Article

Mathematical Models for the Single-Channel and

Multi-Channel PMU Allocation Problem and Their Solution
Algorithms
Nikolaos P. Theodorakatos 1,*, Rohit Babu 2, Christos A. Theodoridis 1 and Angelos P. Moschoudis 3

1 School of Electrical & Computer Engineering, National Technical University of Athens, 157 80 Athens,
Greece
2 Department of Electrical and Electronics Engineering, Alliance University, Anekal, Bengaluru 562106, India;

rohit.babu@alliance.edu.in
3 Department of Electrical and Electronics Engineering, University of West Attica, 122 41 Athens, Greece;

amoschoudis@uniwa.gr
* Correspondence: nikos.theo2772002@gmail.com or nikos277@central.ntua.gr

Abstract: Phasor measurement units (PMUs) are deployed at power grid nodes around the trans-
mission grid, determining precise power system monitoring conditions. In real life, it is not realistic
to place a PMU at every power grid node; thus, the lowest PMU number is optimally selected for
the full observation of the entire network. In this study, the PMU placement model is reconsidered,
taking into account single- and multi-capacity placement models rather than the well-studied PMU
placement model with an unrestricted number of channels. A restricted number of channels per
monitoring device is used, instead of supposing that a PMU is able to observe all incident buses
through the transmission connectivity lines. The optimization models are declared closely to the
power dominating set and minimum edge cover problem in graph theory. These discrete optimiza-
tion problems are directly related with the minimum set covering problem. Initially, the allocation
model is declared as a constrained mixed-integer linear program implemented by mathematical and
stochastic algorithms. Then, the 0⁄1 integer linear problem is reformulated into a non-convex con-
Citation: Theodorakatos, N.P.; Babu, straint program to find optimality. The mathematical models are solved either in binary form or in
R.; Theodoridis, C.A.; Moschoudis,
the continuous domain using specialized optimization libraries, and are all implemented in
A.P. Mathematical Models for the
YALMIP software in conjunction with MATLAB. Mixed-integer linear solvers, nonlinear program-
Single-Channel and Multi-Channel
ming solvers, and heuristic algorithms are utilized in the aforementioned software packages to lo-
PMU Allocation Problem and Their
cate the global solution for each instance solved in this application, which considers the transfor-
Solution Algorithms. Algorithms
2024, 17, 191. https://doi.org/
mation of the existing power grids to smart grids.
10.3390/a17050191
Keywords: phasor measurement unit (PMU); optimization; optimal PMU placement (OPP);
Academic Editor: Binlin Zhang
channel limit capacity; multi-channel PMUs; observability; binary-integer programming; nonlinear
Received: 5 April 2024 programming; algorithms; sufficient conditions for optimality
Revised: 25 April 2024
Accepted: 28 April 2024
Published: 30 April 2024

1. Introduction
Energy management systems (EMSs) play a vital function, serving as the backbone
Copyright: © 2024 by the authors. of smart power grids for monitoring, control, and state estimation [1–4], while Supervi-
Licensee MDPI, Basel, Switzerland. sory Control and Data Acquisition (SCADA) systems are crucial for their implementation
This article is an open access article
of monitoring and control [1–5].
distributed under the terms and
The state of the power system is determined to be the assignment of the phasor volt-
conditions of the Creative Commons
age quantities, electrical current quantities to branches, and power values to generators
Attribution (CC BY) license
and loads. One way to transform the existing power system to a “modern” grid is to meas-
(https://creativecommons.org/license
s/by/4.0/).
ure its state in actual time, which requires continuous monitoring [1–5]. The state of a

Algorithms 2024, 17, 191. https://doi.org/10.3390/a17050191 www.mdpi.com/journal/algorithms

Algorithms 2024, 17, 191 2 of 62

power network is said to be observed in the case of measuring the state variables, such as
the phasors voltage at nodes and phasor currents running in the lines.
Sensors are installed at appropriate nodes in the power network to measure the state
variables quantity, which is telemetered to the central control. The electricity companies
are required to continually observe the state variables [1–6].
In order to observe the power network state, phasor measurement units (PMUs) are
utilized for gathering data and sending them to the system operator [1–4].
The power system can be made fully observable by utilizing solely PMUs. In this
case, the state estimation formulation is linear and solved without the needed iterative
schemes [2,4,6]. PMUs are installed at substations and are crucial devices for system op-
eration and control due to the measuring of the phasor voltages and phasor currents run-
ning the transmission lines. Global time reference can accurately measure these types of
phasors, even from remote physical locations within a designated power system [1–7].
Traditionally, state estimation (SE) is implemented using the SCADA system,
whereas with the utilization of PMUs, hybrid state estimators based on SCADA-PMU and
linear state estimators based on PMUs have emerged [2–4,6].
SE is the most significant tool of the EMS, giving the transmission system operator
an actual-time understanding of the current operation [2–4,6].
In these circumstances, SE helps the system operator maintain the power grid to be
reliable, secure, and economical in its service [1–6]. Furthermore, SE supports the moni-
toring and control application in the EMS [1–6]. SE procedures have undergone a great
improvement since the introduction of the PMU [1–6]. The literature illustrates different
methods to implement and solve the state estimation application [2–4,6].
The SE model is classified as static or dynamic [1–6]. State estimation models based
on a limited PMU number were recently presented in [2–6]. The static algorithm princi-
pally relies on a WLS algorithm, utilizing data calculated at a given time [2–4,6]. By com-
parison, a SE model based on a dynamic response and Kalman filter is presented in [6].
Both SE models are based on a mixed set including conventional measurements as well as
a limited PMU number [1–6].
In a modern transmission system, telemetry relies on the SCADA and the wide-area
monitoring system (WAMS) [1–5]. The WAMS is made of PMUs and phasor data concen-
trators (PDCs) and their communication networks [1–6]. PMUs play an essential part in
the observation of the power transmission network [1–6].
This fact has caused a paradigm shift in the architecture of power grids, increasing
the utilization of PMUs with their installation at selected power grid nodes [1–6]. PMUs
have led to reconsideration of the way the state estimation is executed for reliable moni-
toring of the power grid state [1–6].
PMUs comprise the most commonly obtainable vital measurement infrastructures
that consider synchro phasor technology [1–6]. A PMU enables real-time measuring data,
which are the phasor voltage of the placing node and the phasor currents running in the
branches running from that network bus [1–6].
The real-time data are collected and sent to the phasor data concentrator (PDC), en-
abling it able to undertake applications such as monitoring, control, and network stability
[1].
This measuring fact allows real-time monitoring, which is important for energy man-
agement system applications [1–6].
With the utilization of PMUs, the SE procedure is now able to work more accurately
in large power networks, as it incorporates phasor voltage measurements at network
buses and current measurements through connection lines [1–12].
In the last decade, utilities have made significant investments in enhancing the ac-
tual-time monitoring and power transmission grid control applications [1–12]. All over
the world, many utility industries have installed large numbers of PMUs to render their
optimal deployment at substations [7–11].
Algorithms 2024, 17, 191 3 of 62

The deployment of synchronized sensors is still continuing and somehow is in slow

progress, mainly due to economic and technical restrictions [5].
Due to the high cost of installing a PMU at every power grid node in the WAMS im-
plementation to attain full observability, it is significant to take into account the best PMU
position sites through a solvable OPP problem [7–11].
Hence, the OPP problem still stands, giving a full network observation of the power
grid state while the entire cost investment is minimized [7–11]. A PMU is placed at the
installation bus and computes the phasor voltage at that bus and all the phasor current
lines running from that bus with the assumption that the monitoring device has an un-
limited number of channels [4,6–11].
Over two decades of research, a plethora of significant papers has emerged to tackle
the OPP problem. They aim to ascertain the optimal PMU number needed and strategi-
cally position them to ensure the power grid achieves full observability [4,8–14].
Many of these papers delve into various strategies for solving the OPP problem, with
a primary focus on achieving full observability while satisfying zero–one constraints on a
minimum objective function bound [4,8–30]. Some interesting research studies have been
published recently in the OPP domain [30–67].
The pioneering work is to address the OPP-based model on optimization [28–32] us-
ing a zero–one integer linear programming (ILP) model, and some interesting research
works have been published in the OPP domain in the recent past [33–36].
Mathematical algorithms are utilized in the OPP problem solving [31–42], whereas
heuristic algorithms are also implemented to handle the integer program s inequality to
examine the solution, thereby yielding heuristically sound knowledge [39,43–55].
In the past years, the initial OPP formulation has become more and more attractive in
the incorporation of intellectually demanding tasks under operating conditions [7–11].
These OPP models consider relative constraints to satisfy these issues without sacri-
ficing the entire observability [31–55].
Many research works on OPP rely on a cumulative single objective function that is
minimized subject to a constraint function defined on specific decision bounds [32–41].
On the contrary, a two-product objective minimization model is proposed in [42,47,49,51]
to find those optima points with a maximum system observability index declared in
[33,37,38,41,42,51].
The optimization problem in [37,38,42–45,47,49,51] considers obtaining more than one
objective product. This multi-objective function consists of two conflicting products for
the consideration of those solutions satisfying complete and maximum observability at
the same time.
The impact of ZI buses is considered for further minimization of PMU cost installation
at substations [35,40,46–48]. Despite this, with unknown tap ratios, uniform behavior
among all ZI buses may lead to situations where a topological observability analysis does
not ensure a numerical observability [34,35].
Likewise, it may be the case that the consideration of ZI buses does not necessarily
further minimize the PMU cost allocation at power grid buses [54].
In real-life, each PMU device has a restricted number of channels and thus a more
realistic configuration of the PMUs is taken into consideration with a varying channel ca-
pacity [47,52,56–61,63–67].
Therefore, the restriction in the channel limit capacities is taken into consideration to
implement a more reasonable configuration of PMUs in grids [56–67].
In this study, we consider the extra limitation regarding the channel capacity, and
propose mathematical and heuristic algorithmic schemes to solve the OPP focused on
finding the true solutions including the PMU locations and the branches being observed.
Mathematics, heuristic algorithms, and derivative-free optimizers (DFO) are adopted
to solve the OPP problem for the implementation of the installation of PMUs with a lim-
ited channel capacity per a PMU optimally installed at selected power grid nodes.
Algorithms 2024, 17, 191 4 of 62

As a result, we present the mathematical formulations, the solution algorithms, the

PMU locations, and the branch assignments related to the best PMU sites. All findings are
displayed in tables and plot diagrams to validate our algorithmic approach.

2. Consideration of this Study

The OPP problem delivers an adequate number of PMUs for observing a power grid
and defines the health of a power grid [1–12]. PMUs are placed at power network nodes,
i.e., substations to which transmission lines, generators, and loads are joined [34,35].
Due to lack of a communication facility at some substations [36,39,65], the PMU s cost
placement is further increased and it is forbidden to be placed at substations with a lack
of communication facilities [1,34–36].
Hence, it is still significant to optimize/minimize the number of PMUs placed within
the power grid transmission system [36–38,41,42,51]. The optimization output gives an
appropriate number of PMUs that is sufficient to monitor the power system [37,38,41,51].
In practice, PMU manufacturers yield a few types of PMU, with distinct functions and
varying channel capacities [56,59–61,63–67].
As a consequence, specific algorithms are studied for optimally placing PMUs with a
restricted number of channels at power nodes [60]. We consider an extra constraint func-
tion to the allocation problem related to the number of channels to keep complete network
observability [56,59–61,63–67].
We develop PMU placement models with multiple channels as well as a single-line
placement model to detect PMU sites related to branch assignments. The optimization
model is also interpreted as a minimum edge cover, otherwise known as a node set cov-
ering problem [15]. Single-line PMUs are also known as a unit channel capacity [61].
A single-line PMU is able to measure the voltage phasor of the related bus, and the
phasor current of only one of the branch incidents to the bus where the PMU is placed
[60,61]. So, in the monitoring of the phasor s voltages of grid nodes, PMU measurements
are significant in observing the transmission grid lines [61].
Thus, each PMU enables it to make better line capability observation with PMU meas-
urements [56,58,60,61,64]. As a result, the installation of single-line PMUs could be a
standard practice interpreted as a minimum edge covering problem [28–31,61].
So, the node set covering problem interprets the above strategy in graph theory [28–
31]. In practice, it can be used primarily in industries such as utility companies to construct
a modern grid with the capability of optimal “branch” placement of PMUs [61].
Furthermore, the manufacturers supply PMUs with a varying number of channels
and PMU cost ranges analogous to channel restriction. Some works study this combina-
torial problem considering a restricted number of channels, which is a more realistic OPP
scenario [56–69]. The principal constitution of this study is threefold:
(1) To achieve the proper PMU number by optimizing a linear or a quadratic cost
function under a topological constraint function considering the limitation of the
PMU channel limit capacity.
(2) To accomplish this task, branch-and-bound algorithms [37], sequential quadratic
programming [42], interior-point methods [42], generalized pattern search
algorithms [41,70,71], genetic algorithms [51], and binary swarm optimization [51]
are utilized to achieve optimality.
(3) We focus on finding the proper PMU number with a branch assignment set solution
with a fixed number of channels.
Our aim is to reconsider the ILP-based model under a limited PMU number [56–58]
in the YALMIP environment [37,38]. GA and BSPO are able to handle the ILP model s con-
straints, and finally an equivalent nonlinear model is given to find optimality.
We tested the algorithmic models on the classical IEEE power systems. BBA gives the
solution under a strict optimality warranty followed by evolutionary algorithms solutions
Algorithms 2024, 17, 191 5 of 62

with the same quality and quantity [37,38,41,42,51]. ZI buses are the buses without gener-
ation or load consumption [32–35,39,40].
In actual conditions, buses are mainly substations, as noted in [34,35,54], and the sub-
station s internal consumption is something that cannot be ignored [34,35,39]. Specifically,
the substation itself needs energy to operate [34]. For example, energy is required for the
substation control and protection system [68,69]. Substations also need energy for the cool-
ing fans (ONAN/ONAF) and the on-load tap changer of transformers, uninterruptible
power supplies (UPSs), fire extinguishing systems, lighting fixtures, floodlights, and bat-
tery chargers [68,69]. Moreover, the observability scenario based on a topological ap-
proach, which does not always ensure numerical observability, is also avoided by treating
ZI buses as normal buses [34,35,39,46].
Thus, simulation results are produced by mathematical and heuristic algorithms
when ZI buses are not taken into account for the practical power systems, following re-
cently published papers in the relative research domain [34,37,41,42,51].
Since the ZI s effect will not be included in the PMU arrangement procedure
[34,37,41,51], each power network bus is fully observed, either by a PMU placed at that
node or by other PMUs placed at the incident network nodes [34,37,41,42,51].
To present an OPP model based on a restricted number of channels, ILP [37,38], DFO
solvers [41,51], NLP [41], and a BPSO [51] are utilized to find the exact solution avoiding
near-to-optimal solutions. The solvers produce the PMU number and the branch assign-
ments for a varying number of measurement channels.
The placement problem is optimized in three stages. Integer linear, heuristic, and non-
linear models are simulated using specialized optimization libraries [72–76]. As a guide
in minimizing these placement problems, optimization tactics can be followed by the
reader in previous authors cited works [37,38,41,42,51].
The proposed models are well optimized by BBA, SQP, IPMs, GPSA, Gas, and BPSO
towards an optimum point. We derive not only the PMU number but also the branch as-
signments for ranging channel capacity. The BBA solutions are given within an absolute
zero gap, which means no better solution can be found [37,38,41,42,51].

3. Contribution of this Study

PMU brings a revolution to the manner in which power networks are observed and
operated. Most of the works assume that these monitoring devices have an unrestricted
amount of channel capacity to measure as many phasors voltage and currents running
the lines as are required. In real conditions, manufacturers produce PMUs with different
characteristics.
One characteristic is the channel limit capacity, which allows the PMU s availability
onto the market with a varying cost [52]. Therefore, an optimal investment can be made
based on the selection of appropriate PMU types utilizing efficient algorithms.
This decision leads to a decrease in the cost of hardware needed for the development
of the WAMS system based on synchronized measurements [52].
In this study, we consider that optimal PMU numbers are installed around the power
grid having a restricted number of measurement channels. Although there exist some
works in this domain [56–67], we present a revised optimization model and a number of
algorithms related to its solution. The solution of this model will keep the network com-
pletely observable relying only on synchronized measurements [37].
This study takes into account the realistic scenario where the channel capacity is var-
ied to measure phasors voltage and current running in the connectivity lines and pro-
duces global solutions with strict optimality conditions [13–21].
The optimization study does not consider the existence of ZI buses at this time. This
approach does not take away any truth of models presented in [40,56–67] when ZI, active
injections, and flow measurements are taken into account for further consideration.
Our aim is to give a comprehensive output of our algorithmic approach to the PMU
allocation problem solving based on BBA considering a restricted number of channel
Algorithms 2024, 17, 191 6 of 62

capacities. Furthermore, we examine how the entire observability is preserved with less
measurement redundancy in the state estimation [1–4,6].
To extend the mentioned target, we illustrate a number of algorithmic models to en-
hance our approach. In this direction, we adopt nonlinear, derivative-free, and heuristic
algorithms to find alternative PMU set solutions with the branch assignments covered to
keep the system functional and observable. Our primary outputs are:
a. We create an algorithmic connection between the OPP based-model with a restricted
channel capacity and the optimal solutions produced by the suggested mathematical
and heuristic algorithms.
b. The branch-and-bound algorithm is used in a comparison study with nonlinear and
heuristic algorithms.
c. We present a state-of-the-art solution produced by the 0/1 integer program solving
with a zero absolute gap calculated with zero percentage gap tolerances; this means
that this solution is a global one returned by the solver.
d. We extend the ILP model into a heuristic search of the feasible set to show that evo-
lutionary algorithms can efficiently perform the optimization following the BBA s
solution.
e. Finally, a novel nonlinear program is created and solved by either continuous algo-
rithms or a BBA embedded in the SCIP optimizer performed in MATLAB.

4. Materials and Methods

Despite the fact the PMU s cost was recently reduced, it cannot be workable to place
a PMU at every power grid node in a large-size transmission grid [7–11].
Hence, it has an important value to attain complete observability and maximum re-
dundancy with an optimal number of PMUs placed at buses within the power grid
[33,37,41,43–45,47,49,51].
Many research works have considered the ZI s bus presence to further reduce the
entire PMU number needed to attain power grid observability [32,33,35,36,39,40,43–50,52–
54]. A PMU-PDC monitoring system is designed in [55].
Considering the unknown tap ratios [68,69], having all ZI nodes behave in a similar
manner can provide guidance for schemes where the topological observability may not
secure numerical observability being fulfilled [35,40,46].
Our aim is observability analysis based on a restricted number of channels per PMU
device. Hence, a minimum number of such devices are optimally placed around the
power transmission grid. The whole power network must be fully monitored using an
optimal PMU number [37,38,41,51,56–58].
A PMU device placed at a network bus supplies the system operator with direct
measurements, whereby Kirchhoff s voltage law and Ohm s law are utilized to compute
the pseudo measurements that are essential to maintain the system observability [31].
Optimal PMU placement models are suggested where the PMU number is found
when a linear or quadratic cost function is optimized subject to several topological or nu-
merical observability constraints in [37,38,41,42,51]. The decision vector ranges in either a
binary or a continuous domain [37,38,41,51].
At first, the OPP formulation was suggested in an 0⁄1 integer linear program [31–
33] and a powerful ILP solver was used to obtain an optimal solution [35–38,41,51]. Then,
0/1 semi-definite and nonlinear programs have been proposed to find an optimal solution
from an alternative optimization point of view [37,40].
All optimization models are robust in the implementation, with excellent outcomes
ensuring the topological or the numerical approach in the state estimation tool [30–67].
The high cost of PMUs related to installation limitations and commissioning procure-
ment in substations is still a practical restriction over the wide-area power grid [36–
38,41,51]. The PMU cost is further increased by keeping the same ratio with the number
of channels [36,56,58–61,63–67].
Algorithms 2024, 17, 191 7 of 62

In the bibliography, most articles have supposed an unlimited number of measure-

ment channels for the PMU allocation problem solving [4,7–11].
The restricted number of channels within a PMU monitoring device is an extra con-
straint involved in the optimization model related to finding the best PMUs to ensure the
complete observability scenario [36,56,58].
This realistic assumption about PMU limit capacity can have an effect on the PMU
arrangement in the power grid with or without contingency scenarios [56,58]. Certainly,
if the amount of PMU channel capacity is less than the number of connectivity lines to a
power grid node, a combinatorial number of possible branch phasor current measure-
ments can be found in this realistic scenario [56–58].
In such a realistic scenario, PMU monitoring devices are constructed with a restricted
number of channels. A single PMU can be lost depending on the amount of current chan-
nel capacity embedded in a PMU device [53,54]. Hence, the measurement redundancy
task is very significant for the effective utilization of SE [2–4,6].
We consider mathematical and heuristic models, with the problem solving being pre-
sented using BBA, nonlinear algorithms, derivative-free optimizers, and a BPSO.
This paper completes previous authors papers [37,38,41,42,51], filling the gap relat-
ing to the certainty of reaching a global solution or a near-to-optimum point considering
a restricted number of measurement channels [56–61].

5. Optimization Models Studied for the OPP Model

PMUs are placed at substations in which transmission lines, generators, and loads
are adjoined [29,31,60]. Specific reasons, such as the PMU s high cost and communication
facilities not existing in some substations, increase the entire cost of installation
[7,37,38,41,51,55,60].
Thereby, the power utility industries desire a reasonable number of channels per a
PMU device to determine the entire cost of the investment. In addition to that considera-
tion, an optimal PMU placement is essential in the research domain [37,38,41,42,51].
The initial step is to identify if the solution achieved by the optimization algorithm is
feasible or not. In the case of producing the sequence of points, we first derive a feasible
estimate point in which the objective value is less than the older one on the previous trial
point [13–21].
Such a task between a feasible and an infeasible point is crucial to achieve optimality.
What we desire is to deliver the optimum point over the feasible region constituted by the
objective, the constraint function, and the decision bounds [21–25]. Therefore, the aim is
to identify the feasibility and optimality of the response space solutions. Although there
exist a few works in this domain [56–67], there is no possible combination of the PMU
number and the appropriate locations.
This work proposes an OPP-based model considering the limited number of channels
per PMU device using extended ILP models initially presented in [56–58] and [60] where
a set cover problem is introduced [28–31]. The suggested optimization models have been
solved by BBA, heuristic algorithms, and nonlinear algorithms.
Heuristic algorithms work together with BBA to find a solution of the same quantity
and quality in a minimal running time [25,26]. Furthermore, transforming the binary-in-
teger programming into a nonlinear polynomial model shows the robustness required to
find a proper number of PMUs and their placement sites for a fixed number of channels
[13–24]. The numerical optimization challenge is to run mathematical as well as heuristic
algorithms, including BBA, SQP, IPMs, GPSA, GA, and BPSO, to discover the best combi-
nation between the proper PMU number and the upper bound of channel capacity with-
out increasing the cost of installation [56–60].
 Observability of the graph
Graphs are useful models for many problems considered in combinatorial optimiza-
tion [28–31]. Let 𝐺 = (𝑉, 𝐸) be a graph depicting a power transmission system, where a
Algorithms 2024, 17, 191 8 of 62

vertex depicts a power network node and an edge is a connectivity line by which two
graph vertices are joined [28–31].
Each power network node represents a substation where generators, loads, and
transmission lines are joined. Let 𝛱 be a defined PMU set solution on a graph 𝐺 and 𝛺
the set of monitored vertices [28–31]. The observability of 𝐺is determined by the two rules
relying on electrical laws below [28–31].
Rule 1: If a PMU is installed on an edge{𝑖, 𝑗}, then vertices 𝑖 and 𝑗 are monitored.

{𝑖, 𝑗} ∈ 𝛱 ⇒ 𝑖, 𝑗 ∈ 𝛺 (1)

Rule 2: If a monitored vertex 𝑖 has all its neighbours monitored, save one, then this
vertex is monitored.

𝑖 ∈ 𝛺 and|Γ(𝑖)\Ω| ≤ 1 ⇒ 𝛤(𝑖) ⊆ 𝛺 (2)

By rule R1, the PMU installed at {𝑖, 𝑗} calculates the voltage at 𝑖 and the current on
{𝑖, 𝑗}. Using Ohm s law, we can work out the voltage at the vertex 𝑗. Then vertices 𝑖 𝑎𝑛𝑑 𝑗
are observed [28–31].
By rule R2, if a vertex 𝑖 and all its neighbors 𝑘 ∈ 𝛤(𝑖) are monitored, except a single
vertex 𝑗, then utilizing Ohm s law we define the current on {𝑖, 𝑗} for 𝑘 ∈ 𝛤(𝑖)\ {𝑗}; know-
ing the currents on all {𝑖, 𝑗} (𝑓𝑜𝑟 𝑘 ≠ 𝑗), we can work out the current on {𝑖, 𝑗} utilizing
Kirchhoff s law [28,29,31]. Then, the voltage at 𝑖 and the current on {𝑖, 𝑗} are measured,
and we define the voltage on 𝑗 using Ohm s law. Hence, 𝑗 is observed [28–31].

Definition 1. Statement of the Power Dominating Set Problem:

The power dominating set is the OPP problem which determines an optimal PMU number
with a varying number of channel capacities.

Definition 2. Statement of the Power Edge Set Problem:

The power edge set is the OPP problem which utilizes a single-line PMU and determines a
proper number of them for installation.

5.1. Binary-Integer Programming Model Considering an Unlimited Number of Channel

Capacities
The integer programming models declare the integer variables representing “yes” or
“no” decisions, meaning that they are taking two values, i.e., zero or one [13–16]. Such
integer problems are declared as follows [13–18]:
𝑚𝑖𝑛 𝐽 (𝑥) = ∑ 𝑤 |𝑥 = 𝑡𝑟𝑢𝑒/𝑓𝑎𝑙𝑠𝑒 (3)

𝑡𝑟𝑢𝑒
𝑠𝑢𝑐ℎ 𝑡ℎ𝑎𝑡: 𝛷(𝑥) = 𝛷(𝑥 , 𝑥 , . . . , 𝑥 ) = (4)
𝑓𝑎𝑙𝑠𝑒
where 𝛷(𝑥) is defined as a Boolean constraint, and 𝑛 the number of the decision varia-
bles. The problem s solution is the set of assignment values that produces the best objec-
tive value while the logical constraints are satisfied [13–16].
Consider an event that may or may not happen, and assume that we choose between
these two options. To make such a decision, a binary variable 𝑥 is as follows [13–16]:

1 𝑖𝑓 𝑡ℎ𝑒 𝑒𝑣𝑒𝑛𝑡 𝑜𝑐𝑐𝑢𝑟𝑠

𝑥= (5)
0 𝑖𝑓 𝑡ℎ𝑒 𝑒𝑣𝑒𝑛𝑡 𝑑𝑜𝑒𝑠𝑛′𝑡 𝑜𝑐𝑐𝑢𝑟𝑠

Therefore, a decision is required to solve the optimization problems. In such kinds of

problems, the possible answers are “yes” or “no” [13–16], which are interpreted as zero or
one [16]. This kind of problem is a zero–one linear program in which the variables range
in the decision domain {0,1} [13–18].
Algorithms 2024, 17, 191 9 of 62

Therefore, a BBA is needed that enforces the assignment decision model to result in
a “yes” or “no” output [37,38,41,51]. The power grid is completely monitored if the phasor
voltage is calculated for each network node either in a direct or indirect way [29].
We adopt a logical relationship for declaring those decisions followed by a linear
constraint function [13–16,37,38,41–44,51].
With an unrestricted channel capacity, a PMU is placed at a node that can monitor
all the buses through the connectivity lines between that bus and its neighboring buses
[32]. The OPP model needs a decision value equal to zero or one [16].
Thereby, the OPP problem is originally declared in a 0 − 1 integer program [13–18].
Let us define the programming model related to the power grid with n nodes [31–
38,41,51].
Given an n—bus power network and considering that the PMU channel availability
permits us to calculate the phasor voltage at the installation node and the line phasor cur-
rents emanating from that node [37,38,41,51]. the minimization model is structured con-
sidering an unlimited number of channel capacities [37,38,41,51].

Definition 3. Statement of the decision variables’ nature in the decision-making problem:

The decision variables are stated to be as follows [37,38,41,51]:
1 if a PMU is installed at bus i
𝑥 = (6)
0 otherwise
The binary-integer model is declared as [32,33,37,38,41,51]:
𝑚𝑖𝑛 ∑ 𝑥 (7)

𝑠. 𝑡. 𝐴 ⋅ 𝑥⃑ ≥ 1 (8)
Let us determine the power grid connectivity matrix as [32,33,37,38,41,42,51]:
1 if i=j or buses i and j are connected
𝐴= (9)
0 otherwise

The inequality constraint function is based on a multiple-choice decision and logical

constraints [13–16,37,38,41,51]. The investment must be at least one of the obtainable op-
tions that satisfy the Boolean logic constraint function [37,38,41,51], where 𝑥 =
𝑥 , 𝑥 ,, . . . , 𝑥 is a binary decision variable placement vector whose elements are stated
if the sensor device is placed at a node or not [37,38,41,51].
The objective function (3) represents the entire cost investment. The constraints are
declared in force for each decision variable, and the objective function is minimized sub-
ject to a set of inequality constraints defined on a binary domain [37,38,41,51].
The binary-integer model is solved by a BBA. The focus is to adopt the BBA, which
is utilized to solve relaxed programming problems to achieve bounds on the objective
function value of the original MILP model within a binary domain [13–18].
The BBA solves LP-relaxed problems where the binary restriction on the decision
variables is substituted by the relaxed restriction 0⃗ ≤ 𝑥⃗ ≤ 1⃗ [13–18].
The next steps of the algorithms are to implement and solve the linear relaxed prob-
lems [37,38,41,51]. If the algorithm s output is integer, then the solution is achieved [13–
18]. In a different outcome, two sub-problems are constructed by utilizing a branch meth-
odology on a fractional variable [37,38,41,51]. The gap between the upper and lower
bounds is inside a pre-given tolerance [13–16,37,38,41,51].
The upper bound is the incumbent cost value, and the current lower bound is the
minimum quantity of the lower bounds produced by finding a solution to the candidate
sub-problems [13–16,37,38,41,51]. To run this optimization process, MILP solvers can be
called to achieve an accurate solution point within a range given by the users.

Definition 4. The optimality gap is stated as follows:

Algorithms 2024, 17, 191 10 of 62

𝐺𝑎𝑝 = |𝑢𝑝𝑝𝑒𝑟 − 𝑙𝑜𝑤𝑒𝑟| <∈, ∈ > 0 (10)

Definition 5. Definition of the percentage relative gap:

percentage relative gap = 100 × |Upper − Lower|/(1 + |Upper|) (11)
The termination criterion equal to the optimality gap 𝐺𝑎𝑝 = 0 means that a global
solution has been reached [13–16,37,38,41,51]. The incumbent solution is the best solution
delivered by the repetitive strategy in the BBA search tree [13–16].
The tolerance gap is a need for a successful termination of the BBA [13–
16,37,38,41,51]. The optimization procedure leads to a ϵ-optimality criterion [13–16].
The relative absolute gap is a criterion to measure how close the solution is to being
global [13–16]. The steps of the BBA implementation are outlined in Algorithm 1.

Algorithm 1: Steps of BBA methodology

1. Calculate the lower bound: A procedure is used to achieve the lower bound on the objective
function to solve the given sub-problem and deliver an estimated point.
2. Calculate the upper Bound: An upper bound is evaluated to give the solution.
3. Branching method: The sub-problem is portioned to deliver more than two children by a pro-
cedure strategy
4. Methodology search: A search order is programmed to construct the binary tree.
5. If the stopping criterion is satisfied, then end up the iterative scheme.
6. The BBA ends delivering a solution as an output.

Τhe 0 − 1 integer program can be solved utilizing genetic algorithms and binary
particle swarm optimization [18,51]. They follow a probabilistic search of the feasible re-
gion due to a random selection of the initial population, searching and finding an opti-
mum solution [25,26,51]. Strategies such as special creation, crossover, and mutation func-
tions are utilized to enforce the variables to be declared in the integer problem solving
[72]. GA tries to minimize a penalty function, not necessarily the fitness one, and a penalty
term is added to consider any infeasibility while the feasible region is explored [25–
27,50,51]. The penalty function is performed with a binary tournament choice to find in-
dividuals for subsequent generations [25,26,50,51].
If the member is feasible, the penalty function value of a member included in the
population is the fitness function [51]. Otherwise, the objective function counts the maxi-
mum fitness function among feasible members of the population, plus a sum of the con-
straint violations in case of an infeasible point if the member is found to be infeasible [72].
The steps of the GA code are outlined in Algorithm 2. The flowchart is given in Figure 1.

Algorithm 2: Steps of GA implementation

1 Set 𝑘: = 0; starting from the population 𝑃(0);
2 Estimate 𝑃(𝑘);
3 If the stopping criterion is completed, then end up the iterations;
4 Choose 𝑀(𝑘) from 𝑃(𝑘);
5 Evolve 𝑀(𝑘) to create 𝑃(𝑘 + 1);
6 Set𝑘: = 𝑘 + 1, go to step 2.
Algorithms 2024, 17, 191 11 of 62

k := 0

P(k )

P(k )

M (k )

pc
pc

pm

P (k + 1)

k := k + 1

Figure 1. Flowchart of genetic algorithm.

Particle swarm optimization (PSO) does not utilize mutation and crossover parame-
ters [25,26,39,48,51]. PSO is easily implemented without using encoding or decoding of
the parameters into binary strings, such as those in genetic algorithms which can also uti-
lize real-number strings [25,26,39,48,51].
Algorithms 2024, 17, 191 12 of 62

PSO searches the feasible regions by utilizing trajectories of particles in a quasi-sto-

chastic manner. A deterministic strategy and a heuristic strategy are implemented for the
PSO [25,26,39,48,51].
Each particle is attracted toward the position of the current global best and its own
best location in history, while it has a tendency to move accidentally at the same time.
When a particle detects a location that is better than any previously found locations,
then it updates it as the new current best for the particle 𝑖 [25,26,39,48,51].
There is a current best for all 𝑛 particles at any time 𝑡 during the iterative scheme.
The target is to detect the global best among all the current best solutions so that the ob-
jective will no longer improve after a number of iterations [25,26,39,48,51].
An inertial weight parameter is used to properly tune the optimization mechanism
between the local and global best of the particles at each trial estimated point. This opti-
mization factor is used to give an accurate solution within specific tolerances criteria.
The inertia weight is utilized to balance the global exploration and local exploitation
[51]. We use the linear decreasing inertia weight in Equation (12) [25–27,51].
( )×
𝑊=𝑊 −( ) (12)

Although PSO was proposed to optimize unconstrained optimization problems, it

can also solve constraint optimization problems. The PSO is proposed to locate the opti-
mality handling the constraint function in an appropriate way. A mechanism is utilized
when the particle s best position and the swarm global optimal position are updated [21].
To explore the feasible region, a feasibility restriction is utilized to evaluate the constraint
violation and deliver a feasible and optimal solution [25,27,39,47,48,51].
The PSO code is depicted in Algorithm 3 and the flowchart is given in Figure 2 [25,26].

Algorithm 3: Steps of PSO implementation

Objective function𝑓(→ 𝑦 ), →
𝑦 = [𝑦 , 𝑦 , . . . , 𝑦 ].
Initialize locations 𝑦 and velocity 𝑣 of 𝑝 particles.
Determine 𝐺𝑏𝑒𝑠𝑡 from 𝑚𝑖𝑛 𝑓(𝑦 ), . . . . , 𝑓(𝑦 ) at (𝑚 = 0).
while (criterion)
𝑡 = 𝑡 + 1 (Iteration counter)
For loop over all 𝑝 particles and all 𝑛 dimensions
Produce new velocity 𝑉 = 𝑊 × 𝑉 + 𝐶 × 𝑟 × (𝑃 − 𝑋 ) + 𝐶 × 𝑟 × (𝐺 − 𝑋 )
Compare new locations 𝑋 =𝑋 +𝑉
Assess objective functions at new locations 𝑋
Deliver the current best for each particle 𝑃𝑏𝑒𝑠𝑡
end ε
Deliver the current global best 𝐺𝑏𝑒𝑠𝑡
end while
output the final results 𝑃𝑏𝑒𝑠𝑡 and 𝐺𝑏𝑒𝑠𝑡
Algorithms 2024, 17, 191 13 of 62

Start

Population initialization of particles

with random positions and velocities

Calculate fitness Value of each particle

Compare each particle’s fitness assessment with the

current particle to achieve best Pbest

Compare fitness assessment with the population’s

total previous best to achieve best Gbest

V t +1 =W t ×V it + C 1 × r1 × (Pit − X it ) + C 2 × r2 × (G t − X it )
X it +1 = X it +V i t +1

Is Termination
Criterion
Satisfied?

Stop

Figure 2. The flow chart of particle swarm optimization.

5.2. Constrained Binary-Integer Programming Model with a Fixed Number of Channels

The widespread adoption of PMUs is gradually advancing but faces impediments
from technical limitations and economic factors, thus hindering its pace [56–58].
In actual conditions, every PMU has a limited number of channels, and therefore we
need a more realistic installation of PMUs in a way to further decrease the cost of the
investment [47,56,57,59–67].
Every PMU produced by the industry has a restricted number of channels for instal-
lation at a substation [52,56–60]. The number of branches adjacent to a power grid node
in which a PMU is placed is known as channel capacity [60].
Considering the channel limitation capacities there is a reasonable query: which is
the appropriate set of topological constraints to lead to a least PMU number with an ade-
quate number of observed branches without sacrificing the entire observability at all
[52,56–58,60]?
Here we implement the OPP-based model when the number of measurement chan-
nels is restricted and the zero-injection buses are not assumed in the formulation.
We have taken into account a PMU which has an 𝐿 number of channels and is placed
at a bus 𝑘 within the power grid [47,56–67].
Algorithms 2024, 17, 191 14 of 62

We also consider that the bus 𝑘 is adjacent to 𝑁𝑘 number of buses [56–58]. If the
number of channels 𝐿 is greater than the number of neighbors 𝑁𝑘, then a single PMU
installed at the bus will give the phasor voltages at all its neighbor buses [56–58].
Differently, there will be 𝑟𝑘 combinations of possible channel assignments to
branches adjacent to bus 𝑘 [56–67]:
𝑁𝑘
𝑖𝑓 𝐿 ≤ 𝑁𝑘
𝑟 𝐿 (13)
1 𝑖𝑓 𝑁𝑘 < 𝐿
where 𝑁𝑘 is the number of lines joined to bus k, and 𝐿 is the pre-given channel limit for
the PMUs [56–67]. The number of these combinations is defined by [56–67]:
𝑚=∑ 𝑟 (14)
where:
𝑁𝑘 !
𝑐𝑖 = = (15)
𝐿 !×( )!

𝑁𝑘 is the number of branches, 𝐿 is the number of available channels, and 𝑐𝑖 is the

number of combinations of possible channel arrangements to lines adjacent to the related
bus 𝑖 [56–67].
The channel limit constraint function can be adopted so that a PMU posed at a node
will monitor its adjacent nodes by choosing a suitable combination of L [56–58].
The OPP is declared using the 0 − 1 integer linear programming model based on a
restricted number of channels, and can be declared as follows [56–58]:
𝑚𝑖𝑛𝐽(𝑥) = 𝑤 × 𝑌 = ∑ 𝑤 ×𝑦 (16)

𝑓(𝑌) = 𝐵 × 𝑌 ≥ 1⃗ (17)

𝑌 ∈ {0.1} (18)
where 𝑤 = |1,1, … ,1| and the matrix 𝐵 × consists of rows representing the possible
combinations [56–67].
𝑌 × represents a PMU placement design that gives a combination where one bus
is dominant and observes the neighboring buses [56–58].
This optimization problem is equivalent to the set covering problem [28,29,31], that
is, the problem of seeking out the smallest number of subsets in the family of sets that
constitutes the rows of the matrix 𝐵 [56–67].
Hence, a PMU placement is derived in which each element observes a subset of buses
[60]. Y is a placement vector and is formulated by 𝑌 = {𝑦 , 𝑦, … . . , 𝑦 } ∈ {0,1}, where m is
the number of subsets [56–58]. Each subset is constituted by a set of buses where the center
bus observes some of the neighboring buses. The center bus is the dominant bus and is
the first entry in the subset, as noted in [60].
The Y placement vector s nonzero entries will indicate the rows of related grid nodes,
voltage angles of which can be monitored by these synchronized measurements [56–58].
PMU locations are derived and its bus observes a subset of its neighboring buses with the
branch assignments [60–67].
Thereby, we can derive the PMU locations and the branch assignments, which mean
the lines directly observed by the dominant bus or center bus.
Algorithm 4 presents the MILP solver implementation considering a multi-channel
capacity. We define the objective function, the binary connective matrix, the unity vector
and the lower and upper bound of the decision variables. The MILP routine is invoked to
deliver the upper bound for the minimization problem [13–16,72–81].
Algorithms 2024, 17, 191 15 of 62

Algorithm 4: Steps of MILP Solver implementation

MILP Algorithm: Optimal Placement of PMU with multi-channel capacity
1. 𝑤: The set of implementation costs, 𝑤 = |1,1, … ,1| ;
2. 𝑚: Set of possible combinations including a number of nodes;
3. Define a set of variables y= |1,1, … ,1| , 𝑦 ∈ {0,1} ;
4. Define binary connectivity matrix;
5. 𝐽(𝑦) = 𝑤 𝑦; objective function considering costs;
6. /* solving the problem of mixed integers;
7. programming*/
8. Invoke MILP solver in MATLAB environment;
9. solve placement minimizing using mip;
10. 𝑦 → 𝑜𝑝𝑡𝑖𝑚𝑎𝑙 𝑝𝑙𝑎𝑐𝑒𝑚𝑒𝑛𝑡 𝑜𝑓 𝑐𝑜𝑚𝑏𝑖𝑛𝑎𝑡𝑖𝑜𝑛𝑠 𝑓𝑜𝑟 𝑠𝑒𝑙𝑒𝑐𝑡𝑖𝑛𝑔 𝑃𝑀𝑈𝑠 𝑙𝑜𝑐𝑎𝑡𝑖𝑜𝑛𝑠 𝑎𝑛𝑑 𝑏𝑟𝑎𝑛𝑐ℎ 𝑎𝑠𝑠𝑖𝑔𝑚𝑒𝑛𝑡𝑠
11. Solution 𝑦 (display combinations by which a PMU location is the Center and the other locations representing
the PMUs directly observed);
12. Solve linear relaxation problems of the initial MILP model
13. Tighten LP via cutting planes and domain propagation
14. Search for feasible solutions via primal heuristics for the MILP model
15. Branch on variables
16. Output: a binary-feasible solution to the MILP model
 Exit: MILP status (optimal or infeasible) and solution
 Absolute gap: U – L ≤ 𝐴𝑏𝑠𝑜𝑙𝑢𝑡𝑒𝐺𝑎𝑝𝑇𝑜𝑙𝑒𝑟𝑎𝑛𝑐𝑒
 Relative gap: 100× (U − L)/(abs(U) + 1)≤ RelativeGapTolerance

The MILP solver solves the linear programming relaxations and allows temporarily
the integer decision variables to be continuous [13–18,37,38,41,51,72–81].
Dual-simplex or primal-simplex algorithms finally solve the relaxed problems in the
relaxed decision variable range [0,1] [13–17]. The iterative process is completed within
specific criterions to achieve an accurate solution [19–21]. To conclude, four tasks require
attention when the integer problem is solved by a BBA scheme [14]:
1. Formulation so that the absolute gap is small;
2. Heuristic computations are needed to find a good lower bound;
3. Branching;
4. Node selection.
The MILP solver ends up as the process with an optimality gap, which is given count-
ing the upper bound and the lower bound estimated points [13–18,37,38,41,51,72–81].
This gap is determined as the optimal value where 𝑈 is the best integer solution and
𝑙 is the lower bound. If the absolute gap is found equal to zero, the solver returns a solu-
tion at the globally optimum point [37,38,41,51,72–81].

5.3. Constrained Binary-Integer Programming Model with a Single-Line PMU

Instead of using multi-channel PMUs, single-line PMUs observe a single line by cal-
culating the running current in the line and the phasors voltage at the end point of that
line [61]. The problem is referred to as the minimum edge cover problem in graph theory
and it is directly related to the minimum set covering problem, which is a combinatorial
optimization problem [28–31].
An edge cover of a graph 𝐺 is a subset of edges 𝐹 ⊆ 𝐸(𝐺) such that for any 𝑢 ∈ 𝑉,
there exists 𝑒 ∈ 𝐹 such that 𝑢 is an endpoint of 𝑒. As given in the definition, the power
edge set (PES) problem is as follows [28–31,37,38,41,51]:

Definition 5. Declaration of the edge cover problem:

Given a 𝐺 = (𝑉, 𝐸), find an edge subset 𝛱 ⊂ 𝛦 of minimum cardinality such that a place-
ment of PMUs on edges over 𝛱 yields the full observability of the graph 𝐺 [28,29,31]. It is noticed
Algorithms 2024, 17, 191 16 of 62

that the OPP finds a subset of nodes such that the whole graph is topologically observable similar
to the “set covering problem” [28–31].

The problem solving goal is to monitor a single branch, namely the sending end node
voltage and the sending end current flow towards the receiving end bus [61]. We note that
single-line PMUs are placed at either end of the line since, in either case, both end bus
voltages will be calculated due to the availability of the line current [61].
Thus, the problem is declared in installing PMUs to lines and not to nodes. If a PMU
is considered to be a line, it can be installed arbitrarily at either end of that line without
changing the output of the observability-based model [61,64].
Thereby, the optimal branch phasor current measurement problem is to discover the
minimum set of line phasor currents such that a power node is monitored by, at least and
at most, one line phasor current [28–31,61].
The problem is declared as a binary-integer program, that is, a set covering problem,
or otherwise a minimum edge set problem [28–31,61]. An edge cover is a subset of edges
determined similarly to the vertex cover, that is, a collection of graph edges so as a union
of edge endpoints relating to the vertex set of the graph [28–31,61].
The solution of the edge cover problem gives the smallest possible number of edges
for a specific graph [28,29,31,61]. In this paragraph, we study the OPP model with one
channel, such that to guarantee a complete observability of the whole graph 𝐺 [28–32].
If a PMU is selected for a branch, it can be installed arbitrarily at either end of that
branch without changing the outcome of network observability analysis [31,60,61].
Thus, it confirms the observation that the matching is perfect [28,29]. This is true if it
covers every vertex of the graph, that is, every vertex is within exactly one edge of the
matching where an edge is shown as a set of two different vertices [31,60,61].
This case considers the PMU where the PMU installed at a bus monitors a single line
by computing the voltage phasor and the phasor current running this line [60,61].
A PMU number is computed to maintain the whole observability [60,61]. Let us as-
sume a system consists of 𝑛 buses and 𝐿 branches [78]. The OPP problem is declared in
a binary-integer programming format [61]:
𝑚𝑖𝑛 𝐽 (𝑥) = ∑ 𝑐𝑥
⃗ (19)
𝑠. 𝑡. 𝑇𝑋 ≥ 1
𝑥 ∈ {0,1}
where [31,60,61]:
𝐿: 𝑡ℎ𝑒 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑏𝑟𝑎𝑛𝑐ℎ𝑒𝑠
1 𝑎 𝑃𝑀𝑈 𝑖𝑠 𝑝𝑙𝑎𝑐𝑒𝑑 𝑜𝑛 𝑏𝑟𝑎𝑛𝑐ℎ 𝑖 (20)
𝑥 =
0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

𝑐 : 𝑐𝑜𝑠 𝑡 𝑜𝑓 𝑖𝑛𝑠𝑡𝑎𝑙𝑙𝑎𝑡𝑖𝑜𝑛 𝑜𝑓 𝑎 𝑃𝑀𝑈 𝑎𝑡 𝑏𝑟𝑎𝑛𝑐ℎ 𝑖

1 𝑖𝑓 𝑏𝑟𝑎𝑛𝑐ℎ 𝑗 𝑖𝑛𝑐𝑖𝑑𝑒𝑛𝑡 𝑡𝑜 𝑏𝑢𝑠 𝑖 ⃗ (21)
𝑇𝑖𝑗 = , 1 = [1,1, . . . . ,1]
0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

5.4. Constrained Mixed-Integer Nonlinear Programming Formulation

The minimization problem is stated as a nonlinear model following the binary-inte-
ger programming. A linear cost function is minimized subject to a set of polynomial con-
straints within a binary decision bound. Assume that the optimization problem is declared
as follows [19–26]:
𝐽∗ = 𝑚𝑖𝑛 𝐽∗ = 𝐽(𝑦) = ∑ 𝑦 : 𝑦𝜖 ∈ {0,1} (22)
where:
Algorithms 2024, 17, 191 17 of 62

𝛺 ≔ {𝑦𝜖 ∈ {0.1} : 𝑓 (𝑦) = 0, 𝑖 = 1, … . , 𝑛}, 𝑛≤𝑚 (23)

A combination of buses can be derived by declaring a constraint function as [20]:
𝑓 (𝑦 , 𝑦 , … . , 𝑦 ) = 0
⎧
𝑓 (𝑦 , 𝑦 , … . , 𝑦 ) = 0
⎪
…….
(24)
⎨ ……
⎪ ……
⎩𝑓 (𝑦 , 𝑦 , … . , 𝑦 ) = 0.
where 𝑦 , 𝑦 , … . , 𝑦 are unknown optimization variables and each variable is a possible
subset of buses where a center bus exists monitoring a set of its neighboring buses in the
power transmission grid [60].
The mathematical model is solved by the SCIP solver utilizing LP solvers named
SoPlex and IPOPT for solving the nonlinearities [73,74]. Finally, a BBA tree is constructed
through linearized strategies towards optimality [37,38,41,51,72–76,79,81].
To estimate the quality of the solution, the user examines the optimality gap [81]. The
optimal solution is obtained inside an 0.00% tolerance criterion, which is accurately cal-
culated during the repetitive BBA algorithmic scheme [37,38,42,51]. This tolerance crite-
rion is required to end the termination with a successful output [37,38,42,51].

5.5. Constrained Nonlinear Programming Model

Suppose (P) is the optimization problem 𝐽∗ = 𝑚𝑖𝑛{𝐽∗ = 𝐽(𝑦): 𝑥ϵ𝛺} [19–21]:
𝛺 ≔ {𝑦𝜀𝑅 : 𝑓 (𝑦) ≥ 0, 𝑖 = 1, … . , 𝑛, 𝑛 ≤ 𝑚} (25)
For some 𝐽, (𝑓 ) ∁ 𝑅|𝑦| is generally defined on R [19–21,82]. In mathematical algo-
rithms, optimization is executed relying on gradient-based algorithms for minimizing or
maximizing the objective function in the problem [19,21,41] or building a binary tree by
utilizing BBA as shown in [13–15,37,38,41,42,51].

Definition 6. Statement of the local minimizer of the (P) problem。

We declare that 𝑥 ∗ ∈ 𝛺 is a local minimum of the minimization problem 𝑚𝑖𝑛 ⃗∈ 𝐽 (𝑥⃗) if
∀𝑥 ∈ 𝑁 ⊂ 𝛺, (𝑥 ≠ 𝑥 ∗ ) ⇒ (𝑓(𝑥 ∗ )≤𝑓(𝑥)), where N is a neighborhood of the local minimizer, if there
exists 𝜀 > 0 such that 𝑓(𝑥 ∗ ) ≤ 𝑓(𝑥) whenever 𝑥 ∈ ⊂ 𝛺 satisfying|𝑥 − 𝑥 ∗ | ≤ 𝜀.
The objective value 𝑓 ∗ = 𝑓(𝑥 ∗ ) is a local minimum of the optimization problem where 𝛺 is
the feasible region of the optimization problem [19–21].

Definition 7. Statement of the global minimizer of the (P) problem.

We declare that 𝑥 ∗ ∈ 𝛺 is a strict global minimizer of the minimization problem
𝑚𝑖𝑛 ⃗∈ 𝐽 (𝑥⃗) if ∀𝑥 ∈ 𝑆, (𝑥 ≠ 𝑥 ∗ ) ⇒ (𝑓(𝑥 ∗ ) < 𝑓(𝑥)) [19–21].
The objective value 𝐽∗ = 𝐽(𝑥 ∗ ) is a strict global minimum of the optimization problem, where
𝛺 is the feasible region of the optimization problem [19–21].
The objective function 𝐽(𝑥⃗) can have a strict global minimizer at only one solution point. It
is possible to have non-strict global solution at several points [19–21].
Even though the minimum could be global, there can be more than one minimizer at the least
objective value [19–21].

Definition 8. Statement of the (P) problem.

Problem (P) is declared in general as NP-complete [19–21] and one is in many cases satisfied
with the determination of a local minimum point, which can be reached by calling some nonlinear
algorithms [13–15,19–24].

Such kinds of nonlinear algorithms are based on local search of a neighborhood in

the feasible set constituted by the objective function, the constrained function, and bounds
to hopefully finding a good local minimum point [19,21–24].
Algorithms 2024, 17, 191 18 of 62

Such algorithms estimate a trial point at the current iteration, which is some feasible
point solution belonging to the feasible region [19–24].
Starting from a first estimated solution, a sequence of trial points is generated,
𝑥 , 𝑥 , … , 𝑥 → 𝑥 ∗ , which is hoped to converge to an optimum point [19–24].
That point must satisfy the Karush–Kuhn–Tucker (KKT) optimality conditions
within a given ϵ-tolerance [19–24,41,73,82].
Then, KKT is necessary for determining a local minimum of (P). These optimality
conditions are required for a trial point to be characterized as a local minimum point, and
actually, yet for a stationary point of the Lagrangian function. Hence, any stationary point
must satisfy these sufficient and required conditions for locality [19–24].
In such optimality conditions, the real question if the local minimizer of the cost func-
tion is global at the same time becomes a reasonable query [37,41,42,51].
As the reader may understand, in such NP-complete problems, the difficulty lies in
finding a good approximation of the true optimal solution point [17–19,37,38,41,51]. For
that needed purpose, some sufficient conditions are required to characterize if a solution
point is a global one or not [19–26].
First-order and second-order conditions are able to work together for non-convex
optimization problems by determining local minimum points strongly relying on the se-
lection of the initial point [19–25].
We can use local algorithms such as interior-point methods (IPMs) and sequential
quadratic programming (SQP) methods to find locality, and if we adopt multi-start ini-
tialization techniques [41], a global solution can be reached with high probability [19–25].
SQP algorithms solve a sequence of optimization sub-problems, each of which opti-
mizes a quadratic model of the objective under a linearization of the constraint function.
The nonlinear algorithms can reach a local solution in a reasonable running time [19–
21]. The other option is to adopt a global optimizer routine to determine a global mini-
mum point but at a slower time [19–25,37,38].
Furthermore, derivative-free optimizers (DFOs) are adopted to handle such a kind of
nonlinear programming model [22,41,70,71].
DFO is a totally free direct-search algorithm that calculates the gradient of the objec-
tive and constraint functions per iteration [22,41,70,71].
Direct-search algorithms guarantee global convergence to a local solution with opti-
mality conditions fully satisfied [22,41,70,71]. For that convergence, the Clarke calculus is
adopted for handling non-smooth functions [22,41,70,71].
The generalized pattern search algorithm (GPSA) is a direct-search algorithm per-
fectly suited to PMU allocation problem solving [22,41,70,71]. The search methodology is
implemented to estimate a limited number of trial points [22,41,70,71].
GPSA executes a global exploration of the feasible set and permits us to optimize the
programming model [22,41,70,71]. Τhe aim is to find a local solution point [22].
In the case of failing to find any improvement, the poll suggests trial points in the
neighborhood of the incumbent solution relying on suitable optimality conditions [22,41].
The poll methodology guarantees that the algorithm has a global convergence prop-
erty and that a local minimum can be reached [22,41,70,71].
This optimization problem is smooth but many local minimum points can be found
by gradient-based algorithms or derivative-free optimizers [20,41,70,71]. The nonlinear
optimization algorithms usually present the general iterative scheme for the optimization
problem of the form in Algorithm 5 [19–23,82]:
𝑚𝑖𝑛{𝐽(𝑥): 𝑥𝜀𝛺} (26)

Algorithm 5: Steps of nonlinear algorithms

Step 0: initial estimate 𝑥 𝜀𝛺, we set the iterative number equal to zero, 𝑖 = 0
Step 1: Computation of the search direction ℎ = ℎ(𝑥 )
Algorithms 2024, 17, 191 19 of 62

Step 2: Computation of step-length 𝜆 such as:

𝑧(𝑥 + 𝜆 × ℎ ) = 𝑚𝑖𝑛 {𝑧(𝑥 + 𝜆 × ℎ )}: 𝜆 ≥ 0 , 𝑥 + 𝜆 × ℎ 𝜀𝛺
Step 3: 𝑥 =𝑥 +𝜆 ℎ
Step 4: if‖𝑥 − 𝑥 ‖ ≤ 𝜀, then stop. Otherwise, set 𝑖 = 𝑖 + 1 and go to step 1.

𝐽∗ = 𝑚𝑖𝑛{𝐽∗ = 𝐽(𝑦): 𝑥𝜖𝛺} is solved either with SQP or IPMs [19–21]. The objective
function is declared in a quadratic form and the minimization problem is formulated as
follows [19–25,41]:
𝑚𝑖𝑛𝐽(𝑦) = 𝑦 × 𝑊 × 𝑦, 𝑊 ∈ 𝐼 × (27)

𝑠. 𝑡. 𝑓(𝑦) = 0 (28)

𝑦≥0 (29)
where 𝐽: 𝑅 → 𝑅, 𝑦: 𝑅 → 𝑅 are differentiable functions. Furthermore, nonlinear ine-
quality constraints can be formed in this formulation by utilizing slack variables [19–21].
The IPM strategy achieves a solution by approximately solving a sequence of barrier
problems of the following format [19–21,24,25,41,72,73]:
𝑚𝑖𝑛𝜑 (𝑦) = 𝐽(𝑦) − 𝜇 ∑ 𝑙𝑛(𝑦 ) (30)

𝑠. 𝑡. 𝑓(𝑦) = 0 (31)
where µ is a barrier parameter that is reducing and converges to zero [19–21]. For this
solution, Newton (or quasi-Newton) methods are utilized for convergence [19–21]
whereas some algorithms utilize SQP (or trust-region) methods [19–21].
The Newton methods utilize a search direction strategy from solving a Newton equa-
tion of the perturbed optimality system [19–21]. To obtain a local convergence algorithm
in a fast manner, the barrier parameter needs to be updated carefully [19].
The SQP and IPMs are performed through the Fmincon optimizer function included
in the MATLAB optimization toolbox [41]. The optimizer solver is able to initialize the
iterative process using a multi-start search procedure in order to terminate to a local solu-
tion point, which is approximately the global one with high accuracy [72].
Furthermore, IPMs can be performed by calling the IPOPT solver and reaching a local
point [73]. Both algorithms end the iterative process with specific tolerances [19–21]. The
criteria tolerances are adequate to guarantee convergence to a local point [21].
The objective function is estimated at each iteration taking into account specific tol-
erances such as the |(𝑥 − 𝑥 )| < 𝑆𝑡𝑒𝑝𝑇𝑜𝑙𝑒𝑟𝑎𝑛𝑐𝑒 × (1 + |𝑥 |) and the function toler-
ance |𝑓(𝑥 ) − 𝑓(𝑥 )| < 𝐹𝑢𝑛𝑐𝑡𝑖𝑜𝑛𝑇𝑜𝑙𝑒𝑟𝑎𝑛𝑐𝑒 × (1 + |𝑓(𝑥 )|) [19–21,41,72,73].
Furthermore, the constraint tolerance is considered an upper limit for any constraint
violation at each iteration [21]. The constraint function magnitude must not exceed the
constraint tolerance to avoid the violation of the feasibility [19–21,41,72,73,79].
Considering the importance of the selection option in handling the constraint func-
tion, a variety of mathematical and heuristic algorithms are adopted for solving the OPP-
based model with a limited amount of channel capacity. Our aim is the following:
• Instead of searching local solution points for the optimization models, we examine
all the sufficient and guaranteed optimality conditions for finding global solutions.
• Hence the testing of optimality yields the exact solutions by utilizing the BBA, SQP,
IPMs, GPSA, and BPSO for simulation purposes.
All mathematical algorithms evaluate a linear objective of a quadratic function
through the iterations required to first ensure a feasible trial point, and then optimality,
etc. [13–16,19–21,37,38,41,42,51,82].
Algorithms 2024, 17, 191 20 of 62

The other concern is how to improve the performance of convergence towards an

optimal solution using BBA, nonlinear, and heuristic algorithms [37,38,41,42,51].
Finally, the simulation running of the algorithms reveals the possible combinations
of PMUs, the PMU locations, and the branch assignments [56–61].

6. The IEEE-14-Bus System for Illustrating the Mathematical Models and Numerical
Results Obtained by the Optimization Procedure
In line with the power grid utility industry s concern, the number of constraints is
increasing and it is strongly dependent on the quantity of channel capacity [56–58,60,61].
We first declare the 0 − 1 integer program and the mixed-integer nonlinear program, and
finally obtain the pure nonlinear mathematical models [13–15,23].
The 0 − 1 integer program is declared with a linear objective function to be mini-
mized under specific inequality constraints under a topological analysis.
We consider that the power grid is fully observable with an appropriate PMU num-
ber under the channel restriction capacity. The solution is given by BBA, Gas, and BPSO.
Then, we transform it into a 0 − 1 nonlinear model solved by BBA. Finally, we examine
the conditions under which the pure nonlinear problem is solved using SQP, IPMs, and
GPSA. As a benchmark case study, we use the IEEE-14-bus system and a binary-integer
linear program is firstly studied following the works of Korkali et al. [56–58].
The matrix B, which consists of rows representing the possible combinations, is de-
picted in Equation (32) for the IEEE-14-bus system (Figure 3) [52,56–60,62–64].
1 1 0 0 1 0 0 0 0 0 0 0 0 0
⎡1 1 1 1 1 0 0 0 0 0 0 0 0 0⎤
⎢0 1 1 1 0 0 0 0 0 0 0 0 0 0⎥
⎢ ⎥
0 1 1 1 1 0 1 0 0 0 0 0 0 0
⎢ ⎥
0 1 1 1 1 0 0 0 1 0 0 0 0 0
⎢ ⎥
⎢0 1 1 1 0 0 1 0 1 0 0 0 0 0⎥
⎢0 1 0 1 1 0 1 0 1 0 0 0 0 0⎥
⎢0 0 1 1 1 0 1 0 1 0 0 0 0 0⎥
1 1 0 1 1 1 0 0 0 0 0 0 0 0⎥
𝐵=⎢ (32)
⎢ 0 0 0 0 1 1 0 0 0 0 1 1 1 0⎥
⎢0 0 0 1 0 0 1 1 1 0 0 0 0 0⎥
⎢0 0 0 0 0 0 1 1 0 0 0 0 0 0⎥
⎢0 0 0 1 0 0 1 0 1 1 0 0 0 1⎥
⎢0 0 0 0 0 0 0 0 1 1 1 0 0 0⎥
⎢0 0 0 0 0 1 0 0 0 1 1 0 0 0⎥
⎢0 0 0 0 0 1 0 0 0 0 0 1 1 0⎥
⎢0 0 0 0 0 1 0 0 0 0 0 1 1 1⎥
⎢0 0 0 0 0 0 0 0 1 0 0 0 1 1⎥
Algorithms 2024, 17, 191 21 of 62

19 20
14
12 13

12
11 17

13 18 16
9
6 14
11 10
G
8
15
10
7
8
1 9
5
7
4
1
2 3
4 6

5
2 3

Bus number Line number

Figure 3. IEEE-14-bus system.

Suppose that the channel limit is equal to four, that is 𝐿 = 4, and the possible number
combinations are equal to 18, 𝑚 = ∑ 𝑟 = 18 [56–58,60]. The set of all subsets in the
case of considering four channels is the following [60]:
𝑆 = {1,2,5}, {2,1, 3, 4, 5}, {3, 2, 4}, {4, 2, 3, 5, 7}, {4, 2, 3, 5, 9}, {4, 2, 3, 7, 9}
{4,2, 5, 7, 9}, {4,3, 5, 7, 9}, {5, 1, 2, 4, 6}, {6, 5, 11, 12, 13}
{7,4, 8, 9}, { 8, 7}, {9, 4, 7, 10, 14}, {10, 9, 11}, {11, 6, 10} (33)
{12, 6, 13}, {13, 6, 12, 14}, {14, 9, 13}
Each first entry of the subset is the center bus used to monitor the adjacent buses. For
example, 2 is the center in the second subset, 8 is the center in the 12th subset, 10 is the
center in the 14th subset, and 13 is the center in the 17th set [60].
When 𝐿 = 4, and solving the 0 − 1 ILP model for the IEEE-14-bus network in Figure
3, four combinations are derived, that is, {2, 12, 14, 17}. Assume that 𝑥 ∈ {0,1}; declare a
binary decision variable such that 𝑥 = 1 if a PMU is placed at the center of the subset of
the family S, and 𝑥 = 0 otherwise, where s is an index within the family of the subsets
[60].
The algorithm has chosen as a solution the set {2, 12, 14, 17}. From that subset, and
considering the yes or no decision [15], a PMU placement is obtained in which the center
is declared as 1 in the placement vector and monitors all buses inside that subset [60].
The algorithmic scheme delivers PMUs installed at buses {2, 8,10,13}, which are cen-
ters of the subsets {2,1,3,4,5}, {8,7}, {10,9,11}, {13,6,12,14} , respectively [60]. Thus, the
placement result is a binary vector with yes” or no” decisions [13–18].
𝑥 = [0 1 0 0 0 0 0 1 0 1 0 0 1 0] (34)
Algorithms 2024, 17, 191 22 of 62

Furthermore, the algorithms identify not only the PMU location, but also the branch
assignments considering the extra restriction, which is a fixed amount of capacity. The
solution is adequate to decrease the entire cost without sacrificing observability. For ex-
ample, the PMU at bus 2 monitors buses 2, 1, 3, 4, and 5. The PMU locations are {2, 8, 10,
13}, where the branch assignments are the PMU channels presented in Table 1.

Table 1. Channel capacity, PMU number, PMU locations, and PMU channels.

L nPMU PMU Locations PMU Channels

{2-1, 2-3, 2-4, 2-5}
{8-7}
4 4 2, 8, 10, 13
{10-9, 10-11}
{13-6, 13-12, 13-14}

The above results are produced by executing YALMIP s BBA and the log file is pre-
sented in Algorithm 6 [75,77]. YALMIP s BBA optimizes the 0 − 1 integer model declared
in Equations (33)–(35). The log file is presented in Algorithms 6 and 7 [37,38,72,74–
77,79,80].
The binary decision variable is displayed in Table 1 and it is a placement in which
each element is the 𝑗 − 𝑡ℎ combination related to the center bus 𝑖 [56–61].
The outcome presents a PMU set solution with all network buses observed. One bus
serves as the center interconnected with neighboring buses through branches, such that
all buses are reachable by synchronized measurements and make the power grid fully
observable [56–67].
The set {2, 12, 14, 17} corresponds to PMU locations, namely {2, 8, 10, 13} illustrated
in Figure 3. A PMU placed at bus 2 will monitor buses 1, 2, 3, 4 and 5; the one at bus 8 will
monitor buses 7 and 8; the one at bus 10 will monitor buses 9, 10, and 11; and the one at
bus 13 will monitor buses 6, 12, 13, and 14. Then, the observability is preserved.
Thus, the system observability is preserved with the best trade-off relationship be-
tween the channel restriction capacity and the best PMU locations. Hence, we succeed in
finding PMU set solutions which observe branches with an increased redundancy.
Algorithms 6 and 7 present the output logs produced by BBA optimizing the integer
programs [14–16]. The solution is achieved within a 0.00% optimality criterion [15].

Algorithm 6: Steps of optimizing YALMIP s BBA

+ Solver chosen : BMIBNB
+ Processing objective function
+ Processing constraints
+ Branch and bound started
* Starting YALMIP global branch & bound.
* Upper solver : INTLINPROG
* Lower solver : INTLINPROG
* LP solver : INTLINPROG
* -Extracting bounds from model
* -Performing root-node bound propagation
* -Calling upper solver + Calling INTLINPROG
(Found a solution!)
* -Branch-variables : 0
* -More root-node bound-propagation
Algorithms 2024, 17, 191 23 of 62

* -Performing LP-based bound-propagation

* -And some more root-node bound-propagation
* Starting the b&b process
Node Upper Gap (%) Lower Open Time
+ Calling INTLINPROG
1 : 2.64286E+00 0.00 2.64286E+00 0 0s Poor
lower bound
* Finished. Cost: 2.6429 (lower bound: 2.6429, relative gap 0%)
* Termination with all nodes pruned
* Timing: 27% spent in upper solver (1 problems solved)
* 12% spent in lower solver (1 problems solved)
* 1% spent in LP-based domain reduction (0 problems solved)
* 1% spent in upper heuristics (0 candidates tried)
Columns 1 through 15
0 1 0 0 0 0 0 0 0 1 1 0
1 0 0
Columns 16 through 18
0 0 0
2 10 11 13
+ Solver chosen : BMIBNB
+ Processing objective function

Algorithm 7: Steps of optimizing YALMIP s BBA

+ Solver chosen : BMIBNB
+ Processing objective function
+ Processing constraints
+ Branch and bound started
* Starting YALMIP global branch & bound.
* Upper solver : INTLINPROG
* Lower solver : INTLINPROG
* LP solver : INTLINPROG
* -Extracting bounds from model
* -Performing root-node bound propagation
* -Calling upper solver + Calling INTLINPROG
(Found a solution!)
* -Branch-variables : 0
* -More root-node bound-propagation
* -Performing LP-based bound-propagation
* -And some more root-node bound-propagation
* Starting the b&b process
Node Upper Gap (%) Lower Open Time
Algorithms 2024, 17, 191 24 of 62

+ Calling INTLINPROG
1: 2.64286E+00 0.00 2.64286E+00 0 0s Poor
lower bound
* Finished. Cost: 2.6429 (lower bound: 2.6429, relative gap 0%)
* Termination with all nodes pruned
* Timing: 27% spent in upper solver (1 problems solved)
* 12% spent in lower solver (1 problems solved)
* 1% spent in LP-based domain reduction (0 problems
solved)
* 1% spent in upper heuristics (0 candidates tried)
Columns 1 through 15
0 1 0 0 0 0 0 0 0 1 1
0 1 0 0
Columns 16 through 18
0 0 0
2 10 11 13

An optimal solution was achieved with an absolute gap equal to zero, which means
a better solution cannot be delivered by the iterative process [13,37,38,41,51].
Considering the case of the unlimited PMU channel capacity [31–33,40,46], the best
system observability index was found in [33,37,38,41–45,47,51].
We declare a b-objective optimization model to find solutions with a higher observa-
bility index. The mathematical programming model is stated in a 0 − 1 ILP domain [14]:
1
𝑚𝑖𝑛𝐽(𝑥) = 𝑤 × 𝑌 − ×𝑒 × 𝐵 ×𝑌 (35)
𝑛

𝑓(𝑌) = 𝐵 × 𝑌 ≥ 1⃗ (36)

𝑌 ∈ {0,1} (37)
where 𝑤 = |1,1, … ,1| , 𝑒 = |1,1, … ,1| , 𝐵 × , 𝑌 × [56–58], and 𝑛 is the number of
power grid nodes. Table 1 presents the log file produced by BBA [37,38,41,51,75,77].
For the minimization problem, the BBA builds a sequence of sub-problems to try to
find an optimum point solution [72,79,80]. The solution of the relaxed problems gives the
upper and lower bounds on the objective function solution [72,74–76,79–81].
The first upper bound is any feasible point whereas the first lower bound is the solu-
tion to the relaxation problem. The upper bound is the objective s function bound.
Τhe algorithm constantly estimates the upper bound and lower bound to reach opti-
mality for the minimization problem [17,72,74–76,79–81].
For the minimization problem studied in this paper, the upper bound is assumed to
be the solution to the problem solving. The lower bound is the reason for closing the ab-
solute tolerance gap delivering a global solution point [74,75,79,80].
The solver returns a solution within an 0.00% optimality criterion, proving that no
better optimum can be achieved. Thereby, the solution is a global optimum point.
The optimal result { 2, 10, 11, 13} is interpreted as follows: From Equation (33),
we found the bus combination set solutions:
{2,1, 3, 4, 5}, {6, 5, 11, 12, 13}, {7,4, 8, 9}, {9, 4, 7, 10, 14}. Each first entry in each subset is the
center bus, which is the dominant bus, and that bus is installed as a PMU [60], i.e., the
PMU set solution {2, 6, 7,9} [60]. We definitely achieved a set solution with the maximum
Algorithms 2024, 17, 191 25 of 62

observability index for a fixed channel capacity 𝐿 = 4. The branch assignments are ob-
served by optimal placement of PMUs at power nodes and are as follows:
2 − 1, 2 − 3, 2 − 4, 2 − 5
6 − 5, 6 − 11, 6 − 12, 6 − 13
7 − 4, 7 − 8, 7 − 9
9 − 4, 9 − 7, 9 − 10, 9 − 14
When 𝐿 = 4, solving the 0 − 1 ILP model for the IEEE-14-bus network in Figure 3,
four combinations are derived, that is, {2, 10, 11, 13}. The algorithmic scheme delivers
PMUs installed at buses {2, 6, 7 , 9} of subsets
{2,1,3,4,5}, {6,5,11,12,13}, {7,4,8,9}, {9,4,7,10,14}.
The PMU locations are {2, 6, 7, 9}, where the branch assignments are displayed in
Table 2 [64]. The 0 − 1 ILP model with a single line is optimized with the intlinprog op-
timizer function [37,38,41,51]. Optimal set solutions are derived and presented in Table 3
for the IEEE-14-bus, 30-bus, 57-bus, and 118-bus systems. Each number reflects a line
where a single-line PMU is assigned across the two-points of the line [61].
Two set configurations are derived based on the strategy by which the linear pro-
gramming relaxations are solved. We utilize the primal-simplex algorithm and the dual-
simplex algorithm to deliver a binary feasible solution for the initial model.

Table 2. Channel capacity, PMU number, PMU locations, and PMU channels.

L nPMU PMU Locations PMU Channels

{2-1, 2-3, 2-4, 2-5}
{6-5, 6-11, 6-12, 6-13}
4 4 2, 6, 7, 9
{7-4, 7-8, 7-9}
{9-4, 9-7, 9-10, 9-14}

Table 3. Algorithm BBA: Optimizing the 0 − 1 ILP model with a single line.

IEEE Power PMU Solution Algorithm

Branches
Systems Number Primal Simplex Dual Simplex
14-bus 20 7 1, 6, 10, 14, 17, 18, 19 2, 3, 9, 12, 14, 18, 20
1, 4, 8, 10, 13, 16, 20, 21, 23, 25, 29,
30-bus 41 15 2, 3, 8, 10, 13, 16, 20, 21, 23, 25, 29, 32, 34, 36, 39
32, 34 36, 39
2, 4, 6, 7, 9, 13, 17, 26, 29, 31, 33, 36,
2, 4, 6, 8, 14, 17, 26, 29, 31, 33, 35, 38, 40, 43, 45, 47,
57-bus 80 29 38, 40, 43, 45, 47, 49, 55, 57, 58, 60,
50, 52, 55, 59, 61, 63, 65, 68, 70, 71, 72, 74, 76
62, 64, 68, 70, 71, 73, 76
1, 4, 6, 9, 10, 17, 18, 22, 24, 27, 29, 32,
1, 4, 6, 9, 10, 18, 19, 22, 24, 27, 29, 32, 34, 37, 40, 46, 34 37, 40, 46, 48, 55, 58, 60, 61 64,
48, 55, 58, 60, 61, 64, 73, 76, 79, 81, 83, 91, 93, 95, 73, 78, 81, 83, 87, 90, 94, 96, 99, 101,
118-bus 186 61 96, 103, 105, 111, 113, 114, 120, 122, 126, 129, 132, 105, 111, 113, 114, 118, 122, 126, 129,
134, 137, 140, 144, 147, 152, 157, 159, 162, 165, 170, 132, 134, 137, 140, 144, 147, 153, 157,
172, 173, 176, 177, 179, 182, 183, 184, 186 158, 162, 168, 172, 173, 174 176, 177,
179, 182, 183, 184, 185

We present the PMU numbers, PMU locations, and the related PMU channels for the
IEEE-14-bus system in Tables 4 and 5 [78]. Two instances are studied. The first gives results
produced by minimizing a one-product objective function [56–58].
Algorithms 2024, 17, 191 26 of 62

Table 4. Optimizing the 0 − 1 ILP with multiple lines for IEEE-14 bus.

Primal-Simplex Dual-Simplex
L NPMU
PMU Locations PMU Channels PMU Locations PMU Channels
{2-1} {2-1}
{4-9} {4-3}
{5-1} {6-5}
2, 4, 5, 6, 8, 10, 2, 4, 6, 8, 9, 10,
1 7 {6-12} {8-7}
13 13
{8-7} {9-14}
{10-11} {10-11}
{13-14} {13-12}
{2-3, 2-5} {2-1, 2-5}
{5-1, 5-6} {3-2, 3-4}
2 5 2, 5, 7, 10, 13 {7-4, 7-8} 2, 3, 7, 11, 13 {7-8, 7-9}
{10-9, 10-11} {11-6, 11-10}
{13-12, 13-14} {13-12, 13-14}
{2-1, 2-3, 2-5} {2-1, 2-3, 2-5}
{6-11, 6-12, 6-13} {7-4, 7-8, 7-9}
3 4 2, 6, 8, 9 2, 7, 10, 13
{8-7} {10-9, 10-11}
{9-4, 9-10, 9-14} {13-6, 13-12, 13-14}
{2-1, 2-3, 2-4, 2-5} {2-1, 2-3, 2-4, 2-5}
{7-4, 7-8, 7-9} {8-7}
4 4 2, 7, 11, 13 2, 8, 10, 13
{11-6, 11-10} {10-9, 10-11}
{13-6, 13-12, 13-14} {13-6, 13-12, 13-14}
{2-1, 2-3, 2-4, 2-5} {2-1, 2-3, 2-4, 2-5}
{7-4, 7-8, 7-9} {8-7}
5 4 2, 7, 11, 13 2, 8, 10, 13
{11-6, 11-10} {10-9, 10-11}
{13-6, 13-12, 13-14} 13-6, 13-12, 13-14}

Table 5. Algorithm BBA: Optimizing the 0 − 1 b-objective ILP with multiple lines for IEEE-
14 bus.

Primal-Simplex Dual-Simplex
L NPMU
PMU Locations PMU Channels PMU Locations PMU Channels
{2-3} {2-1}
{5-1} {4-3}
{6-12} {6-5}
2, 5, 6, 7, 9, 10,
1 7 {7-8} 2, 4, 6, 7, 9, 10, 13 {7-8}
13
{9-4} {9-4}
{10-11} {10-11}
{13-14} {13-12}
{2-1, 2-3} {2-1, 2-5}
{6-5, 6-13} {4-2, 4-3}
2 5 2, 6, 7, 10, 13 {7-4, 7-8} 2, 4, 7, 11, 13 {7-8, 7-9}
{10-9, 10-11} {11-6, 11-10}
{13-6, 13-12} {13-12, 13-14}
{2-1, 2-3, 2-5} {2-1, 2-3, 2-5}
{6-11, 6-12, 6-13} {6-11, 6-12, 6-13}
3 4 2, 6, 7, 9 2, 6, 7, 9
{7-4, 7-8, 7-9} {7-4, 7-8, 7-9}
{9-4, 9-10, 9-14} {9-7, 9-10, 9-14}
{2-1, 2-3, 2-4, 2-5} {2-1, 2-3, 2-4, 2-5}
4 4 2, 6, 7, 9 2, 6, 7, 9
{6-5, 6-11, 6-12, 6-13} {6-5, 6-11, 6-12, 6-13}
Algorithms 2024, 17, 191 27 of 62

{7-4, 7-8, 7-9} {7-4, 7-8, 7-9}

{9-4, 9-7, 9-10, 9-14} {9-4, 9-7, 9-10, 9-14}
{2-1, 2-3, 2-4, 2-5} {2-1, 2-3, 2-4, 2-5}
{6-5, 6-11, 6-12, 6-13} {6-5, 6-11, 6-12, 6-13}
5 4 2, 6, 7, 9 2, 6, 7, 9
{7-4, 7-8, 7-9} {7-4, 7-8, 7-9}
{9-4, 9-7, 9-10, 9-14} {9-4, 9-7, 9-10, 9-14}

The second produces results by optimizing a b-objective function under a set of ine-
quality constraints [41,51]. The b-objective is minimized under a certain number of con-
straints so that an optimal point can be determined [38,41,51]. Intlinprog removes the de-
cision bound integrality and permits all variables to be temporarily real [17,18].
The solution was found by solving the relaxed problems in the BBA construction tree
and this solution is feasible and optimal for the initial MILP [13–15,21]. Each instance was
derived within the pre-specified stopping and tolerance criterion [13–16,19,21]. BBA uti-
lizes the optimality gap as a stopping criterion to end the iterative scheme [13–16]. It is
calculated for all instances returning a zero value.
We observe that the placement set gives an optimum point where each center ob-
serves the maximum number of power nodes [78]. Each subset is a connected dominating
set with the characteristic that each center is a vertex reaching any other vertex inside that
subset by a connectivity line [30,60]. Therefore, the observability is attained even in a re-
stricted channel capacity scenario, as noted in [56–62].
In the meantime, we suggest a mixed-integer nonlinear program that leads to opti-
mality and works together with a similar formulation in the continuous domain. We per-
form the 0 − 1 integer nonlinear programming using a global BBA embedded in the
YALMIP toolbox and the log files produce the desired output [37,38,75,77].
The 0 − 1 nonlinear integer programming model is declared using the IEEE-14-bus
system. The polynomial constraint function is written. The objective function is minimized
subject to a set of equality constraints within binary decision variables [19–21,37,38]. The
instance is solved (𝑔𝑎𝑝 = 0) as an empty entry, when the 𝑈𝑏 coincides with the best
value whereas the 𝑙𝑏 closes the optimality gap [37,38,41,51,72,74–76,79,80]. Therefore, the
0 − 1 nonlinear integer programming model is exactly solved by calling the BBA embed-
ded in the SCIP optimizer function [74].

𝑚𝑖𝑛𝐽(𝑦) = ∑ 𝑤 ×𝑦 𝑤 = 1, 𝑖 = 1,2, … 18, 𝑚 = 18 (38)

Algorithms 2024, 17, 191 28 of 62

𝑓(𝑦)
1 − 𝑦(1) × 1 − 𝑦(2) × 1 − 𝑦(9) = 0
⎧
⎪ 1 − 𝑦(1) × 1 − 𝑦(2) × 1 − 𝑦(3) × 1 − 𝑦(4) × 1 − 𝑦(5) × 1 − 𝑦(6) × 1 − 𝑦(7) × 1 − 𝑦(9) = 0
⎪ 1 − 𝑦(2) × 1 − 𝑦(3) × 1 − 𝑦(4) × 1 − 𝑦(5) × 1 − 𝑦(6) × 1 − 𝑦(8) = 0
⎪
⎪ 1 − 𝑦(2) × 1 − 𝑦(3) × 1 − 𝑦(4) × 1 − 𝑦(5) × 1 − 𝑦(6) × 1 − 𝑦(7) × 1 − 𝑦(8)
⎪ × 1 − 𝑦(9) × 1 − 𝑦(11) × 1 − 𝑦(13) = 0
⎪
⎪ 1 − 𝑦(1) × 1 − 𝑦(2) × 1 − 𝑦(4) × 1 − 𝑦(5) × 1 − 𝑦(7) × 1 − 𝑦(8) × 1 − 𝑦(9) × 1 − 𝑦(10) = 0
⎪ 1 − 𝑦(9) × 1 − 𝑦(10) × 1 − 𝑦(15) × 1 − 𝑦(16) × 1 − 𝑦(17) = 0
⎪
1 − 𝑦(4) × 1 − 𝑦(6) × 1 − 𝑦(7) × 1 − 𝑦(8) × 1 − 𝑦(11) × 1 − 𝑦(12) × 1 − 𝑦(13) = 0 (39)
=
⎨ 1 − 𝑦(11) × 1 − 𝑦(12) = 0
⎪ 1 − 𝑦(5) × 1 − 𝑦(6) × 1 − 𝑦(7) × 1 − 𝑦(8) × 1 − 𝑦(11) × 1 − 𝑦(13) × 1 − 𝑦(14) × 1 − 𝑦(18) = 0
⎪
⎪ 1 − 𝑦(13) × 1 − 𝑦(14) × 1 − 𝑦(15) = 0
⎪ 1 − 𝑦(10) × 1 − 𝑦(14) × 1 − 𝑦(15) = 0
⎪
⎪ 1 − 𝑦(10) × 1 − 𝑦(16) × 1 − 𝑦(17) = 0
⎪ 1 − 𝑦(10) × 1 − 𝑦(16) × 1 − 𝑦(17) × 1 − 𝑦(18) = 0
⎪
⎪ 1 − 𝑦(13) × 1 − 𝑦(17) × 1 − 𝑦(18) = 0
⎩ 𝑦𝜀{0.1}

𝑦 is an integer decision variable that takes two values, i.e., 0 or 1 [16]. The value 1
means the 𝑖th PMU combination is installed at the center bus [60]. The 0 − 1 nonlinear
programming is well posed given that the number of equality constraints is fewer than
the number of the decision variables [19–24,70,82]. Using the IEEE-14 bus and the two
instances, two optimal results are derived for a fixed-channel equal to four [64]. The first
model optimizes an objective function with one product and the second a b-objective
(Equation (35)) to find a better redundancy, as declared in [41,51].
Optimizing the augmented objective function, we increase the measurement redun-
dancy and find a better optimal solution. The problem can be written in a compact form.
We combine discrete decision variables, a linear objective function, and non-convexities
in the constraint function and a global optimization execution.
We solve two instances, namely with one objective and a two-product function opti-
mization model. The mathematical model is generally stated as follows [23,24]:
𝑚𝑖𝑛{⌊𝑤, 𝑦⌋: 𝑦𝜖𝛺}
(40)
𝛺 ≔ {𝑦𝜖[0,1]: 𝑓(𝑦) = 0, 𝑦 ∈ 𝑍∀𝑖 ∈ 𝐼}
where y is the objective function and w is a unity vector w = [𝑤 , 𝑤 , … . , 𝑤 ] [56–58].
The 0 − 1 nonlinear program is solved by BBA [37,38,74–76,81]. YALMIP s BBA and
SCIP (Solve Constraint Integer Programming) optimizer tools can solve the 0 − 1 nonlin-
ear program. A solution is found within an optimality gap given by [72,74–76,80,81]:
percentage relative gap = 100 × |primal − dual|/MIN(|dual|, |primal|) (41)
The MILP optimizer routine ends the iterative process with an optimal solution when
the absolute gap between the upper and lower bounds is less by the default value. The
default value can be either 1 × 10−6 or 1 × 10−10. The optimal solution was found within a
minimum value [72,74–76,79–81]. The upper and the lower bounds are computed in this
instance. YALMIP s BBA invokes local and MILP optimizer routines to build the BBA tree
[37,38,74–76,79]. Fmincon delivers the upper bound whereas the lower bound is solved by
the SCIP optimizer through linear programming relaxations [37,38,75]. Furthermore, the
lower bound can be calculated using intlinprog, MOSEK, and Gurobi [72,74–76,79–81].
YALMIP s BBA initially tries to discover the best feasible solution; then, calculations
are implemented by the local [73] and lower solvers [80]. An LP-based domain reduction
Algorithms 2024, 17, 191 29 of 62

computation as well as some needed heuristic computations are performed to obtain the
solution inside the ∈-optimality criterion [13–16,79].
The optimization terminates when the difference between those bounds computed
by YALMIP s BBA routine is within the default value of the absolute gap [37,38,75].
As we observed, the upper bound is the best solution for the minimization process,
whereas the lower bound minimizes the gap (lower bound: 4, relative gap 0%), as dis-
played in the output log in Table 6 [75]. Finally, the termination ends with all nodes
pruned, as displayed in the output log of YALMIP s BBA [37,38,74–77,79].
Tables 6–14 illustrate results towards optimality [37,38]. The number of branch vari-
ables is equal to the number of possible combinations [60].
Further, we can select an upper bound of channel capacity to reduce the cost of in-
stallation (Tables 4 and 5). Recent research reveals that the upper bound of capacity is four
without sacrificing the essential task of observability [56–67].
The ILP solver gives a precise solution inside a 0.00% criterion, meaning that this
solution is the optimal and no better solution exists [13–16]. With this computational
method, YALMIP s BBA identifies two root-nodes delivering a globally solution
[37,38,41,51,74–76,79,80]. We also solve the mixed-integer non-convex nonlinear program
to ε − globality with a spatial branch-and-bound algorithm (s-BBA) [37,38,74].
An s-BBA is performed using an LP solver named SoPlex with the IPOPT optimizer
function to find global optimality within 0.00% [74–76,79–81].
SCIP utilizes SoPlex for solving the linear relaxation sub-problem and IPOPT to solve
the nonlinear problems to determine the optimality [41,73,74]. For the particular instances,
the optimizer function detects a globally optimal solution [37,38,41,51,73–76]. Both in-
stances were solved and the optimal results are presented in Table 10.

Table 6. Algorithm BBA: Steps of optimizing YALMIP s BBA.

* Starting YALMIP global branch & bound.

* Upper solver : fmincon
* Lower solver : SCIP
* LP solver : SCIP
* -Extracting bounds from model
* -Performing root-node bound propagation
* -Calling upper solver (no solution found)
* -Branch-variables : 18
* -More root-node bound-propagation
* -Performing LP-based bound-propagation
* -And some more root-node bound-propagation
* Starting the b&b process
Node Upper Gap (%) Lower Open Time
1 : 4.00000E+00 0.00 4.00000E+00 2 5s Solution found by heuristics
2 : 4.00000E+00 0.00 4.00000E+00 0 6s Poor lower bound | Pruned stack based on new upper
bound
* Finished. Cost: 4 (lower bound: 4, relative gap 0%)
* Termination with all nodes pruned
* Timing: 14% spent in upper solver (2 problems solved)
* 11% spent in lower solver (2 problems solved)
* 8% spent in LP-based domain reduction (36 problems solved)
* 1% spent in upper heuristics (1 candidates tried)
Columns 1 through 15
0 1 0 0 0 0 0 0 0 1 0 1 1 0 0
Columns 16 through 18
Algorithms 2024, 17, 191 30 of 62

0 0 0
2 10 12 13
Linear scalar (real, 18 variables)
Current value: 4
Coefficients range: 1 to 1
Optimal PMU numbers: 4

Table 7. Channel capacity, PMU number, PMU locations, and PMU channels.

L nPMU PMU Locations PMU Channels

{2-1, 2-3, 2-4, 2-5}
{6-5, 6-11, 6-12, 6-13}
4 4 2, 6, 8, 9
{8-7}
{9-4, 9-7, 9-10, 9-14}

As we noticed, the upper bound takes the lead on finding the objective bound, with
an equal lower bound calculated to return an optimal solution within a 0.00% criterion
[75].
The lower bound is found to be equal to four with a 0.00% relative gap, as displayed
in the log output. At the final root-node, the two bounds are found to be equal. Therefore,
the BBA finishes the iterative process with a global solution point [14,37,38,42,51].

Table 8. Algorithm BBA: Steps of optimizing YALMIP s BBA.

+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
* Starting YALMIP global branch & bound.
* Upper solver : fmincon
* Lower solver : SCIP
* LP solver : SCIP
* -Extracting bounds from model
* -performing root-node bound propagation
* -Calling upper solver (no solution found)
* -Branch-variables : 18
* -More root-node bound-propagation
* -Performing LP-based bound-propagation
* -And some more root-node bound-propagation
* Starting the b&b process
Node Upper Gap (%) Lower Open Time
1: 2.64286E+00 0.00 2.64286E+00 2 11s Solution found
by heuristics
2 : 2.64286E+00 0.00 2.64286E+00 0 12s Poor lower bound |
Pruned stack based on new upper bound
* Finished. Cost: 2.6429 (lower bound: 2.6429, relative gap 0%)
* Termination with all nodes pruned
* Timing: 19% spent in upper solver (2 problems solved)
* 3% spent in lower solver (2 problems solved)
* 8% spent in LP-based domain reduction (36 problems solved)
* 1% spent in upper heuristics (1 candidates tried)
Columns 1 through 15
0 1 0 0 0 0 0 0 0 1 1 0 1 0
0
Columns 16 through 18
0 0 0
2 10 11 13
Algorithms 2024, 17, 191 31 of 62

Linear scalar (real, 18 variables)

Current value: 2.6429
Coefficients range: 0.64286 to 0.85714
Optimal PMU numbers: 4
best function value: 2.642857e+00

The YALMIP global branch and bound detects an upper bound with the help of the
local optimizer function, while the lower bound is calculated by solving linear program-
ming relaxations and appropriate branch strategies utilizing an LP solver
[13,14,37,38,75,77].
YALMIP s BBA ends the iterative process when the absolute gap is less than the de-
fault value. As observed, the solution is the upper bound while the lower bound closes
the gap (lower bound: 2.6429, relative gap 0%) [13,14,37,38,75,77]. We have noticed that
the BBA tree ends with two root-nodes; therefore, the solution process is fast. The compu-
tations were performed without computational complexity, as the log files show [17].

Table 9. Channel capacity, PMU number, PMU locations, and PMU channels.

L nPMU PMU Locations PMU Channels

{2-1, 2-3, 2-4, 2-5}
{6-5, 6-11, 6-12, 6-13}
4 4 2, 6, 7, 9
{7-4, 7-8, 7-9}
{9-4, 9-7, 9-10, 9-14}

Table 10. Optimizing mixed-integer nonlinear program with multiple lines for IEEE-14
bus.

SCIP Optimizes the Single Objective Function SCIP Optimizes the b-Objective Function
L NPMU
PMU Locations PMU Channels PMU Locations PMU Channels
{2-1} {2-1}
{4-3} {4-3}
{6-5} {5-6}
1 7 2, 4, 6, 8, 9, 10, 13 {8-7} 2, 4, 5, 6, 8, 9, 10 {6-12}
{9-14} {8-7}
{10-11} {9-14}
{13-12} {10-11}
{2-1, 2-3} {2-3, 2-4}
{4-2, 4-5} {5-1, 5-4}
2 5 2, 4, 7, 11, 13 {7-8, 7-9} 2, 5, 7, 11, 13 {7-8, 7-9}
{11-6, 11-10} {11-6, 11-10}
{13-12, 13-14} {13-12, 13-14}
{2-1, 2-3, 2-5} {2-1, 2-3, 2-5}
{6-11, 6-12, 6-13} {6-11, 6-12, 6-13}
3 4 2, 6, 8, 9 2, 6, 7, 9
{8-7} {7-4, 7-8, 7-9}
{9-4, 9-10, 9-14} {9-7, 9-10, 9-14}
{2-1, 2-3, 2-4, 2-5} {2-1, 2-3, 2-4, 2-5}
{7-4, 7-8, 7-9} {6-5, 6-11, 6-12, 6-13}
4 4 2, 7, 10, 13 2, 6, 7, 9
{10-9, 10-11} {7-4, 7-8, 7-9}
{13-6, 13-12, 13-14} {9-4, 9-7, 9-10, 9-14}
{2-1, 2-3, 2-4, 2-5} {2-1, 2-3, 2-4, 2-5}
{6-5, 6-11, 6-12, 6-13} {6-5, 6-11, 6-12, 6-13}
5 4 2, 6, 8, 9 2, 6, 7, 9
{8-7} {7-4, 7-8, 7-9}
{9-4, 9-7, 9-10, 9-14} {9-4, 9-7, 9-10, 9-14}
Algorithms 2024, 17, 191 32 of 62

Then, we optimize the 0 − 1 nonlinear integer programming in a continuous do-

main where a quadratic objective function is minimized under a set of polynomial con-
straints and a decision variable bound [19–24,41]. The nonlinear programming is declared
in the feasible region Ω as follows [19–21,41,72,73]:
𝑚𝑖𝑛𝐽(𝑦) = ∑ 𝑤 × 𝑦 , 𝑤 = 1, 𝑤 = [𝑤 , 𝑤 … . , 𝑤 ] (42)

𝑠. 𝑡. 𝑓(𝑦) = 0ϵ𝛺 (43)

where 𝛺 ≔ {𝑦𝜖 ∈ [0,1]: 𝑓 (𝑦) = 0, 𝑖 = 1, … . , 𝑛}, n ≤ m [19–21].
The optimization problem is non-convex, for which GPSA, SQP, and IPMs exist for
its solution [19–24,41,70–72]. The GPSA optimizes the nonlinear programming model con-
sidering channel capacity equal to 4, 𝐿 = 4 [64]. GPSA is performed through pat-
ternsearch included in the global optimization library of MATLAB [22,41,70–72].
Patternsearch uses, as a poll method, the gpspositivebasisnp1 strategy, and as the
search method, the GPSPositiveBasis2N methodology [22,41,70–72]. The iterative scheme
ends based on the calculation of the mesh size [22,41,70,71,79].
The optimization ends due to the fact that the mesh size is less than the pre-given
mesh size tolerance [21,22,41,70–72]. Therefore, the local solution satisfies the pre-given
tolerance criteria in order to avoid any constraint violation [41,70–72]. Thus, a feasibility
and local optimality is derived [21]. The log files produced by patternsearch are shown in
Table 11 [22,41,70–72].

Table 11. Algorithm GPSA: Steps of optimizing the GPSA.

Max
Iter f-Count f(x) Constraint MeshSize Method
0 1 11 1 0.25
1 184 5 0 0.009772 Update multipliers
2 473 4.71875 0.09375 0.001 Increase penalty
3 1197 3.99805 0.001953 9.333e-07 Update multipliers
4 2094 3.99934 0.0003319 8.71e-10 Update multipliers
5 3378 4 4.545e-07 8.128e-13 Update multipliers
6 4887 4 0 7.586e-16 Update multipliers

The optimal result is {2, 10, 12, 13}. Each combination is related to the 𝑖 − 𝑡ℎ power
grid node. From Equation (33), we derive the subset combinations
{2,1, 3, 4, 5}, {6, 5, 11, 12, 13}, { 8, 7}, {9, 4, 7, 10, 14}. For a specific number of channel capac-
ity, we derive a set solution including four PMUs optimally placed around the grid [56].
We present a set solution {2, 6, 8, 9} with the branch assignments of the incident lines
to each PMU installed at the related bus.
2 − 1, 2 − 3, 2 − 4, 2 − 5
6 − 5, 6 − 11, 6 − 12, 6 − 13
8−7
9 − 4, 9 − 7, 9 − 10, 9 − 14
The SQP optimizes the nonlinear programming model considering channel capacity
equal to 4, 𝐿 = 4 [64]. The log files produced by fmincon included in the Optimization
Library of MATLAB are shown in Table 12 [41,72]. SQP runs the nonlinear programming
model, spending reasonable function evaluations per iterations [19–21].
The SQP method is invoked through the Fmincon optimizer function to the nonlinear
problem solving [19–21]. A feasible and a local solution are achieved at the final iteration
without any constraint violation (feasibility = 0.000e+00 (Table 12)). The local minimizer
satisfies the constraint function, as shown in the log file produced [20,21].
Algorithms 2024, 17, 191 33 of 62

A possible infeasibility of the unity step-length may be presented close to a local so-
lution point, destroying the super-linear convergence speed [82]. This phenomenon is
called the Maratos effect [82]. As we observed, the step-length 𝜆 = 1 is acceptable for the
line search conditions. Through the simulation results, we prove that 𝑥 = 𝑥 + 𝜆 ℎ
holds eventually. 𝜆 ℎ is the solution of the quadratic programming problem, whereas
the SQP estimates the solution from one trial point to the other [19,20,82]. Despite the fact
that the merit function is a non-smooth function, the step-length is unity and the Maratos
effect is avoided, thus preserving the super-linear convergence towards optimality [82].
The objective function was found to be non-decreasing in any feasible direction based
on the tuning parameters declared by the user s program. The relative first-order optimal-
ity quantity is less than the default value, the maximum constraint violation is zero, and
the local minimum is inside the optimality tolerance [19–21,42].

Table 12. Algorithm SQP: Steps of optimizing the SQP methods.

Norm of First-Order
Iter F-Count f(x) Feasibility Steplength Step Optimality
0 19 1.000000e+01 1.000e+00 2.000e+00
1 38 1.500000e+00 1.000e+00 1.000e+00 3.082e+00 1.000e+00
2 58 2.952500e+00 3.000e-01 7.000e-01 1.050e+00 1.021e+00
3 77 4.250000e+00 0.000e+00 1.000e+00 4.500e-01 1.066e+00
4 96 4.080921e+00 0.000e+00 1.000e+00 2.160e-01 9.631e-01
5 115 4.000027e+00 0.000e+00 1.000e+00 2.845e-01 5.909e-01
6 134 4.000001e+00 0.000e+00 1.000e+00 5.261e-03 2.367e-03
7 153 4.000000e+00 0.000e+00 1.000e+00 1.112e-03 1.108e-03
8 172 4.000000e+00 0.000e+00 1.000e+00 1.691e-05 9.899e-06
9 191 4.000000e+00 0.000e+00 1.000e+00 1.024e-33 0.000e+0

The optimization finishes the iterations due to the fact that the objective function was
found to be non-reducing in any feasible direction [72]. The optimality tolerance and the
constraint violations are found within the pre-given tolerances, showing that optimality
is achieved [72]. Therefore, the solution is feasible and a local minimum at the same time
[21].
The optimal result is {2, 12, 14, 17}. Each combination is related to the 𝑖 − 𝑡ℎ
power node. From Equation (33), the subset combinations are derived as follows:
{2,1, 3, 4, 5}, { 8, 7}, {10, 9, 11}, {13, 6, 12, 14}. This combination of subsets gives the set solu-
tion {2, 8, 10,13} with the following branch assignments:
2 − 1, 2 − 3, 2 − 4, 2−5
8−7
10 − 9, 10 − 11
13 − 6, 13 − 12 13 − 14
IPOPT is a primal-dual interior-point method based on a line search filter strategy to
discover a local solution of the nonlinear program without considering the decision bound
[20,73]. The interior-point methods optimize the nonlinear programming model consider-
ing channel capacity equal to 4, 𝐿 = 4. The log file is displayed in Table 13 [19–21,73]. The
IPOPT: Interior Point NL Solver ends the iterations after spending 21 iterations in a robust
MATLAB code [73]. A restoration phase of the filter process is implemented followed by
some heuristic computations to enhance the entire optimization [20].
The IPOPT optimizer utilizes a feasibility restoration phase in calculating the opti-
mality. The target is to compute a new iteration while the infeasibility is decreased at every
computation step [42,73]. Therefore, a local solution is reached with the calculation of the
infeasibility during the iterative process, with the aim to decrease it. The solution is
Algorithms 2024, 17, 191 34 of 62

returned by the IPOPT solver with a dual infeasibility quantity without giving any con-
straint violation [42,73]. Therefore, a feasible and local solution is delivered [19,20].

Table 13. Algorithm IPOPT: Steps of optimizing the interior-point methods.

Iter Objective inf_pr inf_du lg(mu) ||d|| lg(rg) alpha_du alpha_pr ls

0 9.0000000e+00 0.00e+00 2.00e+00 0.0 0.00e+00 - 0.00e+00 0.00e+00 0
1 2.0000000e+00 1.00e+00 1.00e+00 -11.0 2.00e+00 - 1.00e+00 5.00e-01f 2
2 2.9090909e+00 2.84e-01 5.08e-01 -11.0 5.45e-01 - 1.00e+00 1.00e+00h 1
3 4.1838231e+00 7.36e-02 6.46e-01 -11.0 3.05e-01 - 1.00e+00 1.00e+00h 1
4 5.0510211e+00 1.75e-02 7.35e-01 -11.0 1.85e-01 - 1.00e+00 1.00e+00h 1
5 5.5122440e+00 3.35e-03 1.20e+00 -11.0 1.08e-01 - 1.00e+00 1.00e+00h 1
6 5.5457963e+00 2.60e-03 2.09e+00 -11.0 5.85e-02 - 1.00e+00 1.00e+00h 1
7 5.2306455e+00 1.61e-03 1.51e+00 -11.0 6.58e-02 - 1.00e+00 1.00e+00f 1
8 4.3870988e+00 8.62e-03 1.08e+00 -11.0 3.93e-01 - 1.00e+00 1.00e+00f 1
9 3.9647785e+00 1.07e-02 6.53e-01 -11.0 3.64e-01 - 1.00e+00 1.00e+00f 1
10 4.0022318e+00 2.36e-04 5.75e-02 -11.0 2.86e-02 - 1.00e+00 1.00e+00h 1
11 4.0013014e+00 8.39e-07 5.23e-02 -11.0 2.46e-03 - 1.00e+00 1.00e+00h 1
12 4.0417359e+00 2.28e-07 4.36e+00 -11.0 1.19e-01 - 1.00e+00 1.00e+00h 1
13 4.0012661e+00 3.64e-08 2.84e-01 -11.0 1.16e-01 - 1.00e+00 1.00e+00f 1
14 4.0012370e+00 4.73e-12 9.64e-02 -11.0 1.64e-03 - 1.00e+00 1.00e+00h 1
15 4.0010381e+00 1.49e-12 1.12e+00 -11.0 5.34e-03 - 1.00e+00 1.00e+00f 1
16 4.0005805e+00 1.50e-11 1.21e+00 -11.0 1.25e-02 - 1.00e+00 1.00e+00f 1
17 4.0002220e+00 1.19e-11 7.46e-01 -11.0 1.44e-02 - 1.00e+00 1.00e+00f 1
18 4.0000000e+00 4.46e-12 3.65e+00 -11.0 1.17e-02 - 1.00e+00 1.00e+00f 1
19 4.0000000e+00 2.22e-16 6.79e-04 -11.0 3.78e-06 - 1.00e+00 1.00e+00H 1
20 4.0000000e+00 3.00e-15 4.60e-09 -11.0 6.91e-11 - 1.00e+00 1.00e+00h 1
21 4.0000000e+00 0.00e+00 7.88e-11 -11.0 4.64e-13 - 1.00e+00 1.00e+00h 1

The optimal result is { 2, 11, 14, 17}. From Eq. (33), we derive the following sub-
set combinations: {2,1, 3, 4, 5}, {7,4, 8, 9}, {10, 9, 11}, {13, 6, 12, 14}. This combination gives
the following set solution: {2, 7, 10,13}. The branch assignments are as follows:
2 − 1, 2 − 3, 2 − 4, 2−5
7 − 4, 7 − 8, 7 − 9
10 − 9, 10 − 11
13 − 6, 13 − 12, 13 − 14
GPSA, SQP, and IPMs end the iterations without calculating infeasibility at the final
iteration [42]. The local solution is located and the placement vector is the proper output
of the minimization of the objective function under the related constraints [21].
Heuristic algorithms are also utilized to handle and solve the binary-integer program
[18,25,26,39,44–51]. It is valuable to mention that GAs and BPSO are utilized for PMU al-
location problem solving since these algorithms can produce accurate solutions within a
reasonable time frame [5]. BPSO and GAs optimize the BILP model giving the best PMU
number with appropriate branch assignments to keep the entire system observability
[25,26]. Considering channel capacity 𝐿 = 4, the following results are produced. BPSO
derives the combination { 2, 10, 11, 13} [78,83]. From Equation (33), we derive the
subset combinations {2,1, 3, 4, 5}, {6, 5, 11, 12, 13}, {7,4, 8, 9}, {9, 4, 7, 10, 14}. This combina-
tion gives the following set solution: {2, 6, 7,9}. The branch assignments are:
2 − 1, 2 − 3, 2 − 4, 2 − 5
6 − 5, 6 − 11, 6 − 12, 6 − 13
7 − 4, 7 − 8, 7−9
9 − 4, 9 − 7, 9 − 10, 9 − 14
Algorithms 2024, 17, 191 35 of 62

GA derives the combination: { 2, 10, 12, 13}. From Eq. (33), we derive the subset
combinations [60] {2,1, 3, 4, 5}, {6, 5, 11, 12, 13}, {8,7}, {9, 4, 7, 10, 14} . This combination
gives the following set solution: {2, 6, 8,9}. The branch assignments are:
2 − 1, 2 − 3, 2 − 4, 2−5
6 − 5, 6 − 11, 6 − 12, 6 − 13
8−7
9 − 4, 9 − 7, 9 − 10, 9 − 14

Table 14. Channel capacity, PMU number, PMU locations, and PMU channels.

Obtained results using generalized pattern search algorithm

L nPMU PMU locations PMU channels
{2-1, 2-3, 2-4, 2-5}
{6-5, 6-11, 6-12, 6-13}
4 4 2, 6, 8, 9
{8-7}
{9-4, 9-7, 9-10, 9-14}
Obtained results by using Sequential Quadratic Programming
{2-1, 2-3, 2-4, 2-5
7-4, 7-8, 7-9
4 4 2, 7, 10, 13
10-9, 10-11
13-6, 13-12, 13-14}
Obtained results by using Interior-Point Methods
{2-1, 2-3, 2-4, 2-5}
{8-7}
4 4 2, 8, 10, 13
{10-9, 10-11}
{13-6, 13-12, 13-14}
Obtained results by using Genetic Algorithms
{2-1, 2-3, 2-4, 2-5
6-5, 6-11, 6-12, 6-13
4 4 2, 6, 8, 9
8-7
9-4, 9-7, 9-10, 9-14}
Obtained results by using Binary Swarm Optimization
{2-1, 2-3, 2-4, 2-5
7-4, 7-8, 7-9
4 4 2, 7, 11, 13
11-6, 11-10
13-6, 13-12, 13-14}

The major concern of this research study is to examine the channel capacity s effect
on finding a set of PMUs with the selected number of channels [56–58]. As a result, our
aim is to identify the algorithmic scheme for designing wide-area observability and to
determine the PMU placement set for WAMS application in real time [1–8].
Considering a pre-given number of available channels, we derive not only the PMU
locations but also the branches being measured (Tables 4, 5, 10, 14). Thereby, a number of
measurements are obtained based on PMUs placed at buses with the adjacent current
available but strongly depending on the channel limit [56–58,60,62–64].
Intlinprog optimizes the 0 − 1 integer programming model in two stages [72,79,80].
Initially, we optimize the model (Equations (16)–(18)) and, then, we produce results by
optimizing the mathematical model (Equations (35)–(37)) [37,38,41,51,72,74–76,79,80].
The optimal results are attained inside an absolute as well as relative gap equal to
zero. The solutions are globally found for the IEEE-14-bus system [78,83–85].
Some plots are derived illustrating the best objective function value, the PMU subsets
that are interpreted by the PMU locations, and the covered branches incident to each PMU
Algorithms 2024, 17, 191 36 of 62

placement entry. We show that a solution with the best redundancy is achieved for a fixed
channel capacity 𝐿 = 4 superior to those solutions derived in [64–67].
The plot diagrams displayed in Figures 4 and 5 show the results produced by opti-
mizing the mathematical model (Equations (16)–(18)). The subsets are {2, 12, 14, and 17}
and the PMU placement is {2, 8, 10, and 13}. The branch assignments are {2-1, 2-3, 2-4, and
2-5, 8-7, 10-9, 10-11, 13-6, 13-12, 13-14}. Furthermore, a plot diagram shows the results
produced by optimizing the mathematical model (Equations (35)–(37)) [74–76,79–81,85].
The subsets are {2, 10, 11, and 13} and the PMU placement is {2, 6, 7, and 9} with best
redundancy.
The branch assignments are the following: {2-1, 2-3, 2-4, 2-5, 6-5, 6-11, 6-12, 6-13, 7-4,
7-8, 7-9, 9-4, 9-7, 9-10, 9-14} [60]. That optimal result has the best system redundancy index,
as declared in [33,37,38,41–45]. We succeed in finding the best trade-off relationship be-
tween PMU locations and observed channels for a fixed channel restriction.

Figure 4. Plots derived by the intlinprog optimizer in the MATLAB optimization library.

Figure 5. Plots derived by the intlinprog optimizer in the MATLAB optimization library.
Algorithms 2024, 17, 191 37 of 62

7. Results
Many research articles assume that each PMU device is supplied with unrestricted
channel measurements to observe the phasor currents running all incident connectivity
lines to the bus equipped with a synchronized monitoring device [37,38,41,51].
In the majority of real PMU problem applications, due to economic restrictions,
PMUs are selected with a restricted channel capacity to be placed at substations within
the power grid [56–67]. As shown in the literature, if one supplementary channel is joined
to the commercial PMU, the cost increases by approximately 6.5–20% [7–11].
Our implementation takes into account the bounds on the PMU number channel ca-
pacity [56–67], but these attempts have shown PMU locations without branch assignments
and no optimality was taken into account for achieving a global solution [13,14,16–21,24].
Therefore, a channel limit capacity is taken into account for further consideration and
filling the gap about the uncertainties if the solution is a local or global solution.
We utilize mathematical and heuristic algorithms to determine all the possible com-
binations of buses and how the best number of PMUs and their locations are produced
[37,38,41,51]. The entire observability is ensured by utilizing PMUs with a minimum
measurement channel, which numbers four [56–60,64–67].
In order to be realistic with observability analysis, the OPP problem is revisited tak-
ing a restricted number of channels and solving it from the optimization point of view. It
is noticed that BBA, SQP, IPM, GPSA, GA, and BPSO are able to find the proper number
of PMUs and their related placement sites to monitor the power grid under complete and
maximum observability scenarios [37,41,51].
We utilize the BBA [37], GPSA [41], SQP [41], IPMs [41], GA [51], and BPSO [51] to
determine the proper PMU number and the related branch assignments in case of limited
channel capacity [64]. Simulation results are presented to the optimal PMU placement im-
plementation, either with one-channel or multi-channel capacity [56,61].
To illustrate results in this section, the relative PMU number and associated PMU
placement locations with branch assignments are presented. The results considering the
existence of ZIBs will not be included in this study as presented in [56–58,60,64–67]. How-
ever, this does not take away from the truth of the proposed algorithmic scheme, and the
outputs are the optimum points when ZIBs are not considered [34,37].
Aimed at the minimization problem solving considering a restricted number of meas-
urement channels, minimization modeling is utilized to produce optimal solutions in min-
imal running time.
More exactly, a binary constraint integer program and a nonlinear constraint pro-
gram are solved together towards optimality [42]. They produce suitable cost-efficient so-
lutions based on an appropriate amount of channel capacity per PMU device placed at a
node within the power grid [56–67].
The MILP solver is invoked to deliver the best upper bound to the minimization
problem solving [75,79]. The MILP solver solves the LP-relaxation problems with all inte-
ger variables to be temporarily continuous [13–15,37,38,41,51,72,74,79,80].
If the solution to the relaxed problems entails integers, the solution is the optimum.
Otherwise, the iterative scheme continues until the non-integer variables are eliminated
[13–15,72,74,79,80]. For that, a branching process is followed [79].
Once the process ends for all the decision variables, all the integer solutions are com-
pared. These solutions are achieved by different PMU channels. These solutions are con-
sidered to be the optimal point solutions.
As a result, we present the best PMU number as well as the branch assignments con-
sidering a restricted number of channels. Those solutions are achieved within zero abso-
lute and relative gaps, showing that optimality is reached [37,38,41,51].
Table 15 illustrates the best objective function value for benchmark power networks
considering a varying number of channels [56]. As we observed, the measurement redun-
dancy is maximized for a fixed channel capacity. Mathematical and heuristic algorithms
produce the number and locations of PMUs for a restricted number of channels.
Algorithms 2024, 17, 191 38 of 62

This scenario is more closely linked to the utility industry s attention for installation
at buses/substations within the power grid [56–61]. We use mathematical and heuristic
algorithms as an optimization guide to our previous efforts [37,38,41,51].
The branch-and-bound algorithm [37], nonlinear algorithms [41], derivative-free op-
timizer [41,51], and heuristic algorithm [41] find the optimal solution. PMUs are optimally
selected with fewer channels, which is more economical than utilizing PMUs with variant
channels. As a result, investment can be saved without giving up the power network ob-
servability [35,37,38,41,42,51].
Later, we transform the binary-integer program [56–58] into an equivalent nonlinear
program solved by SQP [41], IPMs [41], and a GPSA [41] to show convergence to desired
local solution points, which are characterized as global solutions.
We utilize a GA [51] and BPSO [51] to handle the 0⁄1 ILP s constraints to show con-
vergence to an optimal solution without violation of the constraint function [21]. GAs and
GPSA are considered derivative-free optimizers [22,41,70] and proper optimizer tools for
optimizing the OPP problem [41,51,72–76,79,80].
The GPSA is a direct-search algorithm for nonlinear problem-solving. The nonlinear
iterative algorithms may be trapped in a local optimum point. The GPSA has the ability
to deliver a global solution with increased accuracy [22,41,70,71].
The solution algorithm is a derivative-free optimizer, and more function estimations
are needed during the iterative scheme towards optimality [22,25,26,41,70,71].
Two steps are needed to start the iterative scheme and achieve some solution. The
poll step is a local descent strategy to compute suitable steps to give a solution that is
relatively close to the locality [19–22,41,70,71,75].
After that, a search step is utilized to search the feasible region in a global way and
thus to avoid being trapped in a local optimum [19–22,41,70,71].
This algorithm can find a global solution with a higher potential action. For that
achievement, the search step is an essential algorithmic application to terminate the itera-
tive with success [19–22,41,70], and redundancy is increased considering multi-channel
PMU capacities [56–61]. We can increase the redundancy of measurements by minimizing
the objective function under the inequality constraints while the decision variables are
binary [13–18]. The maximum measurement values are 19, 52, 72, and 164 for the 14-, 30-,
57-, and 118-bus networks, respectively [33,37,38,41–45,47,51].
The maximum observability is achieved for 𝐿 = 4, 𝐿 = 7, 𝐿 = 6, and 𝐿 = 9 for the 14-
, 30-, 57-, and 118-bus networks [83]. Table 15 illustrates the channel limit capacity, the
PMU number, and the related measurement redundancy for both mathematical models
being optimized in this study. The PMUs with more than four channels increase the cost
installation at selected nodes in the power transmission grid [56–58,60,64–67].

Table 15. Optimal PMU number with a restricted channel capacity for IEEE system.

Objective with One b-Objective

IEEE-Power Channel of
PMUs Number Product Redundancy
Systems PMUs
Redundancy Term Term
1 7 14 14
2 5 15 15
IEEE-14 bus 3 4 14 16
4 4 16 19
5 4 16 19
1 15 30 30
2 11 32 33
IEEE-30 bus 3 10 33 39
4 10 38 46
5 10 40 49
Algorithms 2024, 17, 191 39 of 62

6 10 40 51
7 10 37 52
1 29 58 58
2 19 57 57
3 17 61 62
IEEE-57 bus 4 17 64 68
5 17 69 71
6 17 64 72
7 17 64 72
1 61 122 122
2 41 123 123
3 33 121 123
4 32 134 140
IEEE-118 bus 5 32 148 153
6 32 155 159
7 32 158 162
8 32 161 163
9 32 162 164

As shown, we found the specific channels by which the redundancy is maximized.

We use the IEEE-57 bus system to produce some optimal results with channel restriction
capacity 𝐿 = 3 for a comparative study with those shown in [65,66].
Firstly, we optimize the one-product objective function in binary-integer program-
ming format, and then, the b-objective mathematical model is optimized in a computa-
tional way similar to that presented in [37,38,41,51]. In the second instance, we succeed in
finding an optimal result. The optimal result is superior to those presented in [65,66].
This optimal result is with the maximum observed branches incident to each PMU
optimally selected and well posed at strategically power transmission grid nodes to keep
the entire system observable. The solution is the best found under this contingency.
This solution is achieved when the gaps between the upper and lower bounds on the
objective function are equal so that their difference is zero. The absolute MIP optimality
gap is calculated inside a default value (1 × 10−6) declared in [72,76,79,80].
𝑀𝐼𝑃𝐺𝑎𝑝𝐴𝑏𝑠 = |𝑈 − 𝐿 | ≤ 𝐴𝑏𝑠𝑜𝑙𝑢𝑡𝑒𝐺𝑎𝑝𝑇𝑜𝑙𝑒𝑟𝑎𝑛𝑐𝑒 = 1𝑒 − 6 (44)
𝑀𝐼𝑃𝐺𝑎𝑝𝐴𝑏𝑠 is found with a zero value in conjunction with a relative gap calculated
inside a default value (1 × 10−4) declared in [72,76,79–81,84–88].
|𝑈 − 𝐿 |
𝑅𝑒𝑙𝑎𝑡𝑖𝑣𝑒 𝐺𝑎𝑝 = 100 × ≤ 1𝑒 − 4 (45)
|𝑈 | + 1
Intlinprog optimizer functions solved the 0 − 1 integer program [80]. We detect the
upper bound, which is the incumbent objective value for the minimization problem,
whereas the lower bound closes the gap to find the globally optimal solution [41,74]. In
this case, an optimality certificate is produced by the integer program solver [14].
Furthermore, 𝑀𝐼𝑃𝐺𝑎𝑝𝐴𝑏𝑠 and 𝑅𝑒𝑙𝑎𝑡𝑖𝑣𝑒 𝐺𝑎𝑝 are both found to be equal to 0 [76,80].
The plots present the best objective value within a zero-valued relative gap as well as the
possible combinations connected with the center buses, which are the main entries in the
PMU placement set. Finally, the PMU placement sites are presented in the plots [60].
Figures 6 and 7 illustrate the plot diagrams derived by the entire BBA optimization
executed by intlinprog included in the MATLAB optimization library with relative gap
equal to zero [72,79,80]. Tables 16 and 17 depict the PMU locations and the branch assign-
ments for the IEEE-57 system [83] and channel capacity 𝐿 = 3 [56,57,60,65,66].
Algorithms 2024, 17, 191 40 of 62

Figure 6. Plots derived by the intlinprog optimizer in the MATLAB optimization library. Subsets:
{3, 12, 19, 80, 87, 90, 93, 95, 100, 104, 118, 124, 127, 133, 135, 138 143} [60].

Table 16. Optimal PMU number with channel capacity 𝐿 = 3 with redundancy term equal to 59.

L nPMU PMU Locations PMU Channels

{1-2, 1-16, 1-17
6-4, 6-5, 6-7
9-8, 9-10, 9-12
15-3, 15-13, 15-14
19-18, 19-20
22-21, 22-23, 22-38
25-24, 25-30
27-26, 27-28
1, 6, 9, 15, 19, 22, 25, 27, 32, 36, 41,
3 10 32-31, 32-33, 32-34
44, 47, 50, 52, 54, 55
36-35, 36-37, 36-40
41-11, 41-42, 41-43
44-39, 44-45
47-46, 47-48
50-49, 50-51
52-29, 52-53
54-9, 54-55
57-39, 57-56}
Algorithms 2024, 17, 191 41 of 62

Figure 7. Plots derived by the intlinprog optimizer in the MATLAB optimization library. Subsets:
{3, 13, 38, 81, 87, 90, 94, 97, 98, 100, 104, 112, 118, 126, 130, 137, 143} [60].

The number of the subsets for the IEEE-57 bus is equal to 143 [60]. As we observed,
intlinprog optimizes the b-objective optimization model. Optimal results are derived
where maximum branches are covered and they are adjacent to the center buses [61].
Thereby, we achieved a PMU optimal set solution with an increased measurement
redundancy, which is superior to those found in [65,66], and optimized by a single func-
tion considering all the decision variable weights to be unity [56–67].
With the presentation of the results, it has been shown that the dual bound is as-
sumed to be the optimal objective function bound, whereas the primal bound minimizes
the optimality gap giving a solution within the 0.00% optimality criterion [80].

Table 17. Optimal PMU number with channel capacity 𝐿 = 3 with redundancy term equal to 62.

L nPMU PMU Locations PMU Channels

{1-2, 1-16, 1-17
6-4, 6-5, 6-8
10-9, 10-12, 10-51
15-3, 15-13, 15-45
19-18, 19-20
22-21, 22-23, 22-38
26-24, 26-27
29-7, 29-28, 29-52
1, 6, 10, 15, 19, 22, 26, 29, 30,
3 17 30-25, 30-31
32, 36 38, 41, 46, 49, 54, 57
32-31, 32-33, 32-34
36-35, 36-37, 36-40
38-37, 38-44, 38-45
41-11, 41-42, 41-43
46-14, 46-47
49-13, 49-38, 49-50
54-53, 54-55
57-39, 57-56}

We also use IPOPT, the interior-point optimizer function, which is written in C++ and
released as open-source code under the Eclipse Public License (EPL) [20,42,73].
Algorithms 2024, 17, 191 42 of 62

IPOPT converges rapidly with a super-linear speed. Barrier methods are utilized
with exact Hessian and sparse Newton methodologies that are significantly in problem-
solving [19–21]. Therefore, the second-order optimality conditions are satisfied [19–25].
We test IPOPT on the IEEE-57 bus system with channel capacity equal to four. The
number of subsets for the IEEE-57 bus is equal to 97 [60,65].
We achieve a PMU set solution with the maximum observed branches at once (Figure
8). We present the plot diagram as well as the tabulated optimal results, as follows [73]
(Table 18).

Figure 8. Plots derived by the IPOPT optimizer in the MATLAB optimization library. PMU sub-
sets: {1, 4, 6, 18, 49, 56, 60, 64, 67, 68, 72, 77, 81, 87, 91, 93, 97}.

Table 18. Optimal PMU number with channel capacity 𝐿 = 4 with redundancy term equal to 68.

Interior-Point Methods
L nPMU PMU Locations PMU Channels
{1-2, 1-15, 1-16, 1-17
4-3, 4-5, 4-6, 4-18
6-4, 6-5, 6-7, 6-8
9-8, 9-12, 9-13, 9-55
15-1, 15-3, 15-14, 15-45
20-19, 20-21
24-23, 24-25, 24-26
28-27, 28-29
1, 4, 6, 9, 15, 20, 24, 28, 31, 32, 36,
4 17 31-30, 31-32
38, 41, 47, 51, 53, 57
32-31, 32-33, 32-34
36-35, 36-37, 36-40
38-22, 38-44, 38-48, 38-49
41-11, 41-42, 41-43, 41-56
47-46, 47-48
51-10, 51-50
53-52, 53-57
57-39, 57-56}
Algorithms 2024, 17, 191 43 of 62

We also give the PMU combinations, PMU set solutions, and branch assignments
produced by calling the GA solver included in the global optimization library [72]. We
run the binary-integer program on the IEEE-57 bus system in order to give a derivative-
free approach in optimizing the OPP model based on channel capacity restriction [22]
(Figure 9).
Considering the power network size, we select the population initialized as bit strings
in a double vector matrix. Thus, we avoid premature convergence to a local solution point
and a global one is reached within specific tolerance criteria [51,72] (Table 19).
Current best individual Penalty value

Figure 9. Plots derived by the GA optimizer in the MATLAB optimization library. Subsets: {1, 5,
12, 48, 55, 58, 62, 65, 66, 68, 72, 77, 79, 81, 86, 91, 94}.

Table 19. Optimal PMU number with channel capacity 𝐿 = 4 with redundancy term equal to 65.

Genetic Algoritjms
L nPMU PMU Locations PMU Channels
{1-2, 1-15, 1-17
5-4, 5-6
9-8, 9-10, 9-12, 9-13
15-1, 15-3, 15-13, 15-45
19-18, 19-20
22-21, 22-23, 22-38
26-24, 26-27
29-27, 29-28, 29-52
1, 5, 9, 15, 19, 22, 26, 29, 30, 32, 36, 38,
4 17 30-25, 30-31
39, 41, 46, 51, 54
32-31, 32-33, 32-34
36-35, 36-37, 36-40
38-22, 38-44, 38-48. 38-49
39-37, 39-57
41-11, 41-42, 41-43, 41-56
46-14, 46-47
51-10, 51-50
54-53, 54-55}
Algorithms 2024, 17, 191 44 of 62

We also give the PMU combinations, PMU set solutions, and branch assignments
produced by calling BPSO [51]. We run the binary-integer program on the IEEE-57 bus
system in order to give an optimal solution with a best redundancy index equal to 68
considering the channel restriction 𝐿 = 4 [22]. The optimal result is superior to those pro-
duced by 0/1 SDP in [64,65]. The plot diagram is depicted in Figure 10.

Figure 10. Plots derived by BPSO. PMU subsets: {1, 4, 6, 23, 49, 56, 60, 64, 67, 68, 72, 77, 81, 87, 91, 93,
97}.

The utilization of BPSO results in PMU locations and branch assignments is superior
to those presented in [47]. The PMU locations and the related branch assignments are dis-
played in Table 20. We achieve a set solution with the best redundancy. For 𝐿 = 4, we
increased the number of branches observed by the PMU numbers.

Table 20. Optimal PMU number with channel capacity 𝐿 = 4 with redundancy term equal to 68.

Binary Swarm Optimization

L nPMU PMU Locations PMU Channels
{1-2, 1-15, 1-16, 1-17
4-3, 4-5, 4-6, 4-18
6-4, 6-5, 6-7, 6-8
9-11, 9-12, 9-13, 9-55
15-1, 15-3, 15-14, 15-45
20-19, 20-21
24-23, 24-25, 24-26
28-27, 28-29
1, 4, 6, 9, 15, 20, 24, 28, 31, 32, 36,
4 17 31-30, 31-32
38, 41, 47, 51, 53, 57
32-31, 32-33, 32-34
36-35, 36-37, 36-40
38-22, 38-44, 38-48, 38-49
41-11, 41-42, 41-43, 41-56
47-46, 47-48
51-10, 51-50
53-52, 53-57
57-39, 57-56}
Algorithms 2024, 17, 191 45 of 62

8. Discussion
Keeping in mind the latest advancements in mixed-integer programming, mixed-in-
teger nonlinear programming, and nonlinear programming [13–16,72–76,79,80,84,85], the
targets of this study are (I) to declare the OPP problem model under the consideration of
channel capacity, and (II) to compare the computational and simulation results derived
by different solutions in order to reach a global solution [19–23,26,72–76,79,80,84,85].
The key factor of this study is to identify the channel capacity s effect of a specified
PMU type on the OPP based-model solution for wide-area system observability.
Our goal is to achieve global solutions using BBA, SQP, IPMs, GAs, BPSO, and GPSA.
Initially, a feasible and optimal solution is derived by utilizing the BBA [74–76,79,80].
Then, the binary-integer program is transformed into a heuristically solution through the
utilization of GAs and BPSO [25,26,39,47,48,50,51].
GA utilizes a parameter penalty for approaching and handling the integer program s
constraints [25,26,50,51,72]. Fewer restrictions are needed in the implementation of the
BPSO to find optimality [39,47,48,50,51].
The binary-integer program is transformed into a mixed-integer nonlinear program
towards optimality [51]. Thereby, optimum points are derived with sufficient accuracy;
they are characterized as global points [16,37,38,41,51,72,74–76,79,80].
Finally, some nonlinear algorithms are used to minimize the pure nonlinear pro-
gramming model. The principal core of the proposed modeling formulation is the way it
is defined for the obtainable PMU number channel capacities [56–67].
This implementation does not consider the impact of ZI buses to further decrease the
PMU number, following [39,56–58]. The topological analysis does not always ensure nu-
merical observability taking into account the existence of ZI buses [39,40,46].
The numerical outcome illustrates that the suggested binary-integer model, as well
as polynomial formulations, can be implemented and solve the OPP problem with accu-
racy. The optimization confirms that the OPP-based model on a restricted number of chan-
nels is minimized to render the power system completely observable.
An extensive algorithmic approach was studied and binary-integer programming in
heuristically and nonlinear programming models is presented. Thus, the mathematical or
heuristic models deliver not only the PMU locations but also the branch assignments. We
simulate those programming models with a branch-and-bound algorithm, nonlinear al-
gorithms, derivative-free optimizer, and evolutionary algorithm to find alternative PMU
set solutions with branch assignments [13–17,19–21,23,70,71].
All algorithms are convergent to a desired outcome with a varying number of chan-
nels, including PMU locations as well as current line assignments [57,58,60,64].
The accessibility of the monitoring channels is taken into consideration in the math-
ematical and the heuristic models in this study. Hence, the target is to produce optimal
results that are as realistic as possible by the proposed algorithmic models.
The simulation illustrates that the network observability is achieved with accurate
reliability. Since the evolutionary algorithms do not supply a sufficient optimality gap
between their estimated solution and the true optimum point, the results produced by GA
and BPSO are compared to those found by the 0/1 ILP model [34,47,51]. As the simula-
tion results show, heuristic algorithms guarantee global optimality [25,26,51].
The PMU model based on a restricted number of channels presented in this paper is
a complete programming package; it can produce ILP solutions for the classical IEEE
power systems within a zero optimality gap [13–17].
Based on this feedback, we simulate the robustness of the evolutionary algorithms to
handle the ILP constraints, and thereby produce solutions that are as equally important
and have the same quality as those derived by the 0/1 ILP modeling approach
[25,26,37,38,41,44–47,51].
SQP and IPMs perform the optimization, delivering a local solution point instead of
reaching a global one [19–21,42]. Generalized pattern search and genetic algorithms work
together to deliver optimal solutions for the OPP instances [41,70,71,75].
Algorithms 2024, 17, 191 46 of 62

We extend this BILP model, giving not only the PMU possible combinations but also
the number of PMUs and their locations for a fixed number of channels [64]. Our realistic
algorithmic approach outperforms the SDP approach in producing more than one solu-
tion for a varying number of measurement channels [64–67].
Finally, the proposed nonlinear programming (NLP) model is optimized under a lin-
ear or quadratic objective function subject to a polynomial constraint to find appropriate
local solution points. The execution of the MATLAB programming code is illustrated with
a detailed optimization study based on the specific instances.
An evaluation of the achieved simulation results is presented in the Appendixes A
and B. Finally, all the algorithms result in a global solution in a minimal time.
The objective function is no longer improved for limit channels bigger than four chan-
nels per PMU device [47,56,60,64]. Hence, we reduce the cost installation.
Thereby, we succeed in determining the optimal number of PMUs with the appro-
priate number of channels without sacrificing the system observability.

9. Final Remarks and Future Study

In this study, the OPP problem is reconsidered with the target of considering that the
node voltage phasor and all current phasors running the connectivity lines incident to that
node can be obtainable.
To be accurate in the declaration of this study, the minimization model is rewritten
where a channel limit capacity is considered. Each PMU may vary in terms of the number
of channels to observe the voltage phasor at the node in which it is installed, as well as a
number of branches observed through the channel capacity restriction.
The innovation of this paper lies in using mathematical and heuristic algorithms that
optimize the OPP-based model on a restricted number of channels. Utilizing these algo-
rithms, a global solution is reached by mathematical and heuristic algorithms.
An algorithmic scheme is adopted utilizing the ideas of the mathematical and heu-
ristic algorithms, building up a procedure and a global problem solving. The new algo-
rithmic placement model results in theoretical and calculation improvements in develop-
ing the OPP placement based on a restricted number of channels.
Optimum points are derived, minimizing further the entire cost installation. As we
prove, the optimal number of channels, to achieve complete observability with a reduced
cost, is four. Furthermore, we present solutions with higher or highest observability, but
the quantities highly depend on the number of channels [56–67]. This study presents an
unambiguous combinatorial methodology, splitting into integer linear and nonlinear do-
mains to solve the PMU allocation model, considering the effect of the restricted channel
capacity per monitoring device. We study mathematical models for the varying-channel
OPP based-formulation to derive appropriate and accurate solutions.
We focus on a binary-integer model and an equivalent nonlinear programming
model that produces globally optimal solutions. We illustrate that these optimization
models are easy to implement and solvable by commercial or open-course solvers [72–
76,79–85].
BBA runs the binary-integer program, whereas SQP, IPMs, and GPSA optimize the
mathematical programming models to find the desired outcome in minimal runtime.
Additionally, our numerical results call into question the utilization of the heuristic
algorithms solution for the problem solving for the OPP model based on a restricted num-
ber of channels. BPSO and GA were suggested for the classical OPP problem, with an
unlimited number of channels being solved in [25,26,39,47,48,50,51,72].
GA is known to execute the BIP and generate a global solution within a reasonable
runtime [25,26,51]. Comparatively, BPSO offers simpler implementation than GA and
demonstrates comparable results in terms of both quality and quantity, as evidenced in
[51].
Algorithms 2024, 17, 191 47 of 62

Following the simulation run, PMU site positions were determined along with cor-
responding branch assignments, allowing for the derivation of the number of channels,
PMU assignments, and monitored channels within a fixed channel capacity.
This assessment incorporates considerations such as restricted PMU channel capac-
ity, voltage measurements, power flows across connectivity lines, injections at power grid
nodes, and higher voltage nodes [5]. The forthcoming PMU placement is anticipated to
furnish all requisite information for the state estimation model, ensuring reliable monitor-
ing. In such contexts, leveraging graphics processing units (GPUs) and deep neural net-
works in state estimation applications will prove indispensable, indicative of the growing
adoption of GPU architecture in smart grid applications [3,5].

Author Contributions: N.P.T.; Forming of the algorithmic idea, N.P.T.; Consideration of the Study,
N.P.T.; Mathematical Analysis, N.P.T.; Problem-Solving, N.P.T.; Organization and Integration of the
main body of the research, N.P.T.; MATLAB Implementation, N.P.T.; MATLAB Results Validation,
N.P.T.; Writing—Original Manuscript, R.B.; Software, R.B.; Interpretation of Results, R.B.; Review,
Editing and writing, C.A.T.; Investigation, C.A.T.; Data Analysis, C.A.T.; Engineering Analysis,
A.P.M.; Forming Analysis, A.P.M.; Software, A.P.M.; Engineering Analysis. All authors have read
and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Data Availability Statement: The authors confirm that the data supporting the findings of this study
are cited within the manuscript.
Conflicts of Interest: The authors declare no conflicts of interest.

Appendix A
To illustrate the results, possible combinations, PMU locations, and branch
assignments are derived by mathematical and heuristic algorithms. The optimization
results from the existence of ZIBs will not be accounted for at this time [34,37,38,41,51].
However, this does not take away from the truth of the approach presented in [56–
67] when ZIBs and conventional measurements are taken into consideration to further
reduce the PMU in numbers [40].
Although the formulations in the literature take into consideration zero injection
existing in the power grid, active injections, and flow measurements [64–67], the impacts
of such measurements are not illustrated in this study. However, this does not take away
from the truth of the presented results, and results are feasible and optimal without the
consideration of the above-mentioned measurements. We adopt combinatorial
optimization as well as continuous optimization. BBA, Gas, and BPSO handle decision
variables to take binary values to optimize the mathematical models.
Alternatively, SQP, IPMs, and GPSA handle decision variables to take real values,
rather than discrete variables such as binary. The algorithm, in either the binary or
continuous domain, produces solutions within tolerances and feasibility criteria [13–25].

Appendix B
In this study, we outlined the mathematical and heuristic algorithms used, providing
the exact solutions together with log files [72–76,79,80,83,84]. BBA is the suitable algorithm
for minimizing the 0 − 1 integer program [13–17,37,38,42,51].
Tables A1–A10 illustrate the optimization performed by mathematical and heuristic
algorithms [19–26]. Tables 20, A2, A4, A6, A7, and A8 illustrate the PMU locations with
the related branch assignments for channel capacity 𝐿 = 3, 𝐿 = 4. We optimize the one-
product objective function as well as the b-objective under a set of constraints.
We perform the integer program using the MOSEK optimizer engine (Tables A1 and
A3) [76,81]. The ILP solver optimizes the mathematical model utilizing the 30-bus system
as a benchmark system [83]. We consider as a restriction L = 4 [56–67].
Algorithms 2024, 17, 191 48 of 62

Optimizing the objective function with one product, we found that the optimal PMU
number is 10 together with the branch assignments. Then, we optimize the b-objective
function to find a solution with higher measurement redundancy [37,38].
The 0/1 integer linear problem is divided into two sub-problems, whereas the bi-
nary restriction is relaxed [74–76,79,80]. The sub-problems are repeatedly solved utilizing
the primal-simplex or dual-simplex [15–18,36,37,75,76,84,85].
Non-integer solutions are found by the linear programming relaxations that work for
upper bounds and integer solutions work for lower bound [36,37,41,51].
The optimizer function utilizes a BBA adopting a heuristic approach to determine
feasible solutions and, also, a cut generation is adopted [72]. Additionally, cut linear ine-
quality constraints are adopted so that the function adds to the model solving [79,80].
The inequality constraints restrict the feasible set of the linear programming relaxa-
tions so that their solutions are calculated as integer-valued entries [13–18].
In order to evaluate the quality of any incumbent solution related to its objective
value, one may examine the optimality gap. This optimality criterion calculates how much
the incumbent and the optimum point can deviate in the worst-case scenario [88].
Absolute and relative optimality gaps give valuable feedback to determine termina-
tion criteria for the mixed-integer optimizer functions [81]. The aim is to end the optimi-
zation procedure the moment that the quality of the incumbent solution, calculated in
terms of the absolute or relative gap, is good enough [72,74–76,79,80,84,85].
As observed, YALMIP s BBA solves the 0 − 1 integer program with the help of the
MOSEK solver and finds a solution at one root-node for the two case studies [75,76].
The SCIP optimizer function optimizes the 0 − 1 nonlinear program while the opti-
mal solution is attained inside an optimality gap [13–16,72,74–76,79–81].
The optimality gap is calculated by 100 × |primal − dual|/MIN(|dual|, |primal|)
[74]. We calculate the primal and dual bounds, and their difference leads to a zero opti-
mality gap [15,36,37]. Hence, the 0/1 ILP solver converges to an optimum point with a
zero absolute gap [80]. Tables A1, A3, A5, and A7 illustrate the performance of BBA in
optimizing the mathematical models to locate a global solution [37,38,41,51].
As we noticed, the primal bound takes the lead in the minimization process, with an
equal upper bound evaluated to attain zero-gap tolerance [75,79]. The dual bound mini-
mizes the gap so the solution is attained within the 0.00% criterion [37,38,75,79,81].
Thereby, no better solution can be further derived by the optimization process. This
solution is a global one, as the SCIP solution status declares. The relative output logs are
displayed in Tables A5 and A7, confirming that an accurate solution is achieved [74,75,81].
Algorithms 2024, 17, 191 49 of 62

Table A1. Algorithm BBA: Steps of optimizing YALMIP s BBA.

+ Processing objective function

+ Processing constraints
+ Branch and bound started
* Starting YALMIP global branch & bound.
* Upper solver : MOSEK
* Lower solver : MOSEK
* LP solver : MOSEK
* -Extracting bounds from model
* -Performing root-node bound propagation
* -Calling upper solver + Calling Mosek
(no solution found)
* -Branch-variables : 0
* -More root-node bound-propagation
* -Performing LP-based bound-propagation
* -And some more root-node bound-propagation
* Starting the b&b process
Node Upper Gap (%) Lower Open Time
+ Calling MOSEK
1 : 1.00000E+01 0.00 1.00000E+01 0 0s Solution found by heuristics
* Finished. Cost: 10 (lower bound: 10, relative gap 0%)
* Termination with all nodes pruned
* Timing: 33% spent in upper solver (2 problems solved)
* 16% spent in lower solver (1 problems solved)
* 1% spent in LP-based domain reduction (0 problems solved)
* 3% spent in upper heuristics (1 candidates tried)
ans =
1 5 42 43 58 61 70 75 77 79
Optimal PMU numbers: 10
best function value: 10

The iteration scheme ends at the root-node because the objective value is inside an
absolute gap tolerance of the optimal value; 𝑜𝑝𝑡𝑖𝑜𝑛𝑠. 𝑇𝑜𝑙𝐺𝑎𝑝𝐴𝑏𝑠 = 0 (the default value
by the MIP intlinprog solver of MATLAB) [37,38,41,51,72,74–76,79,80,84,85].
BBA delivers a global solution avoiding being trapped in a sub-optimum point.

Table A2. Channel capacity, PMU number, PMU locations, and PMU channel.

L nPMU PMU Locations PMU Channels

{1-2, 1-3}
{2-5, 2-7}
{6-8, 6-28}
{9-6, 9-10, 9-11}
{10-17, 10-20, 10-21, 10-22}
4 10 1, 2, 6, 9, 10, 12, 18, 23, 25, 27
{12-4, 12-13, 12-14, 12-16}
{18-15, 18-19}
{23-25, 23-24}
{25-24, 25-26, 25-27}
{27-25, 27-28, 27-29, 27-30}
Algorithms 2024, 17, 191 50 of 62

Table A3. Algorithm BBA: Steps of optimizing YALMIP s BBA.

+ Solver chosen : BMIBNB

+ Processing objective function
+ Processing constraints
+ Branch and bound started
* Starting YALMIP global branch & bound.
* Upper solver : MOSEK
* Lower solver : MOSEK
* LP solver : MOSEK
* -Extracting bounds from model
* -Performing root-node bound propagation
* -Calling upper solver + Calling Mosek
(no solution found)
* -Branch-variables : 0
* -More root-node bound-propagation
* -Performing LP-based bound-propagation
* -And some more root-node bound-propagation
* Starting the b&b process
Node Upper Gap (%) Lower Open Time
+ Calling MOSEK
1 : 8.46667E+00 0.00 8.46667E+00 0 0s Solution found by heuristics
* Finished. Cost: 8.4667 (lower bound: 8.4667, relative gap 0%)
* Termination with all nodes pruned
* Timing: 45% spent in upper solver (2 problems solved)
* 21% spent in lower solver (1 problems solved)
* 1% spent in LP-based domain reduction (0 problems solved)
* 2% spent in upper heuristics (1 candidates tried)
ans =
2 4 6 43 58 61 67 70 77 79
Optimal PMU numbers: 10
best function value: 8.466667e+00

YALMIP s BBA ends when the absolute gap is less than a pre-given tolerance [75,77].
Furthermore, the relative gap is given as a certificate of meeting a ε-sub optimality crite-
rion 15]. The optimization ends by giving a global solution within 0.00% optimality [13–
15,37,38,41,51]. Thus, a global solution is achieved [37,38,80].

Table A4. Channel capacity, PMU number, PMU locations, and PMU channels.

L nPMU PMU Locations PMU Channels

{2-1, 2-4, 2-5, 2-6}
{4-2, 4-3, 4-6, 4-12}
{6-2, 6-4, 6-7, 6-8}
{9-6, 9-10, 9-11}
2, 4, 6, 9, 10, 12, 15, 18, 25, {10-17, 10-20, 10-21, 10-22}
4 10
27 {12-4, 12-13, 12-14, 12-16}
{15-12, 15-14, 15-18, 15-23}
{18-15, 18-19}
{25-24, 25-26, 25-27}
{27-25, 27-28, 27-29, 27-30}
Algorithms 2024, 17, 191 51 of 62

Table A5. Solving the mixed-integer-nonlinear program using the SCIP function.
Mixed-Integer Nonlinear Program (MINLP) Optimization
min f(x)
s.t. lb <= x <= ub
cl <= c(x) <= cu
xi = Integer/Binary
Problem Properties:
# Decision Variables: 82
# Constraints: 276
# Bounds: 164
# Binary Variables: 82
# Nonlinear Equality: 30
Solver Parameters:
Solver: SCIP
Objective Gradient: @(x)mklJac(fun,x)
Constraint Jacobian: @(x)mklJac(nlcon,x)
Jacobian Structure: Supplied
time | node | left |LP iter|LP it/n| mem |mdpt |frac |vars |cons |cols |rows |cuts |confs|strbr| dual bound | primal bound | gap
T 0.1s| 1| 0| 0| - |2520k| 0 | - | 690 | 942 | 690 | 637 | 0| 0| 0| -- |6.500000e+001 | Inf
b 0.1s| 1| 0| 0| - |2457k| 0 | - | 690 | 942 | 690 | 637 | 0| 0| 0| -- |1.800000e+001 | Inf
0.1s| 1| 0 | 307 | - |2453k| 0 | 7 | 690 | 942 | 690 | 637 | 0 | 0| 0 |1.000000e+001 |1.800000e+001 | 80.00
0.1s| 1| 0 | 324 | - |2542k| 0 | 13 | 690 | 942 | 690 | 641 | 4 | 0| 0 |1.000000e+001 |1.800000e+001 | 80.00
0.1s| 1| 0 | 327 | - |2612k| 0 | 0 | 690 | 942 | 690 | 643 | 6 | 0| 0 |1.000000e+001 |1.800000e+001 | 80.00
* 0.1s| 1| 0| 327 | - |2620k| 0 | - | 690 | 942 | 690 | 643 | 6| 0| 0 |1.000000e+001 |1.000000e+001 | 0.00
SCIP Status : problem is solved [optimal solution found]
Solving Time (sec) : 0.12
Solving Nodes : 1
Primal Bound : +1.00000000000000e+001 (3 solutions)
Dual Bound : +1.00000000000000e+001
Gap : 0.00
1 2 18 43 58 64 67 71 77 82
Optimal PMU numbers: 10
best function value: 10
BBNodes: 1
BBGap: 0
Time: 0.5361
Algorithm: SCIP: Spatial Branch and Bound using IPOPT and SoPlex
Status: Globally Integer Optimal

Table A6. Channel capacity, PMU number, PMU locations, and PMU channels.

L nPMU PMU Locations PMU Channels

{1-2, 1-3}
{2-1, 2-4, 2-5, 2-6}
{6-2, 6-7, 6-8, 6-28}
{9-6, 9-10, 9-11}
1, 2, 6, 9, 10, 12, 15, 19, 25, {10-17, 10-20, 10-21, 10-22}
4 10
27 {12-13, 12-14, 12-15, 12-16}
{15-12, 15-14, 15-18, 15-23}
{19-18, 19-20}
{25-24, 25-26, 25-27}
{27-29, 27-30}
Algorithms 2024, 17, 191 52 of 62

Table A7. Solving the mixed-integer nonlinear program using the SCIP optimizer.
Mixed-Integer Nonlinear Program (MINLP) Optimization
min f(x)
s.t. lb <= x <= ub
cl <= c(x) <= cu
xi = Integer/Binary
Problem Properties:
# Decision Variables: 82
# Constraints: 276
# Bounds: 164
# Binary Variables: 82
# Nonlinear Equality: 30
Solver Parameters:
Solver: SCIP
Objective Gradient: @(x)mklJac(fun,x)
Constraint Jacobian: @(x)mklJac(nlcon,x)
Jacobian Structure: Supplied
time | node | left |LP iter|LP it/n| mem |mdpt |frac |vars |cons |cols |rows |cuts |confs|strbr| dual bound | primal bound |
gap
T 0.1s| 1| 0| 0| - |2550k| 0 | - | 690 | 942 | 690 | 637 | 0| 0| 0| -- |5.540000e+001 | Inf
b 0.2s| 1| 0| 0| - |2487k| 0 | - | 690 | 942 | 690 | 637 | 0| 0| 0| -- |1.560000e+001 | Inf
* 0.2s| 1| 0| 319 | - |2492k| 0 | - | 690 | 942 | 690 | 637 | 0| 0| 0 |8.466667e+000 |8.466667e+000 | 0.00
0.2s| 1| 0 | 319 | - |2492k| 0 | - | 690 | 942 | 690 | 637 | 0 | 0 | 0 |8.466667e+000 |8.466667e+000 | 0.00
SCIP Status : problem is solved [optimal solution found]
Solving Time (sec) : 0.17
Solving Nodes :1
Primal Bound : +8.46666666666667e+000 (3 solutions)
Dual Bound : +8.46666666666667e+000
Gap : 0.00
2 4 18 43 56 64 67 72 77 79
Optimal PMU numbers: 10
BBNodes: 1
BBGap: 0
Time: 0.3718
Algorithm: SCIP: Spatial Branch and Bound using IPOPT and SoPlex
Status: Globally Integer Optimal

Table A8. Channel capacity, PMU number, PMU locations, and PMU channels.

L nPMU PMU Locations PMU Channels

{2-1, 2-4, 2-5, 2-6}
{4-2, 4-3, 4-6, 4-12}
{6-2, 6-7, 6-8, 6-28}
{9-6, 9-10, 9-11}
{10-9, 10-17, 10-21, 10-22}
4 10 2, 4, 6, 9, 10, 12, 15, 20, 25, 27
{12-13, 12-14, 12-15, 12-16}
{15-12, 15-14, 15-18, 15-23}
{20-10, 20-19}
{25-24, 25-26, 25-27}
{27-25, 27-28, 27-29, 27-30}

For the 30-bus system, we found the set solutions for a fixed channel capacity [65].
Tables A6 and A8 illustrate the best PMU number and the branch assignments related to
each entry included in the PMU placement vector. Considering this, we optimize the two
mathematical models to achieve PMU locations and related branch assignments.
The first optimizes a one product objective, whereas the second optimizes the b-ob-
jective function. We succeed in finding PMU solutions covering branches with higher
Algorithms 2024, 17, 191 53 of 62

redundancy related to previous results presented in [64–67]. Thus, we achieve better re-
sults to those found in [57–67] for a fixed amount of channel capacity [65].

Table A9. Channel capacity, PMU number, PMU locations, and PMU channels.

Binary Swarm Optimization

L nPMU PMU Locations PMU Channels
{1-2, 1-3
{2-1, 2-4, 2-5, 2-6
{6-7, 6-8, 6-9, 6-28
{10-6, 10-17, 10-21, 10-22
{11-9
4 10 1, 2, 6, 10, 11, 12, 19, 24, 26, 29
{12-13, 12-14, 12-15, 12-16
{19-18, 19-20
{24-22, 24-23, 24-25
{26-25
{29-27, 29-30}
{2-1, 2-4, 2-5, 2-6
{4-2, 4-3, 4-6, 4-12
{6-2, 6-7, 6-8, 6-28
{9-6, 9-10, 9-11
{10-17, 10-20, 10-21, 10-22
4 10 2, 4, 6, 9, 10, 12, 15, 18, 25, 27
{12-4, 12-13, 12-15, 12-16
{15-12, 15-14, 15-18, 15-23
{18-15, 18-19
{25-24, 25-26, 25-27
{27-29, 27-30}
Genetic Algorithms
{3-1, 3-4
{6-2, 6-8, 6-9, 6-28
{7-5, 7-6
{10-17, 10-20, 10-21, 10-22
{11-9
4 10 3, 6, 7, 10, 11, 12, 15, 19, 25, 29
{12-4, 12-13, 12-15, 12-16
{15-12, 15-14, 15-18, 15-23
{19-20, 19-18
{25-24, 25-26, 25-27
{29-27, 29-30}
{2-1, 2-4, 2-5, 2-6
{4-2, 4-3, 4-6, 4-12
{6-4, 6-7, 6-8, 6-28
{9-6, 9-10, 9-11
{10-17, 10-20, 10-21, 10-22
4 10 2, 4, 6, 9, 10, 12, 15, 19, 25, 27
{12-13, 12-14, 12-15, 12-16
{15-12, 15-14, 15-18, 15-23
{19-18, 19-20
{25-24, 25-26, 25-27
{27-25, 27-28, 27-29, 27-30}

As observed in Tables A5 and A7, the primal bound on the objective function value
is calculated for the minimization problem. SCIP computes the difference between the
primal and dual bounds on the objective function in the BBA tree implementation [74].
Algorithms 2024, 17, 191 54 of 62

𝐺𝑎𝑝 = 𝑈 – 𝐿 <= 𝐴𝑏𝑠𝑜𝑙𝑢𝑡𝑒𝐺𝑎𝑝𝑇𝑜𝑙𝑒𝑟𝑎𝑛𝑐𝑒 is found to be equal to zero [79,80]. The so-

lution status illustrates that “Globally Integer Optimal” has been detected [74,79].
Tables A9 and A10 presents the optima points derived by mathematical and heuristic
algorithms considering a fixed channel capacity 𝐿 = 3, 𝐿 = 4. Tables A7 and A8 present
the PMU numbers, the PMU locations, and the observed branches obtained by SQP, IPMs,
GPSA, GAs, BPSO, and BBA [41,51,72–76,79,80,84–88].
We derive not only the PMU locations, but also the channels being observed that lead
to the measurement of the redundancy, as illustrated in Table 15. BBA, GA, and BSO op-
timize the integer program, while SQP, IPMs, and GPSA optimize the nonlinear program
to find the best solutions under this contingency [41,51]. SQP and IPMs are local search
algorithms detecting local minimizers at the desired objective function value for the spe-
cific minimization problem [19–21,25,41].
SQP and IPMs are theoretically efficient algorithms, and they are also computational
methods that are not wasteful [19,20,41]. SQP, IPMs, and GPSA emphasize the search of a
neighborhood of a local point, and the solution is not the global optimum in case of min-
imizing non-convex nonlinear problems [19,20,22,25,41,70,71].
All algorithms deliver solutions as the iterations go to the final stage, where the tol-
erance and feasible criteria are fulfilled so as an optimal result is reached [21].
Furthermore, we derive solutions for channel capacity 𝐿 = 3, 𝐿 = 4 utilizing the
COIN-OR Branch and Cut solver (CBC) optimizer routine written in C++ [84].
CBC utilizes a Branch and Cut using CLP with a solution status declared “Integer
Optimal” [84]. Furthermore, the GLPK Simplex Optimizer utilizes a revised simplex algo-
rithm to find an INTEGER OPTIMAL SOLUTION within a 0.0% criterion [26,85]. Finally,
some MILP results are provided by using Gurobi to find global solutions [86–88].
This computational approach permits us to avoid local or sub-optimum points, and
globally optimal solutions can be identified with accuracy under specific tolerances [21].
Mathematical and heuristic algorithms return optimal solutions with the best trade-
off relationship between the optimal PMU locations and the branch assignments.

Table A10. Channel capacity, PMU number, PMU locations, and PMU channels.

Sequential Quadratic Programming

L nPMU PMU locations PMU channels
{1-2, 1-3
{7-5, 7-6
{8-6, 8-28
{10-17, 10-21, 10-22
{11-9
3 10 1, 7, 8, 10, 11, 12, 15, 20, 25, 29
{12-4, 12-13, 12-16
{15-14, 15-18, 15-23
{20-10, 20-19
{25-24, 25-26, 25-27
{29-27, 29-30}
{3-1, 3-4
{5-2, 5-7
{10-17, 10-20, 10-21, 10-22
{11-9
{12-13, 12-14, 12-15, 12-16
4 10 3, 5, 10, 11, 12, 19, 23, 26, 27, 28
{19-18, 19-20
{23-15, 23-24
{26-25
{27-25, 27-28, 27-29, 27-30
{28-6, 28-8, 28-27}
Algorithms 2024, 17, 191 55 of 62

Interior-Point Methods
{1-3, 1-4
{5-2, 5-7
{8-6, 8-28
{10-17, 10-21, 10-22
{11-9
3 10 1, 5, 8, 10, 11, 12, 19, 23, 26, 29
{12-13, 12-14, 12-16
{19-18, 19-20
{23-15, 23-24
{26-25
{29-27, 29-30}
{2-1, 2-4, 2-5, 2-6
{4-2, 4-3, 4-6, 4-12
{6-2, 6-7, 6-8, 6-28
{10-6, 10-17, 10-21, 10-22
{11-9
4 10 2, 4, 6, 10, 11, 12, 19, 23, 25, 27
{12-13, 12-14, 12-15, 12-16
{19-18, 19-20
{23-15, 23-24
{25-24, 25-26, 25-27
{27-25, 27-28, 27-29, 27-30}
Generalized Pattern Search Algorithm
{2-1, 2-4, 2-5
{4-2, 4-3, 4-6
{6-7, 6-8, 6-28
{10-17, 10-21, 10-22
{11-9
3 10 2, 4, 6, 10, 11, 12, 19, 23, 25, 27
{12-13, 12-14, 12-16
{19-18, 19-20
{23-15, 23-24
{25-24, 25-26, 25-27
{27-25, 27-29, 27-30}
{1-2, 1-3
{2-1, 2-4, 2-5, 2-6
{6-7, 6-8, 6-10, 6-28
{10-17, 10-20, 10-21, 10-22
{11-9
4 10 1, 2, 6, 10, 11, 12, 18, 24, 25, 29
{12-13, 12-14, 12-15, 12-16
{18-15, 18-19
{24-22, 24-23, 24-25
{25-24, 25-26, 25-27
{29-27, 29-30}
Genetic Algorithms
{3-1, 3-4
{5-2, 5-7
{8-6, 8-28
{10-17, 10-20, 10-21
3 10 3, 5, 8, 10, 11, 12, 18, 24, 26, 27
{11-9
{12-13, 12-14, 12-16
{18-15, 18-19
{24-22, 24-23, 24-25
Algorithms 2024, 17, 191 56 of 62

{26-25
{27-25, 27-29, 27-30}
Genetic Algorithms
{2-1, 2-4, 2-5, 2-6
{3-1, 3-4
{6-4, 6-7, 6-8, 6-9
{9-6, 9-10, 9-11
{10-9, 10-17, 10-21, 10-22
4 10 2, 3, 6, 9, 10, 12, 15, 19, 25, 27
{12-4, 12-13, 12-15, 12-16
{15-12, 15-14, 15-18, 15-23
{19-18, 19-20
{25-24, 25-26, 25-27
{27-25, 27-28, 27-29, 27-30}
Binary Particle Swarm Optimization
{2-1, 2-4, 2-5
{4-2, 4-3, 4-12
{6-7, 6-8, 6-28
{9-6, 9-10, 9-11
{10-17, 10-20, 10-21
3 10 2, 4, 6, 9, 10, 12, 18, 24, 26, 29
{12-13, 12-14, 12-16
{18-15, 18-19
{24-22, 24-23, 24-25
{26-25
{29-27, 29-30}
{3-1, 3-4
{6-2, 6-8, 6-9, 6-28
{7-5, 7-6
{9-6, 9-10, 9-11
{10-6, 10-17, 10-20, 10-21
4 10 3, 6, 7, 9, 10, 12, 18, 24, 26, 29
{12-13, 12-14, 12-15, 12-16
{18-15, 18-19
{24-22, 24-23, 24-25
{26-25
{29-27, 29-30}
CBC Branch and Cut Algorithm
{2-1, 2-4, 2-5
{4-2, 4-3, 4-6
{6-2, 6-7, 6-8
{9-6, 9-10, 9-11
{10-17, 10-20, 10-21
3 10 2, 4, 6, 9, 10, 12, 18, 24, 25, 27
{12-13, 12-14, 12-16
{18-15, 18-19
{24-22, 24-23, 24-25
{25-24, 25-26, 25-27
{27-28, 27-29, 27-30}
{2-1, 2-4, 2-5, 2-6
{4-2, 4-3, 4-6, 4-12
{6-2, 6-7, 6-8, 6-10
4 10 2, 4, 6, 9, 10, 12, 15, 19, 25, 27
{9-6, 9-10, 9-11
{10-6, 10-17, 10-21, 10-22
{12-13, 12-14, 12-15, 12-16
Algorithms 2024, 17, 191 57 of 62

{15-12, 15-14, 15-18, 15-23

{19-18, 19-20
{25-24, 25-26, 25-27
{27-25, 27-28, 27-29, 27-30}
SCIP Spatial Branch-and-Bound Algorithm
{3-1, 3-4
5-2, 5-7
6-4, 6-7, 6-8
9-6, 9-10, 9-11
10-17, 10-21, 10-22
3 10 3, 5, 6, 9, 10, 12, 15, 19, 25, 27
12-13, 12-15, 12-16
15-12, 15-14, 15-23
19-18, 19-20
25-24, 25-26, 25-57
27-28, 27-29, 27-30}
{3-1, 3-4
5-2, 5-7
6-2, 6-4, 6-7, 6-8
9-6, 9-10, 9-11
10-17, 10-20, 10-21, 10-22
4 10 3, 5, 6, 9, 10, 12, 15, 18, 25, 27
12-13, 12-14, 12-15, 12-16
15-12, 15-14, 15-18, 15-23
18-15, 18-19
25-24, 25-26, 25-27
27-25, 27-28, 27-29, 27-30}
Gurobi: Branch-and-Bound Algorithm
{2-1, 2-5, 2-6
3-1, 3-4
6-7, 6-8, 6-28
9-6, 9-10, 9-11
10-17, 10-20, 10-21
3 10 2, 3, 6, 9, 10, 12, 18, 24, 26, 27
12-13, 12-14, 12-16
18-15, 18-19
24-22, 24-23, 24-25
26-25
27-25, 27-29, 27-30}
{1-2, 1-3
2-1, 2-4, 2-5, 2-6
6-7, 6-8, 6-9, 6-10
9-6, 9-10, 9-11
10-17, 10-20, 10-21, 10-22
4 10 1, 2, 6, 9, 10, 12, 18, 24, 25, 27
12-4, 12-13, 12-14, 12-16
18-15, 18-19
24-22, 24-23, 24-25
25-24, 25-26, 25-27
27-25, 27-28, 27-29, 27-30}
MOSEK: Branch-and-Bound Algorithm
Algorithms 2024, 17, 191 58 of 62

{1-2, 1-3
2-1, 2-5, 2-6
6-4, 6-7, 6-8
9-6, 9-10, 9-11
10-17, 10-21, 10-22
3 10 1, 2, 6, 9, 10, 12, 15, 20, 25, 27
12-13, 12-14, 12-16
15-12, 15-18, 15-23
20-10, 20-19
25-24, 25-26, 25-27
27-28, 27-29, 27-30}
{1-2, 1-3
5-2, 5-7
6-8, 6-28
9-6, 9-10, 9-11
10-17, 10-20, 10-21, 10-22
4 10 1, 5, 6, 9, 10, 12, 18, 23, 25, 27
12-4, 12-13, 12-14, 12-16
18-15, 18-19
23-15, 23-24
25-24, 25-26, 25-27
27-25, 27-28, 27-29, 27-30}
GLPK: Revised Simplex to Build the BBA Tree
{2-1, 2-4, 2-5
4-2, 4-3, 4-6
6-7, 6-8, 6-28
10-17, 10-21, 10-22
11-9
3 10 2, 4, 6, 10, 11, 12, 19, 23, 25, 27
12-13, 12-14, 12-16
19-18, 19-20
23-15, 23-24
25-24, 25-26, 25-27
27-25, 27-29, 27-30}
{3-1, 3-4
5-2, 5-7
6-2, 6-7, 6-8, 6-28
9-6, 9-10, 9-11
10-9, 10-17, 10-20, 10-21
4 10 3, 5, 6, 9, 10, 12, 18, 24, 26, 29
12-4, 12-13, 12-14, 12-16
18-15, 18-19
24-22, 24-23, 24-25
26-25
29-27, 29-30}
GLPK: Revised Simplex to Build the BBA Tree
Optimizing the b-objective function leading to a set with increased redundancy
{2-1, 2-4, 2-5
4-2, 4-3, 4-6
6-7, 6-8, 6-28
9-6, 9-10, 9-11
3 10 2, 4, 6, 9, 10, 12, 15, 20, 25, 27
10-17, 10-21, 10-22
12-4, 12-13, 12-16
15-14, 15-18, 15-23
20-10, 20-19
Algorithms 2024, 17, 191 59 of 62

25-24, 25-26, 25-27

27-25, 27-29, 27-30
{2-1, 2-4, 2-5, 2-6
4-2, 4-3, 4-6, 4-12
6-2, 6-7, 6-8, 6-28
9-6, 9-10, 9-11
10-17, 10-20, 10-21, 10-22
4 10 2, 4, 6, 9, 10, 12, 15, 18, 25, 27
12-4, 12-13, 12-14, 12-16
15-12, 15-14, 15-18, 15-23
18-15, 18-19
25-24, 25-26, 25-27
27-25, 27-28, 27-29, 27-30}
GLPK: Revised Simplex to Build the BBA Tree
Optimizing the b-objective function leading to a set with increased redundancy
{2-1, 2-4, 2-5
4-2, 4-3, 4-6
6-7, 6-8, 6-28
9-6, 9-10, 9-11
10-17, 10-21, 10-22
3 10 2, 4, 6, 9, 10, 12, 15, 20, 25, 27
12-4, 12-13, 12-16
15-14, 15-18, 15-23
20-10, 20-19
25-24, 25-26, 25-27
27-25, 27-29, 27-30
{2-1, 2-4, 2-5, 2-6
4-2, 4-3, 4-6, 4-12
6-2, 6-7, 6-8, 6-28
9-6, 9-10, 9-11
10-17, 10-20, 10-21, 10-22
4 10 2, 4, 6, 9, 10, 12, 15, 18, 25, 27
12-4, 12-13, 12-14, 12-16
15-12, 15-14, 15-18, 15-23
18-15, 18-19
25-24, 25-26, 25-27
27-25, 27-28, 27-29, 27-30}

References
1. Phadke, A.G.; Thorp, J.S. Synchronized Phasor Measurements and Their Applications, 2nd ed.; Springer: New York, NY, USA, 2017.
2. Darmis, O.; Korres, G. A Survey on Hybrid SCADA/WAMS State Estimation Methodologies in Electric Power Transmission
Systems. Energies 2023, 16, 618.
3. Monti, A.; Sadu, A.; Tang, J. Chapter 8—Wide Area Measurement Systems: Applications. In Phasor Measurement Units and Wide
Area Monitoring Systems; Academic Press: Cambridge, MA, USA, 2016; pp. 177–234.
4. Cheng, G.; Lin, Y.; Abur, A.; Gómez-Expósito, A.; Wu, W. A Survey of Power System State Estimation Using Multiple Data
Sources: PMUs, SCADA, AMI, and Beyond. IEEE Trans. Smart Grid 2024, 15, 1129–1151.
5. Varghese, C.H.; Shah, B.; Azimian, B.; Pal, A.; Farantatos, E. Deep Neural Network-Based State Estimator for Transmission
System Considering Practical Implementation Challenges. J. Mod. Power Syst. Clean Energy 2024, early access.
6. Todescato, M.; Carli, R.; Schenato, L.; Barchi, G. Smart Grid State Estimation with PMUs Time Synchronization Errors. Energies
2020, 13, 5148. https://doi.org/10.3390/en13195148.
7. Johnson, T.; Moger, T. A critical review of methods for optimal placement of phasor measurement units. Int. Trans. Electr. Energy
Syst. 2020, 31, e12698.
8. Ahmed, M.M.; Amjad, M.; Qureshi, M.A.; Imran, K.; Haider, Z.M.; Khan, M.O. A Critical Review of State-of-the-Art Optimal
PMU Placement Techniques. Energies 2022, 15, 2125.
9. Biswal, C.; Sahu, B.K.; Mishra, M.; Rout, P.K. Real-Time Grid Monitoring and Protection: A Comprehensive Survey on the
Advantages of Phasor Measurement Units. Energies 2023, 16, 4054.
Algorithms 2024, 17, 191 60 of 62

10. Paramo, G.; Bretas, A.; Meyn, S. Research Trends and Applications of PMUs. Energies 2022, 15, 5329.
11. Menezes, T.S.; Barra, P.H.A.; Dizioli, F.A.S.; Lacerda, V.A.; Fernandes, R.A.S.; Coury, D.V. A Survey on the Application of Phasor
Measurement Units to the Protection of Transmission and Smart Distribution Systems. Electr. Power Compon. Syst. 2023, 1, 1–18.
12. Fotopoulou, M.; Petridis, S.; Karachalios, I.; Rakopoulos, D. A Review on Distribution System State Estimation Algorithms. Appl.
Sci. 2022, 12, 11073.
13. Wolsey, L.A. Integer Programming; John Willey and Sons: Hoboken, NJ, USA, 2020. https://doi.org/10.1002/9781119606475
14. Conforti, M.; Cornuéjols, G.; Zambelli, G. Integer Programming, Graduate Texts in Mathematics; Springer: Berlin/Heidelberg, Ger-
many, 2014; Volume 271
15. Chen, D.S.; Batson, R.G.; Dang, Y. Applied Integer Programming: Modelling and Solution; John Wiley and Sons: Hoboken, NJ, USA,
2010.
16. Sarma, E.; Xedoses, P.; Doksa, H. Springer optimization and its applications (Springer) Literature review. In Multicriteria Port-
folio Construction with Python; Springer: Berlin/Heidelberg, Germany, 2020; Volume 163.
17. Ganian, R.; Ordyniak, S. Solving Integer Linear Programs by Exploiting Variable-Constraint Interactions: A Survey. Algorithms
2019, 12, 248. https://doi.org/10.3390/a12120248.
18. Brajević, I. A Shuffle-Based Artificial Bee Colony Algorithm for Solving Integer Programming and Minimax Problems. Mathe-
matics 2021, 9, 1211. https://doi.org/10.3390/math9111211.
19. Durenberger, D.G.; Yonge, Y. Linear and Nonlinear Programming, 3rd ed.; Springer: Berlin/Heidelberg, Germany, 2008.
20. Necedah, J.; Wright, S. Numerical Optimization; Springer Series in Operations Research and Financial Engineering; Springer:
Berlin/Heidelberg, Germany, 2006.
21. Chinneck, J.W. Feasibility and Infeasibility in Optimization: Algorithms and Computational Methods; International Series in Opera-
tions Research & Management Science; Springer: Cham, Switzerland, 2008.
22. Rios, L.M.; Sahinidis, N.V. Derivative-Free Optimization: A Review of Algorithms and Comparison of Software Implementa-
tions. J. Glob. Optim. 2013, 56, 1247–1293.
23. Sahinidis, N.V. Mixed-Integer Nonlinear Programming 2018; Springer: Berlin, Germany, 2019.
24. Liberti, L.; Maculan, N. Global Optimization: From Theory to Implementation, Nonconvex Optimization and Its Applications; Springer:
Berlin/Heidelberg, Germany, 2006.
25. Yang, X.S. Engineering Optimization: An Introduction with Metaheuristic Applications; Wiley: Hoboken, NJ, USA, 2010
26. Özlü, İ.A.; Baimakhanov, O.; Saukhimov, A.; Ceylan, O. A Heuristic Methods-Based Power Distribution System Optimization
Toolbox. Algorithms 2022, 15, 14.
27. Nardin, A.; D Andreagiovanni, F. A Quantum-Inspired Ant Colony Optimization Algorithm for Parking Lot Rental to Shared
E-Scooter Services. Algorithms 2024, 17, 80. https://doi.org/10.3390/a17020080
28. Diestel, R. Graph Theory; Springer: Berlin, Germany, 2017; Volume 173, pp. 59–172.
29. Haynes, T.W.; Hedetniemi, S.M.; Hedetniemi, S.T.; Henning, M.A. Domination in Graphs Applied to Electric Power Networks.
SIAM J. Discret. Math. 2002, 15, 519–529.
30. Sun, O.; Fan, N. The probabilistic and reliable connected power dominating set problems. Optim. Lett. 2019, 13, 1189–1206.
31. Poirion, P.L.; Toubaline, S.; D Ambrosio, C.; Liberti, L. The power edge set problem. Networks 2016, 68, 104–120.
32. Xu, B.; Abur, A. Observability Analysis and Measurement Placement for Systems with PMUs. In Proceedings of the IEEE PES
Power Systems Conference and Exposition, New York, NY, USA, 10–13 October 2004; pp. 2–5.
33. Dua, D.; Dambhare, S.; Gajbhiye, R.K.; Soman, S.A. Optimal Multistage Scheduling of PMU Placement: An ILP Approach. IEEE
Trans. Power Deliv. 2008, 23, 1812–1820.
34. Pal, A.; Sánchez, A.G.A.; Thorp, J.S.; Centeno, V.A. A Community-based Partitioning Approach for phasor Measurement Unit
Placement in Large systems. Electr. Power Compon. Syst. 2016, 44, 1317–1329.
35. Pal, A.; Vullikanti, A.K.S.; Ravi, S.S. A PMU Placement Scheme Considering Realistic Costs and Modern Trends in Relaying.
IEEE Trans. Power Syst. 2017, 32, 552–561.
36. Ghamsari-Yazdel, M.; Esmaili, M.; Amjady, N. Optimal substation-based joint allocation of PMUs and measuring channels con-
sidering network expansion planning. Int. J. Electr. Power Energy Syst. 2019, 106, 274–287.
37. Theodorakatos, N.P.; Babu, R.; Moschoudis, A.P. The branch-and-bound algorithm in optimizing mathematical programming
models to achieve power grid observability. Axioms 2023, 12, 1040.
38. Theodorakatos, N.P.; Moschoudis, A.P.; Babu, R. Calculating Global Minimum Points to Binary Polynomial Optimization Prob-
lem: Optimizing the Optimal PMU Localization Problem as a Case-Study. J. Phys. Conf. Ser. 2024, 2701, 012001.
39. Pal, A.; Jones, K.D.; Mishra, C.; Centeno, V.A. Binary particle swarm optimisation-based optimal substation coverage algorithm
for phasor measurement unit installations in practical systems. IET Gener. Transm. Distrib. 2016, 10, 555–562.
40. Korres, G.N.; Manousakis, N.M.; Xygkis, T.C.; Lofberg, J. Optimal phasor measurement unit placement for numerical observa-
bility in the presence of conventional measurement using semi-definite programming. IET Gener. Transm. Distrib. 2015, 9, 2427–
2436.
41. Theodorakatos, N.P.; Lytras, M.; Babu, R. A Generalized Pattern Search Algorithm Methodology for solving an Under-Deter-
mined System of Equality Constraints to achieve Power System Observability using Synchrophasors. J. Phys. Conf. Ser. 2021,
2090, 012125.
42. Theodorakatos, N.P.; Lytras, M.D.; Moschoudis, A.P.; Kantoutsis, K.T. Implementation of optimization-based algorithms for
maximum power system observability using synchronized measurements. AIP Conf. Proc. 2023, 2872, 120006.
Algorithms 2024, 17, 191 61 of 62

43. Xia, N.; Gooi, H.B.; Chen, S.X.; Wang, M.Q. Redundancy based PMU placement in state estimation. Sustain. Energy Grids Netw.
2015, 2, 23–31.
44. Ramasamy, S.; Koodalsamy, B.; Koodalsamy, C.; Veerayan, M.B. Realistic method for placement of phasor measurement units
through optimization problem formulation with conflicting objectives. Electr. Power Compon. Syst. 2021, 49, 474–487.
45. Singh, S.P.; Singh, S.P. A Multi-objective PMU Placement Method in Power System via Binary Gravitational Search Algorithm.
Electr. Power Compon. Syst. 2017, 45, 1832–1845.
46. Koutsoukis, N.C.; Manousakis, N.M.; Georgilakis, P.S.; Korres, G.N. Numerical observability method for optimal phasor meas-
urement units placement using recursive Tabu search method. IET Gener. Transm. Distrib. 2013, 7, 347–356.
47. Hanis, N.; Rahman, A.; Zobaa, A.F. Binary PSO Algorithm for Optimal PMUs Placement. IEEE Trans. Ind. Inform. 2017, 13, 3124–
3133
48. Maji, T.K.; Acharjee, P. Multiple Solutions of Optimal PMU Placement Using Exponential Binary PSO Algorithm for Smart Grid
Applications. IEEE Trans. Ind. Appl. 2017, 53, 2550–2559.
49. Dalali, M.; Karegar, H.K. Optimal PMU placement for full observability of the power network with maximum redundancy
using modified binary cuckoo optimization algorithm. IET Gener. Transm. Distrib. 2016, 10, 2817–2824.
50. Müller, H.H.; Castro, C.A. Genetic algorithm-based phasor measurement unit placement method considering observability and
security criteria. IET Gener. Transm. Distrib. 2016, 10, 270–280.
51. Theodorakatos, N.P.; Moschoudis, A.P.; Lytras, M.D.; Kantoutsis, K.T. Research on optimization procedure of PMU positioning
problem achieving maximum observability based on heuristic algorithms. AIP Conf. Proc. 2023, 2872, 120032.
52. Rezaeian Koochi, M.H.; Dehghanian, P.; Esmaeili, S. PMU Placement with Channel Limitation for Faulty Line Detection in
Transmission Systems. IEEE Trans. Power Deliv. 2020, 35, 819–827.
53. Tshenyego, O.; Samikannu, R.; Mtengi, B.; Mosalaosi, M.; Sigwele, T. A Graph-Theoretic Approach for Optimal Phasor Meas-
urement Units Placement Using Binary Firefly Algorithm. Energies 2023, 16, 6550.
54. Rather, Z.H.; Chen, Z.; Thøgersen, P.; Lund, P.; Kirby, B. Realistic Approach for Phasor Measurement Unit Placement: Consid-
eration of Practical Hidden Costs. IEEE Trans. Power Deliv. 2015, 30, 3–15.
55. Singh, S.P.; Singh, S.P. Optimal cost wide area measurement system incorporating communication infrastructure. IET Gener.
Transm. Distrib. 2017, 11, 2814–2821.
56. Korkali, M.; Abur, A. Placement of PMUs with Channel Limits. In Proceedings of the 2009 IEEE Power & Energy Society General
Meeting, Calgary, AB, Canada, 26–30 July 2009.
57. Korkali, M.; Robust wide-area fault visibility and structural observability in power systems with synchronized measurement
units. In Advances in Electric Power and Energy: Static State Estimation; IEEE: New York, NY, USA, 2020; pp. 209–230.
58. Korkali, M.; Abur, A. Impact of network sparsity on strategic placement of phasor measurement units with fixed channel ca-
pacity. In Proceedings of the 2010 IEEE International Symposium on Circuits and Systems, Paris, France, 30 May–2 June 2010;
pp. 3445–3448.
59. Baba, M.; Nor, N.B.M.; Aman Sheikh, M.; Irfan, M.; Tahir, M. A Strategic and Significant Method for the Optimal Placement of
Phasor Measurement Unit for Power System Network. Symmetry 2020, 12, 1174.
60. Fan, N.; Watson, J.P. On integer programming models for the multi-channel PMU placement problem and their solution. Energy
Syst. 2014, 6, 1–19.
61. Emami, R.; Abur, A. Robust measurement design by placing synchronized phasor measurements on network branches. IEEE
Trans. Power Syst. 2010, 25, 38–43.
62. Andic, C.; Ozturk, A.; Turkay, B. Power system state estimation using a robust crow search algorithm based on PMUs with
limited number of channels. Electr. Power Syst. Res. 2023, 217, 109126.
63. Shafiullah, M.; Hossain, M.I.; Abido, M.A.; Abdel-Fattah, T.; Mantawy, A.H. A Modified Optimal PMU Placement Problem
Formulation Considering Channel Limits under Various Contingencies. Measurement 2019, 135, 875–885.
64. Manousakis, N.M.; Korres, G.N. Optimal placement of PMUS considering scada measurements and fixed channel capacity by
semidefinite programming. In Proceedings of the MedPower 2014, Athens, Greece, 2–5 November 2014; pp. 1–6.
65. Manousakis, N.M.; Korres, G.N. Optimal PMU placement for numerical observability considering fixed channel capacity—A
semidefinite programming approach. IEEE Trans. Power Syst. 2016, 31, 3328–3329.
66. Manousakis, N.M.; Korres, G.N. Semidefinite programming for optimal placement of PMUs with channel limits considering
pre-existing SCADA and PMU measurements. In Proceedings of the 2016 Power Systems Computation Conference (PSCC),
Genoa, Italy, 20–24 June 2016.
67. Manousakis, N.M.; Korres, G.N. Optimal PMU arrangement considering limited channel capacity and transformer tap settings.
IET Gener. Transm. Distrib. 2020, 14, 5984–5991.
68. Qi, B.; Yuan, Y.; Yang, Y.; Bu, Q.; Chen, J. Chapter 1—Overview of Smart Substations; Yuan, Y.; Yang, Y., Eds.; IEC 61850-Based
Smart Substations; Academic Press: Cambridge, MA, USA, 2019; pp. 1–24.
69. Li, J.; Huang, Q.; Li, P.; Huang, H.; Zhang, L. Chapter 8—Intelligent Status Monitoring System for Smart Substations; Yuan, Y.; Yang,
Y., Eds.; IEC 61850-Based Smart Substations; Academic Press: Cambridge, MA, USA, 2019; pp. 255–305.
70. Bogani, C.; Gasparo, M.; Papini, A. Generalized pattern search methods for a class of nonsmooth optimization problems with
structure. J. Comput. Appl. Math. 2009, 229, 283–293.
Algorithms 2024, 17, 191 62 of 62

71. Hosseini, S.S.S.; Gandomi, A.H.; Nemati, A.; Hosseini, S.H.S. Reactive Power and Voltage Control Based on Mesh Adaptive
Direct Search Algorithm. In Engineering and Applied Sciences Optimization; Springer: Berlin/Heidelberg, Germany, 2015; pp. 217–
231.
72. The MathWorks Inc. Optimization Toolbox for Use with MATLAB R®. User s Guide for Mathwork. 2016. Available online:
www.mathworks.com (accessed on 15 February 2024).
73. Available online: https://coin-or.github.io/Ipopt/ (accessed on 15 February 2024).
74. Bestuzheva, K.; Besançon, M.; Chen, W.K.; Chmiela, A.; Donkiewicz, T.; van Doornmalen, J.; Eifler, L.; Gaul, O.; Gamrath, G.;
Gleixner, A.; et al. Enabling research through the SCIP optimization suite 8.0. ACM Trans. Math. Softw. 2023, 49, 1–21.
75. Available online: https://yalmip.github.io/_posts/tutorials/2016-09-17-globaloptimization/ (accessed on 15 February 2024).
76. Parametric Fusion (MOSEK 10.1). Available online: https://www.mosek.com (accessed on 15 February 2024).
77. Löfberg, J. YALMIP: A toolbox for modeling optimization in MATLAB. In Proceedings of the 2004 IEEE International Confer-
ence on Robotics and Automation, New Orleans, LA, USA, 2–4 September 2004; pp. 284–289.
78. Zimmerman, R.D.; Murillo-Sánchez, C.E.; Thomas, R.J. MATPOWER: Steady-state operations, planning, and analysis tools for
power systems research and education. IEEE Trans. Power Syst. 2010, 26, 12–19.
79. Available online: https://www.mathworks.com/help/optim/ug/mixed-integer-linear-programming-algorithms.html (accessed
on 15 April 2024).
80. Available online: https://www.mathworks.com/help/optim/ug/intlinprog.html (accessed on 15 April 2024).
81. Available online: https://docs.mosek.com/latest/cxxfusion/mip-optimizer.html (accessed on 15 April 2024).
82. Maratos, N.G. Exact Penalty Function Algorithms for Finite Dimensional and Optimization Problems. Ph.D. Thesis, Imperial
College of Science and Technology, University of London, London, UK, 1978.
83. Available online: https://icseg.iti.illinois.edu/power-cases/ (accessed on 15 February 2024).
84. Available online: https://coin-or.github.io/Cbc/intro.html (accessed on 15 February 2024).
85. Available online: https://www.gnu.org/software/glpk/ (accessed on 15 February 2024).
86. Available online: https://www.gurobi.com/(accessed on 15 April 2024).
87. Available online: https://www.gurobi.com/documentation/current/refman/mipgapabs.html (accessed on 15 April 2024).
88. Available online: https://www.gurobi.com/documentation/current/refman/mipgap2.html (accessed on 15 April 2024).

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual au-
thor(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.