QUICKEST LINE OUTAGE LOCALIZATION UNDER UNKNOWN MODEL

Javad Heydari Zeyu Sun Ali Tajer

ECSE Department School of EECS ECSE Department

Rensselaer Polytechnic Institute Peking University Rensselaer Polytechnic Institute

ABSTRACT data-acquisition and decision-making processes for reliably detect-

Line outage detection and localization play pivotal roles in en- ing and localizing the outage events with the fewest number of mea-
hancing the overall reliability of the electricity grid. The existing surements. This framework relies on knowing the network model
line outage detection and localization techniques often rely on the perfectly. However, while the introduction of phasor measurement
assumption that the information about the grid, such as the topology units (PMUs) enables acquiring real-time synchronized phasor angle
and system parameters, is known perfectly. In practice, however, measurements with a satisfactory precision, accurate power grid pa-
such information bears uncertainties due to inaccuracies in lines pa- rameters are far more difficult to obtain. For instance, reactance val-
rameters and the historical data. This paper studies line outage de- ues of the transmission lines, which play critical roles in outage de-
tection and localization under the assumption that the line reactance tection [2–10], are susceptible to various factors such as unstable fre-
values are known only partially, and aims to find the minimum num- quency, cable quality changes, and electromagnetic fields variations,
ber of measurements to perform outage detection and localization thus deviating the actual values from the nominal values. Therefore,
with target reliability. Specifically, based on the nominal values of in this paper we assume that the actual values of line reactance are
line reactance it proposes a stochastic graphical framework that capi- unknown and different from their nominal values and we analyze the
talizes on the correlation among the measurements generated across impact of such information uncertainties on the optimal line outage
the grid, and designs data-adaptive data-acquisition and decision- detection and localization.
making processes for the quickest localization of the lines in outage. Based on the real-time measurements used for outage detection
The paper also analyzes the sensitivity of the proposed algorithm to and localization, the existing literature can be categorized into two
the changes in the line reactance values and shows its robustness to groups. In one direction, after collecting measurements from the en-
line reactance uncertainties. tire grid, the outage detection is performed by solving different for-
mulations of the problem such as combinatorial optimization [3–5],
1. INTRODUCTION line outage distribution factor [6, 7], graphical model learning [10],
compressive sensing [8], quickest change detection [9], and joint
Reliable delivery of electricity is one of the key goals of system op-
outage detection and state estimation [11, 12]. In order to reduce
erators, and a significant effort is being made to achieve it. Due to
the cost of data collection and processing, in the other direction out-
the large-scale and strong inter-connectivities in the power grid, any
age localization is performed based on the partial observation of the
fault can transcend its realm and disrupt operations in other parts of
grid [13–18]. These studies use pre-specified data collection rules,
the grid. Therefore, real-time monitoring of the grid is of paramount
which, despite their effectiveness, can become inefficient in large-
importance in securing reliable power delivery. Specifically, agile
scale networks that are expanded over large geographic areas. The
detection and localization of system failures facilitate mitigating the
study in [2] has developed a data-adaptive data collection strategy to
disruptive impacts the failure can cause to the network, prevent cas-
circumvent this issue, where it is assuming that all the parameters of
cading failures in larger scales, and reduce the recovery costs.
the grid are known perfectly.
One common type of failure is transmission line outage, which
In this paper, we show that, interestingly, the performance of
occurs due to the transmission lines being constantly exposed to var-
the proposed optimal sampling strategy in [2] displays outstanding
ious sources of disturbance, such as equipment malfunctioning and
resistance towards reactance fluctuations. First, we establish the re-
natural disasters. While the power system is designed to operate re-
silience of the algorithm proposed in [2] against the fluctuations in
liably under single or multiple contingencies, the monitoring task
line reactance values by analyzing its sensitivity. Then, through sim-
should identify those contingencies as quickly as possible to prevent
ulations we show that the degradation in the performance remains
overload in one section of the grid. Detecting such contingencies
marginal as the uncertainty in the values increases.
and localizing them accurately can expedite the process of repairing
the faulty components, speed up restoration of the grid, reduce out- 2. PRELIMINARIES
age time, and improve power system reliability [1]. Hence, outage 2.1. Statistical Model of Bus Measurements
detection and localization have been investigated extensively in the 4
Consider a power grid consisting of N buses collected in set B =
existing literature under different settings and objectives.
{1, . . . , N } and L transmission lines denoted by set E ⊆ B × B,
Motivated by lowering the required communication and reduc-
where (i, j) ∈ E if buses i, j ∈ B are directly connected by a line.
ing the delay of decision-making and its computational complex-
We define θi and pi as the voltage phasor angle and the injected
ity, the study in [2] has developed a stochastic graphical framework
active power at bus i ∈ B, and xij as the reactance of the line con-
for modeling the bus measurements, and has devised data-adaptive
necting buses i and j. From the DC power flow model we have [19]:
This research was supported in part by the U. S. National Science X  θi − θj 
Foundation under Grant ECCS-1455228 and the CAREER Award ECCS- pi = , (1)
1554482. j∈N
xij
i
 4
where Ni = {j ∈ B : (i, j) ∈ E} is the set of neighbors of bus 3. QUICKEST OUTAGE LOCALIZATION
i. It is assumed that line reactance values are not fully known and
In this section, by capitalizing on the discrepancies among the level
there exist some level of uncertainty in their values. Specifically, by
4 of information provided by different buses under different outage
denoting the set of line reactance values by X = {xij : (i, j) ∈ E} events, we formalize a sequential data-acquisition and decision-
we assume that only their nominal values, collected in set X̄ , are making process to collect measurements of voltage phasor angles,
4 4
known. By defining p = [p1 , . . . , pN ]T and θ = [θ1 , . . . , θN ]T , based on which we localize the outage event, when one is deemed to
N ×N
from (1) we have p = Hθ, where H ∈ R is the weighted exist, with the fewest number of measurements. The data-acquisition
Laplacian matrix of the connectivity graph defined as process sequentially collects ` measurements at-a-time from ` differ-
ent buses. The process continues until time τ ∈ N, as the stopping
 P 1
 (i,`)∈E xi`
if i = j time of the process, at which point it terminates and a decision about

4
H[ij] = − x1ij if (i, j) ∈ E . (2) the underlying event is formed. For modeling the dynamic decisions

 about the buses to be observed at time t we define the selection
0 otherwise function ψ(t) ∈ B` as the set of indices of ` buses to be measured
at time t, and denote the vector of measurements collected at time t
Furthermore, from (1) and by accounting for the random distur- by θ(t) ∈ [0, 2π]` . Accordingly, we denote the vector of observed
bances in the system and the uncertainties of load profiles, the aggre- buses and their corresponding measurements up to time t by ψt and
gate injected power at different buses can be modeled as independent θ t , respectively, i.e.,
random variables [20, 21]. Hence, we can show that the statistical
relationship among the measurements collected from different buses 4 4
ψt = [ψ(1), . . . , ψ(t)]T , and θ t = [θ(1), . . . , θ(t)]T . (6)
across the grid can be modeled effectively as [10, 22, 23]
Finally, we define δ ∈ R as the decision rule at the stopping time.
θ|X ∼ N θ̄, (I − R)−1 ,


(3) The quality of decision at the stopping time is captured by the deci-
sion error probability, where by denoting the true event by T ∈ R is
4
where θ̄ is the mean vector for θ, and R is a matrix whose (i, j)-th defined as Pe = P(δ 6= T). Hence, the optimal sampling strategy is
entry is rij which is defined as obtained as the solution to the following optimization problem:
−1
minimize E{τ } Pe ≤ β ,
 X
4 βi 4 1 subject to (7)
rij = , where βi = . (4) τ , δ , ψτ
xij j∈N
xij
i
where β ∈ (0, 1) controls the reliability of the decision.
Distribution of θ given in (3) indicates that, given the line reactance 4. OPTIMAL DECISION RULES
values, the bus measurements form a Gauss-Markov random field
(GMRF) with the same dependency graph as the grid topology, i.e., 4.1. Bus Selection Rule
G(B, E) [10]. We note that, by construction, matrix H is rank- For identifying the buses that should be measured we assign a metric
deficient. By removing the row and column corresponding to one to each bus i ∈ B as follows:
reference bus, the remaining (N − 1) × (N − 1) matrix has full
rank. In the remainder of this paper, when referring to the Laplacian 4 1 X 1
M (i) = max log 2
, (8)
matrix of the network, we always mean the modified full-rank one. U ⊆Ni |U| 1 − r̄ij
j∈U

2.2. Outage Events where r̄ij is calculated based on the nominal reactance values x̄ij .
We only consider the events that keep the underlying post-event Then, at each time t we focus on the buses that are already observed
graph connected and define R = {R1 , . . . , RJ } as the set of such and identify the ones with the largest |θi − θ̄i |, which provides an
events, where Rk ⊆ E contains the lines experiencing outage under estimate of the location of the underlying outage event. Among the
event k ∈ {1, . . . , J}. Additionally, event R0 is reserved to signify neighbors of those buses we select those with the largest metrics
the no-outage event. When an outage occurs, the connectivity pro- M (i). Also, at t = 1, data collection is initialized by selecting `
file of the grid changes. We denote the connectivity graph of the grid buses with the most number of neighbors. The steps of bus selection
under event Rk by Gk (B, Ek ), corresponding to which we define the rule are presented in Algorithm 1.
weighted Laplacian matrix H k similar to (2) except for replacing E 4.2. Stopping Time and Decision Rule
with Ek . Hence, detecting and localizing outage events can be cast
The data-acquisition process is terminated as soon as a decision can
as the following multi-hypothesis testing problem:
be made with the desired reliability. By denoting C as the incident
matrix of the grid topology with c` as its `-th column that corre-
Hk : θ = Bk · p , for k ∈ {0, . . . , J} , (5)
sponds to line ` with reactance x` , and defining
4
where we have defined B k = H −1
( 1 T
k . Under each outage event, θ 4
c θ if ` ∈ Rk
x` `
follows a distinct correlation structure governed by the associated sk [`] = , (9)
topology of the network, which is imposed by matrix B k . Due to 0 Otherwise
the massive scale of power networks, collecting measurements from
at the stopping time τ we have
all the buses incurs prohibitive sensing and processing costs. Also,
uncertainty in line reactance values implies that matrix H k is not ∆θ τ = B τ Csk + B τ n , (10)
completely known. In this paper, we devise a data-adaptive decision-
making framework based on the nominal ractance values that can where B τ is the matrix constructed from H −1 by keeping its rows
form arbitrarily reliable decisions about the state of the grid with the 4
corresponding to set ψτ , and we have defined n = p − p̄ as the
minimal number of measurements.
 Algorithm 1: Data-adaptive bus selection where pij is the power flow from bus i to bus j, ANi ( x1i` ) is the
1 Set t = 1 and compute M (i) for i ∈ B according to (8) average of the inverse reractance values of neighbors of bus i, and
∂θ`
2 ψ(t) ← ` nodes with the largest degree WANi ( ∂x ij
) is the weighted average of partial derivatives of pha-
3 While stopping criterion is not met do sors of buses in Ni with the inverse of the reactances of the lines
4 Take measurements from buses in ψ(t) connected to i as the weights. The sensitivity of node j will be the
5 S ← ψt , t ← t + 1 , ψ(t) ← {} same except for pij and Ni being replaced by pji = −pij and Nj ,
6 While |ψ(t)| < ` do respectively. Furthermore, for any other bus k ∈ B (i.e., k 6= i, j)
7 i ← arg maxj∈S |θj − θ̄j | we have
8 Vi ← Unobserved neighbors of i sorted by decreasing M (·) ∂θk ∂θ`
9 If |Vi | < ` − |ψ(t)| then ψ(t) ← ψ(t) ∪ Vi
= WANk ( ). (15)
∂xij ∂xij
10 Else ψ(t) ← ψ(t) ∪ {Vi (1), . . . , Vi (` − |ψ(t)|)}
Hence, the first term in (14) is propagated to the first neighbors of
11 End if
buses i and j through a weighted average, and it propagates to the
12 S ←S\i
second neighbors by being averaged once more. Therefore, the ef-
13 End while
fect of any change in one line reactance diminishes as we go farther
14 End while
away from the buses connected to that line. It also explains why in
the bus selection rule we have to observe the neighbors of the buses
perturbations in the power injection incurred by an outage, which with the largest changes in their phasor angle values. When an out-
can be modeled as a zero-mean uncorrelated Gaussian random vec- age occurs in line (m, n) for (i, j) 6= (m, n) we have
tor [8]. Since the noise vector B τ n is colored, we include a pre- ∂∆θk ∂θ` ∂ θ̄`
processing whitening stage. By assuming the singular value decom- = WANk ( ) − WAN̄k ( ), (16)
∂xij ∂xij ∂xij
position (SVD) B τ = U ΛV T , and defining
which is negligible since it is the difference between the average sen-
4 −1 T 4 −1 T sitivity of the neighbors of a bus, which are small values. When the
y=Λ U ∆θ τ , and A = Λ U Bτ C , (11)
outage is in line (i, j) the post-outage phasor values will be indepen-
from (10)–(11), corresponding to event Rk we obtain dent of the fluctuations in xij and we have
y = Ask + ñ , (12) ∂∆θi 1 pij ∂ θ̄`
=− − WAN̄i ( ). (17)
∂xij |N̄i | xij AN̄i ( x1 ) ∂xij
i`
where ñ is a white noise vector with covariance matrix I. This leads
to an overcomplete representation of the sparse vector sk by mea- 5.2. Metric Sensitivity
surement vector y given in (12). Therefore, off-the-shelf tools from For analyzing metric sensitivity we follow the same line of argument,
compressed sensing can be applied to find the non-zero elements of and from (8) we obtain
sk to detect and localize the lines in outage. Specifically, we use or-
thogonal matching pursuit (OMP) with the modification that we stop ∂M (m) 1 X −2rmn ∂rmn
= max . (18)
the sampling process when the value of residual, i.e., r = y − Ask , ∂xij U ⊆Nm |U| 1 − rmn
2 ∂xij
n∈U
is smaller than a threshold γ that is selected such that the reliability
By denoting the set of neighbors of bus m that maximizes its metric
constraint is satisfied [2].
by Umax , and replacing the sensitivity of rmn to xij given in (18)
5. SENSITIVITY ANALYSIS we obtain
The analysis in [2] shows that the bus selection rule given in Al- ∂M (m)
=
gorithm 1 combined with the stopping and the final decision rules ∂xij
based on OMP is optimal when X = X̄ and β approaches zero. In
  2r2 2
i=m
2rmn rij

1 ij
P
 − n∈U 2 if
this section we analyze the sensitivity of the proposed algorithm to  |Umax | (1+rij )xij
max
j ∈ Umax

 (1−rmn )xmn
 n6=j
the fluctuations in line reactance values, and show that this algorithm


2
is robust against such fluctuations. To this end, we analyze the sen- −1
P 2rmn rij i=m
2 )x if
sitivity of phasor angles and bus metrics to the line reactance xij for




|Umax | n∈Umax (1−rmn mn j∈ / Umax

each line (i, j) ∈ E. 

0 otherwise
5.1. Phasor Sensitivity By denoting the minimum line reactance value by xmin , we have
From the DC power flow model we have 
−2 i=m
 |Umax ||Nm |(|Nm |2 −1)xmin if j ∈ Umax


∂θ ∂H 
H =− θ = a(ei − ej ) , (13)

∂M (m)

∂xij ∂xij ≤ ,
∂xij i=m
 |N |(|N −2|2 −1)x
 if
where ei is the unit vector with i-th element being 1 and a is defined 

 m m min j∈
/ Umax
4 θ −θ 0 otherwise

as a = ix2 j . Therefore, from (13) we have
ij
which shows that it is proportional to |Nm |−3 . These analyses show
1 ∂θ`
P
∂θi a `∈Ni xi` ∂xij that the selection rule is only loosely sensitive to the uncertainties in
= P 1 + P 1 the line reactance values. In the next section, we verify our analyses
∂xij `∈N xi` `∈Ni xi`
i by numerical evaluation of the changes in the voltage phasers and
1 pij ∂θ` metrics, as well as the overall performance of the proposed localiza-
= · + WANi ( ), (14)
|Ni | xij ANi ( x1 ) ∂xij tion strategy.
i`
 0 100
0% inaccuracy
10% inaccuracy 90
20% inaccuracy
50% inaccuracy 80

70
-5
Phasor angles (deg)

Accuracy (%)
60

50

40
-10
30

20
0% inaccuracy
10% inaccuracy
10 20% inaccuracy
50% inaccuracy
-15 0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 30 50 70 90 110
Bus number Number of measurements

Fig. 1: Sensitivity of phasor values to reactance values Fig. 3: Decision accuracy versus number of measurements.

0.8 100
0% inaccuracy 0% inaccuracy
10% inaccuracy 90 10% inaccuracy
0.7
20% inaccuracy 20% inaccuracy
50% inaccuracy 50% inaccuracy
80
0.6
70
0.5

Accuracy (%)
60
M values

0.4 50

0.3 40

30
0.2
20
0.1
10

0 0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 1 2 3
Bus number Number of lines in outage

Fig. 2: Sensitivity of metric values to reactance values Fig. 4: Decision accuracy versus different number of line outage.

6. SIMULATION RESULTS only marginally within a wide range of reactance uncertainties. The
main reason is that we compare the differences between the pre-
In this section, we evaluate the performance of the proposed algo- outage and post-outage phasor angles, which tend to be larger for
rithm and its insensitivity to the reactance fluctuations on the IEEE the buses close to the outage location, regardless of the impedance
118-bus standard system. The software toolbox MATPOWER is values of the lines connecting those buses. In other words, the data-
used to generate phasor angle measurements as well as the pertinent adaptive approach judiciously takes advantage of the information
power flows under different outage events [24]. It is assumed that provided by phasor angle deviation, which is dominantly determined
transmission line reactance values are uniformly distributed around by the power grid topology rather than the fluctuations of transmis-
their nominal value within a certain range, and all the outage events sion line reactance values.
that cause network islanding are excluded. Figure 4 compares the localization accuracy for multiple line
outages where the lines under outage are in the same locality of the
6.1. Sensitivity of Phasor Values and Metrics grid. Motivated by the observation made in Fig. 3, we set the num-
ber of measurements in all impedance fluctuation levels to 70. It is
First, we assess the sensitivity of phasor angles to various ranges observed that for single and multiple line outage events, the accuracy
of line reactance fluctuations. Figure 1 shows that even with severe degrades only slightly as the reactance uncertainty level increases.
reactance distortion with a fluctuation level as much as 50%, the
phasor angles experience limited variations. Under the same setting,
we assess the reactance fluctuation resistance property of bus metrics 7. CONCLUSION
M (·) in Fig. 2. It is observed that the fluctuation of the metrics is
also marginal. The problem of detecting and localizing line outage events by using
the minimum number of measurements under uncertainty in the line
6.2. Sensitivity of the Localization Algorithm reactance values has been considered. First, by assuming that the
known nominal values of line reactance are accurate, a data-adaptive
In order to numerically demonstrate the inherent robustness of the information-gathering and decision-making process has been pro-
proposed approach, we compare the localization accuracy of the al- posed. Then, through sensitivity analysis it has been shown that this
gorithm for different levels of transmission line fluctuations in Fig. 3. strategy is robust to the fluctuations in the line reactance values. The
It is observed that for any reactance uncertainty level, the recovery results have been verified via numerical evaluations of the proposed
accuracy rises with the increasing number of measurements. Be- rule and its performance in localizing line outage with the minimum
sides, for a fixed number of measurements, the accuracy degrades number of measurements.
 8. REFERENCES Transactions on Power Systems, vol. 30, no. 4, pp. 1659–1668,
Jul. 2015.
[1] Y. Liao, “Fault location utilizing unsynchronized voltage mea-
surements during fault,” Electric Power Components and Sys- [14] S. Maymon and Y. Eldar, “Identification of power line out-
tems, vol. 34, no. 12, pp. 1283–1293, Dec. 2006. ages,” in Proc. European Signal Processing Conference, Lis-
bon, Portugal, Sep. 2014, pp. 1093–1097.
[2] J. Heydari and A. Tajer, “Quickest localization of anomalies in
power grids: A stochastic graphical framework,” IEEE Trans- [15] Y. Zhao, A. Goldsmith, and H. Poor, “On PMU location selec-
actions on Smart Grid, vol. PP, no. 99, pp. 1–1, 2017. tion for line outage detection in wide-area transmission net-
works,” in Proc. IEEE Power and Energy Society General
[3] R. Emami and A. Abur, “External system line outage identi-
Meeting, San Diego, CA, Jul. 2012, pp. 1–8.
fication using phasor measurement units,” IEEE Transactions
on Power Systems, vol. 28, no. 2, pp. 1035–1040, May 2013. [16] Y. Zhao, R. Sevlian, R. Rajagopal, A. Goldsmith, and
[4] J. Tate and T. Overbye, “Line outage detection using phasor H. Poor, “Outage detection in power distribution networks with
angle measurements,” IEEE Transactions on Power Systems, optimally-deployed power flow sensors,” in Proc. IEEE Power
vol. 23, no. 4, pp. 1644–1652, Nov. 2008. and Energy Society General Meeting, Vancouver, Canada, Jul.
2013, pp. 1–5.
[5] ——, “Double line outage detection using phasor angle mea- [17] J. Wu, J. Xiong, P. Shil, and Y. Shi, “Optimal selected pha-
surements,” in Proc. IEEE Power Energy Society General sor measurement units for identifying multiple line outages in
Meeting, Calgary, Canada, Jul. 2009, pp. 1–5. smart grid,” in Proc. Innovative Smart Grid Technologies Con-
[6] T. Guler, G. Gross, and M. Liu, “Generalized line outage distri- ference, Washington, DC, Feb. 2015, pp. 1–5.
bution factors,” IEEE Transactions on Power Systems, vol. 22,
[18] T. Kim and S. Wright, “PMU placement for line outage iden-
no. 2, pp. 879–881, May 2007.
tification via multinomial logistic regression,” IEEE Transac-
[7] J. Guo, Y. Fu, Z. Li, and M. Shahidehpour, “Direct calcula- tions on Smart Grid, no. 99, Mar. 2016.
tion of line outage distribution factors,” IEEE Transactions on
Power Systems, vol. 24, no. 3, pp. 1633–1634, Aug. 2009. [19] A. Abur and A. G. Exposito, Power System State Estimation,
Theory and Implementation. New York: Marcel Dekker,
[8] H. Zhu and G. Giannakis, “Sparse overcomplete representa- 2004.
tions for efficient identification of power line outages,” IEEE
Transactions on Power Systems, vol. 27, no. 4, pp. 2215–2224, [20] J. F. Dopazo, O. A. Klitin, and A. M. Sasson, “Stochastic load
Nov. 2012. flows,” IEEE Transactions on Power Apparatus and Systems,
vol. 94, no. 2, pp. 299–309, Mar. 1975.
[9] Y. Chen, T. Banerjee, A. Dominguez-Garcia, and V. V. Veer-
avalli, “Quickest line outage detection and identification,” [21] A. Schellenberg, W. Rosehart, and J. Aguado, “Cumulant-
IEEE Transactions on Power Systems, vol. 31, no. 1, pp. 749– based probabilistic optimal power flow (P-OPF) with Gaussian
758, Jan. 2016. and Gamma distributions,” IEEE Transactions on Power Sys-
[10] M. He and J. Zhang, “A dependency graph approach for fault tems, vol. 20, no. 2, pp. 773–781, May 2005.
detection and localization towards secure smart grid,” IEEE [22] H. Sedghi and E. Jonckheere, Information and Control in Net-
Transactions on Smart Grid, vol. 2, no. 2, pp. 342–351, Jun. works. Cham: Springer International Publishing, 2014, ch. On
2011. the Conditional Mutual Information in the Gaussian–Markov
[11] Y. Zhao, J. Chen, A. Goldsmith, and H. Poor, “Identification Structured Grids, pp. 277–297.
of outages in power systems with uncertain states and optimal [23] ——, “Statistical structure learning to ensure data integrity in
sensor locations,” IEEE Journal of Selected Topics in Signal smart grid,” IEEE Transactions on Smart Grid, vol. 6, no. 4,
Processing, vol. 8, no. 6, pp. 1140–1153, Dec. 2014. pp. 1924–1933, Jul. 2015.
[12] ——, “Dynamic joint outage identification and state estimation
[24] R. D. Zimmerman, C. E. Murillo-Sanchez, and R. J. Thomas,
in power systems,” in Proc. Asilomar Conference on Signals,
“MATPOWER: Steady-state operations, planning, and anal-
Systems and Computers, Pacific Grove, CA, Nov. 2014, pp.
ysis tools for power systems research and education,” IEEE
1138–1142.
Transactions on Power Systems, vol. 26, no. 1, pp. 12–19, Feb.
[13] J. Wu, J. Xiong, and Y. Shi, “Efficient location identification of 2011.
multiple line outages with limited PMUs in smart grids,” IEEE