Professional Documents
Culture Documents
Abstract—Effective assessment of the dynamic performance of actions a few seconds ahead or display a short list of possible
the power system requires wide-area information from properly preventive measures on the operator’s screen [8], [9].
distributed phasor measurement units (PMUs). However, to maxi- Intuitively, comprehensive monitoring of the grid behavior
mize the information content of the captured signals, the sensors
need to be located appropriately, with due account given to the over time requires instrumentation of all its topological nodes.
structural properties underlying the given system. In this paper, In practice, in addition to being a sufficient but not necessary
two numerical algorithms are proposed to achieve this goal. They condition, this naive view is countered by the curse of di-
aim to maximize the overall sensor response while minimizing the mensionality [10], considering that a standalone medium-size
correlation among sensor outputs so as to minimize the redundant system requires between 500 and 1000 busses for an ade-
information provided by multiple sensors. The sensor responses of
interest are the bus voltage magnitude, and the angle and frequency quate representation. There is also a cost issue: many meters
coherency indexes, which are estimated by means of a statistical means high capital costs in terms of equipment, maintenance
sampling of power system response signals from a transient-sta- and an information management system. Furthermore, any
bility program. Through the “successive addition” scheme, one of single phasor measurement unit (PMU) addition to an already
these algorithms easily incorporates mandatory locations such as well-equipped grid tends to compound the complexity of the
tie-line busses and large generator step-up transformers. The pro-
posed approaches are first illustrated on the Hydro-Québec trans- telecommunication infrastructure, with all the resulting relia-
mission grid and then on a 9-area/67-bus/23-machine test network bility issues. In this context, how can we minimize the number
designed with well-defined geographical boundaries and pre-spec- of PMUs while maintaining “good” monitoring of the system
ified weak interties between electrically coherent areas. dynamic performance? When the state estimation concept was
Index Terms—Digital fault recorders, optimal sensor location, first introduced at the end of the 1960s, Schweppe et al. [1]
phasor measurement unit (PMU), state estimation, wide-area con- immediately emphasized the importance of dispersing the me-
trol, wide-area monitoring. tering devices optimally throughout the grid so as to maximize
the achievable accuracy of the steady-state estimate from a
I. INTRODUCTION least-mean-squares filtering of the available measurements.
A direct synchronous measurement of the complete state
cuit with and without the loss of single or double lines, load dition, is the power spectral density [21] of while
and generation steps, etc.). These simulations were carried out is the cross spectrum of and . Interest-
for various operating scenarios, selected in a way very similar to ingly, the conditional entropy is an effective measure of the
the statistical sampling approach of [17]. Careful analysis of the electrical distance between busses and [20], [22]. It suffices
signal responses allows the candidate busses to be compared in to note that the complex coherency coefficient
a statistical framework according to the relevance and distinc-
tive nature of the information recorded at its location (i.e., over (3)
a significant number of operating scenarios and plausible dis-
turbances). the logarithm of which is integrated between 0 and in order
to obtain the conditional entropy, is equal to 1 whatever the fre-
II. PMU CONFIGURATION CRITERIA quency when . Since its logarithm then is 0, this corre-
Let us assume an -bus network: . sponds to a zero distance between the two busses with identical
We can first define a reduced set of busses, so-called candidate (or proportional) signal responses. By contrast, when the sig-
busses, among which we will select the appropriate PMU sites: nals are completely uncorrelated, their cross spectrum is
. Finally, the optimal sit- nearly 0, which raises and, hence, its integral, the cross
ting of selected among the candidates yields another set: entropy of and , to infinity. The latter measures how far the bus
. In other words, the is electrically from . A good PMU configuration should
problem consists of building the set while fulfilling some include only the busses that are electrically far from each other,
performance and admissibility criteria. The following formal i.e., busses with large cross entropy values.
setting is to be considered here. Let denote a set of
B. Information and Coherency
responses received by the PMU located at the candidate sites, in
a given operating scenario and for a given disturbance A group of busses is said to be coherent with respect to
angle stability when any network disturbance (busbar or line
(1) short-circuit, single- or multiple-line outage, load or generation
where is the time-response of the th bus variable at switching, etc.) applied outside the group results in all incre-
the discrete time-index time while is the number of sam- mental time variations and of the bus angle and
ples. Placement of PMU in the bus set frequency having the same sign within the group [18]. It is
is “optimal” if there is little or no loss of information when intuitively sound to state that the time variation of all phasors in
replacing the signal matrix by . In other a coherent group can be satisfactorily approximated by looking
words, if is a measure of the “information content” of at a single phasor appropriately selected from within the group.
signals associated with a bus set , we should try to ensure that Consequently, it is not necessary to place a PMU at all busses
. belonging to a coherent group. According to [18] and [19], two
busses are coherent for disturbance when the following
A. Information and Entropy criterion is below a specified threshold [see (4) at the bottom of
According to [20] and [21], one measure of the information the page], where is the time variation of variable at bus
content of a signal set is the norm of the corresponding entropy during the th disturbance while is the observation time
matrix, defined as follows [21, p. 604]: frame. Similarly, it is possible to define a disturbance-based
voltage coherency criterion [14]
.. .. .. (2) (5)
. . .
.. .. .. (6)
where is the self-entropy of the th signal while . . .
is the conditional entropy of given the th signal. In ad-
and (4)
KAMWA AND GRONDIN: PMU CONFIGURATION FOR SYSTEM DYNAMIC PERFORMANCE MEASUREMENT 387
It can be observed that all the diagonal elements of this ma- we can achieve if only PMUs are to be allocated. In other
trix are zero, meaning that the coherency of a given bus with re- words, more than busses should be metered in order to fur-
spect to itself is exactly zero. Generalizing this result to , ther reduce the approximation error.
it appears that the coherency is “large” when busses and
pertain to two distinct coherent groups and “small” when both III. SEQUENTIAL PLACEMENT ALGORITHMS
and busses are members of the same coherent group. Clearly,
A. Problem Formulation
the coherency coefficient is as good a measure of the electrical
distance separating two busses as the entropy [18], [22], [25]. The various criteria studied in Section II suggest that a re-
Therefore, as far as possible, a good PMU configuration stricted size of PMU configuration is suitable if it achieves
should avoid including too many busses from the same coherent the same information content as that obtained by monitoring
group; i.e., busses retained in should have large cross co- the full candidate set. Therefore, since the maximum amount
herency indexes . of information available for a given disturbance is , our
objective is to find a set , given the number of PMUs to be
C. Redundancy and Dominant Subspace placed , which minimizes the loss of information,
i.e., the discrepancy between the actual information measure
Instead of seeking a placement set “maximizing the informa-
and the target . This problem is formalized
tion” captured through the bus signal responses, we can look
as follows:
at the problem in a quite complementary way by asking for
a solution “minimizing the correlation” among these signals. Minimize
While still emphasizing busses incorporating original and dis-
tinctive information in their signals, this viewpoint will dis- subject to (9)
card along the way all redundant busses yielding similar in-
where is a measure of the information contained in the
formation in order to relieve congestion of the data communi-
signal response set recorded at the bus set during
cation infrastructure. Let us again consider the time-response
a disturbance . When multiple operating scenarios are consid-
signal matrix whose columns are signal responses of
ered, the signal responses are random variables, making the cost
the corresponding busses. If we have exactly (redun-
function a random variable also. In what follows, we assume
dant) columns of , which can be expressed as a linear
that a random variable can capture the stochastic nature of dif-
combination of the remaining independent columns, these
ferent operating conditions regarding different load levels and
columns together constitute an optimal configuration set ,
network topologies. For each scenario, the same set of dis-
since there is no loss of information when is replaced
turbances is considered over an observation time frame . In
by , i.e.,: . Looking at the problem
addition, let be a set containing all the scenarios considered
this way shows that selecting the appropriate configuration is
and let , , be the probability (weight) associated with
basically a rank reduction of a large matrix. In linear algebra,
scenario . Then (9) becomes
it is well known that singular value decomposition (SVD) [23],
[24] is an ideal tool for this problem, since it allows a vector Minimize
space to be split into two subspaces: one dominant , the
other subdominant , where denotes the sub- subject to (10)
space spanned by the columns of the matrix . That is, in the where
decomposition
(7)
where is a diagonal matrix consisting of the singular It should be noted that any scenario is characterized by a
values of , the subspaces and may performance measure associated with the
be used in a direct sum to reconstitute the subspace given placement . For each scenario, the performance mea-
sure is averaged (expectation operator) over the disturbances
(8)
considered in set . Also, notice that the scenarios are defined
This decomposition of subspaces is equivalent to the least- according to
squares approximation of the -column matrix by a • seasonal load levels, i.e., peak and off-peak load levels for
low-rank -column matrix , which is quite close to our different seasons of the year;
objective of configuring a reduced set of PMUs at busses in • different network topologies for every seasonal load level,
, such that the least-squares errors of the approximation of as a result of forced and scheduled outages of generators,
by is “small.” The “smallness” of the approxi- lines and transformers.
mation error is measured by the corresponding singular values So far, we have not discussed the numerical evaluation of
: if they all equal 0, the approximation error is 0 and matrix norms introduced in (2) and (3) or (8), which constitute
the PMU configuration is therefore strictly optimal. If not, the backbone of the various information metrics in (10). Several
the placement set is only optimal in a loose sense, i.e., the best choices are available, including the following:
388 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 17, NO. 2, MAY 2002
2) For
2) For
1) Build a new PMU configuration set ex-
1) Build a signal matrix cluding the th bus:
2) Compute the SVD of and then its infor-
mation content according to its Gramian
whose first -columns are associated with
the busses already equipped with a PMU
and whose last column consists of the th bus
candidate for a PMU. 3) Compute the loss function for the tentative
2) Compute the SVD of and then its infor- set:
mation content according to its Gramian
3) Sort the loss function values . The bus to
be eliminated for the next PMU placement round is
the one minimizing the information loss from
to , i.e., corresponding to the smallest value of
3) Compute the yield function for the tentative
set: .
3) Sort the yield function values . The bus
retained for the next PMU placement is the one that
maximizes the information improvement from 4) The number of candidate busses for the next step
to , i.e., corresponding to the highest value of is then reduced by 1 while the
number of busses discarded for PMU installment is
increased by 1 . The new candidate
set now excludes the bus .
4) The busses not discarded in the above process are
4) The number of candidate busses for the next step is
those where PMUs should be placed. They are contained
then reduced by 1 while the number
in the set , which is therefore a solution to the
of busses to which a PMU has been allocated is
problem of optimal configuration of PMUs.
increased by 1 . Now define a new
candidate set, which excludes the bus , as well E. Simple Practical Illustration
as a new set of busses equipped with PMU
As a quick check of the above algorithms, we can consider
and the problem of placing PMUs on the 735-kV transmission net-
work of Hydro-Québec’s power system schematized in Fig. 1.
The restriction to the UHV network is necessary to simplify the
5) The busses where PMUs were placed in step 4 are
map and help the reader follow the basic steps of the placement
contained in the set : this is the solution to the problem
procedures.
of optimal configuration of PMUs.
With only 29 substations at 735 kV, our objective will be
roughly to place 10–20 PMUs at selected busses. We consid-
D. Loss Minimization Through Sequential Elimination
ered two equally probable scenarios [ in (10)]: a
Let us assume again that, for each scenario and each strong grid, corresponding to the 1996 winter peak load (36 677
disturbance , a set of time-responses, with the mW generation), and a summer grid (25 000 mW generation),
number of candidate busses in the set , is available. The with the network weakened by the outage of two 735-kV lines
greedy minimization of the information loss proceeds as fol- in the eastern and western corridors. Twelve contingencies were
lows. selected, among those most frequently encountered in operation
1) Initialization Phase 1: Define the set of candidate studies [12], essentially short-circuit faults, cleared in six cycles
busses from which to choose new PMU installations: and followed in a few cases by a single- or double-line outage.
. Then select the All the disturbances for each of these scenarios are simulated
number of PMUs to be placed in the grid. in Hydro-Québec’s stability program (ST600) and
2) Initialization Phase 2: Define the number of busses to be time-responses are recorded, representing signal-response sig-
eliminated, and let . natures of the candidate busses for PMU installation. These re-
390 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 17, NO. 2, MAY 2002
TABLE I
NORMALIZED YIELD AND LOSS VALUES AT THE FIRST ATTEMPT TO ADD OR
ELIMINATE, RESPECTIVELY, A SINGLE BUS
Fig. 4. Sample bus responses to a typical disturbance: bus angle and frequency
deviations are obtained by subtracting their respective overall mean response
averaged over the 67 busses.
mulative information more than two time bigger for a realistic [2] Y. M. Park, Y. H. Moon, J. B. Choo, and T. W. Kwon, “Design of reliable
16-PMU scenario (see Fig. 5) measurement system for state estimation,” IEEE Trans. Power Syst., vol.
3, pp. 830–836, Aug. 1988.
[3] G. N. Korres and G. C. Contaxis, “A tool for the evaluation and selection
of state estimator measurement schemes,” IEEE Trans. Power Syst., vol.
9, pp. 1110–1116, May 1994.
[4] W. A. Mittelstadt, P. E. Krause, P. N. Overholt, J. F. Hauer, R. E. Wilson,
Consequently, the constrained solution is in this example sig- and D. T. Rizy, “The DOE wide area measurement system (WAMS)
nificantly less informative than the unconstrained, globally op- project—Demonstration of dynamic information technology for future
power system,” in Fault and Disturbance Analysis & Precise Measure-
timal solution: this is yet another case where regional require- ments in Power Systems. Blacksburg, VA: Virginia Tech, 1995.
ments may act against global goals. But we believe that global [5] CIGRE Tech. Rep. 34–106, Paris, France, 1998.
configuration without meaningful regional constraints is at risks [6] R. P. Shultz, L. S. Van Slyck, and S. H. Horowitz, “Potential applica-
tions of fast phasor measurements on utility systems,” in Proc. 1989
of working well on network-wide disturbances while being my- IEEE/PES PICA Conf., pp. 49–55.
opic during local disturbances because of information black- [7] R. O. Burnett, M. M. Butts, and P. S. Strelina, “Power system applica-
holes created by the lack of PMUs close to, or inside, the faulted tions for phasor measurement units,” IEEE Comput. Appl. Power, vol.
areas. 7, Jan. 1994.
[8] S. Rovnyak and Y. Sheng, “Using measurements and decision tree
V. CONCLUSIONS processing for response-based discrete-event control,” in Proc. 1999
IEEE/PES Summer Meet., Edmonton, AB, Canada, July 18–22, pp.
The paper describes the development of two new solutions to 10–15.
the optimal PMU placement problem, under the assumption that [9] C. W. Liu and J. Thorp, “Application of synchronized phasor measure-
the installation of PMUs is chiefly motivated by wide-area mon- ments to real-time transient stability prediction,” Proc. Inst. Elect. Eng.,
General Transmission Distribution, vol. 142, pp. 355–360, July 1995.
itoring and control of large disturbances affecting the normal [10] M. H. Sendaula, S. Ye, and W. E. Kazibwe, “Optimal digital fault
operation of the interconnected system. The performance index recorders placement,” in Fault and Disturbance Analysis & Precise
maximized during the configuration of a minimal set of PMUs Measurements in Power Systems. Blacksburg, VA: Virginia Tech,
is consequently the amount of information contained in their 1995.
[11] I. W. Slutsker, S. Mokhtari, L. A. Jacques, J. M. M. Provost, M. B. Perez,
time-response signals under all plausible contingencies and re- J. B. Sierra, F. G. Gonzalez, and J. M. M. Figueroa, “Implementation of
alistic operating scenarios. The yield of a configuration is as- phasor measurement in state estimator at sevillana de electricadad,” in
sessed on the basis of a suitable electrical distance measure Proc. 1995 IEEE PICA Conf., pp. 392–398.
[12] I. Kamwa, R. Grondin, and L. Loud, “Time-varying contingency
(such as the statistical average of the entropy or coherency rep- screening for dynamic security assessment using intelligent-systems
resenting the signatures of the candidate bus set,) which has the techniques,” IEEE Trans. Power Syst., vol. 16, pp. 526–536, Aug. 2001.
property of maximizing the overall sensor responses while min- [13] E. W. Palmer and G. Ledwich, “Optimal placement of angle transducers
imizing their time-space cross correlation. The most computa- in power systems,” IEEE Trans. Power Syst., vol. 11, pp. 788–793, May
1996.
tionally effective scheme studied turns out to be a “greedy” ex- [14] M. M. Begovic and A. G. Phadke, “Voltage stability assessment through
pansion procedure consisting in sequentially adding a single bus measurement of reduced state vector,” IEEE Trans. Power Syst., vol. 5,
to the set of busses already assigned for PMU installation, with pp. 198–203, Feb. 1990.
[15] T. L. Baldwin, L. Mili, M. B. Boisen, and R. Adapa, “Power system ob-
the selected bus at each iteration being the one that maximizes servability with minimal phasor measurement placement,” IEEE Trans.
the incremental yield of information. Power Syst., vol. 8, pp. 707–715, May 1993.
An initial validation on Hydro-Québec’s 735-kV transmis- [16] P. Sadegh and J. C. Spall, “Optimal sensor configuration for complex
systems,” in Proc. 1998 American Control Conf., Philadelphia, PA, June
sion grid provides interesting insights about the proposed place-
23–26, pp. 3575–3579.
ment algorithms, notably the good geographical diversity of the [17] L. Wehenkel, “A statistical approach to the identification of electrical
resulting configurations. However, it is found that regional con- regions in power systems,” in Proc. Stockholm Power Tech Conf., June
straints are necessary on the placement to avoid information 18–22, 1995, pp. 530–535.
[18] S. Sastry and P. Varaiya, “Coherency for interconnected power systems,”
black-holes, i.e., geographical islands deprived of PMUs. In a IEEE Trans. Automat. Contr., vol. AC-26, pp. 218–226, Feb. 1981.
large multiarea system, the simplest form of a constrained op- [19] J. S. Lawler and R. A. Schlueter, “Computational algorithms for
timal PMU placement consisted in making the inclusion of all constructing modal-coherent dynamic equivalents,” IEEE Trans. Power
App. Syst., vol. PAS-101, pp. 1070–1080, May 1982.
inter-tie busses located at the boundaries of electrically coherent [20] P. Lagnotte, J. C. Sabonnadiere, J. Y. Léost, and J. P. Paul, “Structural
areas mandatory, although this approach has the pitfall of re- analysis of the electrical system: Application to secondary voltage con-
quiring prior coherency analysis of the network. Once these trol in france,” IEEE Trans. Power Syst., vol. 4, pp. 479–486, May 1989.
pre-selected busses are assigned to the PMU set, expansion of [21] M. B. Priestley, Spectral Analysis and Time Series. New York: Aca-
demic, 1992.
the latter is straightforward using the proposed sequential ad- [22] S. T. Y. Lee and F. C. Schweppe, “Distance measures and coherency
dition scheme. Comparative assessment on a 9-area/67-bus/23- recognition for transient stability equivalents,” IEEE Trans. Power App.
machine test system pointed to the fact that, while a weak- Syst., vol. PAS-92, pp. 1550–1557, Sept./Oct. 1973.
[23] A. J. Van der Veen, F. Deprettere, and A. L. Swindlehurst, “Subspace-
boundary-constrained solution is, by its very construction, less based signal analysis using singular value decomposition,” Proc. IEEE,
“myopic” about local events, it is about 37% less “informa- vol. 81, pp. 1277–1308, Sept. 1993.
tive” about system-wide phenomena, thus leaving an interesting [24] L. L. Sharf, “The SVD and reduced rank signal processing,” Signal
Process., vol. 25, pp. 113–133, 1991.
tradeoff for the designer to deal with. [25] T. Lie, R. A. Schlueter, P. A. Rusche, and R. Rhoades, “Method of iden-
tifying weak transmission network stability boundaries,” IEEE Trans.
REFERENCES Power Syst., vol. 8, pp. 293–301, Feb. 1993.
[26] A. Conejo, T. Gomez, and J. I. de la Fuente, “Pilot-bus selection for
[1] F. C. Schweppe and E. J. Handschin, “Static state estimation in electric secondary voltage control,” Eur. Trans. Elec. Eng., vol. 5, pp. 359–366,
power systems,” Proc. IEEE, vol. 62, pp. 972–982, July 1974. Sept./Oct. 1993.
394 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 17, NO. 2, MAY 2002
[27] I. Kamwa, R. Grondin, and Y. Hebert, “Wide-area measurement based Robert Grondin (S’77–M’80–SM’99) received the B.A.Sc. degree in electrical
stabilizing control of large power systems—A decentralized/hierar- engineering from the University of Sherbrooke, PQ, Canada, in 1976, and the
chical approach,” IEEE Trans. Power Syst., vol. 16, pp. 136–153, Feb. M.Sc. degree from INRS Energie, Varennes, PQ, in 1979.
2001. He joined the Hydro-Québec Research Institute/IREQ, Varennes, where he
is involved in power system monitoring, modeling, and identification and in
the development of real-time computer-based systems applied to power system
control and protection.
Innocent Kamwa (S’83–M’88–SM’98) received the B.Eng. and Ph.D. degrees Mr. Grondin is a member of CIGRÉ, and is a Registered Professional Engi-
in electrical engineering from Laval University, PQ, Canada, in 1984 and 1988, neer in the province of Québec.
respectively.
Since 1988, he has been with the Hydro-Québec Research Institute/IREQ,
Varennes, PQ, Canada. Currently, he is a Senior Researcher in the Power System
Analysis, Operation, and Control Department. He is also an Associate Professor
of Electrical Engineering at Laval University.
Dr. Kamwa is a Registered Professional Engineer.