people have
The purpose of a static state estimator is summarized in
contributed to this evolution as indicated by the papers inFig. 2. The following definitions provide a basis for the subsecluded in the bibliography. An interesting survey instigated
by Ariatti within the CIGRE Study Committee 32, Working quent discussions (numerical values are indicated onlyto give
Group 01,contains the opinions
of more than thirty utilities on a feel for the nominal values and many exceptionsexist).
Tim:
on-line powersystem state estimation [30].The authors
of [ 141,
t denotes continuous time; t,, n = 1, 2,
, denotes dis[20], [26],
[37], 1421-[44], [52] discuss static state estimacrete time.
tion plans of various companies both in Europe and North
America. Other utilities are actively developinga static state
System Conditions:
Transient condition: Systembehaviorimmediately
estimation capability. However, there was and is no univerfolsal acceptance of the concept ([45]is a published critique). lowing major disturbance. The transient conditions time conThere has been some confusion on what a static state esti- stant TT is the effective time duration of transients. For line
switching: 0.1 ~ 5 ~ s~
mator is and what its purposes are.
; for1 loss
5 of generators: 2 s 5~~1.5
The words state estimation are used in modern system
min.
(control)theoryto
refer to a very wide class of problems
Static condition: Quasi steady-statesystembehavior
which are discussed in many papers and books ([MI is one where generators are responding normally to
loadchanges.
The static condition time constant TS is the amount of time
example which at least one of the authors of this paper is
prejudiced in favor of). In this general sense, state estimation before the system changes an appreciable amount: seconds
concepts can be applied in many power system areas ranging 5 TS <many minutes.
from short-range load forecasting to power plant work (see
Independent Varkbles (Unknown):
for example
[47], [48]).
The term static state estimaNo& injections Q ( t ) : Real and reactive power injections
tion is used in this paper to refer to the particular applicainto nodes (or buses) of the network. Complex, nonstationary,
tion indicated inFig. 1.
environmentally dependent vector valued stochastic
process
Three possible approaches to obtaining the crucial data
which is well behaved in static conditions. The static conbase of Fig. 1 are 1) data acquisition with no processing; 2) dition time constant T S is determined by the rate of change
data acquisition with digital computer used for data logging, of Q ( t ) .
limit checking, andsimple logic comparisonsbetween reTrue structure [ ( t ) : Interconnection of transmission lines
dundant measurements of essentially the same variable; and
and transformers and location
of meters. Binary type vari3) data acquisition with a digital computer used in a system- ables; mostly constant except a few changes a day.
atic way, based on a mathematical model, to clean up the
True parameter values p ( t ) : Values which determine indidata (by treating small random errors, bad data,
modeling vidual transmission lines, transformers, and meter charactererrors,andparametererrors),andtocompute(estimate)
tics exhibit smallslow variations with time.
quantities and variableswhich are not directly measured. The
Statc Variabks ( Unknown):
third approach is static state estimation. The first approach
State vector x @ ) : Voltage at all network nodes. During

- -

[MI,

974

PROCEEDINGS OF THE IEEE, JULY

the static condition is

it assumed that

where f is a nonlinear function based on Ohms and Kirchhoffs laws. Dimension x(&,) is 2N-1 where N is number of
nodes. (Nmay range anywhere from
10 to 1OOO.)
LoadfEow: An algorithm for computing x(L) given Q(t,,)
((L), e(&,) by solving (3.1) using aniterativetechnique.
(Often some combination of injection and voltage magnitude
is specified instead of just P.)
Mefeting System:
Meter time constant SM: 0.1 s SZM 1 s.
Meter scan: Operation of reading all metersonceand
telemeteringthedatatothecentralcomputer.Usually
t,,
denotes the time instantof one scan.
Scan time A: Time between two complete meter scans:

A = t , , - L l . 1.0s < A s M s .
Snapshot s c a ~All meters are read at the same instant
h).
(samevalue of t betweenand
Sequential scan: Meters are read sequentiallyat different
time instants in the interval
St
Redundancy 7 : Ratio of number of meters to number of
states 1.1
53.0.
Local redundancy a: Redundancy at node k considers
only node k and all other nodes not more than two nodes
away fromnode k.
Lost data: Failure of meter and/or communication link

st,,.

which is known by the central computer.

Known Qwnfitics:
Measurements z(t,,): Telemetered line flows, bus injections, bus voltages, possibly currents, obtained during scanat
time t,,. Dimension of z is 7 times dimensionof x.
Pseudo measurements: Nontelemeteredinformationon
line flows, injections, voltage magnitudes (such as load distributionfactors,
knowledge of t a p changingtransformer
control logics, etc.). Henceforth, z refers to both telemetered
and pseudo measurements.
Parameter wlues #(t,,): Estimate of e(t,,) obtained from
system design data,
Structure s(t,,): Estimate of structure [ ( t i ) obtained from
telemetered or telephoned switch and circuit breaker
positions and meter status duringa scan at time t,,.
Modcl for Measurements z(fl):
For a snapshot scan

h [ ] is a nonlinear function based on Ohms and Kirchhoffs

laws. Dimension of z is greater thandimension of x .
Measurementerror u&): Randomvectorrepresenting
thenormal
difference betweentheactualobservation
z(&,) and h [ ] caused by: 1) nonzero meter time constant,
SM, combined with noninfinite static condition time constant
7s; 2) difference between the model h [ * ] in (3.2) and the
actualthree-phase power system; 3) meter nonlinearities,
miscalibrations,and noise; and 4) A/D conversion errors
and/or communication noise errors. Time structure of u*(t,.)
is very complex and effectively unknown.
Bias error 9: A major component of measurement errors
are biases which are constant in time (provided the state
itself does not change appreciably). I n this paper, biases are
modeled as zero mean random variables (constant in time) so
that u*(t.,) is zero mean. However, the bias can be modeled as

1974

the unknown mean value of uz(t,,), in which case i t must be

estimated (see Section IX).
Measurement error covariance matrix R :

~ [ v s ( t , > v s ~ (A
~ >Rtme(tn)
l

(3.3)

where E denotes expectation. The components of R-(f,,)

are included in e&). (In general, multiplicative errors make
R as function of z, but this dependence is not indicated to
simplify the notation.)
Bad datab(t,,):
Unusuallylargemeasurementerrors
added to u,(t) and caused by: 1) observations during transient
swings, 2) majormeter-communicationfailures
which are
not known by the central computer.b(t,) is usually zero with
large magnitude components occumng with unknown time
structure.
Model for Parameter Values p:

IN#)= e(k)

vp(C).

(3.4)

The parameter value error up(&,)is small with unknown time

structure. # contains R which is Rt, plus errors.
Mo&l for Structure s(h):

The structure error c(t) contains binary-type variables which

have nonzero components only very rareiy. Time structure is
completely unknown.
Static State Estimator:
Hypothesize m o d c l : Given s(t,,) and e(k) determine

h(x) = h b , S(tn), e(t.)ll~=a

e P

(3.6)

and the valuesof R.

State estimate ;(f,,): Value of x which minimizes the residual r(L)

in some Sense during static conditions.

Estimation: Process of computing ;(t,,).
Detection: Test to determine whether bad datab and/or
structural errors c are present in z(t,,) or s(t,,).
Identifiation: Process todetermine
which measurement(s) is (are) bad or which part of the structure is wrong.
Estimate updufe time, T:Time interval between instants
when a new state estimate is computed.
A < T <many minutes
to hours (A is scan time).
B. Ovmall Approach
The four-step conceptof Fig. 3-1) hypothesize structure;
2) estimate; 3) detect; and4) identify-is an excellent general
way to view the static state estimation dataprocessing problem. Fig. 3 applies directly only to the processing of a single
measurementscan,butitsstructure
also appliesin more
general situations. The static state estimator
of Fig. 2 includes an estimator as shown in Fig. 3.
Sections IV through VI11 discuss the four steps of Fig. 3,
which handle measurement errors u,, bad data b, and most
structural errors c. Parameter errors up are discussed separately in Section IX. Overallsystem design considerations
are discussed in Section X.

915

SCHWEPPE AND HANDXHIN: STATIC STATE ESTIMATION I N POWER SYSTEMS

V. ESTIMATION:
TIMEINVARIANT CASE
F a r e t e r ic.o,ntion
p
Structural informtion a

The second step of Fig. 3 is Estimation. In this section

the estimationusing only a single snapshot scan is considered.
Dropping the time dependence from the notation, the model
is given by

z = h(x)

E [ v . v , ~=
] R.

V,

(5.1)

A . Weighted Least Squares ( W L S ) Criteria

With the WLS approach [1]-[4], the state estimate
i is
defined as thatvalue of X which minimizes the weighted
squared residuals r = z - h(x); i.e.,

J ( x ) = [z - h(x)]TR-[z- h ( x ) ] .

(5.2)

WLS estimation iscommon in many fields and has manygood

properties. If u, is Gaussian, then i is the maximum likelihood
estimate; if u, is sufficiently small then i has a minimum error
(see, for example, [54]formore
covariancematrix,etc.
generaldiscussions).
Differentiation of (5.2) shows that x must satisfy the nonlinear equation

Fig. 3.

Four basic operations of static s t a t e estimation.

where

H ( i ) = dh/dx 1 x 4
C. Limitations of Paper
I t is not possible in a paper of this length to discuss all
aspects of static state estimation.A very important aspect of
static state estimation is the problemof computer hardwaresoftware organization and the organization of the data base.
This is notdiscussed as itis a complex subject worth a special
paper.Inactualimplementation,astaticstateestimator
would also include prefilters (data screening logics based on
limit checking, curve
fitting
algorithms,
etc.)
between
Measurements and Estimation inFig. 3. Such preprocossingmethodsarenotdetailedinthispaper.Inmost
cases,
electric power systems are interconnected to their
neighbors
so thetotalsystemcan
be dividedintotwoparts,Own
System and the Interconnected System
(i.e., the rest of
the network). Some interconnected system issues are briefly
addressed later in the paper.
IV. HYPOTHESIZE
MODEL
Conceptually the role of Hypothesize Model in Fig. 3 is
very simple. I n practice, i t is nontrivial to obtainh(x) for some
given structure s(L) as there are a large number of possible
switch-circuit breaker-meter locationcombinations. Careful
logical programming based on network theory is required to
obtain a useful algorithm. This is an important part of static
state estimation, but explicit design techniques are not discussed here. (See [IS],[52], [53].)
The state vector x consists of the N complex node voltages which can be expressed either as magnitude and phase
angle or real and imaginary components. The
choice of coordinate system depends on the application.
The value of R depends on many diverse factors related
to the metering system, the network, and the value
of the
measurements themselves. There is no one set of universal
formulas,buttheactualcomputation
effort withanyapproach will be small.

is the Jacobian matrix. In theory there may

be several X which
satisfy (5.3). However, themultiple-minimaproblemdoes
not appear important in practice and is ignored in all further
discussions.
For an N-node network, (5.3) constitutes a set of 2N-1
nonlinear equations which must be solved for the 2N- 1 unknowns i.The equations arewell behaved, so any of a wide
variety of iterativealgorithmscan
be used. Variousapproaches will be discussed.
Minimization of J ( x ) of (5.2) can be viewed as a sequential unconstrained minimization technique (SUMT) problem
of nonlinear programming [22],
Hence still other types
of algorithms can be found in appropriate textbooks on optimization, such as

[HI.

[SI.

B . Basic W L S Algorithm
One way to solve (5.3) is the iteration

Gi[Xi+l - xi] = H(xi)R-[z - h(xi)],

for i = 0, 1, 2,

(5.4)

where x i is value of X at ith iteration, x 0 an initial guess, and

Gi is a gain matrix. If convergence occurs, (5.3) is solved as
xi+l = X , = ; ,
independent of the value of Gi (provided Gi is of
full rank). Two issues associated with (5.4) are 1) the choice
of gain matrices Ci and 2) the choice of method to solve the
linear equation

Gd=b

(5.5)

where d = x ; + l - x i and b=H(xi)2P1[z-h(xi)].

Linearization with a Taylor series expansion of h ( 3 ) about
xi gives h(;) =h(x;) + H ( x i ) [ ; - x i ]
andsubstitutinginto
(5.3) assuming H ( x i ) = H ( i ) yields:

HT(Zi)R-H(Xi);[ - x i ]

HT(x~)R-[z
- h(xi)]. (5.6)

976

PROCEEDINGS OF TEE IEEE, JULY

1974

does not vary appreciably for the range

of x usually of interest

This implies that

a HT(Xi)R-H(Xi)

PI.

P-8, Q-V Decoupling: Consider the four variables: P: real

power; 8: voltage phase angle; Q: reactive power; V : voltage
is a good choice from a convergence point of view. Z; is called magnitude. For most operating conditions, the P-0 pair and
the information matrix. Note that Gi in (5.7) depends on x i . Q-V pairsare closely coupled withlimited cross coupling.
A direct method to solve (5.5) for d is to use square-root By suitable ordering of the variables, a P-0 and Q-V deby
factorization (Choleskeys method) where the-gain matrix compositioncan be used toreducecomputerstorage
P-V and
G isfactorized as G = @:.\/@ such that 4 G is a lower obtaiiing a block diagonal gain matrix with the
Stability in the sense
triangular matrix (assuming G is positive definite). d i s ob- 0-Q terms in Z equal to zero [4], [a].
tained by forward and backward substitution; i.e., .\/G y = b of (5.9) is maintained.
Tuning: The preceding gain? are based on modifying the
is solved for y and then d @ d = y is solved for d. (See, for
informationmatrix Z. However,thisisnot
necessary. For
example,[59], 1601.)
The dimensions of z and x can be very large. Thus from a example, any positive definite full rank G canbe used
computer implementation point of view, the following facts where the values of G are tuned (adjusted) so that (5.9) is
satisfied and good convergence occurs.
are important: 1) the Jacobian matrix H ( x ) is, very sparse;
I n general, the use of G; =Zi as in the basic algorithm of
2) elements of H ( x ) and h(x) have very similar equations so
the computations used in obtaining h(x) can be reused to (5.8) results in the fewest number of iterations but requires
obtain H ( x ) ; 3) the information matrixZ is sparse (it is roughlythe most number of calculations per iteration and the most
preceding alternativesrequire less
twice as full as H). Because of these facts, direct implemen- computerstorage.The
possible
tation is practical for large dimensional problems, provided storage andless computer time per iteration but make
sparsityprogrammingis
used. Optimalordering of the more iterations.
A true Newtons method algorithm results by choosing
equationimprovesthe effectiveness of thealgorithm [MI,
the gain Gi to be the Hessian matrix d*J(x)/dx* [22]. How[65 I.
For the purpose of this paper the basic WLS algorithm ever, the information matrix Zi can be viewed as an approximation to the Hessian matrix where from an implementation
is defined as:
point of view, the use of Z is superior.
Gs = zi

(5.7)

D . Choice of Method to Solve Linear Equations

Direct methods are not the only
way to solve the linear
equation (5.5). Indirectiterativemethodssuch
as Gaussoptimal .ordering.
Seidel,Gauss-Jordan, etc., also exploit the sparsity of matrices [59].A Gauss-Seidel method will alwaysconverge if
G is positive definite. SuFh indirect methods have low comThis basic algorithm works very well and can be implemented puter storage requirements and can
be used with any gain.
in a straightforward fashion. Provencomputer
codes for
A direct solution using square-root factorization as in the
square-root factorization, etc., are readily available in various basic algorithmrequires
more sophisticatedprogramming
languages [59], [60]. The basic algorithm also provides a and more computer storage. There is no definite answer as to
solid basis for bad data, structural error detection-identifica- which approach requires the least amount of computer time,
tion, and for performance analysis, as will be discussed later. but direct methodsseem to be preferable.
However, other algorithms might be better for some applications because of computer time and storage considerations.
E . Sequential Algorithms
Some of these alternatives are now discussed.
An iterative technique such Gauss-Seidel
as
can be applied
directly to (5.3). This will sequentially try to satisfy each
C. Choice of Gains G
of the 2N-1 equations in (5.3). Such an algorithm will have
The iteration (5.4) will yield an i which satisfies (5.3) for close to the smallest possible storage requirements. However,
any full rank gain G provided convergence occurs. A linear- convergence properties and total computer time may not
be
ized stability analysis [SI, [%] shows that the iteration tends as favorable.
to be stable (i.e., to converge) for anypositive definite Gi such
Debs et al. [21] describe an algorithmwhich can be viewed
that
(roughly) as a combination of an approximation to the inforGi 2 $1;
(5 * 9) mation matrix, a tuning matrix to obtainconvergence, and
a Gauss-Seidel iteration. This algorithm processes each measurement
sequentially. I t has low storagerequirements. I t
so there is a lot of freedom in the choice of Gi. Various apevolved fromthe simplified dynamic model approachto
proaches will be outlined.
be discussed in Section VI-C.
BandwidthReduction: The information matrix Z is bandwidth limited (provided thenodes are appropriately ordered).
Storagerequirementscan
be reduced by using a G with a F. P-8, Q-V Decomposition
The P-8, Q-V decomposition as discussed in Section V-C
smallerbandwidth,obtainedby
prespecifying the desired
bandwidth. Any element I j k outside this band is modified as for simplifying thegain matrix can also be used to completely
follows: G j j = I j j 1 I j k I ; G k k = I k k f 1 I j k 1 ; and Gjk = 0 ; while separate the state estimation problem into two parts wherein
for all other cases G j k = I > I . Stability in the sense of (5.9) is the 0 states are estimatedusing onty the P measurements
(see [SI) and the V-states using just V and Q measuremaintained.
Constant Gain: A constant gain matrix G such as G = Z ( X O ) ments. These separated solutions may be iterated back and
is a veryreasonable way tosavecomputertime
as I(%) forth.
Solve (5.4) with Gi=Zi of (5.7) usingsquare-

SCHWEPPEANDHANDSCHIN:STATICSTATEESTIMATION

911

IN P O W E R SYSTEMS

Thistype of decomposition reduces computerstorage

requirements. However, i t does not actually minimize (5.1)
and its accuracy can degenerate in
cases of very heavy line
loading (power flows).

formally rewritten as
xi+l

- xi

= Wi[z

- h(xi)]

with

Wi = (Zi)-H(xi)Z?.

An approximate WLS solution can conceptually be obtained

by using the iteration

G . Trunsformation of Variables (Line-Only Algorithm)

X+
- xi = Wo[z - h(Xi)].
The following approach, like P-8, Q-V decomposition,
is based on the physical nature of the problem, i.e., the struc- This is not desirable because W Ois a full matrix. However,
ture of h(x) and does not actually minimize (5.1). I t is often many of theelements of W areverysmall
so a practical
called the line-only algorithm because i t was developed to algorithm is
process measurements of line power flows only (see [ 1 1 1-[ 161).
Xi+ - xi = @O[Z - h(x)]
(5.14)
The longitudinal voltage y across a network element is
given by the node voltages x and the corresponding lineflows where
= W Owith small elements set equal to zero [4].
Zas
The advantage of (5.14) is that f t o can be computed offY = dx, 4 .
(5. lo) line and the on-line computer time and storage requirements
are very small. The disadvantageof (5.14) is that there is no
The advantage of the nonlinear transformation (5.10) is the guarantee that theresulting
will be close to the minimum
property that between y and x the following linear relation- WLS value underall conditions. Also, the matrix fro has to be
ship holds: y = Bx, where B isthenode-elementincidence
recomputed after a structure change.
matrix [56]-[58]. The estimation criterion (5.1)is changed to
I . Load- Flow Techniques
J(x) =
- BxITRV:b - Bx]
(5.11)
Consider (5.1) with z partitioned as
when y instead of z is interepreted as measurement. Differentiation of (5.11) yields

Ignoring the termdy/dx in (5.12) yields with (5.10)

[WRY-B]jZ= PR,-g(;,

where the dimension of 11 equals the dimension of x . One way

to estimatex is to solve the equation[lo]
2 =

hl(X).

(5.16)

2).

In the case where z1 corresponds to the bus injections P the

solution of (5.16) corresponds to the load-flow problem for
which many very efficient and fast algorithms have been deWRu-BXi+l = B T R g ( X i , 2).
veloped [56]-[!%].
A common algorithmis the Newton iteraI t is essential that measurements of lines of different length tion which corresponds to (5.8) when E P 1 exists. Ways to use
are weighted appropriately. Therefore, the iteration
theremaining z2 toimprovetheestimate
by WLStype
techniques have been studied [17], [18].
BTDBX+l
P D g ( X i , Z)
(5.13)
The advantage of this approach is the possibility of using
is the
is used where the elements of the diagonal matrixD are given already developed algorithms. A major disadvantage
by D u = 1 T k k l * with Tu the complex admittance of that line difficulty of choosing the subvectorz1 and theeffects of errors
on which the kth line flow measurement is taken. The dis- in z1 [16]. The approach can be especially vulnerable to bad
cussion of Section V-D on solving (5.5) applies equally well data and structural errors. The authors of this paper do not
recommend this approach.
to (5.13).
I t is important to distinguish between a load-flow based
Oneadvantage of (5.13) is that BTDB isindependent
(as discussedhere)
andanon-line
load
of the state x as in the constant gainbasic algorithm of stateestimator
Section V-C (bothalgorithmsrequirerecalculation
of gain flow. In this paper, an on-line load flow is defined as a diagnostic tool used to determine the line flows that will result
matrices after a structure change). Another advantage (not
apparentfrom (5.13)) is thatestimates for the real and from some hypothesized set of bus injection B (or combination of injectionsandvoltagemagnitudes)andnetwork
imaginary components of x can besolved separately. (The
structure F;. Usually, the hypothesized injections correspond
P-B, V-Q decomposition algorithm has a similar property.)
to some predicted future condition and/or the hypothesized
A potential disadvantage of the line-only algorithm is that
there is no guarantee t h a t the resulting i will be close to the structure corresponds to some planned or possible network
switching. On-lineload
flows andstateestimationserve
minimalWLSvalue;asthecriterionisdifferentandthe
dy/dx term in (5.12)is neglected. The restriction on the types different functions. On-line loadflows were in operation severof measurements which can be processed may be a disad- al years before the first stateestimator (see, for example,
vantage; for example,pseudo measurements on load injections [67
cannot be handled. However, the concept can be extended to
J . Spatial Decomposition
include other measurements [31], 1631.From the bad data
point of view, the basic WLS algorithm offers more flexibility.
The natureof the equationh(x) is such that isitreasonable
to spatially decompose (or quantize) z and X into subvectors
H. Weighting Matrices
andthenprocessthemindividuallyto%patiallysweep
AnotherapproximateWLSsolutionisobtained
using around the nodes of the network with a recursive algorithm.
weightingmatrices.
The basic algorithm of (5.8) can be Because of the nonlinearity several such sweeps may be re-

A solution can be obtained with the iteration

I).

978

PROCEEDINGS OF TEE IEEE, JULY 1974

quired. The decomposition may follow geographical borders

of individualutilities,thevoltage
levels withinone power
system, etc. (see [4], [25], [39], [a]).
For a single system this has the advantage
of reducing
the number of equations which have to be solved at any one
point in the computation and
hence reduces computer storage
requirements.
However,
programming
complexity
inis
creased. Computation time advantages or disadvantages are
not known. For aninterconnectedsystemoperation,this
spatial decomposition formsthe basis forinformation exchange between control centers.

A . Tracking Estimator
Consider a time sequence of snapshot scans so that

z(tn) = h[x(fn)]

Uz(tn),

n = 0, 1, 2, .

(6.1)

with E [vz(t,,)vZT(t,l)]
= R. If models are not available for the
time behavior of either the state X(&,) or the measurement
errors vr(t,,),the best estimate of x(tn)will be ;<tm) as computed
from z(t,) using the ideasof Section V.
Assume a new estimate is to be computed every scan so
that the scan time A = t , , - L l equals the estimate update
time T .
K . Performance (Error) Analysis
Theiterativesolutions
of Section V requireaninitial
guess xo of the value of x which minimizes J ( x ) . A reasonable
An estimate may not
be satisfactory because of an inadequate metering system or too much lost data. Therefore, initial guesswhenprocessing
t ( t , ) toget
2(tn) is to use
i t is important to havea performance or error analysis which i(t-1).
If T=A=t,-t,l
issmallcomparedtothestatic
provides a quantitative evaluation of the estimation process. condition -time constant 7 8 , then i ( t , ) -;(t,,->) will be small
Such an error analysis is straightforward when the WLS enough so that only one single iteration of (5.4) is required
criterion of Section V-A is used. Let xtNe denote the true but which leads to the tracking algorithm[ 5 ]
unknown value of X . Equation (5.3) can be written as

H(i)R-'[R(xt,.) - h ( i )

G(L)[i(tn) - i(fn-1)]

+ uz] = 0.

The linearization of h ( x h e ) using a Taylor series expansion

about 3 yields

X,,

so thatthestateestimationerrorcovariancematrix
defined as
2,

A E [ ( i - xtme)(; - true)^]

is given approximately by the inverse information matrix

8 , = I-' = [HT(j3>R1H(i)]-'.

(5.17)

Usually, one is more interested in the error of some function

~(13)
of 2 such as line flows calculated from the states. The
error covariance matrixof

i is

approximately:

H(i(fn-I))R-'[Z(tn)- h ( i ( f n - ~ ) ) ](6.2)
.

All the discussions of Section V on the choice of the gain

matrix G(tn), methodstosolvelinearequations,etc.,
also
apply to (6.2). An obvious generalization of (6.2) is to allow
the number of iterations per scan to varyas needed.
The simplicityof this tracking concept leads to
a straightforwardimplementation. I t isnot necessary to makeany
assumptions on the time structure
of either x(tn) or v,(t,). The
beauty of not needing to model time structure is paid for by
ignoring "information" contained in the past observations.
The preceding concept can be extended to the case where
the static state is not updated each (i.e.,
scanT > A ) . A variety
of adaptive logics can be developed to vary the update time
depending on how fast the system is changing
(i.e., T is a
function of 7s) so that update is done "as needed" [SI, [12].
Such logic can result in sizable computer time savings.

B . Sequential Scan
In thecase of a sequential scan, all the meters are not read
where Y = d y / d x is the Jacobian matrix of the transformation
at the same instantof time (simultaneously read meters with
~ ( ([54]
i ) contains more extensive discussions).
sequential data transmission is still a snapshot). The impor8. is a full symmetric matrix. However, because of spartance of the difference between snapshot and sequential scans
sity and the nature of the equations, the full 8, need not be
depends on the relative values
of the static condition time
computed and stored to get 8,. Use of-square-root factorizaconstant 7 8 , the meter time constant
T M , the scan time A,
tion of Z and sparsity programming makes the on-line cometc. For example, if 7 s >>A or if T J >>A
~
a sequential scan is
putation of 8, quite feasible. One of the advantages of the
effectively equivalent to a snapshot scan. If there is an imbasic algorithm (5.8) is that the extra computer programming
portant difference, a sequentialscancan
be treated as a
effort needed to obtain Z, is relatively small.
snapshot scan where the difference is accounted for by inThe residual vectoris
i A z - i = h ( x ) + v , - h ( ; ) = Wv.
creasing R by some amount [38].
where the residual sensitivity matrix W is given by
Situationswhereitisnotreasonableto
consideraseW = Z - HI;+HTR'.
(5.19) quential scan in terms of some equivalent snapshot scan do
not seem to occur in practice so further discussions are not
Performance (error) analysis equations such as the above given.
also can be applied directly to on-lineload flows to determine
the effect of errors in the hypothesized future injections.
C . Simplified Dynamic Models
Valid mathematical models for the time structure of x(tn)
VI. ESTIMATION:
TIMEVARYING CASE
and u,(t,) are not available. However, one approach is to use
Section V considered "Estimation" (the second stage of the following simplified dynamic models for the state X(tn)
Fig. 3) for the time invariant case, Methods to handle time
and measurement erroru.(t,) :
varyingstates x andmeasurementerrors
u. are now discussed.

979

SCHWEPPE AND HANDSCHIN: STATIC STATE ESTIMATION IN POWER SYSTEMS

Q L

large errors to occurin some lines. As discussedin Section

V-K,the
basic algorithm of (5.8) hastheadvantage
of
E [ ~ ( t n ) ~ ~ ( t m )=
] RLn
(6.4) automatically providing such performance (error) analysis.
Now consider the time-varying case and assume the trackwhere &,,,=l if m - n , Ln=0if m # n ; i.e., both w(t,) and ingalgorithm of (6.2) is used. If the update time is small
uz(tn) areassumedto
bewhite-noise sequences (see [17]enough, and if no major system change has occurred,
(6.2) can
[20]). The assumptions of a dynamic model (6.3) and (6.4) providesatisfactoryperformancewithasimple
gain (even
provide, in amathematical sense, a method to handlesediagonal as in [SI). In such a case, use of one of the more
quential scans withoutviewing them as snapshot scans.
sophisticated gains (such as in the basic algorithm of (5.8))
For (6.3) and (6.4) theKalman-Bucy
filtering theory is a waste of computer time.
(combined with linearization) provides algorithms for calcuOne reasonable overall approach (suggestedin [4]) is to
lating ;(tn) where ;(tn) now explicitly depends on z ( t n ) , have several different algorithms available where the system
z(t,J,
. . . . Implementation of theexact Kalman-Bucy operator (or computer itself) chooses the one best suited a t
equations involves the manipulation of full matrices and the any given time.One of many possible combinations is as
choice of Q is not obvious. However, the computer time re- follows. Code the basic algorithm of (5.8) so that the imporquirements become practical by diagonalizing the appropriate tant performance (error) analysis
is available. For iterative
matrices a t each time step. A search for a value of Q which solutions after a major system change,
use either the basic
makestheoverallalgorithm
work (tuningthe Q matrix), algorithm or one of its variations where the gain is some apcan be done.
proximation to the information matrix. For normal tracking
The simplified state model (6.3) is at best a very crude
operations use a very simple gain.
representation of the actual state behavior while the simpliVII. DETECTION
fied observation noise model (6.4) is definitely wrong as much
of the measurement noise is bias and correlated type.errors.
The estimatorsdiscussed in Sections V and VI to compute
Thus it is the personal opinion of the authors of this paper
i are based on the assumption that no bad data b or structhat the basic concept of using a simplified dynamic model tural errors c are present. The third stageof Fig. 3 is Detecand Kalman-Bucy theory is definitely inferior to the concept tion which tests whether b or c are actually present. This
of combining the tracking ideas of Section VI-A with a suitcan be viewed as a hypothesis testing problem with the two
able algorithm of Section V. However, the resulting algorithm hypotheses:
has been successfully implemented. One version of i t [21] was
XO:no bad data b or structural errors c are present:
discussed in Section V-E for the time-invariant case.
XI: X0 is not true.
D . Choice of Estimation Algorithm
as those of
The following discussions arenotascomplete
Consider first the time-invariantcase of Section V. Assume Sections V and VI because there exists a companion paper
a good initial guess is not available and that the iterations are[9] which considers detection(andidentification)inmuch
to be continueduntil convergenceoccurs. The basic WLS more detailand which uses notationandterminologyas
algorithm of (5.8) has the highest storage requirements,while defined in the present paper. Thediscussion here is restricted
the least storage results either
by using a diagonal gain matrix to the time-invariant system as in Section V. The residuals
(obtained by bandwidth limiting or tuning) or by applying i are defined by i = z - h ( ; ) when i satisfies (5.3).
Gauss-Seidel iteration directly to (5.3). Comparison of comOne way to test the hypothesisX0 is the sum of squared
puter time requirements is complicated by the tradeoff beresiduals or /(;)-test
[SI,[4]. If J ( ; ) = i T F 1 ; ; then
tween time per iteration and number of iterations, but the
accept X.
weighting matrixalgorithm of SectionV-Hcanhavethe
J(;) <
smallest on-line time requirements, as the weighting matrices
>y,
reject Xo.
(7.1)
can be completed off line. One measure of the performance
of an algorithm is the accuracy of the estimate (assuming no The prespecified parameter y is the thresholdof the test.
A second type of testistheindividual
residuals r or
baddata,structuralerror,
or parametererror).The
WLS
r-test using i = D; where D is a diagonal weighting matrix.
algorithms of Sections V-B-V-E all have the same potential
accuracy as they all minimize (5.1). The algorithms of Sec- In this test each component 6 of i is considered individually,
and if
> y forsome k = 1, 2, . . . , thenreject Xo. Two
tions V-F-V-H will tendto
giveless accurateestimates.
different individual residual tests are obtained byusing differHowever, mostpresent-dayapplicationsinvolve
system
security functions and high accuracy is less important than ent weighting matrices D.The rw-test uses D = (diag
The rN-test uses D = (diag Xv)-l. The rN-test is superior
providing a reliable estimate (Le., onewithoutundetected
it is hardertocomputethe
weighting
large errors). Thus the performance
of the detection and iden- to therw-testbut
tification steps of Fig. 3 (to be discussed in Sections VI1 and matrix D.
For both types of tests the choice of the threshold 7 deVIII) are more important than accuracy. Since the detectiontermines P,,the probability of rejecting 3& when X0 is true.
identification logics to be discussed are based on residuals
resultingfrom minimizing (5.1), theyarebestmatched
by Assuming uz is Gaussian (and 3 is small), the distribution of
the basic algorithm of ( 5 . 8 ) or one of the variations of Sec- J ( ; ) and i (when X0 is true) can be evaluated and P, computed (or bounded) for any y. In the case of one (or several)
tions V-C-V-E. The fact that these algorithms are also the
most accurate may become important in later years if more bad data, the distribution of J ( ; ) and can be computed as a
function of the size of the bad data. This allows the evaluaaccuracy-dependent applications are introduced.
Knowledge of theaccuracy of theestimates (or rather tion of the probability of choosing X1(detection) when X1
line flows) can be very important as certain combinations of is true, i.e., the power function of the previously mentioned
tests.
lost data, network switching, and injections may cause very

E [ ~ ( t n ) ~ ~ ( t m=
)]

980

PROCEEDINGS OF THE IEEE, JULY

Computer implementation of any of these detection tests

is trivial except for the computation of D in the rN-test which
corresponds t o doing a n error analysis as discussed in Section

V- K.
VIII. IDENTIFICATION

1974

tion concepts of Section VIII. Since the parameters are effectively constant in time itis reasonable to use a time series of
z(tn), n = 1, 2,
, to get an estimate. Let p1denote the subset of p whose values are to be estimated. The lackof mathematical models for the time structure of x ( t n ) and v,(tn) can
be overcome by defining

If Detection indicates the existence of bad d a t a or structural errors, the logic of Fig. 3 proceeds to Identification
whose functionistodecide
which measurementscontain
bad data and/or what part of the structure is wrong. As in
Section VII, the discussion here isbrief as [9] contains many
more details.
I n general, i t isreasonabletofirsttrytoidentifybad
N
data andif that fails to try tofind structure errors.
J W =
J(tw PI)
(9.2)
Onebasicapproachtoidentification
of bad d a t a is to
-1
employ logical search techniques based on the residuals(rw or
TN). Anotherseeminglydifferent
butactuallyrelatedapover PI. Relatedalgorithmscan
be foundin [?I,[8], and
[62]. One explicit algorithm is to minimize / ( P I ) with respect
proach is to use estimators with nonquadratic error criteria
[a], [23]. The computational requirements of either approach to p1 for fixed i ( t n ) ,use this new value of p1 to compute anew
canincrease thecomputertimerequirementscomparedto
set of i(tn),
etc. (This is a special case of the algorithm of
one WLS estimationby a factor of slightly oyer oneto several [611.)
In the case of interconnectedsystemoperation,
orders of magnitude depending on the nature and location of
i t is
desirable for each own system to havea network equivalent
the bad data and the meter configuration.
The identification of a structural error is in general more
for the outside world. Simulation studies have shown that it
difficult than bad dataidentification. One approach is to con- is possible t h a t identification of the parameters of the networkequivalents(real
power) can be computedfromthe
vert the structural error problem into an equivalent bad
data problem in the parameter p [7]. Brute force search for time series z(tn), n = 1, 2, . . * , of measurements made only
structuralerrorscan
be aidedby
thesystemoperator
.if in the own system [8].
sufficient computer input-output facilities are provided.
Inthispaper,
v,(t.) has beenmodeled as a zeromean
As shown in [9], the effects of bad data areonly noticeable random vector so that the effects of biases @ are incorporated
in the neighborhoodof the bad data (this should also occur in R. An alternative approach (mentioned in Section 111) is
to consider the biases @ to be unknown constants which are to
for the structural errors). I t does not spread over the entire
be estimated. With such an approach, the value
of R is renetwork. Therefore, i t is possible to decompose the network
duced accordingly, while @ is included in PI, the parameters
into a contaminated part and a healthy part. Estimation can continue to update the system state in
a tracking t o be estimated. If biases are viewed asparametererrors
mode for the healthy part, while off-line techniques are used which are not estimated, the use of a large R can be viewed
as a way to compensate for them. Obviously, the detection
t o identify the nature of the problemin the contaminated
tests of Section VI1 will be more sensitive if the biases are
part.
estimated and a smaller R can be used.
ERRORS
IX. PARAMETER
The logic of Fig. 3 ignores the difference v, between the
parameter vector p obtained from system design
data and
the true valuee. Such parameter errors arenow discussed.
Consider
I = h(x, s,

e)

+ u.,

wherep = e

Then a Taylor series linearization

yields

z = h(x, s, p )

+ C,

+ v,.

of h(x, s, 8 ) about

with

e=p

dh

v = uz + - v,.
dP

If the parameter error v, is small, its effect can be partially

modeled by assuming p is error-free and increasing the
covariance matrix R of the observation error v, a corresponding
amount. It is of course very difficult to decide how much R
should be increased butfortunately,theWLSestimation,
detection,andidentification
logicsdescribed in thispaper
tend to be insensitive t o errors in R. A paper by Stuart and
Herget [35] contains. an excellent discussion on t h e effect of
parameter errors with many numerical results (see also [37],
[381,[631).
If there are a few parameters whose values are badly in
error i t may be possible to estimate themusing the identifica-

X. OVERALL SYSTEM CONSIDERATIONS

A . Meter Configuration
In designing the meter configuration one has to guarantee
that each component of the state vector X can be estimated.
This does not require measurements fromall nodes. However,
every nonmeasured node has t o have an adjacent node where
measurements are taken. One typeof minimal cost criterion
is satisfied whenall quantities of a selected node are measured
because communication links tend to be more expensive than
measurement devices. I t is reasonable to first select all important generator and load
buses and then add additional
buses as needed. Meter configurations can be compared using
the error covariance matrix Z, as a criterion [2]. The design
of the meter configuration should also consider bad data and
structuralerrordetectionandidentificationperformance.
I t is also necessary t o consider how lost data affect the degradation of performance. I n this context, the design
of the
corresponding communication system is equally, if not more,
important.
The redundancy q alone is not sufficient to describe the
quality of the metering system. Reichert et al. [X] show how
the trace of 8, can be reduced by a factor of 4 using different
meteringconfigurationshaving
all thesame q . Handschin

SCHWEPPE AND HANDSCHIN: STATIC STATE ESTIMATION IN POWER SYSTEMS

98 1

H. M. Memll and F. C. Schweppe, Bad data suppression in power

IEEETrans.PownApp.Syst.,
systemstaticstateestimation,
vol. PAS-90, pp. 2718-2725, Nov./Dec. 1971.
-, On-line system model error correction, presented at IEEE
Winter Power Meet., New York, N. Y., 1973. Paper C 73 106-2.
T. DeVille and F. C. Schweppe, On line identification of interconB . Implementation
nected networkequivalentsfromoperatingdata,
presented a t
IEEE Summer Power Meet., San Francisco, Calif., 1972. Paper C 72
The design of a static state estimator shouldbe tailored to
464-6.
particular needs of the situation and there is no one best
E. Handschin, F. C. Schweppe, J. Kohlas, andnA. Fiechter, Bad
approach. Designs can range from a basic simple design redata analysis for power system state estimation, t o be presented a t
IEEE Summer Power Meet., Anaheim, Calif., 1974.
quiring relatively little computer capacity to highly sophistiG. W. Stagg, J. F. Dopazo, 0. A. Klitin,andL.
S. VanSlyck,
cated systems. I t is also possible to start with an initial simple
Techniques for the real-time monitoring of power system operadesign which can be added to as needed.
tions, IEEE Trans. P w n App. Syst., vol. PAS-89, pp. 545-555,
Aor. 1970.
The first stage of implementingrequires a considerable
[ l l f T.-F. Douazo. 0. A. Klitin. G. W. Starm. and L. S. VanSlvck.
- . State
amount of off-line studies,hardwareinstallations
for both
Llculatibn of power systems from liGflow measurements, I E E E
Trans. Pourer App. Sysf., vol. PAS-89, pp. 1698-1708, Sept./Oct.
computerandtelemeteringequipment,displaytechniques,
1970.
and software development. An initial implementation should
J. F. Dopazo, 0. A. Klitin, and L. S. VanSlyck, State calculation of
be limited to j u s t data processing and display where the estipower systems from line flow measurements, part 11, IEEE Trans.
Power App. Sysf.,vol. PAS-91. pp. 145-151, Jan./Feb. 1972.
mator serves the operator as an information system and
all
A. M. Sasson, S. T.Ehrmann, P. Lynch,and L. S. Van Slyck.
commandsarestill
given by theoperator.Having
gained
Automatic power system network topology determination, IEEE
Trans. Pown App. Syst.,vol. PAS-92, pp. 610-618, Mar./Apr. 1973.
operating experience,morecomplex
functions can be autoJ. F. Dopazo, S. T. Ehrmann, 0. A. Klitin,and A. M.Sasson,
mated such as security monitoring
[49], economicload disJustification of the AEP real timeload flow project, IEEE Trans.
patch [SO], etc.
POWWApp. Sysf.,V O ~ .PAS-92, pp. 1501-1509, Sept./Oct. 1973.
J. F. Dopazo, 0. A. Klitin, and A. M. Sasson, State estimation for
power systems: Detection and identification of gross measurement
XI. CONCLUSIONS
errors, in Proc. 8th P I C A Conf. (Minneapolis, Minn., June 1973),
On-linestaticstateestimation
provides a systematic
pp. 313-318.
J. F. Dopazo, Conventional load flow not
L. S. VanSlyckand
procedure for building up a reliable and complete data base
suited for real time power system monitoring,n in Proc. 8th P I C A
from raw measurements. I t enables the full exploitation of the
Conf. (Minneapolis, Minn., June 1973), pp. 369-376.
R. E. Larson, W. F. Tinney, and J. Peschon, State estimation in
redundancy of theexpensivemetering-communicationsyspower systems, part I : Theory and feasibility, IEEE Trans. Pcwcr
tems at a cost in software sophistication and computer hardApp. Syst..vol. PAS89, pp. 345-352, Mar. 1970.
ware which can be adapted to meet the needs of each indiR. E. Larson, W. F. Tinney, L. P. Hajdu, and D. S. Piercy, State
estimation in power systems, part 11: Implementation and applicavidualutility.Theconcepthasalready
been successfully
tions, IEEE Trans. Power App. Syst., vol. PAS-89, pp. 355-359,
implemented by several power companies and other installaMar. 1970.
tions are in various stages of completion. At the present time,
A. S. Debs and R.E. Larson, A dynamic estimator for tracking the
state of a power system, IEEETrans. Power App.Syst., vol.
i t seems appropriate to classify static state estimation as a
PAS-89, pp. 1670-1678, Sept./Oct. 1970.
proven technique. Even more effective (in terms of computer
A. S. Debs and W. H.Litzenberger,Experimentalevaluation
of
tracking
state estimation, inProc. 7 f hPICA Conf. (Boston, Mass.,
time-storage) numerical algorithmsmaking full use of the
1971).
particularnature of theunderlyingequationsshould
defiA. S. Debs, R. F. Larson, and L. P. Hajdu, On-line sequential state
nitely appear (consider, by analogy, the growth in load-flow
estimation for power systems, presented at 4th PSCC, Grenoble,
France, 1972. Paper 3.3/7.
technology over the years). Computing time and storage reE. Handschin, Real-time data processing using state estimation in
quirements will provideone basis for comparingdifferent
electric power systems,in Real Time Control of Electric Power
approaches [ 3 6 ] . There is a definite need to establish some
Systems. Amsterdam, The Netherlands: Elsevier, 1972, ch. 2.
J. Kohlas, On bad data suppression in estimation, IEEE Trans.
standard test cases so that such comparisons can be made.
Aufomaf. Contr. (Tech. Notes and Corresp.),
vol. AC-17, pp. 827828, Dec. 1972.
ACKNOWLEDGMENT
E. Handschin and F. D. Galiana, Hierarchial state estimation for
real-time monitoring of electric power systems, in Proc. 8th PICA
A copy of an earlierversion of this paper was sent to many
Conf. (Minneapolis, Minn., June 19731, pp. 304-312.
of the individuals who have contributed to the state estimaK. Reichert, E. Handschin, J. Caneel, and H. Asal, On-line Lastflussberechnung(stateestimation)inelectrischenEnergievelsortion field. Some very helpful suggestions were received and
gungssystemen, Grundlagen, Elcktrotech. 2. Ausg. A , 1973.
incorporated into the final paper.
K. Reichert, J. Cancel,H. A d , and E. Handschin,Die Ueberwachung eines Teils des Schweizerischen Hochspannungsnetzes
xnit State-Estimation, Prozessaufbau and praktische Erfahrungen,
REFERENCES
SEV-Bulkfin (Special Issue), Zurich, Switzerland, 1978.
data processing for
E. Handschinand J. Kohlas,Onstatistical
(Attempts to provide a complete bibliography rarely sucpower systemoperation,in
Proc.4thPSCC
(Grenoble,France,
ceed. Apologies are made to authors
whose work inadvertently
1972). Paper 3.3/5.
was overlooked or misrepresented.)
F. Ariatti, Selection of telemetered quantities foron-line supervision
and control of an electric power system with the aid of a real-time
[l] F. C. Schw;ppe, J. Wildes, and D. Rom, Power system static state
computer, presented a t Operational Research Symp. (Varna, Bulestimation, Power Syst. Eng.Group,M.I.T.Rep.
10, Nov. 1968.
garia, May 1970).
[2] F. C. Schweppe and J. Wildes, Power system static-state estimaF. Ariatti, V. Castagnoli,L.Marzio,and
P. Ricci, Methods for
tion,part I : Exact model, IEEETrans. Pourer App. Syst., vol.
electric power system stateestimation:Comparativeanalysis
of
PAS-89, pp. 120-125, Jan. 1970.
comoutine simulations based uDon neneralized load flow. trackine
[3] F. C. Schweppe and D. B. Rom, Power system static-state estimaand leastsquare methods, in kroc.-4fh PSCC (Grenoble, France,
tion, part 11: Approximate model, IEEE Trans. Pown App. Sysf.,
1972). Paper 3.3/10.
vol. PAS-89, pp. 125-130, Jan. 1970.
F. Ariatti, Answers to questionnaire on on-line power system state
[4] F. C.Schweppe,Powersystemstatic-stateestimation,part
111:
estimation, Document of CIGRE Study Committee 32, Working
Implementation, IEEETrans. Pown App.Sysf., vol. PAS-89,
Group 01, 1973.
pp. 130-135, Jan. 1970.
[5] R. D. Masiello and F. C. Schweppe, A tracking static state estiS. L. Johnson,Analgorithm for state estimationin power sysmator, IEEETrans. Power App.Syst., vol. PAS-90, pp. 1025tems, in Proc. 8 f h P I C A Conf. (Minneapolis, Minn., June, 1973),
1033, Mar./Apr. 1971.
pp. 319-326.

et at. [9] discuss the fact that thelocal redundancy Qk is much

more important than 7 for the data detection-identification
performance.

982
[32] 0. J. M. Smith,Powersystemstateestimation,
ZEEETrans.
Pourer A p p . Syst., vol. PAS89, pp. 363-379, Mar. 1970.
[33] E. J. Wolters, Results of a numerical and a hybrid simulation of a
power system tracking state-estimator with
a n emphasis on linecurrent measurements, in Proc. 4th PSCC(Grenoble, France, 1972).
Paper 3.3/3.
[34] A. 2. Gamm, Power system instantaneous
state estimation as a
problem in nonlinear programming, Electr. Technol. USSR, vol. 3,
pp. 89-105, 1972.
[35] T. A. Stuart and C. J. Herget, A sensitivity analysis of weighted
leastsquares state estimationfor power systems, IEEE Tram.
Power A p p . Syst., qol. PAS-92, pp. 1696-1701, Sept./Oct. 1973.
[36] F. Aschmoneit, &jt- undSpeicherbedarfder
exakten on-line
Lastflussberechnung, Elektrotech. Z. Ausg. A , vol. 94, no. 8 , pp.
499-500, 1973.
[37] B. Poretta and R. S. Dhillon, Performance evaluation of state estimation from line flow measurements on Ontario hydro power systern, ZEEE Trans. Pmer App. Syst., vol. PAS-92, pp. 1702-1712,
Sept./Oct. 1973.
[38] K.A. Clements, 0. J. Denison, and R. J. Ringlee, The effecta of
measurement, nonsimultaneity bias, and parameter uncertainty on
power system state estimation, in Proc. 8th PICA Cmf. (Minneapolis, Minn., June 1973), pp. 327-332.
[39] -,
A multiarea approach to state estimation in power system
networks, presentedat I E E E Summer Power Meet.,San Francisco,
Calif., 1972. Paper C72 465-3.
[40] K. Uemura, State estimation of large scale electric power systems
by decomposition methods, in Proc. 5th IFAC World Coup. (Paris,
France, 1972). Paper 7.5.
[41] J. Bubenko and L. 0. Petterson, State estimation in electric power
systems using filtering techniques, in Proc. 4th PSCC (Grenoble,
France, 1972). Paper 3.3/6.
state estimation,Dept. Elec.
[42] L. 0. Petterson,Powersystem
PowerSystemsEng.,RoyalInst.
Technol.,Stockholm,Sweden,
1973.
[43] J. Svcen, S. A. Fismen, H. H. Fanes, and A. Johannessen, The
online d d - l o o p approachforcontrol
of generation and overall
protection a t Tokke power plants, presented at Int. Conf. Large
High Tension Electric Systems, CIGRE, Paris, France, 1972. Paper
32-06.
[44] P. Goeminne, P. Van Miegroet, and J. Waha, Descarte project,
contribution to theimprovement of security, in transportnetworks,
in Proc. 4thPSCC (Grenoble, France, i972). Paper 3.3/9.
(451 D. F. Koenig and J. P.OMalley,
Stateestimation: Anon-line
luxuryfortheelectric
utilities industry?, presented a t IEEE
Summer Power Meet.,San Francisco, Calif., 1972. Paper C 72 463.8.
[46] H.Kobayashi, S. Ninagawa, K.Tsujiyama, Y. Tamura,and S.
Narita, Dynamic state estimation techniques applied to security
Proc. 4th PSCC (Grenoble,France, 1972).
controlproblems,in
Paper 3.3/2.
[47] F. Galiana and F. Schweppe,Aweatherdependentprobabilistic
model for short term forecasting,presented at IEEE WinterPower
Meet., New York, N. Y., 1972. Paper C72 171-2.
(481 R. L. Moore and F. C. Schweppe, Adaptive coordinated control for
nuclear power plant load changes, in Proc. 8th P I C A Conf. (Min-

PROCEEDINGS OF THE IEEE, JULY

1974

neapolis, Minn., June 1973), pp. 180-186.

[49] F. D.Galiana, F. C. Schweppe, and H. Glavitsch,Synthesis of
securitymonitoringschemes
in power systems,in Proc. ZFAC/
ZFIP Conf. Digital Computer Applicatwns to Prouss Control (Zurich,
Switzerland, Mar. 1974).
[SO] H. Glavitsch, Economic load dispatching and corre?
reschedulingusingon-lineinformation
of thesystemstate,in
Proc. 8th
P I C A Conf. (Minneapolis, Minn., June 1973), pp. 412-420.
[51] E. Handschin, Theorie und Anwendung von Estimationsmethoden
fur dieon-line Ueberwachung undFiihrungelectrischerEnergiesysteme,
Habilitationsschrift,
Eidgenkische Tech.
Hochsch.,
Ziirich, Switzerland, 1973.
[52] R.G. Rice and G. W. Stagg, Applicatio:program design criteria for
the Southern Company control system, in
Proc.8th P I C A Conf.
(Minneapolis, Minn., June 1973).
[53] T. DyLiacco, K. Mamarao, and A. Weiner, Network status analysis
for real time systems,in Proc. 8th P I C A Conf. (Minneapolis, Minn.,
June 1973).
[S4] F. C.Schweppe, UncertainDynamic Syskms. EnglewoodCliffs,
N. J.: Prentice-Hall, 1973.
New York:
[55] D. A. Pierre, OptimizationTheorywithApplications.
Wiley, 1969.
(561 G. W. Stagg andE. H. El-Abiad, Computer Methods in PownSystem
Analysis. NewYork:McGraw-Hill, 1968.
[Sf] J. R. Neuenswander, Modern Power Systems. Sa-anton, Pa.: International Textbook Co., 1971.
[58] 0.I. Elgerd, Electric Energy Systems T h y . New York: McGrawHill, 1971.
[59] G. Forsythe and C. Mohler, Computer Solutions of L i w Algebraic
Systnns. E n g l e w d C M s , N. J.: Prentice-Hall, 1967.
[60] J. H. Wilkinson and C. Reinsch. Linear Algebra (handbook for automaticcomputation, vol. 11), F. L.Bauer et d.,Eds. New York:
Springer, 1971.
[61] R. D. Masiello and F. C. Schweppe,Multi-machineexcitation
stabilization via systemidentification,submitted
toIEEE-PAS
1974 Summer Pow& Meet.
[62] A. S. D e p , Estimation of steady-state power systemmodelparameters, presented at IEEE-PAS 1974 Winter Power Meet., New
York, N. Y., Jan. 1974, Paper T 74,150-9.
[63] K. Srinivasan and Y. Rogichaud, A dynamic estimator for complex
bus voltage determination, presented
a t IEEE-PAS 1974 Winter
Power Meeting, New York, N. Y., Jan.1975. Paper T 74 153-3.
[64] W. F. Tinney and C. E. Hart, Power flow solution by Newtons
method, ZEEE Trans. Power A p p . Syst., vol. PAS-86, pp. 14491460, Nov. 1967.
[65] T. A. Stuart, Optimal ordering of network buses for weighted leastsquares state estimation,presented at IEEE-PAS 1974 Winter
Power Meet., New York, N. Y., Jan. 1974. Ptper C 74 152-5.
[66] R. Yokoyama, Y. Igarashi, and Y. Tamura, On detection and location of changes in power system configuration, presentedat CIGRE
Meet., Study Committee 32, Melbourne, Australi? Mar. 1973.
[67] J. M.Undrill, 0. J. Denison, and D.Hfyward,Dispatchersload
flow for theREMVECdispatchcenter,presented-at
I E E E 1969
SummerPowerMeet.,Dallas,Tex.,June
1969. Paper 69 CP660PWR.