Upload
mines
View
0
Download
0
Embed Size (px)
Citation preview
1
Abstractâ Smart grid (SG) technology transforms the traditional
power grid from a single layer physical system to a cyber-
physical network that includes a second layer of information.
Collecting, transferring, and analyzing the huge amount of data
that can be captured from different parameters in the network,
together with the uncertainty that is caused by the distributed
power generators, challenge the standard methods for security
and monitoring in future SGs. Other important issues are the
cost and power efficiency of data collection and analysis, which
are highlighted in emergency situations such as blackouts. This
paper presents an efficient dynamic solution for online SG
topology identification and monitoring by combining concepts
from compressive sensing and graph theory. In particular, the
SG is modeled as a huge interconnected graph, and then using a
DC power-flow model under the probabilistic optimal power
flow, topology identification is mathematically reformulated as a
sparse recovery problem. This problem and challenges therein
are efficiently solved using modified sparse recovery algorithms.
Network models are generated using the MATPOWER toolbox.
Simulation results show that the proposed method represents a
promising alternative for real time monitoring in SGs.
Index Termsâ Smart grids, topology identification and fault
detection, cyber-physical systems, compressive sensing (CS),
sparse recovery.
I. INTRODUCTION
N intelligence revolution is happening in the
technology of power grids, supporting the development
of so-called smart grids (SGs). Roughly speaking, âthe
term SG refers to electricity networks that can intelligently
integrate the behavior and actions of all parameters and users
connected to themâ [1]. This revolution transforms a
traditional power grid from a single-layer physical system into
a huge cyber-physical network, using a layer of information
that flows through the system. This information includes the
status of several parameters in the network such as bus
voltages, active and reactive powers, phase angles, branch
currents, and load consumption, in addition to the actuating or
controlling commands fed back to the network from
controllers and decision making units [1]-[4], [22]-[23].
Another important feature of next generation SGs is the
aggregation of renewable energy micro-sources in distributed
generation technology [5]. The nature of the injected power
from renewables as wind turbines and solar panels is
inherently random. This randomness, along with the
uncertainty in the load, lead to probabilistic behavior in the
load-power flow at buses. The probabilistic optimal power
flow (P-OPF) problem has been fully discussed in [6]-[8] and
references therein.
Collecting, transferring, and analyzing the huge amount of
data flowing through the information layer of the SG, together
with the uncertainty caused by distributed power generators
and load characteristics, challenge the standard methods for
security, monitoring, and control. Identifying changes of lines
is particularly critical for a number of tasks, including state
estimation, optimal power flow, real-time contingency
analysis, and thus, security assessment of power systems [7].
Other important issues are the cost and power efficiency of
real-time data collection and analysis [10]. The Supervisory
Control and Data Acquisition (SCADA) system and Wide
Area Mentoring System (WAMS) provide voltage and power
data for each local system in near real time [2]-[4], [22]-[23].
The most popular sensing technology used widely in SG data
collection systems is the synchronous Phasor Measurement
Unit (PMU) [4]. The system data exchange (SDX) module of
the North American Electric Reliability Corporation (NERC)
can provide grid-wide topology information on an hourly basis
[10]; however, near real-time monitoring of transmission lines
is mandatory in order to change the grid into a smart system.
The power network (PN) topology identification (TI)
problem has been addressed in [9] based on a consuming or
economic approach. Some related work has addressed the
problem of fault or power line outage identification (POI)
using PMU data in a power grid; see [7], [10]-[14] and
references therein. In this work, we introduce an efficient and
low cost solution for TI (that can also be used to solve the POI
problem) using a few measurements of the system parameters.
Our work combines sparse sampling theory from the field of
Compressive Sensing [18] with concepts borrowed from graph
theory [15]. Most of the previous approaches (which are
mainly focused on POI) rely on a DC linear power flow model
[7], [10]-[14]. Using this approximated model, the PN is
modeled as a graph and the POI problem is expressed in terms
of certain matrices defined by this graph. In some recent
works, POI is formulated as a combinatorially complex
(integer programming) problem, which is computationally
tractable only for single or, at most, double line changes (refer
to [10]). Due to the discrete formulation of the problem, the
optimal solution can be found through an exhaustive search
(ES). However, the computational complexity of such an ES is
high, and thus it is inappropriate from a time cost perspective.
Moreover, coping with multiple line changes is critical in the
face of cascading failures in recent blackouts, especially in
case of large networks. This has motivated several existing
works [7], [10] and [13], which consider how to handle
multiple line outages. A recent alternative approach to line-
change identification adopts a GaussâMarkov graphical model
of the PN and can deal with multiple changes at affordable
complexity; the Gauss-Markov approach is based on a
stochastic characterization of power flows that models the
busesâ phase angles as Gaussian random variables and
comprehensively addresses their dependencies [7]. An
ambiguity group-based location recognition algorithm has
been also proposed for recognizing the most likely outage line
locations in multiple outages. Another new method for solving
the POI problem has been developed based on the theory of
quickest change detection. Also recently a distributed
framework to perform the identification locally at each phasor
Compressive Sensing-Based Topology
Identification for Smart Grids
A
2
data concentrator has been presented. Moreover, another
approach considered the same problem under scenarios with a
limited number of PMUs. Most recently, a sparsity-based
method has been developed in order to address the bad PMU
data records and their imperfections in power outage detection
and localization. For more information on these approaches,
please refer to [10]-[14] and references therein.
In [10] and [21], the authors formulate POI as a sparse
recovery problem (SRP). Their method leverages the fact that
the number of outage lines after a failure represents a small
fraction of the total number of lines. In [11], a global
stochastic optimization technique based on cross-entropy
optimization (CEO) is presented. The initial idea of this
method is based on the formulation presented in [10]. It has
been shown that CEO has better performance compared to
[10] in general. However, both of these methods require
hourly basecase grid topology information in addition to the
system parameters to be able to find the outage lines in the
grid. In [10], the measurement or sensing matrix of the SRP
includes the weighted Laplacian matrix đ” (also called the
topology, structure, or nodal-admittance matrix) of the
corresponding graph of the power network, which is
dependent on the previous structure of the network and needs
to have the topology information of the graph delivered in an
hourly manner. In [11], the measurement vector is calculated
from multiplication of the same Laplacian matrix with the
phase angle vector, so again hourly basecase topology
information from the grid is needed. Moreover, the
performance of most of the aforementioned methods is highly
related to the level of noise. It has been shown in the literature
([10] and [11]) that the recovery performance of these
methods significantly decreases as the level of noise increases,
so in both of the approaches in [10] and [11], a pre-whitening
procedure is applied before further analysis. Another
important issue to be considered is the rank deficiency of the
matrix đ”. In order to avoid the problems arising from this rank
deficiency, one bus is normally considered as the reference
bus and its corresponding row and column are removed from
the matrix đ”, so that the matrix becomes full rank. However,
due to cyber-attacks and other precision deficiency issues,
whenever the records from the reference bus are affected by
bad data, the overall procedure can be corrupted. Another
important problem is with the switching events that are
mistakenly unreported. These kind of errors can change the
structure of the nodal-admittance matrix đ”. Thus, if the last
status of the topology of the network has not been reported
within a suitable time interval the final results can be affected.
On the other hand, to the best of our knowledge, most of the
methods mentioned above do not account for the random
behavior applied by renewable energy sources and the
uncertainty in loads. Finally, the mathematical formulation of
some aforementioned methods is dependent on the
connectivity of the power grid graph, where they assume that
the power outages may not be allowed to result in islanding in
any part of the network [10].
Our work is based on the probabilistic power flow model.
Our approach is completely different from previous
formulations, in that the presented reformulation relies only on
the measurements from system parameters and does not need
any a priori information about the topology of the network. In
fact, we reformulate the TI problem in such a way that the
output of the optimization problem is the structure or topology
matrix đ” of the network itself. As a result, this framework is
capable of dealing with the aforementioned imperfections.
Case studies using standard IEEE test-beds [16] show that the
proposed method represents a promising new strategy for
topology identification, line change, fault detection, and
monitoring issues in SGs.
The rest of this paper is organized as follows. Section II
discusses the model of a PN as an interconnected graph and its
corresponding mathematical representation using a DC power
flow model. Section III presents the proposed sparse topology
identification technique. Moreover, it contains a short review
on CS and a short description of the SRP solvers to be used in
this paper. Simulation results of a study of the sensing matrix
coherence (Section III.A), the effectiveness of the algorithm,
and the results from applying a model of structured sparsity
are discussed in Section IV. Finally, Section V summarizes the
final conclusions of this paper.
II. POWER NETWORK MODEL AND CORRESPONDING GRAPH
In this section, the graph representation of a PN, its
corresponding matrices, and its connection with the DC power
flow model are described. Moreover, we discuss a random
Gaussian distribution assumption for the system (bus power
and voltage) parameters in a large SG with a considerable
amount of renewable energy sources.
A. Graph of the Power Network
We represent the PN as a graph đș(đđ , đđž), consisting of a
set of đŸ nodes đđ = {1, . . . , đŸ}, where each node đ represents
an arbitrary bus of the PN, and a set of đż edges or lines đđž =
{đđ,đ: đ, đđđđ}, which connect certain nodes and form the
general structure or topology of the grid. Following the
approaches in [7], and almost all of examples from [10], we
assume that the data from the nodesâ parameters (such as
power and phase angles) are obtained from a network of
sensors. The main goal of this paper is to identify the structure
or topology of the PN in near real time using a small set of
data, without any a priori knowledge about the structure of the
network. Nowadays, a vast amount of work is focusing on
power sensor technology especially on software-based PMUs
with much lower costs per unit; [25] and the references
therein.
B. DC Power Flow Model and Corresponding Graph of PN
The AC power-flow model, also known as the AC load-
flow model, is a numerical technique in power engineering,
applied to a power system in order to analyze the behavior of
different forms of AC power under normal steady-state
operation. In such a model two distinct sets of non-linear
equations (called power balance equations) indicate the
relationship between active and reactive power and the voltage
magnitude and phase angle as follows [17]:
đđđ = đđ,đđđ2 â đđ,đđđđđđđđ (đđ â đđ) + đđ,đđđđđđ đđ(đđ â đđ) (1)
đđđ = đđ,đđđ2 â đđ,đđđđđđđđ (đđ â đđ) + đđ,đđđđđđ đđ(đđ â đđ). (2)
Here đđđ and đđđ are the active and reactive powers injected
from node đ to node đ, respectively; đđ and đđ represent the
amplitude and phase of the voltage on node đ, respectively;
3
đđ,đ is the real part of the admittance of the line đđ,đ (or the
conductance); and đđ,đ is the imaginary part of the admittance
of line đđ,đ (or the susceptance). It has been shown that when
the system is stable, under a normal steady-state condition, the
phase angle differences are small, meaning that đ đđ(đđ â đđ ) â đđ â đđ. Using this simplification and the DC power
flow method discussed in [17], the power injected to a distinct
bus đ follows the superposition law (in practical cases equation
(4) cannot be exactly validated):
đđ = â đđđđ = â đđ,đđđđźđ
(đđ â đđ) (3)
đđ = â đđđđ = â đđ,đđđđźđ
(đđ â đđ) . (4)
Here đźđ is the set of neighbor buses connected directly to bus
đ. It is useful to rewrite these summations in a matrix-vector
format, where we have:
đ = đ”đ đđđ đŸ (5)
đ = đ”đŁ đđđ đŸ . (6)
In this notation the voltage amplitude and voltage phasor angle
values from all of the nodes are collected in two vectors
đŁ, đ đđ đŸ , respectively. Also, the active and reactive power
values of all of the nodes are stored in the vectors đ and đ,
respectively. The matrix đ” â đ đŸĂđŸ is called the nodal-
admittance matrix, describing a power network of đŸ buses,
and its elements can be represented in the following format:
đ”đđ = {
âđđ,đ , đđ đđ,đđđđž
â đđ,đ , đđđźđđđ đ = đ
0, đđĄâđđđ€đđ đ
} (7)
The nodal-admittance matrix of the IEEE Standard-118 Bus
is shown in Fig. 1. We see that this matrix has a sparse
structure. It has been mentioned in the literature that this
matrix can be viewed as a weighted version of the Laplacian
matrix of the graph đș(đđ , đđž) [17] and [15]. Moreover, since
the Laplacian matrix has a direct relationship with other
important structural matrices of a graph, the nodal-admittance
matrix đ” can give a full description of the power network
structure; see [17] for more information on graphs and their
corresponding matrices. As a result, our goal is to design an
appropriate and fast method to determine the structure of the
matrix đ” from measurements of the PN parameters.
C. Gaussian Distribution of Parameters
In [6]-[8] it has been thoroughly discussed that due to the
load uncertainty in large-scale transmission networks and the
increasing contribution of distributed renewable sources in
smaller scale PNs, injected active and reactive powers can be
modeled as random variables. Moreover, since âthe injected
power models the superposition of many independent factors
(e.g., loads) it can be modeled as a Gaussian random variable
[6]â. The linear relationship in (3) suggests that the difference
of phasor angles (đđ â đđ) for each sample of time across a bus
can be approximated by a Gaussian random variable as well
[7]. Moreover, allowing for the fixed phasor at the slack bus,
under steady-state conditions individual phasor angle
measurements đđ can be modeled as Gaussian random
variables. Consequently, we will assume that during the
observation window of a TI problem, the measurements of the
phase angle of node đ can be well approximated by Gaussian
random variables. As we will discuss in Section III, this
random behavior has an impact on the structure of the sensing
matrix in the TI problem, and it is beneficial to the
performance of sparse recovery techniques in Section IV.
Fig.1. Corresponding normalized Laplacian matrix of the standard IEEE
118 Bus.
III. SPARSE TOPOLOGY IDENTIFICATION
We now focus on the problem of recovering the topology
matrix đ” from vectors of measurements đ and đ (see (5)), or
equivalently, đ and đŁ (see (6)). Our key assumption here is
that the PN is a sparse interconnected system. This assumption
is based on a survey of articles and standard power network
models found in software and toolboxes such as
MATPOWER. We observe that the regular maximum
connectivity level of a bus in a network is typically less than
5-10%, especially in very large power transmission systems.
For example, the highest connectivity level of a bus in IEEE
30 bus is 6 (10%), in IEEE 118 bus is 9 (7%) and in IEEE
2383 is 9 (0.3%). This means that each column in the nodal-
admittance matrix satisfies the definition of a sparse signal
(Section III-B). This sparsity helps us formulate TI as a sparse
recovery problem, which as supported by the theory of
Compressive Sensing, can be solved with a small set of
measurements in a fast and accurate way using SRP
algorithms. If the sparsity assumption is violated, however, the
performance of the algorithm will suffer. The formulation of
TI as a SRP is presented in Section III-A. Section III-B gives a
short background on CS. Finally, in Section III-C we
introduce the types of the SRP solvers used in this work.
A. Sparse Setup of Topology Identification Problem
Given an interconnected power network of đŸ nodes or
buses, let the measurements of the active power and phase
angle of node đ be associated with the following time series of
đ sample times:
đđ(đĄ) đđđ đĄ = 1,2, . . . , đ (8)
đđ(đĄ) đđđ đĄ = 1,2, . . . , đ . (9)
As mentioned before, in the DC model, the (active or
reactive) power injected into a distinct bus đ follows the
superposition law (3) and (4). Since (4) cannot be validated
properly in some practical situations, we prefer to use (3) to
formulate the sparse TI problem. This means that, for each
node đ in the network and at each sample time đĄ, we have
4
đđ(đĄ) = â đđđđ (đĄ) = â đđ,đđđđźđ
(đđ(đĄ) â đđ(đĄ)) . (10)
As mentioned before, đđ,đ is the susceptance along the line
đđ,đ under the DC model, and đźđ is the set of neighbor buses
connected directly to bus đ. If we write đđ,đ = 0 for đ â đźđ, we
can extend (10) as follows:
đđ(đĄ) = â đđđđ (đĄ) = â đđ,đđđđđŸ
đ (đđ(đĄ) â đđ(đĄ)) + đąđ(đĄ) + đđ(đĄ), (11)
where đđŸđ is the set of all nodes in the network except node đ
(so the set đđŸđ contains đŸ â 1 nodes), đąđ is a possible leakage
of active power in node đ itself, and đđ is the measurement
noise. Since we assume data is collected at đ sample times,
we can use matrix-vector notation to write
đđ = đŽđđŠđ + đąđ + đđ, (12)
where, đđ , đąđ , đđđđ đ. In equation (14) we have
đŽđ = [đ1,đđ , ⊠, đđâ1,đ
đ , đđ+1,đđ , . . . , đđŸ,đ
đ ] đ đ đĂđŸâ1 (13)
đđ,đđ = (đđ(đĄ) â đđ(đĄ)) đđ đ đđđ đĄ = 1,2, . . . , đ (14)
đŠđ = [đđ,1, ⊠, đđ,đâ1, đđ,đ+1, . . . , đđ,đŸ]đ
đđ đŸâ1. (15)
Each column of the matrix đŽđ represents the difference of
phase angles between node đ and one node đđđđŸđ in the network
across đ samples of time. As discussed in Section II-C, this
difference can be modeled as a Gaussian random variable. The
vector đŠđ is a sparse vector, with all of its elements equal to
zero except a small portion, which are located in those
positions corresponding the (few) neighbors of node đ. Summing đąđ and đđ (which can also be modeled as a vector of
white Gaussian noise), we end up with the following equation
for each node:
đđ = đŽđđŠđ + đđ . (16)
The TI problem can be viewed as the estimation of all
vectors {đŠđ}đ=1đŸ that best match the observed measurements
{đđ}đ=1đŸ . In order to solve such a problem, we can define the
following optimization problem:
đđđ{ïżœÌïżœđ}đ=1
đŸâ âđŽđïżœÌïżœđ â đđâ2
2đŸđ=1 . (17)
Problem (17) can be solved individually for each node đ. If đŽđ is full rank, the number of measurements đ exceeds the
number of unknowns đŸ â 1, and the variance of the noise
vector, đđ, is small enough, this optimization problem can be
solved using Least Square (LS) techniques. Since the value of
đŸ depends on the number of buses in the grid, a large number
of measurements đ will be needed to solve this problem in the
case of large power grids. In order to avoid this problem, we
suggest using sparse recovery techniques [18] to solve for the
vectors {đŠđ}đ=1đŸ . Since each vector đŠđ is sparse, we can view
the goal in (16) as recovering an S-sparse vector đŠđđđ đŸâ1
from a set of observations đđđđ đ, where đŽđđđ đĂđŸâ1 is the
sensing matrix, đđđđ đ is a vector of white Gaussian noise, and
đ is the number of nodes which are directly connected to the
one individual node đ for which we are solving the problem. In
general we want đ âȘ đŸ so the problem can be solved using a
reasonable set of observations in huge power grids. After
solving this problem for each sparse vector đŠđ corresponding
to each node đ, we can concatenate all of the sparse vectors
together, form the weighted Laplacian matrix đ”, and the
process is completed. Because đđ(đĄ) â đđ(đĄ) = 0 for đ = đ, đđŸđ
should include đŸ â 1 nodes; in other words, we should keep
the node đ out of đđŸđ since it produces a vector of zeros in the
corresponding column of the matrix đŽđ. This means that we
are not able to find the value of the parameter đ”đđ from the
recovered vector đŠđ; however, according to the definition of
the nodal-admittance matrix đ”,
đ”đđ = â đđ,đđđđźđ . (18)
Thus, after recovering the vector đŠđ , đ”đđ can be easily
calculated. In the next section, we briefly give a summary of
the required CS and SRP background and concepts used in
interpreting and solving (16) as an SRP.
B. CS Background and Related Concepts
Definition 1 (Sparse Signal): An S-sparse signal đŠđđ đŸ is a
signal of length đŸ with đ non-zero entries where đ < đŸ (in
many cases (đ âȘ đŸ).1
Intuitively, the SRP is an optimization problem in which the
goal is to recover an S-sparse signal đŠđđ đŸ from a set of linear
observations (measurements) đ = đŽđŠ đđ đ where đŽđđ đĂđŸ is
the sensing matrix with đ < đŸ (in many cases đ << đŸ).
Typically in CS, the number of measurements đ will be
proportional to the sparsity level đ of the signal đŠ to be
recovered. For example, it is known that when the sensing
matrix đŽ is drawn randomly from a suitable distribution (i.e.,
the measurement protocol is random), with high probability
any đ-sparse signal đŠ can be recovered from the measurements
đ = đŽđŠ if the number of measurements đ is merely
proportional đ đđđ (đŸ
đ). This number of measurements can be
significantly smaller than the overall signal length đŸ, and is
only greater than the fundamental limit đ by a logarithmic
factor; one can interpret this logarithmic factor as the price
that should be paid for not knowing the locations of the sparse
coefficients in advance. Remarkably, assuming the signal đŠ to
be exactly sparse and the measurements đ = đŽđŠ to be noise-
free, this recovery is exact. If the signal to be recovered is
nearly sparse (compressible2) or if the measurements are
noisy, this recovery is provably robust. Compressible signals
appear in variety of applications such as when CS-based
techniques are used for MRI; in many such applications,
approximate signal recovery is sufficient. However, in our
work each column of the matrix đ” is exactly sparse.
Due to the underdetermined nature of this recovery
problem (i.e., since đ < đŸ), the null space of the sensing
matrix đŽ is non-trivial; therefore, we have infinitely many
candidate solutions to the system of equations đ = đŽđŠ.
However, under certain conditions on the sensing matrix đŽ,
CS-based recovery methods can be guaranteed to efficiently
1 The notation in this Section matches that in Section III-A, except for
convenience we suppose the unknown vector đŠ has length đŸ rather than đŸ â1.
2 The nearest đ-sparse signal to đ„, đ„đ , can be obtained simply by keeping
the đ largest entries of đ„ and setting all remaining entries to 0. If đ„ is exactly
đ-sparse, then đ„ = đ„đ . If the distance from đ„ to đ„đ is small (but not
necessarily zero), đ„ is said to be compressible.
5
find the candidate solution that is sparse enough. The
following definitions and theorems provide examples of such
conditions; for more discussion and technical proofs and
details please refer to [18]-[19] and references therein.
Definition 2 (Coherence): The coherence of an đ Ă đŸ
matrix đŽ with normalized columns đ1, đ2, ⊠, đđŸ is defined as
follows:
đđŽ = max1â€đ,đâ€đŸ,đâ đ
|âšđđ, đđâ©| . (19)
The coherence of an đ by đŸ matrix đŽ is a value in the interval
[1
âđŸ, 1].
Definition 3 (Restricted Isometry Property): An đ by đŸ
sensing matrix đŽ satisfies the Restricted Isometry Property
(RIP) of order đ if
(1 â đżđ)âđŠâ22 †âđŽđŠâ2
2 †(1 + đżđ)âđŠâ22 (20)
holds for all coefficient vectors đŠ with âđŠâ0 †đ where
đżđđ(0,1) is known as the isometry constant of order đ.
Theorem 1 (Coherence-based recovery guarantee): Suppose
đŠ is S-sparse and let đ = đŽđŠ. If
đđŽ <1
2đâ1 , (21)
then ïżœÌïżœ = đŠ is the unique solution to đ = đŽïżœÌïżœ having sparsity đ
or less. The smaller the coherence of đŽ, the larger the
permitted value of đ, and the broader the class of signals that
can be recovered. Roughly, what these conditions require is
that any two subsets of columns of the sensing matrix đŽ must
be almost orthogonal to each other.
Theorem 2 (RIP-based recovery guarantees): Suppose đŽ
satisfies the RIP of order 2đ with đż2đ < 0.4651. Let đ =đŽđŠ + đ be noisy measurements of a vector đŠ. Then for any S-
sparse vector đŠ, (22) correctly returns ïżœÌïżœ = đŠ given đ:
ïżœÌïżœ = đđđđđđïżœÌïżœ
âïżœÌïżœâ1 đ đąđđđđđĄ đĄđ âđ â đŽïżœÌïżœâ2 < í. (22)
Moreover, for any vector đŠ, if í â„ âđâ2, then the solution ïżœÌïżœ to
(22) obeys:
âđŠ â ïżœÌïżœâ2 †đ¶1âđŠâđŠđâ1
âđ+ đ¶2í , (23)
where đ¶1 and đ¶2 depend only on the matrix isometry constant.
The vector đŠđ is the closest đ-sparse approximation to đŠ. The
optimization (22) is widely known as Basic Pursuit Denoising
(BPDN) and is an SRP. In the next section, we briefly
introduce several SRP solvers which are used in this work. By
generalizing these methods, we will introduce new adaptable
strategies in order to address the practical challenges involved
in solving the sparse TI problem.
C. Sparse TI Structure and Data Correlation Issues
In general, the standard techniques for solving SRPs can be
categorized into two major groups: greedy algorithms and
algorithms based on convex optimization. In this work we will
use the popular greedy Orthogonal Matching Pursuit (OMP)
algorithm (see Algorithm 1) and the convex optimization
algorithm known as LASSO. The LASSO estimator is defined
as follows:
ïżœÌïżœ = đđđđđđïżœÌïżœâđ â đŽïżœÌïżœâ22 + đâïżœÌïżœâ1 . (24)
This can be viewed as the Lagrangian form of the BPDN
problem (24).
Algorithm 1 Orthogonal Matching Pursuit - OMP require: matrix đš, measurements đ , stopping criterion
initialize: đđ = đ, đđ = đ, đ = đ, đșđŒđ·đ
= â repeat
1.match: đđ = đšđ»đđ 2.identify support indicator:
đđđđ
= {đđđđđđđ|đđ(đ)|}
3.update the support:
đșđŒđ·đ+đ = đșđŒđ·đ âȘ đđđđ 4.update signal estimate:
đđ+đ
= đđđđđđđ: đđđđ(đ)âđșđŒđ·đ+đâđ â đšđâđ
đđ+đ = đ â đšđđ+đ
đ = đ + đ Until stopping criterion met
output: ïżœÌïżœ = đđ
a. Structured Sparsity
By inspection of the columns of the Laplacian matrices of
standard IEEE test-beds (Fig.1 and Fig.2) we see that the
columns tend to exhibit a clustered sparse structure: the
column vectors not only contain few nonzero elements (i.e.,
they are sparse), but the nonzero elements tend to occur in
nearly adjacent positions. In the case of the IEEE Standard-14
Bus or IEEE Standard-30 Bus, columns 2, 5 and 6 or columns
6 and 10 can be taken as clear examples, respectively. Since in
a real PN, nodes and lines are ordered based on their position
in the network, and commonly the neighbor nodes are
connected through neighbor lines (except in special cases), it
is realistic in practice to assume that nodes have an ordering
that lends itself to clustered sparsity.
Fig.2. Corresponding Laplacian matrices of two of the standard IEEE test-
beds (Scaled between [-1, 1]).
Definition 4 (Clustered sparsity): We refer to a signal đŠ âđ đŸ
as (S;C)-clustered sparse if the total number of nonzero
coefficients is đ and all nonzero coefficients are distributed
within đ¶ disjoint clusters (the locations and sizes of the
clusters are arbitrary).
Figure 3 shows the difference between the general structure
of an đ-sparse signal (Fig. 3a) and a clustered-sparse signal
(Fig. 3b).
Modified Clustered OMP
Clustered OMP is a generalized version of the traditional
greedy method OMP. COMP modifies the support
identification step of OMP in order to improve the SRP results
whenever we expect clustered sparsity in the signal to be
recovered [20]. In Section IV, we show how COMP offers
improved recovery of clustered sparse signals. The step-wise
description of COMP (from [20]) is given in Algorithm 2.
Laplacian Matrix of Graph # 14
2 4 6 8 1012 14
2
4
6
8
10
12
14 -1
-0.5
0
0.5
1Laplacian Matrix of Graph # 30
10 20 30
10
20
30 -1
-0.5
0
0.5
1
Laplacian Matrix of Graph # 57
10 20 30 40 50
10
20
30
40
50
-1
-0.5
0
0.5
1Laplacian Matrix of Graph # 118
20 40 60 80 100
20
40
60
80
100
-1
-0.5
0
0.5
1
6
Algorithm 2 Clustered Orthogonal Matching Pursuit- COMP require: matrix A, measurements đ , maximum cluster size m, stopping criterion
initialize: đđ = đ, đđ = đ, đ = đ, đșđŒđ·đ
= â repeat
1.match: đđ = đšđ»đđ 2.identify support indicator:
đđđđ
= {đđđđđđđ|đđ(đ)|}
3.extend support
đŹđźđ© Ì đ = {đđđđ â đ + đ, ⊠, đđđđ, ⊠, đđđđ + đ â đ} 4.update the support:
đșđŒđ·đ+đ = đșđŒđ·đ âȘ đŹđźđ© Ì đ 5.update signal estimate:
đđ+đ
= đđđđđđđ: đđđđ(đ)âđșđŒđ·đ+đâđ â đšđâđ
đđ+đ = đ â đšđđ+đ
đ = đ + đ Until stopping criterion met
output: ïżœÌïżœ = đđ
In order to improve the handling of border effects, we make a
slight modification to Step 3 of COMP as follows:
3.1. finding the boundaries of the support
Upper Bound = đđđ((đđđđ + đ â đ), đ))
Lower bound = đđđ(đđđđ â đ + đ), đ”)
3.2. modified extend support:
đđđđ = {đłđđđđ đđđđđ , ⊠, đđđđ, ⊠, đŒđđđđ đ©đđđđ }
We call the resulting algorithm Modified COMP (MCOMP).
b. Data Correlation Issue
Even though graph data analytics are essential for gaining
insight into big networks, large-scale interconnected network
processing is complex because of its graph-specific
challenges, including complicated correlations among data
entities, highly skewed distributions, and so on. Considering
large-scale interconnected power networks, the possible
correlation among the recorded data from the nodes
(especially from neighboring nodes) is an issue in sparse TI
since it may result in highly coherent sensing matrices đŽ. As
noted in Theorem 1, high coherence can have an undesirable
effect on the performance of SRP solvers. Recently, two new
techniques named Bandâexclusion (B) and Local
Optimization (LO) have been developed in [22] in order to
help address the high coherence issue in sparse signal
recovery. The BLO technique known as BLOOMP can be
interpreted as adding an additional constraint on the support
selection step in OMP to prevent the algorithm from picking
highly correlated indexes in the estimated support of the
sparse signal [22]. Within this algorithm a new concept termed
the coherence band is defined. The coherence band (đđ) and
related concepts are defined as follows: for some tolerance
đŒ â„ 0,
đ”đŒ(đ) = {đ| |âšđđ, đđâ©| > đŒ} (25)
đ”đŒ(đŹ) =âȘđđđŹ đ”đŒ(đ) (26)
đ”đŒ(2)(đ) = đ”đŒ(đ”đŒ(đ)) =âȘđđđ”đŒ(đ) đ”đŒ(đ) (27)
đ”đŒ(2)(đŹ) = đ”đŒ(đ”đŒ(đŹ)) =âȘđđđŹ đ”đŒ
(2)(đ). (28)
However, there is an issue with applying such a method in the
sparse TI problem. Due to nature of this method, whenever the
correlation between two columns of the sensing matrix is high
one of the corresponding positions will not be selected within
the true support. On the other side, as it has been mentioned
before, in TI the positions of the nonzero elements of a signal
vector are typically clustered. Since correlation in a sensing
matrix tends to be highest for columns corresponding to
neighboring nodes, this means that applying the BLOOMP
method may cause us to miss some nonzero elements of the
signal vector. In order to deal with this problem, we suggest an
algorithm composed of BLOOMP and MCOMP that we refer
to as BLOMCOMP (Algorithm 3); this algorithm adds the pre-
knowledge of clustered sparsity to BLOOMP. This
modification prevents us from losing nonzero neighbor
elements within a cluster. In addition to band-exclusion, LO is
used in BLOMCOMP as a residual-reduction technique (for
more information please refer to [22]).
Algorithm 3 Band-excluded Locally Optimized MCOMP
(BLOMCOMP) require: matrix đš, measurements đ , Coherence band đ¶ > đ, maximum
cluster size đ, stopping criterion
initialize: đđ = đ, đđ = đ, đ = đ, đșđŒđ·đ = â
repeat
1.match: đđ = đšđ»đđ 2.identify support indicator:
đđđđ
= {đđđđđđđ|đđ(đ)|} , đ â đ©đ¶(đ)
(đșđŒđ·đâđ)
3.update the support:
Upper Bound = đđđ((đđđđ + đ â đ), đ))
Lower bound = đđđ(đđđđ â đ + đ), đČ)
đđđ Ì đ = {đłđđđđ đđđđđ , ⊠, đđđđ, ⊠, đŒđđđđ đ©đđđđ } đșđŒđ·đ+đ = đłđ¶(đșđŒđ·đ âȘ {đđđ Ì đ}) 4.update signal estimate:
đđ+đ
= đđđđđđđ: đđđđ(đ)âđșđŒđ·đ+đâđ â đšđâđ
đđ+đ = đ â đšđđ+đ
đ = đ + đ Until stopping criterion met
output: ïżœÌïżœ = đđ
Local Optimization Procedure require: đš, đ , Coherence band đ¶ > đ, đșđŒđ·đ = âȘ {đđđ Ì đ} đđđ đ = đ: đ+1
repeat: for đ = đ: đ+1
1. đđ
= đđđđđđđ:đđđđ(đ)=(đșđŒđ·đâđ \{đđđ Ì đ}) âȘ {đđ},đđđđ©đ¶({đđđ Ì đ})âđ â đšđâđ
2. đșđŒđ·đ = đđđđ(đđ)
output: đșđŒđ·đ+đ
Fig.3 (a) Structure of a 6-sparse signal. (b) A (25;5)-clustered-sparse signal.
(c) The average value of Coherence đđŽ vs # of measurements đ over all 118
corresponding sensing matrices 'đŽ' for all of the nodes in the IEEE 118-Bus.
0 20 40 60 80 100 120 140 160 180 2000
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Mag
nitu
de
0 20 40 60 80 100 120 140 160 180 2000
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Index
Mag
nitu
de
a) b)
c)
7
IV. SIMULATION RESULTS AND DISCUSSION
A. Setup
In this section, we test the proposed method for recovering
the topology of an SG using compressive observations (đ <đŸ) collected from the system parameters. We see that the
probability of successful recovery varies for different nodes in
the network based on the local pattern of sparsity. In these
simulations, we use the IEEE Standard-30 Bus, 118 Bus and,
2383 as case studies. These three PNs include 30 Buses and 47
power transmission lines, 118 Buses and 186 lines, and 2383
Buses and 2896 lines, respectively, and their detailed
specifications have been fully described in the MATPOWER
toolbox [16]. To generate the data, we first feed the system
with Gaussian demand and simulate the PN. The
MATPOWER [16] toolbox is used for solving the power flow
equations in various demands and the resulting phase angle
and active power measurements are applied as the input to the
SRP. Finally, white noise has been added to measurement
vectors to set the SNR in the range of 10-20 dB randomly.
Each column of the matrix đ” represents one of the sparse
vectors đŠđ which we attempt to recover by solving an SRP. For
the IEEE Standard-118 Bus, Fig. 3c shows how the coherence
(19) of the corresponding sensing matrix đŽ changes as the
number of measurements đ increases; curves are averaged
over 1000 realizations of the network and over all 118 nodes.
This coherence measure appears to approach a non-zero
asymptote as the number of measurements increases. As has
been mentioned before, the smaller the coherence metric, the
larger the sparsity đ of signals that can be recovered. One
class of matrices well known to have low coherence are
random Gaussian matrices. As mentioned in Section II-C and
in the literature such as [6]-[8], due to the load uncertainty and
the aggregation of the renewables in SGs, the difference of the
phase angles across a bus can be approximated by a Gaussian
random variable. Moreover, in the SRP formulation in Section
III-A, the sensing matrix is formed by the variation in bus
phase angles. Thus, the closer the behavior of the phase angles
is to being Gaussian, the better one might expect the sparse
recovery algorithms to perform.
B. Results
Within the network graph, SRP solvers exhibit different
recovery performance on different nodes for a given number
of measurements, mainly because of the following two factors.
The first factor is the sparsity level of the signal đŠđ (or in our
PN, the in-degree, or the number of connected links to a bus),
and the second factor is the presence of a possible structure in
the position of the non-zero components. Figs.4-7 show the
recovery performance of the OMP (green), LASSO (blue) and
MCOMP (red) algorithms over 1000 realizations of three
IEEE networks used as case studies. In the first case, the IEEE
standard-30 Bus system has been used. In Figs.4a and 4b,
recovery results are illustrated for nodes 3 and 10
(corresponding to the easiest and the hardest cases in the IEEE
30 bus network), respectively. At each specific number of
measurements đ (đ đ {4, ⊠,60}), the vertical axis shows the
percentage of trials in which perfect recovery is achieved, over
1000 realizations of the network. Nodes 3 and 10 are
distinguished from each other by their in-degrees. We see that
node 3, with in-degree 2, is more frequency recovered
correctly than is node 10, which has in-degree 6. This is to be
expected: in the corresponding SRPs, the signal đŠ3 has a
sparsity level of 2, while the signal đŠ10 has sparsity 6; signals
with higher sparsity levels are, in general, more difficult to
recover.
Fig.4. Recovery rate comparison of nodes 3 (a) and 10 (b) in the IEEE 30 bus system. These nodes have in-degree 2 and in-degree 6, respectively.
Another important factor, which affects the SRP solverâs
performance, is the existence of a possible clustered structure
in the sparse signals to be recovered. As we have mentioned,
MCOMP is an extended version of OMP, which has been
adapted to exploit the knowledge that the non-zero entries of
the signal appear in clusters. The structured sparsity in the
columns of the matrix đ” can be used to improve the recovery
performance. Figs.4a and 4b show how the presence of
clustered sparsity helps the MCOMP algorithm to outperform
the LASSO and OMP, especially in the case of node 10, where
đŠ10 is a (6, 4)-clustered sparse signal and the in-degree of the
node (or the number of nonzero elements of the signal to be
recovered) is larger.
In order to compare the performance over more complicated
cases, the IEEE Standard-118 Bus and 2383 Bus systems have
also been used for modeling the PN. In the IEEE 118 bus, we
focus on recovering the connections to nodes 74 (which has
in-degree 2), 49 (in-degree 9), and 59 (in-degree 6). In the
IEEE 2383 bus, we focus on nodes 3 (in-degree 3) and 1920
(in-degree 9). From the sparsity level viewpoint, đŠ49 and đŠ1920
correspond to the most complicated signals đŠđ to be recovered
in the two networks, respectively. Recovery results are again
shown over 1000 realizations of the corresponding network.
Figs.5a-c show the recovery results for the IEEE 118 Bus
network. As might be expected, we see that node 74 (with in-
degree 2) is more likely to be recovered from a given number
of measurements than is node 59 (in-degree 6). Node 49 (with
in-degree 9) is the most difficult to recover. These plots also
show how MCOMP outperforms LASSO and OMP in the case
a)
b)
8
of node 49, where đŠ49 is a (9, 5)-clustered sparse signal and
the in-degree of the node (or the number of nonzero elements
of the signal to be recovered) is larger. On node 59 (where the
corresponding signal đŠ59 is a (6, 3)-clustered sparse signal)
MCOMP again outperforms OMP and LASSO. On node 74
(where đŠ74 has no clustered sparsity), LASSO is the best
performing algorithm.
Fig.5. Recovery rate comparison of nodes 74, 59, and 49 in the IEEE 118-Bus,
with in-degree 2, 5, and 9 respectively.
Moreover, Fig.6.a and b are representing the recovery
performance of each of 3 SRP solvers for 3-sparse node 3 and
9-sparse node 1920 in IEEE 2383 Bus network. Due to the
larger scale of this network, we generally need more
measurements to achieve the perfect recovery for any
individual node compared to smaller networks like the IEEE
118 Bus. As it has been mentioned before, the number of
measurements needed for full recovery depends on both the
sparsity level and the original dimension of the signal đŠ (here,
the dimension đŸ equals the number of buses in our sparse TI
problem); specifically, đ must be at least proportional to
đ đđđ(đŸ/đ). As a rough rule of thumb, in many applications,
when the sensing matrix has low coherence, one observes
perfect recovery when đ â 4đ. However, as the dimension đŸ
grows, it becomes necessary to take more measurements. For
example, in the case of the IEEE 118 Bus, node 49 has in-
degree 9, and perfect recovery is possible when đ â 50. In
the case of the IEEE 2383 network, node 1920 again has in-
degree 9, but perfect recovery is not possible until đ â 80.
Fig.6. Recovery rate comparison of nodes 3 and 1920 in the IEEE standard 2383 network. Nodes 3 and 1920 have in-degree 3 and 9, respectively.
Comparing the results from different size networks it is
clear that number of measurements required for perfect
recovery is affected less by the network size than by the
sparsity level of the corresponding signal đŠđ to be recovered.
Taking fewer measurements reduces the recovery performance
of any SRP solver; however, the LASSO and MCOMP
generally show better recovery performance than OMP.
Moreover, we cannot recover the nodes with higher in-degrees
(such as node 49, where đŠ49 is a (9, 5)-clustered sparse signal)
until đ is large enough that the PN coherence metric reaches a
suitably small level, and if such a level were never to be
reached, we might be limited in the degree of nodes that we
could recover with these techniques.
Finally, in order to highlight the effect of the data
correlation issue in addition to the impact of clustered sparsity,
we have examined the recovery results for two other clustered
sparse signals in the IEEE 118 bus: signal đŠ37 which is a (6,
2)-clustered sparse signal, and signal đŠ80 which is an (8, 4)-
clustered sparse signal. In general, as one should expect đŠ37 to
be easier to recover compared to đŠ80. In order to evaluate the
results we define the interconnection order (đŒđ) for node đ as
follows:
đŒđ(đ) = â đđ â đđđđđđ(đ)đđđźđ (29)
Node 37 has 6 direct neighbors and đŒđ(37) = 17; node 80
has 8 direct neighbors and đŒđ(80) = 22. The interconnection
order can be interpreted as a factor that can reflect the level of
data correlation within a specific bus versus the rest of the
network; however, it will not be the only factor. Fig.7
0 10 20 30 40 50 60 70 80 90 1000
10
20
30
40
50
60
70
80
90
100
Recovery Performance of 3 SRP solvers over 1000 realization of the network for node 74
Number of Measurments "M"
Reco
very
Perf
orm
an
ce (
%)
LASSO
MCOMP
OMP
0 10 20 30 40 50 60 70 80 90 1000
10
20
30
40
50
60
70
80
90
100
Recovery Performance of 3 SRP solvers over 1000 realization of the network for node 59
Number of Measurments "M"
Rec
ov
ery
Per
form
ance
(%
)
LASSO
MCOMP
OMP
0 10 20 30 40 50 60 70 80 90 1000
10
20
30
40
50
60
70
80
90
100
Number of Measurments "M"
Rec
ov
ery
Per
form
ance
(%
)
Recovery Performance of 3 SRP solvers over 1000 realization of the network for node 49
LASSO
MCOMP
OMP
a)
b)
c)
b)
a)
OMP
9
illustrates how BLOMCOMP outperforms MCOMP, OMP
and LASSO in the case of node 80, where we face a more
complicated and interconnected local topology and a higher
coherence in the corresponding sensing matrix.
Fig.7. Recovery rate comparison of nodes 37 and 80 in the IEEE 118-Bus.
An initial value of m = 4 is chosen in MCOMP and BLOMCOMP.
Figs.8a-c illustrate the recovery performance for 3 SRP
solvers bus-by-bus (i.e., for each of the columns of the nodal-
admittance matrix đ”) over 100 realizations of the IEEE 118-
Bus network. The vertical axis represents the number of
measurements đ taken for sparse recovery of each sparse
signal from the following set: {10, 20, 30, 40, 50}. The
horizontal axis shows the bus number from 1 to 118, or
equivalently, the column number from the nodal-admittance
matrix. The recovery performance is illustrated by a color
spectrum ranging from dark blue (corresponding to 0%
recovery) up to dark red (corresponding to 100% recovery).
As can be seen, almost all of the algorithms can reach 100%
recovery performance with 50 measurements. However,
LASSO and MCOMP generally exhibit better performance
when the number of measurements is lower, and MCOMP
generally has better performance than LASSO, especially in
the case of clustered sparse signals. Table I compares the
recovery performance of 3 SRP solvers using 30
measurements for some of the most significant structured
columns in the IEEE 118-Bus system.
Finally, Fig.9 illustrates the whole network topology
recovery performance for 3 SRP solvers for the standard IEEE
118 Bus system, over 1000 realizations of the network. The
vertical axis shows the percentage of trials in which all 118
columns of the nodal-admittance matrix đ” (and, as a result, the
whole network topology) are perfectly recovered. This figure
also shows that the whole network topology can typically be
recovered from 50 or fewer measurements per bus.
Fig.8. Recovery performance for 3 SRP solvers over all of the buses of the
standard IEEE 118 Bus system: (a) OMP, (b) LASSO, (c) MCOMP.
Table. I. Recovery performance of 3 SRP solvers using 30 measurements for
the most significant clustered sparse buses in the IEEE 118-Bus system.
Bus (Column) # OMP (%) LASSO (%) MCOMP (%)
5 90 91 100
11 92 100 98
12 52 54 51
15 85 93 100
17 74 71 88
56 89 92 100
62 92 99 99
85 74 92 99
100 31 32 78
105 90 90 100
Fig.9. Whole network topology recovery performance for 3 SRP solvers
for the standard IEEE 118 Bus system over 1000 realizations of the network.
0 10 20 30 40 50 60 70 80 90 1000
10
20
30
40
50
60
70
80
90
100
Recovery Performance of 3 SRP solvers over 1000 realization of the network for node 37
Number of Measurments "M"
Rec
ov
ery
Per
form
ance
(%)
LASSO
MCOMP
OMP
BLOMCOMP
0 10 20 30 40 50 60 70 80 90 1000
10
20
30
40
50
60
70
80
90
100
Recovery performance of 3 SRP solvers over 1000 realization of the network for node 80
Number of Measurments "M"
Rec
ov
ery
Per
form
ance
(%
)
LASSO
MCOMP
BLOMCOMP
OMP
a)
b) OMP
LASSO MCOMP
BLOMCOMP
b)
c)
a)
10
V. CONCLUSIONS
In this study, we explore the topology identification
problem for smart grids. The DC power flow model, graph
theory, and sparse recovery techniques from compressive
sensing are composed in order to address the following
challenges in topology identification and line outage
localization of transmission lines: 1) the computational
complexity of data collection and analysis procedures in large
scale SGs; 2) the effect of distributed generation systems such
as wind turbines or PV cells; and 3) the uncertainty of system
states and parameters caused by the nonlinear and stochastic
behavior of loads. The proposed method models the SG as an
interconnected graph. Next, using an approximated version of
the AC power-flow model (a DC model) under the
probabilistic power flow regime, the TI problem is
mathematically formulated as an optimization problem.
Finally, due to the sparse nature of this optimization
problem the TI can be interpreted as an SRP, which can
efficiently be solved using SRP solvers such as greedy (OMP,
MCOMP) or optimization based (LASSO) algorithms. We
have discussed important concepts from CS such as coherence
and its behavior in a sparse representation of the TI problem
for IEEE standard networks. Network models have been
generated using the MATPOWER toolbox and standard IEEE
test-beds. The main advantage of the presented method is that
this sparse reformulation of the TI problem relies only on the
measurements from the system parameters and does not need
any a priori information about the topology of the network. In
fact, in our technique, the output of the SRP solver methods is
the structure or topology matrix of the network itself. Case
studies using standard IEEE standard test-beds have shown
that the proposed method represents a promising new strategy
for topology identification, fault detection, and monitoring
using only a small set of observations from some of the bus
parameters. This paper shows that the recovery performance
of the SRP solvers is mainly dependent on the in-degree
(number of lines connected) of each bus in the network.
Moreover, the sparse signals to be recovered often exhibit a
clustered-sparse structure, so it is possible to improve the
recovery results using SRP solvers that encourage structured
sparsity. We have also addressed the problem of correlation in
the sensing matrix among columns that are likely to belong to
the signal support.
REFERENCES
[1] X.Y.Yu, C. Cecati, T. Dillon, and M. G. SimĂ”es, âThe new frontier of
smart grids,â IEEE Ind. Electron. Mag., vol. 5, no. 3, pp. 49â63, Sep. 2011.
[2] V. Gungor, D. Sahin, T. Kocak, S. Ergut, C. Buccella, C. Cecati, and G.
Hancke, Smart Grid Technologies Communications Technologies and Standardsâ IEEE Transactions on Industrial Informatics, vol.7, no. 4,
pp. 529-539, Sept. 2011.
[3] Weilin Li, Mohsen Ferdowsi, Marija Stevic, Antonello Monti, and Ferdinanda Ponci, âCosimulation for Smart Grid Communicationsâ,
IEEE Transactions on Industrial Informatics, vol. 10, no. 4, Nov 2014.
[4] X. Fang, S. Misra, G. Xue, and D. Yang, âSmart grid-The new and improved power grid: A survey,â IEEE Commun. Surveys Tutorials, vol.
14, no. 4, pp. 944â980, Fourth Quarter 2012.
[5] Harirchi, Farnaz; Simoes, M.Godoy, Al Durra, Ahmed and Muyeen, S.M. "Short transient recovery of low voltage-grid-tied DC distributed
generation". ECCE, 2015 IEEE, 20-24 Sept. 2015, pp: 1149 - 1155,
Montreal, QC, Canada
[6] H. Sedghi, E Jonckheere. Information and Control in Networks. In
Giacomo Como, Bo Bernhardsson, Anders Rantzer, editors, Lecture Notes in Control and Information Sciences, vol. 450, 2014, Springer
Cham Heidelberg New York Dordrecht London, Ch. 9, pp 277-297.
[7] M. He and J. Zhang. âA dependency graph approach for fault detection and localization towards secure smart gridâ. IEEE Transactions on
Smart Grid, no. 2, pp. 342â351, June 2011.
[8] A. Schellenberg, W. Rosehart, and J. Aguado, âCumulant-based probabilistic optimal power flow (P-OPF) with Gaussian and Gamma
distributions,âIEEE Trans. on Power Systems, vol. 20, pp. 773â781,
May 2005. [9] V. Kekatos, G. B. Giannakis, and R. Baldick, âGrid topology
identification using electricity prices,â in Proc. IEEE Power & Energy
Society General Meeting, National Harbor, MD, Jul. 2014. [10] H. Zhu and G. B. Giannakis, âSparse Overcomplete Representations for
Efficient Identification of Power Line Outagesâ IEEE Trans. on Power
Systems, vol. 23, pp. 1644-1652, Oct. 2012. [11] J.C. Chen, W.T Li, C.K. Wen, J.H. Teng, P. Ting. âEfficient
identification method for power line outages in the smart power gridâ.
IEEE Trans. Power Systems, vol. 29, pp. 1788â1800, 2014. [12] Zhao and W.-Z. Song, âDistributed power-line outage detection based
on wide area measurement system,â Sensors, vol. 14, no. 7, pp. 13 114â
13 133, Jan. 2014. [13] J. Wu, J. Xiong, and Y. Shi, âEfficient location identification of multiple
line outages with limited PMUs in smart grids,â IEEE Trans. Power
Syst., vol. 30, no. 4, pp. 1659-1668, Jul 2015. [14] Wen-Tai Li, Chao-Kai Wen, Jung-Chieh Chen, Kai-Kit Wong, Jen-Hao
Teng, and Chau Yuen. âLocation Identification of Power Line Outages Using PMU Measurements with Bad Dataâ. arXiv:1502.05789v1
[cs.SY] 20 Feb 2015
[15] D. B. West, Introduction to Graph Theory, 2nd ed., Prentice Hall, 2001. [16] R. D. Zimmerman, C. E. Murillo-Sanchez, and R. J. Thomas, âSteady-
state operations, planning and analysis tools for power systems research
and education,â IEEE Trans. on Power Systems, vol. 26, no. 1, pp. 12â19, Feb. 2011.
[17] A. J. Wood and B. F. Wollenberg, Power Generation, Operation, and
Control, 2nd ed. New York: Wiley, 1996. [18] E.J. Candes and M.B. Wakin, âAn introduction to compressive sam-
`plingâ, IEEE Signal Processing Magazine, vol. 24, pp. 21â30, Mar
2008.
[19] M. B. Wakin. Compressive Sensing Fundamentals. In M. Amin, editor,
Compressive Sensing for Urban Radar, pp. 1-47. CRC Press, Boca
Raton, Florida, 2014. [20] B.M. Sanandaji, T.L. Vincent, and M.B. Wakin. âCompressive
Topology Identification of Interconnected Dynamic Systems via
Clustered Orthogonal Matching Pursuitâ, in Proceedings of the 50th IEEE CDC-ECC 2011, Orlando, Florida, USA, Dec. 2011, pp. 174â180.
[21] M. Babakmehr, M.G. SimÔes, A. Al Durra, F. Harirchi and Qi. Han.
âApplication of Compressive Sensing for Distributed and Structured Power Line Outage Detection in Smart Gridsâ, IEEE, ACC 2015,
Chicago, Illinois, USA, July 2015. p. 3682-3689.
[22] A. Fannjiang, W. Liao. "Coherence PatternâGuided Compressive Sensing with Unresolved Grids". SIAM J. IMAGING SCIENCES, vol. 5,
no. 1, pp. 179â202, 2012.
[23] Debomita Ghosh, T. Ghose, and D. K. Mohanta. âCommunication Feasibility Analysis for Smart Grid with Phasor Measurement Unitsâ.
IEEE Transactions on Industrial Informatics, vol. 9, no. 3, pp 1486-
1496, Aug 2013.
[24] Ayman I. Sabbah, Student Member, IEEE, Amr El-Mougy, and
Mohamed Ibnkahla, A Survey of Networking Challenges and Routing
Protocols in Smart Grids, IEEE Transactions on Industrial Informatics, vol. 10, no. 1, pp 210-221, Feb 2014.
[25] David M. Laverty, Robert J. Best, Paul Brogan, Iyad Al Khatib, Luigi
Vanfretti, and D. John Morrow, âThe OpenPMU Platform for Open-Source Phasor Measurementsâ IEEE Transactions on Instrumentation
and measurements, vol. 62, no. 4, Apr 2013.