You are on page 1of 7

1042 IEEE Transactions on Power Apparatus and Systems, Vol. PAS-104, No.

5, May 1985
NETWORK OBSERVABILITY: THEORY

*
A. Monticelli Felix F. Wu
Department of Electrical Engineering and Computer Sciences
and the Electronics Research Laboratory
University of California, Berkeley CA 94720

Abstract -A complete theory of network observability is (1) Are there enough real-time measurements to make state estima-
presented. Starting from a fundamental notion of the observability of a tion possible?
network, a number of basic facts relating to network observability, (2) If not, which part or parts of the network whose states can still be
unobservable states, unobservable branches, observable islands, estimated with the available measurements?
relevancy of measurements, etc. are derived. Simple and efficient algo-
rithms can be developed based on these basic facts to (i) test network (3) How to estimate the states of these observable islands?
observability, (ii) identify observable islands and (iii) place measure- (4) How to select additional pseudo-measurements to be included in
ments for observability. the measurement set to make state estimation possible?
(5) How to guarantee that the inclusion of the additional pseudo-
measurements will not contaminate the result of the state estima-
I. INTRODUCTION tion?
State estimation processes a set of redundant measurements to
estimate the state of the power system. The result of state estimation The analysis which lead to the answers to these questions may be called
forms the basis for all real-time security analysis functions in a power observability analysis. The analysis includes observability test,
control center [1]. There are three types of real-time measurements. identification of observability islands, and measurement plicenient. It
(i) The analog measurements that include bus voltage magnitudes, real should be performed prior to the state estimation. In practice most of
and reactive power injections, and real and reactive power flows. (ii) the time the system is observable. The observability analysis becomes
The logic measurements which consist of the status of switches and valuable only in those rare situations when the system becomes unob-
breakers. (iii) The pseudo-measurements that may include forecasted servable.
bus loads and generations and zero-injections in passive nodes. Analog In this paper a complete theory of network observability is
and logic measurements are telemetered to the control center. Logic developed, which provides a theoretical foundation for the answers to
measurements are used in Topology Processor to determine the system all the questions raised above. A number of basic facts are derived.
configuration. The State Estimator uses a set of analog measurements, Based on these facts, algorithms can be developed for
along with the system configuration supplied by the topology processor,
network parameters such as line impedances, and perhaps some * testing observability
pseudo-measurements as its input. If the set of measurements is * identification of observable islands
sufficient in number an well-distributed geographically, the state estima-
* measurement placement for observability.
tor will give an estimate of the system state. When there is enough
redundancy in measurements, the state estimator will be able to process The design and testing of the algorithm, including computational
bad data [2]. considerations, are presented separately [3]. The algorithms are charac-
In the design stage, the following questions concerning the meas- terized by
urement set arise naturally.
(1) Are there sufficient measurements to make state estimation possi- * being extremely simple
ble? * using subroutines already in a state estimation program
(2) If not, where additional meters should be placed so that state esti- * incurring very little extra computation.
mation is possible?
Some algorithms for testing observability based on heuristic
If the set of measurements is sufficient to make state estimation possi- approaches were proposed earlier [4-71. Clerments, Krumpholz, Davis
ble, we say the network is observable. Observability depends on the in a series of papers [8-11] proposed a graph-theoretic foundation for
number of measurements available and their geographic distribution. network observability. The algorithms that have resulted from their
The first question raised here is concerned with the test of theory are combinatoric. These algorithms seem to be computationally
observability. The second question is meter or measurement placement complex.
for observability.
The organization of this paper is as follows. The state estimation
Usually a system is designed to be observable for most operating models are reviewed in Section II. A fundamental definition of net-
conditions. Temporary unobservability may still occur due to unantici- work observability is introduced in Section'III. Equivalent statements
pated network topology changes or failures in the telecommunication of network observability are then derived. Section IV discusses how to
systems. The following questions emerge naturally in conjunction with test observability. Section V deals with the questions of how to identify
state estimation in system operation. the observable islands and how to estimate the states of the observable
islands. The measurement placement problem is studied in Section VI.
A mathematical format (theorems, proofs) is adopted for the develop-
ment of the theory in this paper for the purpose of maintaining rigor
On leave from Departamento de Engenharia Eletrica, UNICAMP, Campinas, S. and precision. Physical interpretations of the results are provided
P., Brazil.
throughout.

II. STATE ESTIMATION MODELS


84 SMI 581-5 A paper recommended and approved In this section we review the models of standard WLS state esti-
by the IEEE Power System Engineer-ing Committee of- mation [12] and the model-decoupled state estimation [13]. Similar to
the IEEE Power Engineering Society for presentation Clements et al. [8-11], we use the linearized state estimation model for
at the IEEE/PES 1984 Summer Meeting, Seattle, the network observability theory.
Washington, July 15 20, 1984. Manuscript submit-
-

ted February 2, 1984; made available for printing 1. WLS State Estimator
June 6, 1984. The non-linear equations relating the measurements and the state
vector are

z = h(x) + w (1)
0018-9510/85/0005-1042$01.00©1985 IEEE
1043
where z is the (mxl) measurement vector, h(-) is the (mxl) vector of {i= 0,k; 0 k = voltage angle at bus k.
non-linear functions, x is the (2nxl) true state vector, w is the (mxl)
measurement error vector, m is the number of measurements, and n is Ui = Qkim/ Vk; Qkcm
reactive power flow from bus k to m.
=
the number of buses.
The estimate of the unknown state vector x is designated byx and Ki = Qk/ Vk; Qk = reactive power injection into bus k.
is obtained by minimizing the weighted least squares function E, = Vk; Vk = voltage magnitude at bus k.
J(x) = [z-h(x)] TW[z-h(x)] (2) The state vector x is given by

where W is a diagonal (mxm) matrix whose elements are the measure- x= (O,V) (12)
ment weighting factors.
The condition for optimality is that the gradient of J vanishes at where V is is a (nxl) vector whose elements are the bus voltage mag-
the optimal solution x i.e., nitudes, and 0 is the (nxl) vector whose elements are the bus voltage
angles, including the angular reference (or angular references).
H T (x)W[z-h(x)] = 0 (3) The Jacobian matrix is

where the Jacobian matrix -HPQ (x) is


HpQ (x) = h (4) HpQ - ah l (13)
Ox
As for computing the estimate x the method that has received where
ax H
Qr QV

wide acceptance is the iterative method which computes the corrections


Oxk at each iteration by solving eq. (5). Hpg = ahp/aO,Hpv = ahp/8V,HQO = ahQ/IO,
GPQ(Xk)8Xk = HJQ(Xk)W(Z-h(Xk)) (5) and HQV OhQ/QV.
=
Xk+± = Xk + 8Xk (6)
By applying the decoupling principle to the matrix HpQ, we
obtain the decoupled gain matrix
for k = 0,1,2,... until appropriate convergence is attained. Here the Ge
gain matrix GpQ is
(14)
GPQ = OG
GpQ (Xk) = HJQ (Xk) WHpQ (Xk) (7) II
GV

In general, sparsity-oriented LDU decomposition is used in solving eq. where


(5).
The gain matrix GpQ in (7) can be written as Go = HTWPHpH (15)

GpQ = T W HPQ=-m [xhik


HPQ W [ Oh1i
i Gv= HQvWQHQV (16)

Here the matrices Wp and WQ contain the weighting factors


corresponding to the measurements zp and zQ, respectively.
where is the ith row of the Jacobian matrix HpQ, and w, is the Similar to (7) the matrices Go and G v may be formed by process-
weighting factor associated with measured zi. The above equation sug- ing the measurement one at a time.
gests that in forming the gain matrix one can process the measurement
one at a time. 3. Linearized (DC) State Estimator
In this subsection we derive a linearized (DC) state estimator that
2. Decoupling has the same features as the DC load flow (BO = P)
Equation (1) can be rewritten as The same approximations used in obtaining the matrix B' of the
fast-decoupled load flow (or the matrix B of the DC load flow) can be
Zp = hp(x) + wp (8) used to simplify the Jacobian matrix Hpq as well as the gain matrix GO:
a) Flat voltage profile, i.e., V = 1 p.u. and 0 = 0.
zQ = hQ(x) + WQ (9) b) Line susceptances approximated by l/x, where x is the line reac-
tance.
where zp is the (mpxl) vector of the "real" measurements (real power Let us call G° the resulting gain matrix.
flows, injections, and voltage angles), and zQ is the (mQxl) vector of The DC state estimator computes the bus voltage angles by solv-
the "reactive" measurements (reactive power flows, injections, and vol- ing the equation:
tage magnitudes):
Go0=HPh Wpzp (17)
Zp= (T,I,O) (11)
which basically corresponds to performing the first 0-iteration of the
fast model-decoupled state estimator [131.
zQ = (U,K,E) The linearized state estimator (17) is equivalent to a linear least
square problem [14, pp. 208-249]. Let us write H = WJ/2Hp0 and
Here the components of vectors T, I,1, U, K, and E are respectively: C = Wj/2zp. Then (17) is equivalent to the solution of the following
least square problem: determining a vector 0 that minimizes the sum
T,= Pkin' Vk;Pk,, = real power flow from bus k to m. of squares of the residual vector.
I= Pk/ Vk;Pk = real power injection into bus k.
r= C- HO (18)
1044
Equation (17) becomes Intuitively we call a network observable if any flow in the network
can be observed by some sort of indication in the set of measurements.
In other words, whenever there is any nonzero flow in the network, at
HTHO = HTC (19) least one of the measurements should read nonzero. This is equivalent
to saying that a network is observable if, whenever all measurements
The network observability theory we are going to present is based are equal to zero implies that all flows are zero. When a network is not
on the linearized state estimator (17). We may make the same approxi- observable, it means that it is possible to have all measurements zero,
mations to HQV as well as the gain matrix Gv in the reactive power yet there still are nonzero flows in the network. In such a case, those
miodel to obtained similarly a linearized Q-V state estimator model. branches having nonzero flows will be called unobservable branches.
These definitions will be made formally below.
Definitions. A network is said to be observable if for all 0 such that
III. NETWORK OBSERVABILITY H0= 0, AT0 = 0. Any state 0 for which HO = 0,AT X . 0, is
We shall use the linearized state estimator model to develop a called an unobservable state. For an unobservable state 0, let
theory of network observability. For ease of presentation, the real 8 = ATo@, if 87; 0, then we call the corresponding branch an
power model is used. The development is very similar for the reactive unobservable branch.
power model, however, whenever the difference becomes important, it
will be pointed out. It should also be pointed out here that in the fol- Theorem 1 below follows immediately from the definitions. The
lowing development, the vector 0 represents a true state vector. theorem supplies us with two other equivalent statements of observabil-
For network observability, we are concerned with the power flows ity.
in the network and the measurements made on the network. Let us Theorem 1. Assume that there is no voltage measurement, then the
first elaborate on the flows and the measurements. Given is a state following statements are equivalent.
vector 0, the power flow through the branch connecting buses k and m
is equal to i (ok - m). For observability, we will only be concerned (i) The network is observable.
Xi (ii) Let H be obtained from H by deleting any column, then H is of
with the fact whether the flow is zero or not, not the actual numerical full rank.
value of the flow when it is nonzero. Therefore for simplicity let us set (iii) The triangular factorization reduces the gain matrix G = H TH
xi= 1 and call the "flow" to be the same as the angle difference into the following form:
8i = Ck-Om. Using the (unreduced) network incidence matrix A the
set of "flows," 8, can be written as
8 = ATE (20)

Thus,
k m (26)
81= 1 0. (21)

On the other hand, given the state vector 0, the set of measurements is
written as
where the shaded area corresponds to possibly nonzero elements.
0= HO (22)
Proof. (i) <=# (ii). Because ATo = 0 if and only if 0 = a 1 where
For the real power model there are two types of measurements 1 = (1,...1) and a is any real number, we have (i) <=> (a)HO = 0
(i) line flow. If measurement i is the line flow from bus k to bus m, iff 0 = a 1. Now we show (a) <= (ii).
then
(a) => (ii). Let H be obtained from H by deleting the k-th column h.
k m Suppose HO = 0. Let 0 = (-,0). We thus must have 0 = 0, which
says H is of full rank.
h. -h. 0 (23)
fii) => (a). Since the column sum of H is always zero, i.e.,
HI = -h. Thus (HjTjj)-1fjTi= -1. Now suppose HO = 0 or
(ii) injection. If measurement i is the injection at bus k, where there HO+hOk = 0. Then 0 =-(H HT)-l HT0k = lOk.
are branches connecting bus k to buses m, n, l, then
(ii) <=# (iii). Let H = (H,h), we have
_ k rm n
(i= 1-hz -hm -h 0 (24) G= HTH= |THH jThi
[hTR hThJ

where S = hm + hn + hi. Note that HT-


-
T is nonsingular if and only if the triangular factorization
reduces it to a triangular matrix.
For the reactive power model there is an additional type of meas- U
urement, which is the voltage-magnitude measurement. The
corresponding one in the real power model would be the voltage-angle Comments
measurement. Even though the voltage angle measurement is not 1) Statement (ii) of Theorem 1 relates the intuitive concept of
available in real life, it is still helpful to include this in the considera- observability to the solvability (existence of a unique solution) of
tion. Later in the paper we will show that a key result in our theory of the state estimation problem. As a matter of fact, the solvability
network observability may be interpreted as adding 0-pseudo measure- of state estimation has been used as the definition of observability
ments. in the literature. To see the relation, note that H is of full rank if
(iii) voltage. If measurement i is the voltage-angle at bus i, then and only if (HTf) is nonsingular, this is exactly the condition
i that is needed for the state estimator (19) to have a unique solu-
tion. Thus Theorem 1 implies that a network is observable if and
only if the state estimation problem can be solved with a unique
I
I (25) solution.
1045
2) Statement (iii) of Theorem 1 provides a numerically stable
method for testing observability based on the triangular factoriza-
tion of the gain matrix. As a matter of fact the triangular
factorization of the gain matrix will also provide the information
about the observable islands as will be shown in the subsequent
sections. (29)
3) When there are voltage measurements, statement (ii) will be
replaced by: H is of full rank.

IV. DETERMINATION OF UNOBSERVABLE STATES


When the network is not observable, we proceed to find an unob-
servable state, which is a solution of HO = 0. The solution of HO = 0
is sensitive to the numerical values of the elements of H, as well as the
errors introduced during the solution process [14, pp. 317-318]. This is Proof. Let H = [Hlh2H3] where h2 is a column.
highly undesirable. Lemma 1 below provides an alternative to solving
HO = 0.
Lemma 1. HO = 0 if and only if (HTH)O = 0 ITHH H/Th2 HITH3
Proof. (=-) Premultipl, HO = 0 by HT. G = HTH = h2TH1 h2Th2 h2/H3 (30)
(<== ) Premultiply (H H)O = 0 by 0T results in|11011= 0, which H/TH1 H3Th2 H3TH3
implies HO = 0.
The triangular factorization reduces G to
When the given network is not observable, the triangular factori-
zation with complete pivoting reduces G = H TH to the following form
[14, pp. 126-127]:

(9a 0. (27)

where
For any arbitrary O,, for example, Ob = (0,1,2,...) T, solving the upper
half of the above equation (27) yields a Oa) then (0a,O,,) is an unob-
p = h2Th2-h2TH1(H/THI)-1H/Th2 (31)
servable state.
An alternative way to obtain the same unobservable state (Oa,Ob) q= h2TH3 - h2TH1 (HlTH)-'HTH3 (32)
is (i) replacing the diagonal elements of the lower right matrix by 1's,
and (ii) replacing the corresponding right-hand side of (27) by
(0,1,...) T, (iii) solving the resulting equation The situation here is that we have p = 0. The following claim is first
proved.
HIH1l nonsingular and
p =
h2Th2-h2THI(H/TH1)-lHITh2= 0
ea 0 if and only if
(28) columns of H1 are linearly independent and
h2 = H1a for some vector a.

By substituting h2 = H1a into the expression for p we obtain 0.


Note that (28) is identical to the state estimation equation (19) with
the pseudo-measurements of the voltage angles at buses corresponding
Suppose the columns of (HI h2) are linearly independent, then
the matrix
to Ob present and all other measurements set to zero. This important
observation will be used again.
Triangular factorization with complete pivoting involves permuta-
[H/TH H h2T
(33)
tion of rows and columns of G and the corresponding reordering of the |TH h/h2
vector 0. Since for the solution of large systems by triangular factoriza-
tion, the ordering of the matrix is done based on sparsity considera- is nonsingular. The determinant of matrix (33) is equal to the
tions, it is desirable not to have two different orderings. Theorem 2 product det(H/THI)det (p). Since (H17) is nonsingular, p #: 0.
below shows that the reordering for complete pivoting above is actually We reach a contradiction. Therefore the claim is proved. Using
not necessary for observability test for obtaining an unobservable state. the above result that p 0 implies h2 Hla, we substitute
= =
This result greatly simplifies computation. h = Hla into (32) and obtain q = 0.
Theorem 2. In the triangular factorization of the gain matrix G, if a 2
zero pivot is encountered, then the remaining row and column are all
zeros, i.e., G is reduced to the form:
1046
Comment. The implication of Theorem 2 is that for any given ordering Theorem 3 below asserts that as far as the subnetworks a, /3, and
of the gain matrix G whenever a zero pivot is encountered in the y are concerned, the given unobservable state 0 is observable, hence
process of triangular factorization, the corresponding 0 belongs to 0b the subnetworks are candidates for observable islands. The reaon that
and it can be assigned an arbitrary value in obtaining an unobservable they are merely candidates is because the given 0 is only one unobserv-
state. Or equivalently, one adds a 0-pseduo-measurement for that
able state, there may be other states that make these subnetworks
node. In other words, the zero pivot in G is replaced by a 1, and the
unobservable. An observability test on each subnetwork is still
corresponding right-hand side 0 is replaced by the value assigned to the required.
0-pseudo-measurement. The triangular factorization may then con- Theorem 3. 0a is not an unobservable state for the subnetwork ax with
tinue. A network is observable if and only if there is only one zero measurements Ha, Similarly for 0 and 0,.
pivot, which necessarily happens at the end, whereas when the network
is not observable one encounters more than one zero pivots in the tri- Proof. Equations (34) and (35) give us: *a is a solution of
angular factorization of G.
HaOa = 0 and AaT0a = 0.
Subnetworks a,43, and y are candidates for observable islands.
V. IDENTIFICATION OF OBSERVABLE ISLANDS The measurement sets that are relevant at this time in the determina-
When the network is not observable, we would like to know tion of observable islands are Ha,Hg, and H.. Subsets of them will
eventually be used in the state estimation of the observable islands.
which part or parts of the network whose states can be estimated by
processing the available measurements. These subnetworks will be Therefore the measurements in Ha,y are irrelevant as far as state esti-
called observable islands. Two questions are answered in this section: mation of the observable islands is concerned. We shall call these
measurements in Hasy irrelevant measurements. They are discarded for
(a) How to identify the observable islands. further analysis of observability.
(b) How to estimate the states of the observable islands. Theorem 3 suggests an iterative scheme to identify observable
We start from a given unobservable state 0 obtained from the islands. Since each subnetwork a identified in (35) is a candidate for
solution of GO = 0. Let us arrange 0 so that the components having observable island, we may test its observability with respect to the
identical values are grouped together. For example, suppose there are measurement set H,. If the answer is yes, we find an observable
three groups of values Oa, HP,0. in 0, we have 0 = (0a,"p,0y) where island. If not, we proceed to find an unobservable state in subnetwork
O= (Oa,O,a,..Ga)5 etc. Subnetwork a consists of nodes in Ga a and further decompose it by discarding unobservable branches.
together with the branches connecting them. Similarly for subnetworks After observable islands are identified, we use the relevant meas-
,8 and y. urements to estimate the states of the observable islands. Note that in
Let us further group together (i) line flow measurements for lines the process of identifying observable islands, we have introduced a set
in the same subnetwork, and (ii) injection measurements for which all of 0-pseudo-measurements in (28). A question arises as to whether
the nodes connected to the injection node belong to the same subnet- these 0-speudo-measurements would affect the final state estimation
work. It can easily be shown that it is impossible to have line flow results. Theorem 4 below guarantees that they would not.
measurements for lines connecting different subnetworks, for if this
were true the 0 values for these two components would have to be the
same. Therefore, the remaining measurements are (iii) those injections Theorem 4. Consider the state estimation model
for which the nodes connected to it belong to different subnetworks. (36)
According to this grouping the matrix H becomes (rearrange rows): . C =HO +r
Suppose that the measurement set consists of the 1-pseudo-
r- 0
measurements O' introduced in (28), and all other measurements equal
to zero. Then the residuals r = 0.
ytt Proof. We can renumber the network nodes and the measurements
such that
1. (34)

I'
1
0 regul ar
measurements
m-Q

2
(37)
Os m-Q+l
. 0-pseudo
measurements
The rows of Ha, Hp, and H. correspond to the measurements (i) and m
(ii) of subnetworks a, /3, and y respectively, and the rows of Hap,P
correspond to the measurements (iii). Also if we group branches in the
same network together, AT becomes:
I
voltage angle
at other nodes
Oat n-Q.

[S 0 0= (38)
n-)+l voltage angle at
(35) as nodes with 0-pseudo
n
measurements

Accordingly, the gain matrix and the Jacobian matrix can be partitioned
as follows:

where A,, Ap, and A., correspond to the branches in the subnetworks
a, /, and y, respectively, and AaB correspond to the branches con-
necting different components. In other words, the rows of Aa#v,,
correspond to the set of unobservable branches.
1047
* would the additional measurements contaminate the state estima-
.
tion of the observable islands?
Theorem 5 below can be proved mathematically. Due to space
limitation the proof is omitted here.
Hr H*a
Theorem 5. If a minimal set of additional non-redundant (pseudo)-
m-Q measurements is so selected that they make the network barely observ-
able, then the estimated states of the already observable islands will not
H= m-Q+l be affected by these pseudo-measurements.
It is clear that the candidates for these additional measurements
I. are (i) injection pseudo-measurements for which the nodes it is con-
m nected to belong to different observable islands, and (ii) line flow
pseudo-measurements for lines connecting different observable islands.
The second type of pseudo-measurements are seldom available. We
shall concentrate on the first type.
This set of pseudo-measurements may be obtained sequentially by
HTH
rr r HTrtt adding a pseudo-measurement from the candidate list one at a time.
The additional pseudo-measurement will make some unobservable
I branches observable, thus coalescing observable islands into larger
ones. Computationally this selection process can be accomplished by a
0) scheme which adds a pseudo-measurement from the candidate list and
then recalculates the observable islands at each iteration until the whole
network becomes one observable island.
.HTHIir
VII. CONCLUSION
A complete theory of network observability is presented. The
major points are summarized below:
where II is the unit matrix of order 1. Performing Gaussian elimina- * A network is said to be observable if, whenever there is any
tion in (40) we get nonzero flow in the network, at least one of the measurement
shuld read nonzero. In other words, a network is observable
when, all measurements are zero implies that all flows are zero.
* A network is observable if and only if the state estimation can be
solved with a unique solution.
* An observability test can be performed based on the triangular
factorization of the gain matrix. It goes as follows. In carrying
out triangular factorization of the gain matrix G. whenever a zero
pivot is encountered, we add a 0-pseudo-measurement for the
n-Q corresponding node, the effect of that is that the 0 is replaced by a
n-k+l I in the triangular factor. The network is observable if only one
iero pivot is encountered in the process and not observable if
more than one zero pivots are encountered.
n * By setting the 0-pseudo-measurements of the nodes of zero pivots
to different values and all the other measurements to zero, the
solution of the state estimation equations yields an unobservable
where state of the system.
* Consider the subnetwork formed by the set of nodes having
A = HITHa - HaTHr (HrTHrY)71HrJRa =0 identical value in the unobservable state, together with the
The estimate is given by
branches connecting them. For the measurement set, consider
the line flow measurements for which all the nodes connected to
Go = HTC the injection node belong to the same subnetwork. This subnet-
work along with the measurement set is a candidate for an observ-
Introducing (37)-(42) into (43), and inverting G-1, we get: able island.
O= (HjTHr-HHaOS (44) * Branches between two candidates of observable islands are unob-
servable branches.
#4 =es (45) * Those injection measurements for which there is an unobservable
branch connected to the injection, are irrelevant as far as observa-
and from (36) we get: bility of the network is concerned and may be discarded for the
purpose of identifying observable islands.
rt = [ Hr(HrTHr)-IHrTHa + HJIs
- (46)
* Using the relevant measures and the 0-pseudo-measurements, the
states of the observable islands may be estimated. The 0-pseudo-
rS =0 measurements do not affect the state estimation results on line
flows.
Now consider * If a minimal set of additional non-redundant (pseudo) -

lkrll = (ir) (rr)


T (47) measurements is so selected that they make the network barely
observable, then the esitmated states of the already observable
0 islands will not be affected by these pseudo-measurements.
By substituting (42), (46) into (47), we obtainlkrll = 0. Hence r =

VI. MEASUREMENT PLACEMENT The actual design and testing of efficient algorithms based on

Given an unobservable network, we would like to know these results are presented in a companion paper [3].
* what is a minimal set of additional measurements whose inclusion
will make the network observable?
1048
ACKNOWLEDGEMENTS [8] G. R. Krumpholz, K. A. Clements, and P. W. Davis, "Power Sys-
Research sponsored by the Electric Power Research Institute, tem Observability: A Practical Algorithm Using Network Topol-
Power System Planning and Operations Program, under contract RP ogy," IEEE Trans. Power Apparatus and Systems, vol. 99, pp.
1999-6. 1534-1542, July/Aug. 1980.
[9] K. A. Clements, G. R. Krumpholz and P. W. Davis, "Power Sys-
tem State Estimation Residual Analysis: An Algorithm Using Net-
REFERENCES work Topology," IEEE Trans. Power Apparatus and Systems, vol.
[1] T. E. Dy Liacco, "State Estimation in Control Centers," Electric 100, no. 4, pp. 1779-1787, April 1981.
Power and Energy Systems, vol. 5, no. 4, pp 218-221. [10] K. A. Clements, G. R. Krumpholz and P. W. Davis, "Power Sys-
[21 A. Garcia, A. Monticelli, and P. Abreu, "Fast Decoupled State tem State Estimation with Measurement Deficiency: An Algo-
Esimtation and Bad Data Processing," IEEE Trans. Power rithm that Determines the Maximal Observable Subnetwork,"
Apparatus and Systems, vol. PAS-98, no. 5, Sept./Oct. 1979, pp. IEEE Trans. Power Apparatus and Systems, vol. 101, no. 9, pp.
1645-1652. 3044-3052, September 1982.
[3] A. Monticelli and F. F. Wu, "Network Observability: Identification [11] K. A. Clements, G. R. Krumpholz and P. W. Davis, "Power Sys-
of Observable Islands and Measurement Placement," submitted tem State Estimation with Measurement Deficiency: An
for presentation at IEEE PES Summer Meeting, Seattle, WA, Observability/Measurement Placement Algorithm," IEEE Trans.
1984. Power Apparatus and Systems, vol. PAS-102, pp. 2012-2020, July
[4] K. A. Clements and B. F. Wollenberg, "An Algorithm for Obser- 1983.
vability Determination in Power System State Estimation," Paper [12] F. C. Schweppe, J. Wildes, and D. P. Rom, "Power System Static
A75 447-3, IEEE PES Summer Power Meeting, San Francisco, State Estimation," Parts I, II, and III, IEEE Trans. Power
CA, 1975.
Apparatus and Systems, vol. 89, pp. 120-135, January 1970.
[5] J. S. Horton and R. D. Masiello, "On-Line Decoupled Observabil- [13] A. Monticelli and A. Garcia, "Reliable Bad Data Processing for
ity Processing," PICA Proc., Toronto, Ontario, pp. 420-426. Real-Time State Estimation," IEEE Trans. PAS, voL. 102, no. 5, pp.
[6] J. J. Allemong, G. D. Irisarri, and A. M. Sasson, "An Examina- 1126-1139, May 1983.
tion of Solvability for State Estimation Algorithms," Paper A80 [14] G. W. Stewart, Introduction to Matrix Computations, Academic
008-3, IEEE PES Winter Meeting, New York, NY, 1980. Press, 1973.
[7] E. Handschin and C. Bongers, "Theoretical and Practical Con-
siderations in the Design of State Estimators for Electric Power
Systems," in Computerized Operation of Power Systems, S. C.
Savulescu, Editor, Elsevier, pp. 104-136, 1972.

Discussion A. Monticelli and F. F. Wu: We thank Prof. Alvarado for his comments.
F. L. Alvarado (University of Wisconsin, Madison, WI): This work is We do not see much relation between our paper and Ref. [Al]. Ref. [Al]
important, comprehensive and well presented and is bound to become has two parts in it. The first part is an attempt to derive network obser-
an important reference. A few years ago, Fetzer and Anderson [Al] vability condition (H full rank) from the observability condition in Linear
presented a paper on dynamic observability. The static case of [Al] System Theory for linear dynamical systems. The derivation in [Al] is
resulted in not only topological but quantitative measures of degree of not correct. The steady-state of a dynamical system x = Ax + f is reach-
observability of a system with a given set of measurements, and was based ed when x = 0, which does not imply A = 0, at the starting point in
on a HHT matrix much like the gain matrix G. Can the authors com- [Al]. The second part suggests the use of "well-conditioning" of the
ment on the relationship (if any) between the two papers? gain matrix as the interior for selecting the "best" set of measurements.
The relation between this criterion and the objective of state estimation
is questionable. There are other proposed criteria for measurement place-
REFERENCE ment cited in Refs. [10-15] in the companion paper [3].
[Al] E. E. Fetzer and P. M. Anderson, "Observability in the State
Estimation of Power Systems," IEEE Trans. on Power Apparatus and Manuscript received December 26, 1984.
Systems, vol. PAS-94, no. 6, Nov./Dec. 1975.
Manuscript received July 26, 1984.

You might also like