See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/367483917

A Scalable PDC Placement Technique for Fast and Resilient Monitoring of Large
Power Grids

Article in IEEE Transactions on Control of Network Systems · December 2023

DOI: 10.1109/TCNS.2023.3240200

CITATION READS
1 69

5 authors, including:

Md Zahidul Islam Shamsun Nahar Edib

New York University University of Massachusetts Lowell
12 PUBLICATIONS 19 CITATIONS 7 PUBLICATIONS 39 CITATIONS

SEE PROFILE SEE PROFILE

Vinod Vokkarane Xiaoyuan Fan

University of Massachusetts Lowell Pacific Northwest National Laboratory
201 PUBLICATIONS 4,092 CITATIONS 52 PUBLICATIONS 229 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Md Zahidul Islam on 05 February 2023.

The user has requested enhancement of the downloaded file.

 This article has been accepted for publication in IEEE Transactions on Control of Network Systems. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCNS.2023.3240200

A Scalable PDC Placement Technique for

Fast and Resilient Monitoring of
Large Power Grids
Md. Zahidul Islam, Student Member, IEEE, Shamsun N. Edib, Student Member, IEEE, Vinod
M. Vokkarane, Senior Member, IEEE, Yuzhang Lin, Member, IEEE, and Xiaoyuan Fan,
Senior Member, IEEE

Abstract— The Wide Area Measurement System (WAMS) phasor data concentrator (PDC), and associated communication
is a key enabler of real-time monitoring of power grids. The network (CN). With the help of Global Positioning System
essential goals of WAMS design are fast and resilient data (GPS), PMUs installed at substations measure the real-time
transfer from phasor measurement units (PMU) to phasor
operating state of the power grid in a synchronized manner. The
data concentrators (PDC). We propose a scalable two-stage
PDC placement technique for minimizing the end-to-end measurement data are then transmitted via the CN to PDCs
delay while maintaining resiliency. In the prescreening located at central substations or control rooms, which align and
stage, the plausible candidates of PDC configurations are process the PMU measurements according to the GPS time
identified based on a graph theory-based multi-median stamps. Typically, for larger power grids, there may be multiple
function (MMF). A computationally efficient meta-heuristic regional PDCs and one super PDC in the control center.
algorithm is used to address scalability. In the candidate
The deployment of WAMS in power grid requires significant
selection stage, two different algorithms, namely
Suurballe’s and Dijkstra’s, are employed to identify the best amount of capital investments. Over the years, researchers have
of those plausible PDC configurations as the final design. investigated the optimal placement of WAMS components in
This technique not only minimizes the hop paths between the power grid. Various methods have been presented in the
PMUs and PDCs, but also ensures network resiliency literature regarding the optimal PMU placement in power grid,
against single PMU, PDC, or communication link failure by such as integer linear programming (ILP), greedy algorithm,
incorporating the roles of PMUs in power grid observability genetic algorithm, tabu search, and binary particle swarm [3-5].
into routing policy. Simulation results on the IEEE 57-bus They aim at minimizing the PMU placement cost while
test power system and the 2000-bus test power system guaranteeing the full observability of the power grid based on
demonstrate the effectiveness and scalability of the
resiliency requirements [6] [7]. However, PMUs along cannot
proposed technique.
make the system observable, as it requires associated CN to
Index Terms—Wide Area Management System (WAMS), transfer the measurements to the PDCs [8]. Hence, concurrent
Phasor Data Concentrator (PDC), Phasor Measurement Unit optimal PMU and CN placement has drawn the attention of the
(PMU), power grid monitoring, smart grid, communication researchers in recent years [9-11]. Ref [10] proposes a
network, resiliency. simultaneous PMU and CN placement method incorporating
the bandwidth requirements. Ref [11] takes one step further to
I. INTRODUCTION incorporate the impact of PDC location in the optimal PMU-

I N recent years, electric power grids are becoming more CN placement problem and minimizes the overall cost by
employing the binary imperialistic competition algorithm and
diversified with renewable energy generations, energy
Dijkstra’s algorithms. However, it considers only one PDC in
storage systems, and controllable loads, which require real-time
the system, whereas, large-scale power grids are likely to have
monitoring functions with high resolution. To enhance the grid multiple PDCs [12] [13].
operator’s situational awareness, the wide area measurement As a pivotal component in WAMS, PDC performs many
system (WAMS) are being deployed in the power grid [1] [2]. fundamental yet crucial tasks such as time-alignment of the
The WAMS comprises of phasor measurement unit (PMU), measurements, data compression, bad data rejection, and local
This work relates to Department of Navy award N00014-20-1-2858 data recording [14-16]. PDC locations also have great impact
issued by the Office of Naval Research. The United States Government on the communication network latency, reliability, bandwidth,
has a royalty-free license throughout the world in all copyrightable and traffic rate [17] [18]. However, unlike the optimal PMU
material contained herein.
M.Z. Islam, S.N. Edib, V.M. Vokkarane, and Y. Lin are with the and CN placement problem, the PDC placement problem has
Electrical and Computer Engineering Department, University of not been sufficiently investigated in the literature. Recently,
Massachusetts Lowell, Lowell, MA 01854 USA (e-mail: Ref. [17] presents a PDC placement technique which places the
[email protected]; PDCs at optimal locations over the grid in order to minimize
[email protected]; [email protected];
[email protected]) traffic in the network. However, this technique is not resilient
X. Fan is with Pacific Northwest National Laboratory, Richland, WA against any device failure and does not provide any detailed CN
99352 USA (e-mail: [email protected]) configuration to demonstrate how the installed PMUs and

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIV OF MASS-LOWELL. Downloaded on January 31,2023 at 00:02:02 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Control of Network Systems. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCNS.2023.3240200

PDCs communicate. In [18], PDCs are regionally placed such 3) Resilient against failure of any single component in
that the relative locations of installed PMUs and PDCs WAMS (PMU, communication links, PDC);
maximize the reliability of the communication network 4) Facilitative to the timely and feasible execution of state
considering PMU and power line failure. However, the estimation for real-time power grid monitoring.
technique applied brute-force approach in placing PMUs and The rest of the paper is organized as follows: the problem
PDCs in the network and does not consider data transfer delay, description is given in Section II. The proposed PDC placement
an important parameter in the network while placing the PDCs. technique and its complexity analysis are described in Section
In addition, existing techniques take into account the resiliency III and IV, respectively. The case study results are presented
only in terms of the availability of the devices, but do not and discussed in Section V. Section VI concludes this paper.
recognize the actual effects of devices’ availability on power
grid observability. Furthermore, most of the methods use ILP II. PROBLEM STATEMENT
for the PDC placement and therefore, are not scalable to large Based on the motivation described in Section I, the PDC
power grids. placement problem addressed by this paper is specifically
For the dynamic monitoring and control of power grids, it is defined as follows.
vital to run the state estimation function periodically to track the Input: Suppose the power grid topology is given. A PMU or
state variables of the system [19] [20], and the execution of the a PDC can be installed at any buses in the power grid. The PMU
synchrophasor data based linear state estimation function configuration, i.e., the set of buses that are equipped with
requires real-time measurements from PMUs that can make the PMUs, is given, which makes the power grid observable even
power grid observable in high resolution [21] [22]. For timely under any PMU failure. The potential communication network
completion of state estimation, PDCs buffer measurements configuration, i.e., the set of pairs of buses between which a
from multiple PMUs with the same time stamp, up to a certain communication link can be installed, is given.
time (known as PDC wait-time) [16]. When the buffer is full or Objective: Determine a PDC configuration, i.e., a set of
the wait-time is over, PDCs forward the received measurements buses to be equipped with PDCs, which allows building a
to the state estimation function. If the delay in the data transfer communication network topology that 1) minimizes the
from a PMU to the PDC exceeds the PDC wait-time, the maximum and average hop count between PMUs and PDCs; 2)
measurement data will be dropped immediately at the PDC keeps the power grid observable by the PMUs remaining
without being used by the state estimation function [23] [24]. connected to PDCs under the loss of any single PMU,
The loss of PMU data may make the system unobservable, communication link, or PDC.
disabling the execution of state estimation. Therefore, fast data An effective method should be developed to solve the PDC
transfer between PMUs and PDCs is necessary to ensure the placement defined above for real-world-size power grids with
observability crucial for power grid operation. One strategy to tractable computational efficiency.
facilitate fast data transfer is to minimize the hop count between
the source and the destination, because the end-to-end delay is III. PROPOSED PDC PLACEMENT TECHNIQUE
positively correlated to the hop count [25]. In WAMS, the
average hop count between all PMUs and PDCs should be The proposed PDC placement technique constitutes several
minimized to reduce the overall delay in data transfer. As a PDC steps as shown in Fig. 1. First, the power grid buses having
attempts to wait for the data from all PMUs, it is also necessary higher potential for PDC placement are identified by employing
to receive data from the farthest PMU within a certain delay, graph centrality, where the power grid is represented as a bi-
which can be fulfilled by minimizing the maximum hop count directed graph. At the same time, the optimal PMU
among all PMUs and PDCs. Thus, the optimal PDC placement configuration is obtained to ensure the full observability of the
based on average hop and maximum hop counts among the power grid ensuring resiliency requirements. Then, possible
PMUs and PDCs is worth further investigation. PDC configurations are evaluated based on a multi-median
In this paper, we propose a resilient PDC placement function (MMF) by applying a meta-heuristic for plausible
technique to facilitate data transfer between PMUs and PDCs PDC configurations (prescreening stage). Finally, two shortest
by minimizing both the average hop and the maximum hop path algorithms, Suurballe’s and Dijkstra’s, are employed to
between them. More specifically, we first screen preferable find detailed hop metrics for each plausible PDC configuration,
buses for the placement of the PDCs using graph theory-based and the optimal configuration is selected based on some hop
metrics. Then, we employ a multi-median function to identify a specified criteria to ensure fast observability (final selection
set of plausible candidates of PDC configurations. In this stage, stage). During this process, PMUs are considered differently
we apply a meta-heuristic to evaluate the multi-median based on their roles in power grid observability, which allows
function, wherein the brute-force approach would be infeasible resiliency against various types of component failures. Note
to implement especially in large systems. Finally, the best that multiple PDC placement problem is combinatorial in
candidate is obtained by using two shortest path finding nature and the possible PDC configurations to be evaluated in
algorithms. In this process, different PMUs are considered the prescreening stage increase substantially with the increase
differently based on their roles in power grid observability of power grid buses. Therefore, a meta-heuristic i.e., Genetic
analysis. Compared with existing literature, our main Algorithm (GA) is used in the prescreening stage instead of the
contributions are summarized as below: brute-force approach to explore the solution space more
1) Applicable to real-world cases of multiple PDCs; thoroughly to find plausible PDC configurations. The details of
2) Scalable to very large power grids where the solution the steps are described below.
space becomes prohibitively high-dimensional;

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIV OF MASS-LOWELL. Downloaded on January 31,2023 at 00:02:02 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Control of Network Systems. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCNS.2023.3240200

Power grid In order to choose the PDC locations, the nodes having
topology
Grid connect.
higher degree and closeness are more desirable. A threshold for
Graph, G(V, E) matrix, A degree and closeness centrality should be set to control the
Candidate PDC location
PMU placement and role
number of candidate nodes selected for the PDC locations. The
screening based on graph
centrality
selection threshold for the degree centrality can be set to one, because the
PDC
network will be less reliable if a PDC is placed at a node with
ePMU, rPMU
locations, V’ only one edge. The threshold for closeness should be set as per
Generation of
Genetic the desirable size of the candidate set. A lower threshold close
Algorithm (GA)
Initial PDC to the minimum closeness of the graph will generate a larger
configurations candidate set, whereas a higher threshold does the opposite. The
CN connect.
matrix, D
computational efficiency increases with smaller candidate set,
Crossover,
but at the risk of missing good candidates. After this screening
New configurations
Fitness evaluation
based on (6)
Mutation, Elite step, a set of candidate PDC locations, V’ is obtained from the
children full set of nodes V of the graph G.
No B. PMU Role Selection
Stopping
Selection
criteria In a power grid, the voltage phasors of all the buses are
Configurations selected
Yes
by GA
defined as state variables. Once they are obtained, all other
variables of the system can be computed. The sufficiency of
Plausible configurations measurements to recover all the state variables is defined as
Prescreening screening with minimal power grid observability. PMUs can be placed over the power
Stage fitness
grid to collect enough measurements to achieve the
Plausible configurations observability. As installing a PMU at a bus makes the bus and
Final Selection Stage Resilient its incident buses observable, it is unnecessary to place PMUs
routing policy at all the buses to obtain full observability [28]. As mentioned
Final configuration
selection based on hop
Hop metrics
Calculation of detailed hop
metrics using Dijkstra’s
earlier, extensive research has been performed for optimal PMU
metrics and Suurballe’s algorithm. placement to achieve full observability of the grid. In this
process, PMU failures must be taken into consideration. To
Final PDC make the system observable under single PMU failure,
configuration
redundant number of PMUs should be placed over the system
Fig. 1: Structure diagram of the proposed resilient PDC placement technique. such that each bus is observed by more than one PMU [29].
For a power grid having n buses, the well-known PMU
A. Candidate PDC Location Screening placement problem can be formulated as follows [28-29]:
n
As depicted in Fig. 1, the first step of the PDC placement minimize ∑c x , i i
(3)
technique is the filtering of candidate power grid buses for PDC i =1

placement. To start with, the power grid will be represented as subject to Ax ≥ b,

a bi-directed graph G (V, E), where V, the set of nodes, where x = [x1, x2, …, xn] is a binary decision vector, where an
represents the buses of the power grid and E, the set of edges, entry is 1 when there is a PMU installed at the corresponding
denotes the branches of the grid. bus and otherwise 0, ci is the cost associated to the PMU
After creating the graph G, the candidate PDC locations (i.e., installment at the ith bus, and b =2 ⋅ 1n×1 , which ensures that each
buses in the power grid) can be screened based on several bus is observed by at least two different PMUs, the entries of
centrality measures of the graph. In graph analytics, centrality the grid connectivity matrix A are defined as follows:
is used to measure the importance of various nodes (i.e., buses
1, if i = j;
in the power grid) in a graph. Two important centrality 
measures, degree and closeness, can be used for screening Aij = 1, if bus i and bus j are connected ; (4)
candidate PDC locations in power grids [26]. The degree 0, otherwise.

centrality is the number of edges incident to a node, and the
The generated PMU configuration by (3) is redundant
closeness centrality of a node is the inverse sum of the shortest
because it considers the situation when a PMU fails.
paths to all other nodes [27]. The degree, CD(i) and closeness,
Given the PMU configuration obtained from (3) or using the
CC(i) of a node i can be expressed by (1) and (2), respectively.
n
predetermined PMU configuration in the power grid, the roles
CD (i ) = ∑ eij , (1) of PMUs in power grid observability can be further
j =1 distinguished. Among the placed PMUs, a minimum subset of
j ≠i
PMUs, which are enough to achieve the full observability of
n −1 power grid are defined as the essential PMUs (ePMUs). The
CC (i ) = , (2)
∑ remaining PMUs are defined as the redundant PMUs (rPMUs)
n
d
j =1 ij
that provide backup measurements to achieve observability
where eij denotes the number of the edges between node i and j, when a single ePMU fails. To obtain the ePMU and rPMU sets
dij is the physical distances between node i and j, and n is the from the installed PMUs, an ILP can be formulated as follows:
total number of nodes.

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIV OF MASS-LOWELL. Downloaded on January 31,2023 at 00:02:02 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Control of Network Systems. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCNS.2023.3240200

m j. The MMF for evaluating the fitness of each configuration can

minimize ∑c y ,
i =1
i i
(5) be developed as below [30]:
S0r = {dl1r , dl2 r ,..., dlq r },
subject to A red y ≥ 1,
where y = [y1, y2, …, ym] is a binary decision vector, where an 1
=
entry is 1 when an ePMU is installed at the corresponding bus
S1 ∑ min(S0r ) + λ ⋅ ∀rmax
| ePMU | r∈ePMU ∈ePMU
(min( S0r )),
and otherwise 0, 1 is a column vector of all ones with m 1
=
elements. The reduced connectivity matrix Ared represents the S2 ∑ min(S0r ) + λ ⋅ ∀rmax
| rPMU | r∈rPMU ∈rPMU
(min( S0r )), (6)
connectivity of the buses having a PMU installed. Ared can be
obtained by selecting the rows of A corresponding to the PMU 1
=
installed buses. Though the PMU placement problem (3) is a
S3 ∑ min(S0r ) + λ ⋅ ∀rmax
| ePMU | r∈ePMU 2 nd
(min( S0r )),
∈ePMU 2 nd

well-studied problem, their role selection problem (5) is a Sc= α S1 + β S 2 + (1 − α − β ) S3 ,

unique contribution. It will be utilized to make the system
resilience against any single component failure in WAMS (i.e., where S0 is the set of shortest hop counts from a PMU installed
r

PMU, PDC, communication link), and thus is part of the at bus r to q number of PDCs installed at buses l1, l2, …, lq; r is
proposed PDC placement technique. the index of PMU buses; S1 sums the average and the scaled
The ePMUs and rPMUs are not only used to achieve maximum of the minimum of S0r for all ePMUs, and the
resiliency against single PMU failure, but are also utilized to average and the maximum hop of P1
ePMU can be minimized by
boost the system’s resiliency against single PDC or reducing S ; similarly, the average hop and maximum hop of
1
communication link failure. To make the system resilient P
rPMU and P ePMU can be minimized by reducing S2 and S3,
2
against any single device failure, we will consider two shortest r
disjoint paths, a primary path (P1 ) and a backup path respectively; The min( S0 ) term in S3 finds the second
ePMU 2 nd
(P2ePMU), which connects an ePMU to two different PDCs, and minimum of S0r that represents the shortest hop count of
one shortest path (PrPMU) from the rPMU to a single PDC. The
backup path P2ePMU; λ is a scaler between 0 and 1 to determine
detailed process of obtaining the communication paths will be
the contribution of maximum hop to S1, S2, and S3; the weights
described in Section III-E. Before proceeding to the path setup,
(α, β) determine the contribution of P1ePMU, P2ePMU, PrPMU to the
the PDC configuration (i.e., combination of PDC locations) for
MMF, Sc. α should be set closer to 1 and β should be set closer
gathering the PMUs measurements over the grid should be
to 0, which set the priority of all the paths to the MMF in the
known. For the placement of q PDCs in a system with n buses,
order of: first P1ePMU, second PrPMU, and finally P2ePMU.
the total possible PDC configurations will be n!/((n-q)!q!),
As the MMF only estimates the fitness of each configuration
which increases drastically with the increase of n and q. Due to
to minimize the average and maximum hop paths over P1ePMU,
this combinatorial nature of the PDC placement problem, the
P2ePMU, and PrPMU, overall, it may not result in the exact best
number of solutions to be enumerated for selecting the best
configuration. However, with careful setting of the parameters
configuration is intractable even for very powerful computers.
α and β, the MMF can find a set of plausible PDC
In order to efficiently find the best configuration in large power
configurations among which the best configuration may lie. The
grids, we design a prescreening stage, which will be discussed
accurate determination of the best configuration will be carried
in the following two subsections.
out in our final selection stage using both Suurballe’s and
C. Candidate PDC Configuration Screening Dijkstra’s algorithms in Section III-E. The implementation of
In this subsection, we will describe the prescreening stage of MMF is to drastically reduce the computational burden as it is
the proposed technique: screening plausible PDC more efficient to perform. Without this prescreening stage,
configurations from all possible configurations by employing storing the results of Suurballe’s and Dijkstra’s algorithms over
an efficient multi-median function (MMF). The configurations an astronomical number of configurations require prohibitively
having the best MMF fitness are strong candidates to be high memory space and computational time, which can be
considered in the final selection stage. observed from (6). Direct evaluation of each possible PDC
In the screening process, the MMF should be developed such configuration using Suurballe or Dijkstra requires us to store a
that it can enumerate the average hop and maximum hop along total of six hop count values (average hop and maximum hop)
the paths P1ePMU, P2ePMU, and PrPMU for each configuration. The and detailed communication paths over P1ePMU, P2ePMU, and
contribution of the three paths to the MMF should be PrPMU. In addition, the stored average and maximum hop counts
determined based on the need of real-time power grid of all possible configurations need to be compared to find the
monitoring. In a power grid, the priority is to achieve the best configuration and corresponding CN; whereas the
ePMUs data via P1ePMU, which alone make the system fully proposed MMF only needs to compute and store a single fitness
observable under the normal operating condition. However, as value of each configuration without detail communication
there can be a failure in an ePMU or a link P1ePMU, the average paths. Therefore, filtering the potential candidates based on this
and the maximum hop of the P2ePMU and PrPMU should also be reduces the time and space complexity of the proposed
accounted for in MMF. technique.
Based on the abovementioned policy, for developing the The time complexity for finding shortest paths from source
MMF, we will first construct an n-by-n symmetric CN node to all other nodes using Dijkstra’s algorithm is O(nξlogn),
connectivity matrix, D from graph G. Each entry of the matrix where n is the number of nodes and ξ is the number of edges
D, dij represents the shortest hop count between node i and node connected to each node of the graph G, and as the Suurballe’s

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIV OF MASS-LOWELL. Downloaded on January 31,2023 at 00:02:02 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Control of Network Systems. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCNS.2023.3240200

algorithm requires two iterations of Dijkstra, the same time ranks the population based on the ascending order of their raw
bound applies to the Suurballe. In the prescreening stage, the fitness value.
CN connectivity matrix D is constructed using Dijkstra for the 3) Parent Selection: Parent selection is the process of selecting
n nodes for only once, so the time complexity for the PDC configurations from the current generation. The PDC
construction of D is O(n2ξlogn). In the MMF, as the locations of the parents would be updated to create child PDC
corresponding shortest hop counts between each PMU-PDC configurations for the next generation. Parent selection is very
configuration are selected from D, the time complexity of the crucial to the convergence rate of the GA. The stochastic
MMF is O(nq), where q is the number of PDCs to be placed. universal sampling selection process is used in this paper,
Therefore, the total time complexity of the prescreening stage where parents are created using roulette wheel and uniform
will be O(n2ξlogn+κnq), where κ is the total number of times sampling based on the fitness value and the number of parents
the MMF runs. to be selected. This process is chosen because it gives lower
Although the MMF-based prescreening algorithm is much ranked configurations a chance to be chosen as parents, which
more efficient than direct application of Suurballe’s and increases the diversity of the population [32].
Dijkstra’s algorithms, it is infeasible to run an exhaustive search 4) Crossover: After selecting the parent PDC configurations,
among all possible configurations when the size of the power the PDC locations between two configurations can be combined
grid is large. In order to apply exhaustive search, κ would be to generate a new child PDC configuration, which is known as
equal to the total number of PDC configurations n!/((n-q)!q!), crossover. There are various methods for crossover operation
which grows drastically with the number of PDCs and power such as single point crossover, multipoint crossover, and
grid buses. Such increase in κ results in an exponential increase scattered crossover. Among them, the scattered crossover is
in the time complexity of MMF as the size of the grid gets large. used in this paper because it generates promising PDC
Therefore, a brute-force approach to find the PDC configuration configurations for the proposed technique. In this crossover, the
with minimum MMF fitness is infeasible in large power grids. PDC locations from two parent configurations are combined
One practical approach is using meta-heuristic to explore the based on a random binary vector. The crossover selects the PDC
solution space more thoroughly to find near-optimal locations from the first parent where the binary vector is 1,
configurations. otherwise selects PDC locations from the second parent to
generate the child.
D. Meta-heuristic for Efficient Evaluation of MMF
5) Mutation: Mutation refers to occasional random changes in
The PDC placement problem is a combinatorial optimization any PDC locations of a parent PDC configuration. It is used to
problem where PDCs are placed in the selected buses among all maintain the diversity and explore a broader search space,
the buses. GA is an efficient meta-heuristic for solving large which is necessary for the convergence of the GA. One of the
combinatorial problems [31]. GA is more effective than random commonly used mutation operators (i.e., Gaussian mutation) is
searching and more efficient than exhaustive search for the used in this paper, where chosen PDC locations from the parent
PDC placement problem in very large-scale power grid in the configuration are changed to a random integer between the
real world [32]. The GA is a stochastic search algorithm that upper and lower bound [33]. The random integer is obtained
mimics the process of natural evolution of species. As species from Gaussian distribution with zero mean and a decreasing
evolve, new genetic information is evolved in the chromosome standard deviation with the increasing number of generations.
by exchanging genes (crossover) between the parent 6) Creating Next Generation: The next generation consists of
chromosome as well as by mutation. Starting with a random children obtained by crossover and mutation operation and a
generation of initial population, the similar procedure of few elite (unchanged parents) children from the current
reproduction, crossover, and mutation as species evolution is generation. After getting the next generation, the whole process
followed in GA until the best generation of population is of the GA would be run iteratively until the change in the best
obtained. The detail steps of GA in the context of PDC fitness function value is lower than a tolerance up to a certain
placement are described below. iteration. The iteration process of GA is also shown in Fig.1.
1) Population Initialization: To start with the GA, a population
set (i.e., a set of PDC configurations) need to be initialized. The E. Decision Making Based on Optimal Communication
initial generation of population can be created by selecting Design
random bus locations from all the candidate buses selected from In this subsection, we will describe the final configuration
Section III-A. Here, the PDC configurations resemble the selection stage of the proposed technique, where both
chromosome and the PDC locations of each configuration Suurballe’s and Dijkstra’s algorithms will be employed to find
resemble the genes of genetic evolution. the optimal PDC configuration among the plausible PDC
2) Fitness Function: For each candidate PDC configuration, the configurations. In this process, the communication network
fitness function would evaluate how ‘fit’ the configuration is between the PMUs and the PDCs will be designed such that the
for placing the PDCs in the system. In the proposed technique, system becomes resilient against single PMU, PDC or
the best-fit PDC configuration will have the lowest average and communication link failure.
maximum hop paths between the installed PDCs and the PMUs. To illustrate the concept of resiliency against single PMU,
The fitness function for evaluating the PDC configuration is PDC, or communication link failure with optimal design, first
described in (6) in Section III-C. The fitness value evaluated by consider a simple communication network with several routers
the MMF is referred as the raw fitness, which is converted to a as shown in Fig. 2. There are twelve routers in this network with
range of fitness that is suitable for the parent selection process available communication links between them and another
in GA. The scaling used in the paper is rank scaling, which virtual node that will be used for finding two disjoint paths

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIV OF MASS-LOWELL. Downloaded on January 31,2023 at 00:02:02 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Control of Network Systems. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCNS.2023.3240200

Table I. Risk Scenarios and Coverage.

R2 PDC1
R6
Risk Scenarios Covered by
Single ePMU failure rPMUs
Single link failure of P1ePMU P2ePMU
R1 R3
R7 R10 Primary PDC failure Backup PDC
ePMU
Single rPMU failure
No need to cover, as power grid
Single link failure of P2ePMU or PrPMU
Virtual observability is not affected
R4 RV
Node Backup PDC failure
R8 R11 PDC2

find the shortest path between the rPMU and the virtual node.
After finding the path, the virtual node can be discarded to find
R5
PrPMU. The concept of virtual node is illustrated in Fig. 2. In the
R9 R12
rPMU
PDC3
figure, each of the three PDCs (PDC1, PDC2, and PDC3) is
Primary Path Backup Path
connected to the virtual node via a virtual link. After adding the
virtual node RV, two shortest disjoint paths can be found as (R1,
Fig. 2: PMU-PDC connectivity using the concept of virtual node and obtaining
resiliency against single PDC or communication link failure. R2, R6, RV) and (R1, R4, R8, R12, RV). Now, by discarding the
virtual node RV and the associated virtual links, the primary
between the PMUs and PDCs. Suppose an ePMU and a rPMU path P1ePMU and backup path P2ePMU can be obtained. Similarly,
are installed at those buses with router R1 and R5, respectively, PrPMU can be found as (R5, R9, R12).
and three PDCs (PDC1, PDC2, and PDC3) are placed at buses To find two disjoint paths, Dijkstra needs to be run twice. In
with routers R6, R11, and R12, respectively. To make the the first run, the shortest path can be found. Then, in the second
system resilient against single PMU, PDC, or communication run, a shortest disjoint path can be obtained by removing the
link failure, two shortest disjoint communication paths can be first path from the graph. On the other hand, Suurballe’s
set from an ePMU to two PDCs and one shortest path can be set algorithm finds the disjoint paths simultaneously such that total
from a rPMU to single PDC. The path with fewer hops in the hop count of the primary and backup paths is minimum [34].
two disjoint paths between an ePMU and two PDCs is called Note that, in Fig. 2, the Suurballe’s and Dijkstra’s algorithms
the primary path, and the other is called the backup path. The generate similar disjoint paths. However, as Dijkstra considers
primary path connects an ePMU to the primary PDC, and the only one shortest path at a time, it generates a shorter primary
backup path connects the ePMU to the backup PDC. For path than Suurballe at the expense of a longer backup path.
instance, in Fig. 2, the ePMU has a primary path P1ePMU and a By using the aforementioned network design heuristics, the
backup path P2ePMU, to the primary PDC (PDC1) and the backup final configuration will be selected from the plausible PDC
PDC (PDC3), respectively. Consequently, if a single PDC fails configurations to make the power grid observable with
(either PDC1 or PDC3), the ePMU measurements will be minimum delay even if a device fails. As discussed earlier, to
rerouted to the alternative active PDC. Similarly, if any link in achieve the power grid observability, all ePMUs must become
P1ePMU fails, P2ePMU will transfer the measurements to the available to the PDCs within specified limit of wait-time. There
backup PDC. Note that the primary path P1ePMU has shorter hop is no need to enforce the same requirement for rPMUs;
count (two hops: R1, R2, R6) than backup path P2ePMU, as however, if more rPMU measurements reach the PDCs within
shown in Fig. 2. In contrast to the ePMU, the rPMU needs only the same wait-time limit, the accuracy and bad data detection
one path (PrPMU), because the power grid remains fully capability of the state estimation function will be improved with
observable by the ePMUs when any rPMU or a link in PrPMU the help of the redundant information [35].
fails. In summary, the resiliency against any single device Based on the needs of power grid monitoring, the best PDC
failure is achieved as follows: single ePMU failure is covered configuration is chosen based on three criteria with different
by the other active PMUs; single link failure in P1ePMU by levels of priority, from high to low:
P2ePMU; single primary PDC by backup PDC. Meanwhile, there Criterion 1. Select the candidates with lowest maximum hop of
is no need to cover the failure of backup devices as the primary P1ePMU. This will ensure that all the ePMUs data become
devices along can ensure the observability of the system. The available to the PDCs within a given wait-time limit under
risk scenarios considered in this paper and their corresponding normal operating conditions.
backups are listed in Table I. Criterion 2. If there are multiple candidates with similar
In order to find two disjoint paths (P1ePMU and P2ePMU) from performances under Criterion 1, select the candidates having
an ePMU to the primary and backup PDCs, a virtual node can lowest maximum hop of PrPMU and P2ePMU. The reason is that
be added in the existing graph, G. The virtual node will be when an ePMU or a link of P1ePMU fails, the corresponding
directly connected to all PDCs in the system with virtual rPMUs or P2ePMU will be responsible for maintaining the
communication links (new zero-cost edges in graph G). After observability of the system, respectively. In this case, the
adding the virtual node, we will run the shortest path algorithms maximum hop of PrPMU and P2ePMU will be of concern for the
(Suurballe’s and Dijkstra’s) to find two disjoint paths from the observability requirement.
ePMU to the virtual node. Then, the virtual node and virtual Criterion 3. If there are multiple candidates with similar
links will be discarded, and the remaining parts of paths are the performances under Criterion 2, select the candidate with the
disjoint paths from the ePMU to the primary and backup PDCs. lowest average hop for all paths. This will minimize the overall
To find the shortest path between rPMU-PDC (PrPMU), the delay in the network and improve the performance of state
virtual node can also be considered. In this case, the algorithms estimation.

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIV OF MASS-LOWELL. Downloaded on January 31,2023 at 00:02:02 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Control of Network Systems. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCNS.2023.3240200

IV. COMPLEXITY ANALYSIS Table II. IEEE 57-Bus system: set of ePMUs and rPMUs.

The complexity analysis of the proposed technique is ePMU Set rPMU Set
described for the two stages of the proposed technique below: 1, 4, 9, 15, 20, 24, 25, 28, 29, 32, 36, 2, 6, 12, 19, 22, 27, 31, 33, 34, 37,
1) Computational complexity of the prescreening stage: In the 38, 41, 47, 51, 54, 57 43, 45, 46, 50, 53, 56
prescreening stage, the possible PDC configurations are
parameters are set as below: population size being 100, function
evaluated thoroughly by applying GA based on MMF fitness
tolerance being 10-6, maximum stall generation being 50, elite
described in (6), and the top several configurations are selected
children being 5% of the population, crossover children being
as the plausible ones. As the performance of the GA is measured
80% of the population, and mutation children being from the
by the total number of evaluations of the fitness function during
remaining population. For the MMF, because it is noticed that
the runtime, the complexity of GA can be simplified to the
the maximum hop along each path is almost double to the
complexity of the MMF fitness [36]. However, a CN
corresponding average hop for standard test cases, λ is set to 0.5
connectivity matrix D needs to be constructed once before
so that both average hop and maximum hop equally contribute
running the prescreening stage as mentioned in Section III-C
to S1, S2, S3; α and β are set to 0.7 and 0.2, respectively, so as to
and its complexity can be added to this stage. Therefore, the
prioritize the contribution of the paths to the MMF in the order
computational complexity of the prescreening stage Tc1 is
of P1ePMU, PrPMU, and P2ePMU. The rationale behind the choice of
computed as: Tc1 = Complexity for the construction of matrix D
the parameters will be validated later. The simulation results are
+ Complexity of the genetic algorithm = O(n2ξlogn) +
carried out by using MATLAB R2021a software in a personal
O((Population size, p) × (No. of Generations, g) × O(MMF
computer with 8GB RAM and Intel Core i5-8265U processor.
fitness)) = O(n2ξlogn) + O( p × g × (No. of PMUs, n) × (No. of
PDCs, q)), which implies the following expression: A. Concept Validation on Small Scale Test Case
= Tc1 O ( n 2ξ logn ) + O ( pgnq ) . (7) In this subsection, we will demonstrate the effectiveness of
2) Computational complexity of the final selection stage: In GA in achieving the plausible PDC configurations based on the
the final selection stage, both Suurballe’s and Dijkstra’s MMF. As discussed in Section III-C, the exhaustive search (i.e.,
algorithms are employed to find the optimal PDC configuration brute-force approach) guarantees to find the configurations with
among the plausible PDC configurations found in the the optimal MMF fitness, but it will be infeasible when the
prescreening process. In this process, we find the detailed hop solution space becomes very large. Therefore, the GA is used
metrics among PMUs-PDCs for each plausible PDC to evaluate the MMF and find a set of plausible PDC
configuration using the algorithms and select the best configurations. However, as the GA cannot guarantee
configuration based on hop metrics. Therefore, the optimality as exhaustive search, so we will first evaluate the
computational complexity of the final selection stage Tc2 is performance of the GA with respect to exhaustive search in a
computed as the summation of the two terms: Tc2 = Complexity small test system where exhaustive search is feasible. If the
for finding hop metrics for plausible PDC configurations + MMF fitness values of the plausible configurations obtained by
Complexity for the selection of best configuration = O((No. of the GA are close to those of the configurations obtained
plausible PDC configurations, ρ) × O(shortest hop findings exhaustive search, then the performances of the GA in MMF
among all nodes in a graph) + O(No. of plausible PDC fitness optimization will be verified.
configurations, ρ)), which implies the following expression: For the concept validation purpose described above, the
simulation case of placing two PDCs in IEEE 57-bus test
= Tc 2 O ( ρ n 2ξ logn ) + O ( ρ ) . (8)
system will be carried out. We will first represent the power
Therefore, the overall time complexity of the proposed PDC grid as a bi-directed graph G and compute the degree and
placement technique is Tc1 + Tc2. However, as the selected closeness measures of each node. Based on the degree and
plausible configurations 𝜌𝜌 for the final stage are very few, the closeness, all nodes pass the screening except Node 33. The
complexity of the proposed technique is mostly dominated by screened PDC locations will be, V’= {all nodes from 1 to 57
the complexity of the prescreening stage. except 33}. Before evaluating the MMF, we will generate
optimal PMU locations using (3) and (5). The ePMUs and
V. CASE STUDY rPMUs for the 57-bus system are shown in Table II. After
In this section, we will first evaluate the performance of the obtaining the locations of ePMUs and rPMUs, both the
GA to prescreen plausible PDC configurations based on MMF exhaustive search and the GA are performed to find plausible
on IEEE 57-bus system. Then, we will verify the effectiveness PDC configurations with optimal MMF fitness. The five PDC
and scalability of the proposed technique on a very large power configurations with the best MMF fitness found by both
grid: a synthetic 2000-bus system [37]. For designing the algorithms are shown in Table III. In this case, the
communication network, it is assumed that communication configurations obtained by the GA exactly match those
links can be installed between two nodes only when there is obtained by exhaustive search.
power grid branch between the corresponding buses. The The above analysis is repeated for different numbers of PDCs
proposed technique can also be applied when the (from 3 to 6) and the plausible PDC configurations found by the
communication network topology is different than the power two algorithms are listed in the same table. It can be seen from
grid topology. As two disjoint paths are considered from a Table III that when the number of PDCs are small (2 and 3),
single ePMU to two different PDCs, when installing PMUs in both exhaustive search and the GA yield very similar results.
the system, we prioritize those buses with at least two power When the #PDCs increase (4 and above), there are some
grid branches to avail the disjoint paths. For the GA, the tuning discrepancies between the results of those two algorithms. To

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIV OF MASS-LOWELL. Downloaded on January 31,2023 at 00:02:02 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Control of Network Systems. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCNS.2023.3240200

Table III. Prescreening stage of PDC configurations using both Exhaustive search and GA while placing different number of PDCs in IEEE 57-bus system.
MMF based #PDCs=2 #PDCs=3 #PDCs=4 #PDCs=5 #PDCs=6
Evaluation Ex. search GA Ex. Search GA Ex. Search GA Ex. Search GA Ex. Search GA
Config.1, {13,24}, {13,24}, {9,25,38}, {9,25,38}, {9,25,29,38}, {9,25,29,38}, {6,9,24,32,38}, {6,9,24,32,38}, {9,14,19,27,31 {9,14,19,27,31
Fitness 5.219 5.219 4.409 4.409 3.752 3.752 3.40 3.40 ,37}, 2.925 ,37}, 2.925
Config.2, {24,49}, {24,49}, {13,25,38}, {13,25,38}, {13,25,29,38}, {9,24,32,38}, {9,15,24,32,38}, {9,15,24,32,38}, {9,14,18,27,31 {9,14,19,27,31
Fitness 5.365 5.365 4.480 4.480 3.756 3.759 3.407 3.407 ,37}, 2.938 ,39}, 3.046
Config.3, {11,24}, {11,24}, {9,30,38}, {9,24,38}, {13,29,32,38}, {9,25,28,38}, {3,9,28,32,38}, {4,13,28,32,37}, {9,14,19,28,30 {4,13,29,31,38
Fitness 5.411 5.411 4.493 4.510 3.757 3.764 3.409 3.410 ,37}, 2.954 ,40}, 3.168
Config.4, {9,38}, {9,38}, {9,24,38}, {11,25,38}, {9,24,32,38}, {13,28,31,38}, {3,9,29,32,38}, {6,9,25,32,38}, {9,14,19,29,30 {4,13,24,29,32
Fitness 5.615 5.615 4.510 4.581 3.759 3.794 3.409 3.413 ,37}, 2.960 ,38}, 3.169
Config.5, {9,37}, {9,37}, {11,25,38}, {12,25,38}, {15,25,29,38}, {7,9,32,38}, {4,13,28,32,37}, {3,9,24,32,38}, {15,20,30,49,52 {9,18,25,29,32
Fitness 5.645 5.645 4.581 4.611 3.763 3.794 3.410 3.421 ,56}, 2.965 ,38}, 3.176
Mean Fitness
7.571 6.509 5.808 5.288 4.876
of all Config.
Table IV. Selection of the Final configuration based on Dijkstra’s algorithm (Final Configuration Selection Stage) from Table III.
#PDCs=2 #PDCs=3 #PDCs=4 #PDCs=5 #PDCs=6
Comparison Scope
Ex. search GA Ex. Search GA Ex. Search GA Ex. Search GA Ex. Search GA
{4,13,28,32 {4,13,28,32 {9,14,19,27 {9,14,19,27
Best configuration {24,49} {24,49} {9,24,38} {9,24,38} {13,29,32,38} {13,28,31,38}
,37} ,37} ,31,37} ,31,37}
Max hop counts over
4,5,9 4,5,9 4,5,8 4,5,8 3,4,6 3,4,6 3,3,4 3,3,4 2,3,5 2,3,5
(P1ePMU, PrPMU, P2ePMU)
Total average hop of all
12.842 12.842 9.213 9.213 7.639 7.882 6.743 6.743 6.566 6.566
paths
3 2 1
quantify the degree of discrepancy, we also listed the average
MMF fitness of all possible configurations at the bottom of the
4 17 16
table. Clearly, the MMF fitness of the plausible configurations 15
yielded by the two algorithms are much better than the average, 14 13
and the relative differences between them are almost negligible. 5
18 45 12
6 10
Therefore, though GA may not provide exactly the optimal 19 44
46
49
results as exhaustive search in this case, the suboptimal 47
50
8 7 20
candidates it yields have comparable performances to the 38
48
optimal ones obtained by exhaustive search. 21
To find the best configuration among the plausible 29 51
22 57 11
configurations shown in Table III, the final candidate selection 28
37
39
stage is carried out by using only Dijkstra’s algorithm, as 23
36 43
described in Section III-E. The results are shown in Table IV. 24 40 56
In all of the cases, the GA leads to exactly the same final PDC 27
35
configuration as exhaustive search except when the number of 52 34
PDCs is 4. In this particular case, the optimal configuration by 26
25 42 41
exhaustive search is {13,29,32,38} while the suboptimal 53
configuration by GA is {13,28,31,38}. The MMF fitness of the 30 31 32 33
two configurations are 3.757 and 3.794, respectively, and the 54
average MMF fitness of all possible PDC configurations is 55 9
5.808, as listed in Table III. It is observed that the MMF fitness
of suboptimal configuration is very close to the fitness of the
ePMU rPMU PDC
optimal one. Therefore, the suboptimal configuration is
Power trans. line
expected to have comparable performance to the optimal one.
Comm. link with Power trans. line underneath
Comparing the performance metrics (the max hop counts over
Fig. 3. PMUs to PDCs (No. of PDCs=3) connectivity using the proposed PDC
P1ePMU, PrPMU, P2ePMU and the total average hop of all paths) of placement technique for IEEE 57-bus system.
the two configurations in Table IV, we can observe that both
configurations have the same maximum hop (3,4,5) over the As an illustrative example, the communication paths for
paths (P1ePMU, PrPMU, P2ePMU). The total average hops of all placing 3 PDCs in the 57-bus system obtained using the
paths of the two configurations are 7.639 and 7.882, proposed technique are shown in Fig. 3. The PDCs are placed
respectively, which are very close. Therefore, we can conclude at buses {9,24,38}. The communication links are shown as solid
that GA may generate suboptimal configurations, and they line in the figure. It can be seen that each of the ePMUs has two
perform very closely to the optimal ones attained by exhaustive shortest disjoint paths to two PDCs and each rPMU has one
search. Note that the exhaustive search generates accurate shortest path to one PDC. For instance, an ePMU at bus 57
results at the expense of very high computational time, whereas needs 3 hops (57,39,37,38) over primary path to access the
almost the same solution can be obtained by the GA with much primary PDC at bus 38, and 4 hops (57,56,41,11,9) over backup
lower cost. The scalability advantage of the GA will be further path to access the backup PDC at bus 9. Whereas, an rPMU at
discussed in the next subsection. bus 31 has only one path with minimum 3 hops (31,31,25,24)

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIV OF MASS-LOWELL. Downloaded on January 31,2023 at 00:02:02 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Control of Network Systems. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCNS.2023.3240200

Table VII. Performance metrics of first five plausible PDC configurations obtained in the Prescreening stage as shown in Table VI.
Primary Paths of ePMU Paths of rPMU Backup Paths of ePMU
#PDC Fitness based
Average Hop Maximum Hop Average Hop Maximum Hop Average Hop Maximum Hop
Configuration on MMF (Sc)
Suurb. Dijk. Suurb. Dijk. Suurb. Dijk. Suurb. Dijk. Suurb. Dijk. Suurb. Dijk.
Config.1 9.798 5.002 4.932 9 9 4.885 4.885 10 10 7.587 7.836 14 14
Config.2 10.079 5.226 5.071 10 9 4.958 4.958 10 10 8.253 8.581 15 15
Config.3 10.318 5.156 5.056 10 10 4.984 4.984 10 10 7.865 8.151 14 15
Config.4 10.363 5.226 5.129 10 10 5.09 5.09 10 10 7.932 8.18 16 16
Config.5 10.398 5.251 5.108 10 10 4.964 4.964 10 10 8.315 8.66 15 15

Table V: Performance comparison of the proposed and baseline methods when Table VI. Plausible five PDC Configurations (No. of PDCs=10) for synthetic
the CN topology is different than the power grid topology. ACTIVSg2000 test case (Prescreening stage).
Maximum hop required for Maximum hop required for PDC
Methods Bus Number
100% coverage of ePMUs 100% coverage of all PMUs Configurations
Proposed 3 6 Config.1 {4014,4017,4106,5063,5077,5120,5164,6162,7004,7420}
Degree-based 6 9
Config.2 {3058,4040,4057,5179,5449,5464,6070,6212,7237,7379}
Closeness-based 6 8
Config.3 {3041,3054,4056,4073,5041,5464,6003,6075,6225,7094}
to access the PDC at bus 24. When using Suurballe instead of Config.4 {3037,4014,4020,5047,5073,5143,5164,6096,7002,7420}
Dijkstra for finding the disjoint backup paths, it generates very Config.5 {3058,4064,4073,5097,5164,5179,5464,6165,7231,7356}
similar communication paths for this case, thus they are not
shown here. techniques, we consider the nodes with the highest degree and
Though in the above analysis, it is considered that the closeness centrality as the PDC locations, respectively.
topology of the power grid and CN are identical, which may not To start with, the synthetic ACTIVSg2000 test case [37] is
be the case in practice. Therefore, a test case is simulated for represented by a bi-directed graph and the plausible PDC
the 57-bus system when the power grid and CN topologies are locations set V’ are obtained by screening the nodes based on
different. In this case, a realistic CN topology is synthesized degree and closeness measures as discussed in Section III-A. In
from the given power grid topology by rewiring the this process, 456 nodes having a degree of one and two are
connectivity among power grid nodes without changing its excluded, and 1544 nodes are kept. This process will improve
topological characteristics [26]. In the synthetic CN topology, the computational time by shrinking the solution space of the
it is considered that each node in the power grid has an attached next steps. As a result, the plausible PDC locations will be, V’=
corresponding node (router) in the CN, but the connectivity {2000 nodes} - {456 nodes with lowest degree and closeness}.
between nodes can be different in the power grid and in the CN Subsequently, PMUs are optimally placed in the 2000-bus
[38]. Apparently, when the topology of the power grid and CN system to obtain the full observability with resiliency against
are different, the PMU configuration for ensuring the grid single PMU failure using (3) and (5). It places a total of 1354
observability mentioned in Section III-B needs to be obtained PMUs in the system. Among them, there are 518 ePMUs and
from the power grid topology and the paths among PMUs- 836 rPMUs. After obtaining the locations of ePMUs and
PDCs described in Section III-E need to be obtained from the rPMUs, several plausible PDC configurations are found by the
digraph representing CN topology. The performance GA as described in Sections III-C and III-D. Herein, the results
comparison of the three methods is listed in Table V in terms of for placing 10 PDCs over the grid are shown, as they are
maximum required hop for all the ePMUs/PMUs to access the obtained in a large solution space and can be used to show the
PDCs. The better performance is achieved when 100% scalability of the proposed technique. Several plausible PDC
ePMUs/PMUs access PDCs with lower hop count and it is clear configurations with the optimal MMF fitness are listed in Table
from Table V that the proposed method outperforms the VI.
baselines even if the CN topology is different than to power grid Finally, the average hop and the maximum hop of all the
topology. It is noted that other than the one-to-one node three paths (P1ePMU, PrPMU, and P2ePMU) are calculated using both
connectivity, there can also exist multiple-to-one, one-to- Suurballe’s and Dijkstra’s algorithms. The results for each
multiple connectivity between the two networks [38]. However, plausible PDC configuration are shown in Table VII. The best
this is out of the scope of this paper and would be considered in configuration from these candidates will be selected as the final
future studies. design based on the selection criteria described in Section III-
E. Based on this Criterion 1 (lowest maximum hop of P1ePMU),
B. Scalability Evaluation on 2000-bus Test System while using Suurballe, only Config. 1 will be chosen as best
In this subsection, the proposed PDC placement technique candidate because it has the lowest hop 9 over the path P1ePMU.
will be applied on a synthetic 2000-bus test system [37] and the However, for Dijkstra, both the Config. 1 and Config. 2 have
performance of the technique will be compared with two the lowest maximum hop (9 hops). Therefore, Criterion 2
baseline PDC placement techniques that are solely based on the (lowest maximum hop of P2ePMU and PrPMU) should be used for
degree centrality and the closeness centrality. As the PDC further discrimination between these candidates. Based on this
placement problem with the similar goal as ours is not well- criterion, Config. 1 has a shorter backup hop path than Config.
studied in literature, the above two baseline techniques are used 2. Therefore, Config. 1 is chosen as best configuration using
for comparison. As discussed earlier in Section III-A, the Dijkstra. In fact, we evaluate hundreds of top configurations,
degree and closeness measures are two important parameters to and find that the best configuration always lies within the first
identify the most important nodes in a graph, thus are natural five configurations. For the baseline techniques, the degree-
factors to be considered for PDC placement. In the two baseline based configuration is {1071, 4089, 5033, 6239, 7104, 7130,

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIV OF MASS-LOWELL. Downloaded on January 31,2023 at 00:02:02 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Control of Network Systems. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCNS.2023.3240200

10

Fig. 5. Comparison between Suurballe’s and Dijkstra’s algorithm at finding

primary and backup paths in the proposed technique when pacing 10 PDCs in
synthetic 2000-bus test system.
available paths. On the other hand, the degree and the closeness-
based techniques require 18 and 20 hops, respectively, for the
coverage of all PMUs. Moreover, it can also be observed from
Fig. 4(b) that higher numbers of PMUs access the PDCs in our
technique than the other two techniques within the same hop
count. For instance, using our technique, around 13% more
coverage than degree-based and 27% more coverage of PMUs
than closeness-based techniques are obtained when the hop
Fig. 4. Comparison results among proposed, degree-based, and closeness- counts are 6 and 7. Therefore, the proposed technique not only
based techniques when 10 PDCs are installed in synthetic 2000-bus test transfers PMUs data to PDCs with lowest hops, but also
system. (a) only #ePMUs (%) get access to PDCs through P1ePMU (b) all maximizes the coverage of the PMUs at every hop count both
#PMUs (%) get access to PDCs through P1ePMU, PrPMU, and P2ePMU.
during normal and single component failure conditions.
7159, 7186, 7346, 7366} and closeness-based configuration is The performance comparison between Suurballe’s and
{3048, 5045, 5120, 5239, 5260, 6003, 6045, 6107, 6210, 6292}. Dijkstra’s algorithms for the proposed technique is shown in
For the degree- and closeness-based PDC placement, the Fig. 5. The figure shows the ePMUs coverage along the primary
configurations are selected because they have highest degree and the backup paths. As can be seen, Dijkstra provides better
and closeness measures, respectively. coverage than Suurballe over the primary path, whereas
In order to comprehensively evaluate the performance of the Suurballe provides better coverage over the backup path. This
proposed technique, we calculate the PMU coverage metric, is what we expect from the principles of the two algorithms.
which is the cumulative number of PMUs accessing the PDCs Therefore, if we prioritize the lower primary path, then Dijkstra
within a given hop count. The more the PMUs transfer their should be considered in CN design. However, if we desire
data to the PDCs within a given hop, the higher the coverage lower hop paths overall (both primary and backup), then
will be. Therefore, a PDC configuration providing higher PMU Suurballe provides better CN design.
coverage using less hop count has better performance. In Fig. Finally, in Table VIII, the computational time to run the
4(a), the ePMUs coverage over the primary path is shown; in exhaustive search and GA for evaluating the prescreening stage
Fig. 4(b), all PMUs’ coverage over all available paths are based on MMF for different number of PDCs in several test
shown. The corresponding results for the degree centrality- cases is compared. As seen from the table, the time required to
based configuration and the closeness centrality based evaluate the MMF using exhaustive search while placing 2
configurations are also presented for comparison. As can be PDCs in IEEE 57-bus system, is only 0.087 second. However,
seen in Fig. 4(a), all the ePMUs access the PDC with fewer than if we want to place 10 PDCs in the same system (IEEE 57-bus),
9 hops by using the proposed configuration, whereas the degree then the total PDC configurations will be 3.318×1010 and the
and the closeness-based configuration require 11 hops and 14 required time to run MMF is around 308.89 hours (1.111×106
hops, respectively. Moreover, it can also be observed from the seconds). Again, if we want to place the same number of PDCs
figure that with the increase of each hop, higher numbers of (10 PDCs) in a system with 2000 buses [37], there will be
ePMUs access the PDCs in our technique compared with the 2.759×1026 possible PDC configurations that will take
other two techniques. For instance, in Fig. 4(a), when the hop approximately 1.02×1015 years (3.23×1022 seconds) to evaluate.
count is 5, more than 60% ePMUs have the coverage, whereas Therefore, a brute-force approach to find plausible PDC
in the other techniques, the coverage of ePMUs is around 40%. configurations based on MMF becomes infeasible with the
Therefore, during the normal operating conditions without any increase of the solution space. On the other hand, the GA based
faulty devices, our technique facilitates the ePMUs-PDC data approach, generates efficient plausible PDC configurations
transfer by reducing the maximum hop and the average hop. with much lower computational time. Even for placing 10
However, if any single ePMUs or links in primary path or PDCs on the largest test case (2000-bus), the GA takes around
primary PDC is faulty, the backup paths of ePMUs and the 20 seconds to process the prescreening stage of our technique.
paths of rPMUs to PDCs are required to obtain full The time complexity of the proposed technique mostly depends
observability. Fig. 4(b) shows that our technique needs 14 hops on the prescreening stage and the time complexity of other
to connect all ePMUs and rPMUs to the PDCs using all remaining stages is negligible. In fact, although GA has higher

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIV OF MASS-LOWELL. Downloaded on January 31,2023 at 00:02:02 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Control of Network Systems. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCNS.2023.3240200

11

Table VIII. Time (sec) to run the Exhaustive search and the GA for the prescreening stage based on MMF for different number of PDCs in different test cases.
#PDCs=2 #PDCs=3 #PDCs=4 #PDCs=5 #PDCs=6 #PDCs=10
Test Cases
Ex. search GA Ex. search GA Ex. search GA Ex. search GA Ex. search GA Ex. search GA
IEEE 57-Bus 0.087 1.808 0.916 2.160 10.862 2.762 120.207 3.804 1048.252 3.196 1.111×106 4.184
IEEE 118-bus 0.216 2.167 6.791 2.569 199.663 4.250 4243.734 4.114 74972.64 3.638 1.71×109 9.451
IEEE 300-bus 1.032 2.178 80.508 4.015 2474.566 5.908 1.123×10 5
4.408 4.232×10 6
6.028 2.09×1012 6.218
Synthetic 2000-
139.452 6.771 7.117×10 5
8.861 2.760×10 7
7.408 8.51×10 9
9.971 2.18×10 12
8.296 2.41×10 21
20.177
bus

ePMUs Coverage (100%) All PMUs coverage (100%) further discriminating the plausible configurations based on
20 Suurballe’s and Dijkstra’s algorithms. In the proposed
technique, the resiliency against any single device failure is also
Required Hop Count

15
ensured by setting up a routing policy based on the PMUs’ roles
10
in power grid observability. From the simulation, it is found that
the proposed technique is scalable to cases with multiple PDCs
5 in very large modern power grids. It is also verified that the
applied GA is effective and computationally efficient in
0 identifying plausible candidates in prescreening stage where the
0 0.2 0.4 α 0.6 0.8 1 exhaustive search becomes intractable. Moreover, the
Fig. 6. Parameter α vs. required hop counts for ePMUs coverage and all PMUs performance of the technique is validated by comparing with
coverage, when β is set to a lower value of 0.2 and λ to 0.5. baseline techniques based on degree and closeness centrality
complexity as it explores the search space more thoroughly to metrics only. We find significant improvements in both average
find results closer to the global minimum, the simulation cases hop and the maximum hop to facilitate for data transfer between
given in Section V-B are large enough for representing today’s PMUs and PDCs, which ensures timely execution of various
large-scale modern power system that shows the sufficient functions for real-time monitoring and control of the power
scalability of the proposed technique for practice. Therefore, grid. In the future work, multiple independent component
the proposed technique is scalable to multiple PDCs placement failures will be considered to further improve the resiliency of
in very large modern power grids. the design.
For the parameters (α, β, λ) selection, assuming that average
hop count and maximum hop count are equally important for REFERENCES
the user when selecting plausible PDC configurations, λ is [1] S. Song, P. Wu, Y. Lin, and Y. Chen, "A general dynamic state estimation
tuned such that both average hop and maximum hop equally framework for monitoring and control of permanent magnetic
synchronous generators-based wind turbines," IEEE Access, vol. 9, pp.
contribute to S1, S2, S3 in (6). For the simulation, λ is set to 0.5 72228-72238, 2021.
as it is noticed that the maximum hop along each path is almost [2] Z. Fang, Y. Lin, S. Song, C. Li, X. Lin, and Y. Chen, "State Estimation
double the corresponding average hop in the standard test cases for Situational Awareness of Active Distribution System With
(refer to the average hop and maximum hop counts in Table Photovoltaic Power Plants," IEEE Transactions on Smart Grid, vol. 12,
no. 1, pp. 239-250, Jan. 2021.
VII). Similarly, as it is mentioned parameters (α→1, β→0) [3] F. Aminifar, M. Fotuhi-Firuzabad, A. Safdarian, A. Davoudi, and M.
screen plausible PDC configurations that can connect the PMUs Shahidehpour, "Synchrophasor measurement technology in power
with lower hops in the order of P1ePMU, PrPMU, and P2ePMU, systems: panorama and state-of-the-art," IEEE Access, vol. 2, pp. 1607-
different values of α are chosen for a fixed lower value of β and 1628, 2014.
[4] N. M. Manousakis, G. N. Korres, and P. S. Georgilakis, "Taxonomy of
the corresponding required hop counts are recorded for ePMUs PMU placement methodologies," IEEE Transactions on Power Systems,
coverage and all PMUs coverage, respectively, as shown in Fig. vol. 27, no. 2, pp. 1070-1077, May 2012.
6. As noticed in the figure, parameter α in the range of 0.6-0.8 [5] N. M. Manousakis, G. N. Korres, and P. S. Georgilakis, "Optimal
provides lower hop coverage for both ePMUs and all PMUs, placement of phasor measurement units: a literature review," 2011 16th
International Conference on Intelligent System Applications to Power
which is expected according to the above discussion. Therefore, Systems, 25-28 Sept. 2011, pp. 1-6.
α and β are set to 0.7 and 0.2, respectively to prioritize the [6] S. Chakrabarti, E. Kyriakides, and D. G. Eliades, "Placement of
contribution of the paths to the MMF in the order of P1ePMU, synchronized measurements for power system observability," IEEE
PrPMU, and P2ePMU. Transactions on Power Delivery, vol. 24, no. 1, pp. 12-19, Jan. 2009.
[7] F. Aminifar, M. Fotuhi-Firuzabad, M. Shahidehpour, and A. Safdarian,
"Impact of WAMS malfunction on power system reliability assessment,"
VI. CONCLUSION IEEE Transactions on Smart Grid, vol. 3, no. 3, pp. 1302-1309, Sep.
2012.
In this paper, a PDC placement technique is discussed to [8] Y. Seyedi, H. Karimi, C. Wetté, and B. Sansò, "A new approach to
facilitate fast and resilient data transfer between PMUs and reliability assessment and improvement of synchrophasor
PDCs. The technique mainly contains two stages: the communications in smart grids," IEEE Transactions on Smart Grid, vol.
prescreening of plausible PDC configurations, and the final 11, no. 5, pp. 4415-4426, Sep. 2020.
[9] M. Shahraeini, M. S. Ghazizadeh, and M. H. Javidi, "Co-optimal
selection. Auxiliary steps such as candidate PDC location placement of measurement devices and their related communication
screening, PMU placement and role selection, parameters infrastructure in wide area measurement systems," IEEE Transactions on
selection in GA are also included. In the prescreening stage, the Smart Grid, vol. 3, no. 2, pp. 684-691, July 2012.
plausible PDC configurations are identified from all possible [10] X. Zhu, M. H. F. Wen, V. O. K. Li, and K. Leung, "Optimal PMU-
communication link placement for smart grid wide-area measurement
configurations based on MMF by employing the GA. In the systems," IEEE Transactions on Smart Grid, vol. 10, no. 4, pp. 4446-
final selection stage, the best PDC configuration is obtained by 4456, July 2019.

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIV OF MASS-LOWELL. Downloaded on January 31,2023 at 00:02:02 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Control of Network Systems. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCNS.2023.3240200

12

[11] M. Beg Mohammadi, R. Hooshmand, and F. Haghighatdar Fesharaki, "A [33] K. Deb and D. Deb, "Analysing mutation schemes for real-parameter
new approach for optimal placement of PMUs and their required genetic algorithms," Int. J. Artif. Intell. Soft Comput., vol. 4, no. 1, pp. 1–
communication infrastructure in order to minimize the cost of the 28, 2014.
WAMS," IEEE Transactions on Smart Grid, vol. 7, no. 1, pp. 84-93, Jan. [34] J. W. Suurballe and R. Tarjan, "A quick method for finding shortest pairs
2016. of disjoint paths," Networks, vol. 14, pp. 325-336, 1984.
[12] M. Shahraeini, M. H. Javidi, and M. S. Ghazizadeh, "Comparison [35] Y. Lin and A. Abur, "A highly efficient bad data identification approach
between communication infrastructures of centralized and decentralized for very large scale power systems," IEEE Transactions on Power
wide area measurement systems," IEEE Transactions on Smart Grid, vol. Systems, vol. 33, no. 6, pp. 5979-5989, Nov. 2018.
2, no. 1, pp. 206-211, Mar. 2011. [36] Fernando G. Lobo, David E. Goldberg, and Martin Pelikan, "Time
[13] F. Ye and A. Bose, "Multiple communication topologies for PMU-based complexity of genetic algorithms on exponentially scaled problems," 2nd
applications: introduction, analysis and simulation," IEEE Transactions Annual Conference on Genetic and Evolutionary Computation, Las
on Smart Grid, vol. 11, no. 6, pp. 5051-5061, Nov. 2020. Vegas, Nevada, 2000, pp. 151–158.
[14] A. Phadke and J. Thorp, Synchronized phasor measurements and their [37] A. B. Birchfield, T. Xu, K. M. Gegner, K. S. Shetye, and T. J. Overbye,
applications. New York, NY, USA: Springer, 2008, pp. 97–98. "Grid structural characteristics as validation criteria for synthetic
[15] A. Armenia and J. H. Chow, "A flexible phasor data concentrator design networks," IEEE Transactions on Power Systems, vol. 32, no. 4, pp. 3258-
leveraging existing software technologies," IEEE Transactions on Smart 3265, Jul. 2017.
Grid, vol. 1, no. 1, pp. 73-81, 2010. [38] Y. Chen et al., "Cascading Failure Analysis of Cyber Physical Power
[16] "IEEE Guide for Phasor Data Concentrator Requirements for Power System With Multiple Interdependency and Control Threshold," IEEE
System Protection, Control, andMonitoring," IEEE Std C37.244-2013, Access, vol. 6, pp. 39353-39362, 2018.
pp. 1-65, 2013.
[17] M. H. F. Wen and V. O. K. Li, "Optimal phasor data compression unit Md. Zahidul Islam (S’22) received his B.Sc. and M.Sc. degrees in Electrical
installation for wide-area measurement systems—an integer linear and Electronic engineering from Bangladesh University of Engineering and
programming approach," IEEE Transactions on Smart Grid, vol. 7, no. 6, Technology, Dhaka, Bangladesh, in 2017 and 2021, respectively. He has
pp. 2644-2653, Nov. 2016. been pursuing his Ph.D. in the Department of Electrical and Computer
[18] F. Haghighatdar Fesharaki, R. Hooshmand, and A. Khodabakhshian, Engineering at the University of Massachusetts, Lowell, MA, USA since
"Simultaneous optimal design of measurement and communication 2021. His research interests include modeling, analysis, and resilience of
infrastructures in hierarchical structured WAMS," IEEE Transactions on cyber-physical power systems.
Smart Grid, vol. 5, no. 1, pp. 312-319, Jan. 2014.
[19] A. Abur and A. Gómez-Expósito, Power system state estimation: theory Shamsun Nahar Edib (S’20) received her B.Sc. degree in Electrical and
and implementation. New York, NY: Marcel Dekker, 2004. Electronic Engineering from Bangladesh University of Engineering and
[20] G. Cheng, Y. Lin, Y. Chen, and T. Bi, "Adaptive state estimation for Technology, Dhaka, Bangladesh, in 2017. She has been a Ph.D. student in
power systems measured by PMUs with unknown and time-varying error the Department of Electrical and Computer Engineering at the University of
statistics," IEEE Transactions on Power Systems, pp. 1-1, 2021. Massachusetts, Lowell, MA, USA since 2019. Her research interests include
[21] S. Soltan, M. Yannakakis, and G. Zussman, "Power grid state estimation modeling, analysis, and optimization of cyber-physical power systems.
following a joint cyber and physical attack," IEEE Transactions on
Control of Network Systems, vol. 5, no. 1, pp. 499-512, Mar. 2018. Vinod M. Vokkarane (SM’ 09) is a Professor in the Department of Electrical
[22] G. Wang, H. Zhu, G. B. Giannakis, and J. Sun, "Robust power system and Computer Engineering at the University of Massachusetts Lowell. He
state estimation from rank-one measurements," IEEE Transactions on received the B.E. degree with Honors in Computer Science and Engineering
Control of Network Systems, vol. 6, no. 4, pp. 1391-1403, Dec. 2019. from the University of Mysore, India in 1999, the M.S. and the Ph.D. degree
[23] K. Zhu, A. T. Al-Hammouri, and L. Nordström, "To concentrate or not to in Computer Science from the University of Texas at Dallas in 2001 and
concentrate: performance analysis of ICT system with data concentrations 2004, respectively. His primary areas of research include design and
for wide-area monitoring and control systems," IEEE Power and Energy analysis of architectures and protocols for ultra-high speed networks, smart
Society General Meeting, 2012, pp. 1-7. grids, and green networking. Dr. Vokkarane is the co-author of a book,
[24] R. Kateb, P. Akaber, M. H. K. Tushar, A. Albarakati, M. Debbabi, and C. “Optical Burst Switched Networks,” Springer, 2005. He is currently on the
Assi, "Enhancing WAMS Communication Network Against Delay Editorial Board of IEEE/OSA Journal of Optical Communications and
Attacks," IEEE Transactions on Smart Grid, vol. 10, no. 3, pp. 2738- Networking and Springer Photonic Network Communications Journal. He
2751, 2019. has co-authored several Best Paper Awards, including the IEEE
[25] B. Yang, K. V. Katsaros, W. K. Chai, and G. Pavlou, "Cost-efficient low GLOBECOM 2005, IEEE ANTS 2010, ONDM 2015, ONDM 2016, and IEEE
latency communication infrastructure for synchrophasor applications in ANTS 2016.
smart grids," IEEE Systems Journal, vol. 12, no. 1, pp. 948-958, Mar.
2018. Yuzhang Lin (M’18) is an assistant professor in the Department of Electrical
[26] X. Fan et al., "Automated realistic testbed synthesis for power system and Computer Engineering at the University of Massachusetts, Lowell, MA,
communication networks based on graph metrics," 2020 IEEE Power & USA. He obtained his Bachelor and Master’s degrees from Tsinghua
Energy Society Innovative Smart Grid Technologies Conference (ISGT), University, Beijing, China in 2012 and 2014, respectively, and Ph.D. degree
Feb. 2020, pp. 1-5. from Northeastern University, Boston, MA, USA in 2018. His research
[27] W. Wang and C. Y. Tang, "Distributed estimation of closeness centrality," interests include modeling, situational awareness, and cyber-physical
2015 IEEE Conference on Decision and Control (CDC), Dec. 2015, pp. resilience of smart grids. He is a recipient of NSF CAREER Award.
4860-4865.
[28] S. N. Edib, Y. Lin, V. M. Vokkarane, F. Qiu, R. Yao, and D. Zhao, Xiaoyuan Fan (S'12–M'16-SM'19) received the Ph.D. degree in electrical
"Optimal PMU restoration for power system observability recovery after engineering from the University of Wyoming, Laramie, WY, in 2016, and
massive attacks," IEEE Transactions on Smart Grid, vol. 12, no. 2, pp. M.S. and B.S. degrees in electrical engineering from Huazhong University of
1565-1576, Mar. 2021. Sciences & Technology, Wuhan, China, in 2012 and 2009, respectively. He
[29] F. Aminifar, A. Khodaei, M. Fotuhi-Firuzabad, and M. Shahidehpour, is currently a senior research engineer with Pacific Northwest National
"Contingency-constrained PMU placement in power networks," IEEE Laboratory (PNNL) located in Richland, Washington, he is also Team
Transactions on Power Systems, vol. 25, no. 1, pp. 516-523, Feb. 2010. Leader for Power Electronics Team. Serving as a project manager, PI/co-PI
[30] S. L. Hakimi, "Optimum distribution of switching centers in a and key contributor, Xiaoyuan has been managing and supporting multiple
communication network and some related graph theoretic problems," research projects funded by the federal agencies and industrial
Operations Research, vol. 13, no. 3, pp. 462-475, 1965. collaborators. His research interests focus on data analytics for power
[31] C. Blum and A. Roli, "Metaheuristics in combinatorial optimization: system reliability, wireless communication, multi-discipline resilience
overview and conceptual comparison," ACM Comput. Surv., vol. 35, no. analysis, and high-performance computing. Xiaoyuan is a Senior Member of
3, pp. 268–308, 2003. IEEE, and serves as a volunteer reviewer of 20+ top-level journals and
[32] B. Dengiz, F. Altiparmak, and A. E. Smith, "Local search genetic conferences in power systems and signal processing. Xiaoyuan is the
algorithm for optimal design of reliable networks," IEEE Transactions on recipient of a 2021 Federal Laboratory Consortium for Technology Transfer
Evolutionary Computation, vol. 1, no. 3, pp. 179-188, Nov. 1997. Award, and three PNNL Energy and Environment Directorate Outstanding
Performance Awards.

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIV OF MASS-LOWELL. Downloaded on January 31,2023 at 00:02:02 UTC from IEEE Xplore. Restrictions apply.
View publication stats