You are on page 1of 6

Real-Time Monitoring of Distributed Generation

Based on State Estimation and Hypothesis Testing


Antonio Simões Costa, Fellow, IEEE and Mariana Carneiro dos Santos

Abstract – This paper presents an application of State Es- This paper extends conventional state estimation in or-
timation to the problem of monitoring Distributed Genera- der to: (a) identify among the DG sources connected to a
tion in distribution systems. The proposed method provides distribution network those which are currently active, and
a tool to check whether the current levels of distributed
generation installed on a distribution feeder is consistent (b) estimate the power outputs of each active DG source.
with the values expected by the system operator. If that This is accomplished by formulating distribution state esti-
is not the case, statistical hypothesis tests are performed mation as a constrained optimization problem in which the
to identify the generating units whose outputs are incon-
sistent with the measurements taken along the feeder. In unknown DG sources are modeled as equality constraints.
addition, the method can also provide estimates for the un- A hypothesis testing procedure to be performed on the La-
known generated power outputs. Simulations carried out grange multipliers associated with those constraints is then
with the IEEE 34-bus test feeder are used to illustrate the devised in order to identify whether actual DG outputs
proposed method’s performance.
Keywords – Distributed Generation; Power System State comply with the levels assumed by the system operator. If
Estimation; Distribution System Real-Time Monitoring. inconsistencies are detected, the method provides estimates
for the power outputs of the DG sources by processing the
available measurements taken throughout the distribution
I. I  feeders, as well as forecasted load data. Methods based on
The increasing difficulties to build large generation and similar tools have been successfully applied to the identifi-
transmission facilities and the incentives provided by the cation of inadvertent bilateral transactions [2] and topology
restructured power markets to independent power pro- errors [3] in transmission networks.
ducers have fostered the widespread dissemination of Dis- In Section 2 of this paper, the proposed extended state
tributed Generation (DG). As a consequence, the number estimation formulation for distribution networks is intro-
of small to medium size generating units connected at the duced. Section 3 describes the hypothesis testing procedure
sub-transmission and distribution levels is growing fast on to identify active DG sources. Simulation results obtained
a worldwide scale. A particularly appealing aspect of DG for an IEEE test feeder are presented and discussed in Sec-
is its ability to prevent the negative effects of large central- tion 4, which is then followed by the concluding remarks.
ized generation and the high costs of transmission network II. S E
 D   N 
expansion [1]. On the other hand, the emergence of DG has
brought about new challenges to distribution system oper- Unlike state estimation applications at the transmission
ators. In the case of wind power based DG sources, unit level, real-time measurement redundancy for distribution
operating cycles are to a large extent dictated by wind networks is usually low [4]. To make up for that, addi-
speed, so that power output may be intermittent during tional non-telemetered information on system quantities,
unfavorable periods. Even with more conventional tech- known as pseudo-measurements, are employed to reinforce
nologies, it is not uncommon to find situations in which the distribution system metering scheme in order to en-
DG operating decisions are taken by the power producer sure proper performance of the state estimator. This is
himself. Such uncertainty concerning the actual output the case, for instance, of forecasted load data [4]. For the
values of DG sources connected to the distribution network same purpose, the proposed distribution system state esti-
may pose difficulties to the system operator in his task of mator is so devised as to allow the processing of a priori
maintaining secure and reliable operating conditions. information on the state variables [5], as described next.
The need then arises for reliable monitoring tools to as- Also, the state estimator should be capable of processing
sist the distribution system operator in identifying, among current magnitude measurements, since such information
the various DG sources whose output is unknown, those is often available at the distribution level [4]. As shown in
which are in fact injecting power into the network. In ad- [6], however, current magnitude measurements are not al-
dition, for those which are actually in operation, it is de- ways effective to ensure network observability, so that they
sirable to estimate the amount of the corresponding power are considered in this paper mainly as a means to improve
outputs. Conventional state estimators are not well suited redundancy.
to handle such uncertainties about the current statuses of In addition to the use of pseudo-measurements and a pri-
ori state information along with real-time measurements,
multiple generating units. the proposed state estimator is able to account for struc-
Antonio Simões Costa and Mariana Carneiro dos Santos are with
tural and operational constraints. Structural constraints
Federal University of Santa Catarina, Florianópolis, Santa Catarina, represent the system physical configuration and model in-
Brazil. E-mails: simoes@labspot.ufsc.br, mariana@labpsot.ufsc.br. formation such as zero-injection buses and the definition
of the reference bus. On the other hand, operational con- This sets up an iterative process which goes on until the
straints represent system relationships that change with vector of state corrections becomes negligible.
the operating condition. They are thus well suited to model Vector λ contains the Lagrange multipliers associated
buses which DG sources are connected to, as proposed in to measurement, structural, and operational constraints.
this paper. As remarked in [8], normalized Lagrange multipliers can
To summarize, distribution system state estimation is be effectively used to detect and identify inconsistencies in
formulated in this paper as a constrained minimization the mathematical model of the state estimation constraints.
problem, where the objective function is the weighted sum The Lagrange multipliers are normalized as
of squared residuals plus a term which takes into account
the a priori state information. The equality constraints λN
i = √λWi (6)
comprise the measurement model and the structural and ii
operational constraints, as follows [5]: where Wii is the i-th diagonal element of the Lagrange mul-
Minimize rmT Rm−1rm + 12 (x̂ − x̄)T P −1(x̂ − x̄)
1 tiplier covariance matrix, W. This matrix can be obtained
2
Subject to rm = zm − hm (x̂) from the inverse of the coefficient matrix in Eq. (2), as
hs (x̂) = 0
(1) discussed in [3] and [8].
ho (x̂) = 0 III. H  T

where rm is the residual vector, Rm is the measurement er- According to the proposed procedure, in the absence of
ror covariance matrix, x̂ is the vector of state estimates, zm precise information about the current status of a given
is the measurement vector, hm(x̂) is the vector of nonlin- number of DG sources, the operator assumes that all of
ear functions relating the measured quantities and states. them are off. Such assumption is considered the null
In addition, hs(x̂) and ho (x̂) are vectors representing the hypothesis of the hypothesis test, and is represented as
structural and operational constraints, respectively. In the H0 = {0 0 . . . 0}, where the zero in the i-th position of the
objective function, x̄ is the vector of the available a pri- set means the i-th DG source is off. The corresponding null
ori information on the states and P is the corresponding DG injections are mathematically represented by means of
covariance matrix. P is usually considered as a diagonal operational constraints h (x̂). On the other hand, the set of
matrix whose (i, i) entry is the assumed variance σ̄i2 for alternative hypotheses H , i = 0, considers that at least one
o

the a priori state i [3]. DG source is on (which is indicated by “1” in the proper lo-
From the Lagrangean function of Problem (1), it is pos- cation), and contains all possible combinations of unknown
sible to obtain a set of nonlinear equations that represent DG source statuses.
the first order condition to find the optimal solution. By A Bayesian Hypothesis Testing (BHT) procedure is pro-
applying Newton’s method, the following matrix equation posed in this paper to identify the actual status of DG
results [3], [5], [8]: sources whose operating conditions have not been reported
   to the distribution system operator. The purpose of the
−P −1 H T ∆x̂ − P −1 (x̄ − x̂k )
= z − h(x̂k ) (2) BHT procedure is to determine which hypothesis is best
H R λ supported by the information currently available to the state
estimator. Therefore, if BHT rejects the null hypothesis in
where favor of an alternative hypothesis H , then there are DG
i

sources injecting power into the distribution network, and


z=∆ z 0 0 T
  the operational constraints corresponding to the genera-
m tors pointed out as active in H must be relaxed (since
∆  T T i

h(x̂) = hm (x̂) hs (x̂) ho (x̂)


T T (3) they impose zero outputs for those generators). The BHT
∆  λT λT λT T procedure is preceded by the λ -test, whose purpose is to
N

λ= m s o evaluate the null hypothesis and to select suspect opera-



∆ ∂h(x̂)  T  ∂hs (x̂) T  ∂ho(x̂) T T tional constraints.
H= = ∂hm (x̂
∂ x̂
)
∂ x̂ ∂ x̂
∂ x̂ A. The λ -Test
N

and, if Ro is the diagonal covariance matrix of the opera- Before applying hypothesis testing, operational con-
tional constraints, straint Lagrangean multipliers corresponding to the DG
⎛ ⎞ buses of unknown status should be normalized as [8]:
Rm 0 0
R=⎝ 0 0 0 ⎠ (4) λN =
λo,i
(7)
0 0 Ro o,i
Wo,ii
By solving Eq. (2), a vector of corrections on the cur- where λo,i is the Lagrangean multiplier for the DG source
rent state variables is obtained, so that the latter can be i, λN
o,i
is its normalized value and Wo is the covariance ma-
updated as: trix of λo . If the value of a given normalized Lagrangean
x̂(k+1) = x̂(k) + ∆x̂ (5) multiplier exceeds a specified threshold λt , then: (a) the
null hypothesis is rejected, and (b) the corresponding oper- the largest conditional probability value. The next subsec-
ational constraint is considered suspect, that is, contrary to tion is devoted to the description a computational proce-
the initial assumption, the corresponding DG source may durefor that purpose.
be injecting power into the network. Threshold λ can be t

obtained from a fixed false alarm probability and the fact C. Computational Issues
that, in the absence of modeling errors, λ is normally dis- The computational burden to determine the largest a
tributed [8]. In short, the λ -test is used to detect a false
N
posteriori probability can be greatly reduced by noticing
null hypothesis and also as a screening procedure to reduce that probability density function (pdf) terms appear both
the number of hypotheses submitted to the more elaborate in the numerator and the denominator of Eq. (8). To take
Bayesian hypothesis tests outlined below. advantage of that, let us express all determinants |Ωi| in
B. Bayesian Hypothesis Testing
terms of |Ω0| as
The final conclusion on which hypotheses are actually |Ωi| = νi|Ω0|, i = 1, . . . , (NC − 1) (10)
true is based on the computation of a posteriori conditional
probability values of each hypothesis H given the infor- where νi is a positive real number. Also, define the
i

mation (real-time measurements, pseudo-measurements, a quadratic form in the exponent of f (z|Hi) as


priori data) currently available to the state estimator. This
probability can be obtained from Bayes’tTheorem as [10]: αi = (z − H x̄) Ω− i (z − H x̄) (11)
T 1

P (Hi |z ) =  fC −z1|Hf (z|HP jH)P (Hj )


N
( i ) (
(8) Efficient methods to compute both νi and αi have been
i )

previously developed and are reported in the related liter-


j=1
ature [2], [9]. In the present context, it is useful to express
where f (z|Hi ) is the conditional probability density func- νi as
tion for z if Hi is true and P (Hi ) is the a priori proba- νi = eln νi (12)
bility of Hi . In this paper, we assume that all alternative
hypotheses have the same a priori probability. If NDG Substituting Eqs. (11) and (12) into Eq. (9), we have
stands for the number of DG units under analysis, let NC
K
be the number of possible status (on/off) combinations of f (z |Hi ) = 2π− 2 |Ω0 |− 2 e−ε
1
(13)
i

those sources. Clearly,


where the exponent εi is given by
NC = 2NDG
1
and, since hypothesis H0 has been rejected, εi = (ln νi + αi )
2 (14)
P (Hi ) = NDG
1 We can now use the simplified form of the conditional pdf
2 −1 given by Eq. (13) to determine the largest a posteriori
On the other hand, assuming that the measurement errors probability. When we substitute Eq. (13) into Eq. (8), we
are normally distributed, then the conditional probability notice that the first two terms in the l.h.s. of the former
density function f (z |Hi ) is also normal and given by [10]: equation do not depend on i, so that they can be factored
K2 1 {− 1 (z −Hx̄)T Ω−1 (z −Hx̄)}
out of the summation in the denominator of Eq. (8) and

f (z |Hi ) = 2π |Ωi | e− 2 2 i (9) eventually cancelled out with the same term in the numer-
ator, yielding
where = 1
i ( C − 1) Ωi is the covariance matrix
,..., N ,

of vetor given that hypothesis Hi is true, and is the


z K
P (Hi |z ) =
e−ε i
(15)
number of network buses plus the number of operational N C −1 −εj
j =1 e
constraints For consistency, we also define Ω0 as the covari-
.

ance matrix of vetor considering that the null hypothesis Let imax be the value of i corresponding to the maximum
z

H0 is valid. P (Hi |z ), that is


Efficient techniques to compute both the determinant
|Ωi| and the exponent ( − )T Ω−i 1 ( − ) in Eq. (9) P (Himax |z) > P (Hi|z), i = 1, . . . , (NC − 1), i = imax (16)
z H x̄ z H x̄

have been previously developed for applications to prob-


lems at the transmission level [2], [9], so that a posteriori Since the denominator of Eq. (15) is the same for all i, we
probabilities P (Hi |z) can be explicitly obtained. However, conclude that imax can be obtained from
some characteristics peculiar to distribution networks, such
as high r/x ratios, tend to degrade the numerical perfor- εimax = min{ε1 , ε2 , . . . , ε(NC −1) } (17)
mance of those techniques. Those difficulties can be cir-
cumvented once one recognizes that it is not really neces- Therefore, the largest a posteriori probability P (Hi |z) is
sary to explicitly compute the value of P (Hi|z ). What is determined by searching for the minimum value of εi as
relevant is to know which alternative hypothesis exhibits given by Eq. (14) over the set of all alternative hypotheses.
Fig. 1. IEEE 34-bus test system with 4 DG sources

D. Constraint Relaxation flow measurements. Current magnitude measurements are


Once Himax is identified, the next step is to reformulate represented by filled circles on the main feeder and later-
the state estimation problem by revising the set of opera- als, whereas active and reactive power flow measurements
tional constraints. Since the original operating constraints are taken solely on the main feeder represented by the red
are defined by the null hypothesis, it is now necessary to horizontal line in Fig. 1. In addition, we assume that
make them consistent with the selected alternative hypoth- load forecast based bus injection pseudo-measurements are
esis in order to obtain state estimation results for the actualavailable for all load buses, except those with DG sources.
DG status configuration. However, reducing the number of Also, voltage magnitude is measured only at the substation
constraints would require the solution of a different state node.
estimation problem. To avoid that, we propose to sim- Several distinct combinations of DG unit statuses are
ply relax the constraints related to the DG sources pointed considered. In all cases, it is assumed the true DG statuses
out as active by hypothesis Himax . This is accomplished by have not been reported to the distribution system opera-
assigning large variance values (or, equivalently, very low tor. Therefore, the procedure starts by assuming that all
weights) to those constraints. The net result is that their generators are out of service, that is, the null hypothesis
influence on the outcome of the state estimation becomes is H0 = {0 0 0 0}. State estimation is then performed, fol-
negligible. lowed by the λN -test, which confirms or rejects the null
The above strategy is implemented through modifica- hypothesis. In case of rejection, the test also determines
tions on the covariance submatrix Ro defined by Eq. (4). the set of DG sources whose assigned status is considered
We then make use of techniques to solve modified linear suspect. This set is then submitted to the BHT proce-
systems [2] in order to obtain the solution under hypothesis dure outlined in Section III. The threshold value to be
Himax as a modification of the solution previously obtained employed in the λN -tests is λt = 3.0, which corresponds to
under H0. a false alarm probability of approximately 1% [3].
Since the modified problem no longer imposes hard con- A. Case A: No Active DG Source
straints on the power delivered by DG units considered We initially consider the case in which the outputs of
active in Himax , the state estimator is able to obtain valid all four DG sources depicted in Fig. 1 are zero, so that
estimates for those power outputs, as illustrated in the case their actual operating condition confirms null hypothesis
studies presented in the following section. H0. After performing state estimation through the sparse
tableau algorithm described in Section II, the normalized
IV. S R

Lagrangean multipliers for the operational constraints are
The IEEE 34-bus test feeder [11] modified through the computed through Eq. (7) and those corresponding to the
addition of four DG sources and considering all laterals as four DG sources are shown in column “A” of Table I. It can
three-phase is used in this paper to illustrate the applica- be seen that their values are well below the λN -test thresh-
tion of the proposed approach. Fig. 1 shows the test sys- old, so that no operating constraint is selected as suspect.
tem, along with the corresponding current magnitude and This result validates the null hypothesis and consequently
TABLE I C. Case C: Four Active DG Sources
N
L         
In Case C, we consider that all four DG sources are ac-
λN
i tive. The outputs of generators at buses 10 and 18 are
DG source bus A B C as in Case B, whereas generators at buses 24 and 33 are
10 1.67 −96.5 −119.8 both injecting 120 kWinto the distribution network. As
18 −0.70 −110.6 −164.9 in the previous cases, the operator is not aware that the
24 −0.61 −92.6 −148.5 DG sources are on, so that their output is assumed as zero.
33 −0.12 −22.4 −129.4 The results reported in column “C” of Table I show that
the normalized Lagrange multipliers computed under the
null hypothesis are orders of magnitude larger than the de-
the state estimation results, as expected. fined threshold, so that all operating constraints are again
selected as suspect. The BHT procedure is then applied to
B. Case B: Two Active DG Sources determine which alternative hypothesis exhibits the best
adherence to the measurement values. As shown in Table
This case considers that, unlike the zero values assumed II, the outcome points out to the correct combination of
by the distribution system operator, the actual outputs of active DG sources. Finally, that result is used to redefine
DG sources at buses 10 and 18 are 100.0 kW and 150.0 kW, the operational constraints, which enables the state estima-
respectively. State estimation is performed considering the tor to provide fairly accurate estimates for the DG source
actual bus power injections, and again the normalized La- outputs, as shown in the second section of Table II for case
grange multipliers are computed. The results are displayed C.
in column “B” of Table I. As one can readily verify, the λN
values for all constraints corresponding to the DG sources D. Importance of DG Monitoring for Operating Decisions
are much larger than the threshold, so that all of them are To underline the importance of a fairly accurate knowl-
selected as suspect. edge of DG source outputs for the distribution system op-
The next step is then to apply the BHT procedure of erator decision-making process, Fig. 2 presents the voltage
Section III to determine which of the 24 − 1 = 15 alter- profiles of the main feeder and laterals considering the DG
native hypotheses exhibits the largest a posteriori condi- configurations described in cases A, B and C above. From
tional probability given by Eq. (8). As also described in the plots, it is possible to notice the significant influence
Section III, in order to do so it is not necessary to explic- of diverse DG source outputs on the voltages along the
itly compute all probability values, but rather identify from feeder, which should be taken into account as the opera-
partial results related to Eq. (8) which hypothesis is best tor defines voltage control strategies to be implemented on
supported by the measurement values. For Case B, the re- the distribution network. The presence of DG sources can
sults are summarized in the corresponding column of Table bring about other accompanying effects, such as reversing
II. In spite of the large values of all normalized Lagrange the direction of power flows. Power flow reversals in fact
multipliers associated to operational constraints, the BHT occur on some feeder sections for cases B and C, but space
procedure does not confirm that all DG source statuses are limitations prevent a more detailed discussion in this paper.
wrong. Instead, it is found that only DG sources at buses
10 and 18 are active.
After the correct hypothesis is identified, the constraints
that incorrectly impose null outputs to actually in service
DG sources are relaxed and the state estimation can pro-
ceed. No further errors are detected by re-applying the λN -
test. The estimated outputs for the DG sources at buses
10 and 18 are in agreement with the corresponding simu-
lated values, as shown in the second section of column B
in Table II.
TABLE II
H  T
R  
 E  DG O  

Case B Case C
Hypothesis of max P( H | ) {1 1 0 0} {1 1 1 1}
i z

Estimated DG
Fig. 2. Feeder voltage profile for the three case studies

Buses Outputs ( )
kW

10 100.0 100.0 V. C


18 150.0 150.3
This paper introduces a method for real-time moni-
24 − 118.2
toring of distributed generation connected to distribution
33 − 121.1 networks based on power system state estimation and a
Bayesian hypothesis testing algorithm. Simulation results [6] A. Abur, A. Gomez Expósito, “Detecting Multiple Solutions in
conducted on a modified version of the IEEE 34-bus test State Estimation in the Presence of Current Magnitude Mea-
surements”, IEEE Transactions on Power Systems, v. 12, n. 1,
feeder indicate that the proposed method is able to cor- Feb. 1997, pp. 370-375.
rectly identify which DG sources are injecting power into [7] A. Abur, A. Gomez Expósito, “Bad Data Identification when Us-
the network. In addition, the method also provides good es- ing Ampere Measurements”, IEEE Transactions on Power Sys-
tems, v. 12, n. 2, May 1997, pp. 831-836.
timates for the corresponding power outputs. The method [8] K. A. Clements, A. Simões Costa. “Topology Error Identification
has been tested by using other test feeders, with consis- Using Normalized Lagrange Multipliers”, IEEE Transactions on
tently encouraging results. Therefore, it qualifies as the Power Systems, v. 13, n. 2, 1998, pp. 347-353.
[9] E. M. Lourenço. “Observability Analysis and Topology Error
basis for an effective monitoring tool to assist system op- Identification in the Generalized State Estimation”. Ph.D. the-
erators in identifying the causes of unaccounted flows and sis, Electrical Engineering Graduate Program, Universidade Fed-
voltage profile deviations due to the unrevealed connection eral de Santa Catarina, Brazil, (in Portuguese) 2001.
[10] A. Papoulis. “Probability Random Variables and Stochastic Pro-
of DG sources to the distribution network. cesses”, 3rd. ed, Mc Graw Hill, 1991.
[11] IEEE Distribution Planning Workgroup Report. “Radial Distri-
A
 bution Test Feeders”. IEEE Transactions on Power Systems, v.
6, n. 3, Aug. 1991, pp. 975-985.
Antonio Simões Costa thanks the financial support of the
Brazilian National Research Council (CNPq), and Mariana
Carneiro dos Santos acknowledges the financial support of
the Brazilian Agency for the Improvement of High Educa-
tion (CAPES).
R
[1] E.R. Brown and L.A. Freeman, “Analyzing the Reliability Im-
pact of Distributed Generation”, Proc. of the IEEE PES Summer Antonio Simões Costa(F’06) received his degree in Electri-
Meeting, Vancouver, Canada, 2001, pp. 1013-1018. cal Engineering from Federal University of Pará, Brazil, in 1973,
[2] K.A. Clements, A. Simões Costa and A. Agudelo, “Identification and the M.Sc. and Ph.D. degrees in Electrical Engineering from
of Parallel Flows in Power Networks through State Estimation Federal University of Santa Catarina, Brazil, in 1975, and Uni-
and Hypothesis Testing”, International. Journal on Electrical versity of Waterloo, Canada, in 1981, respectively. Since 1975 he
Power and Energy Systems, Vol. 28, 2006, pp. 93-101. is with Federal University of Santa Catarina, Brazil, where he is
[3] E. Lourenço, A. Simões Costa and K.A. Clements, “Bayesian- a Professor in the Department of Electrical Engineering. His re-
Based Hypothesis Testing for Topology Error Identification in search interests are concerned with computer methods for Power
Generalized State Estimation”, IEEE Transactions on Power Systems Operation and Control.
Systems, v. 19, n. 2, May 2004, pp. 1206-1215.
[4] M.E. Baran, “Challenges in State Estimation on Distribution Mariana Carneiro dos Santos received her degree in Elec-
Systems”, Proc. of the IEEE PES Summer Meeting, Vancouver, trical Engineering and her Master’s degree both from Federal Uni-
Canada, 2001, pp. 429—433. versity of Santa Catarina, Brazil, in 2003 and 2006, respectively.
[5] M. Carneiro dos Santos, “Real-Time Monitoring of Distributed Since 2005 she is with Tractebel Energia, Florianópolis, Brazil.
Generation in Distribution Systems”, Master’s Thesis, Electrical Her research interests are in the area of real-time monitoring of
Engineering Graduate Program, Universidade Federal de Santa Distributed Generation and load forecasting methods.
Catarina, Brazil, (in Portuguese), 2006.

You might also like