This article has been accepted for publication in IEEE Access.

This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3387704

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.

Digital Object Identifier 10.1109/ACCESS.2017.Doi Number

A comparative review of current optimization

algorithms for maximizing overcurrent relay
selectivity and speed
Sethembiso Nonjabulo Langazane1, and Akshay Kumar Saha2, (Senior Member, IEEE,
Member IEEE HKN)
Discipline of Electrical, Electronics, and Computer Engineering, University of KwaZulu-Natal, Durban 4041, South Africa
Corresponding author: Akshay Kumar Saha (e-mail: [email protected])

ABSTRACT An exponential growth and complexity in diverse distribution systems have contributed to
protection coordination challenges. Initially, protection coordination schemes were achieved by means of
conventional techniques; however, the utilisation of such methods is based on trial-and-error principles and
laborious. Consequently, current studies have adopted the utilisation of particle swarm optimization, artificial
intelligence models, and genetic algorithms to optimise overcurrent relay selectivity and operational speed.
Particle swarm optimization, artificial intelligence, and genetic algorithms are optimization techniques that
at times converges prematurely due to poor selection of control parameters and lack of optimal values, which
results in increased computational time. Therefore, this paper presents a comprehensive review of recent
developments in terms of parametric sensitivity analysis, selection of artificial intelligence models based on
data availability, and the likelihood of solving overcurrent relay coordination problems. The reviewed
literature shows that particle swarm optimization performance is greatly influenced by inertia weight and
swarm size, while the number of iterations has insignificant effect. The findings also indicate that crossover
rate, mutation probability, and population size affect genetic algorithms behaviour. Artificial intelligence
models lack sensitivity study for parametric tuning, that is, number of hidden layers, membership functions,
epsilon in support vector machine, and number of fuzzy rules affects the models’ performance.

INDEX TERMS Adaptive neuro-fuzzy inference system, artificial intelligence, artificial neural networks,
control parameters, genetic algorithms, overcurrent relay, particle swarm optimization, power system
protection, protection coordination, selectivity, sensitivity analysis, speed.

ABBREVIATIONS GA Genetic Algorithms.

ABC Artificial Bee Colony. IDMT Inverse Definite Minimum Time.
AC Alternating Current. IEC International Electrotechnical
Commission.
AI Artificial Intelligence.
MAPSO Modified Adaptive Particle Swarm
ANN Artificial Neural Network.
Optimization.
ANFIS Adaptive Neuro-Fuzzy Inference
MFs Membership Functions.
System.
OBL Opposition Based Learning.
CTI Coordination Time Interval.
PSO Particle Swarm Optimization.
DE Differential Evolution.
PSM Plug Setting Multiplier.
FO Fractional Order.
TMS Time Multiplier Setting.

VOLUME XX, 2017 1

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3387704

S.N. Langazane, A.K. Saha: A comparative review of current optimization algorithms for maximizing overcurrent relay selectivity and
speed

I. INTRODUCTION converging prematurely. Kennedy et al. [16] categorised

During last decades, power systems engineers and researchers velocity equation into three segments for successful
relied on traditional optimization methods for coordinating controlling of particles’ previous velocity to current
overcurrent relay. The drawback of the techniques is that momentum. Reference [17], studied the behaviour of PSO
solutions are based on trial-and-error, and the process is control parameters, it was discovered that both cognitive and
laborious as well as time-consuming [1], [2]. Therefore, ref. social components contained in the velocity update equation
[3] and [4] advocated the importance of adopting evolutionary must not be greater than 4 [17]. A further study in [18]
algorithms to overcome disadvantages presented by traditional presented hypothetical assessment when both cognitive and
optimization techniques. Authors in [5] also mentioned the social components exceed 4, the particles experienced higher
need to utilise meta-heuristic algorithms to eliminate oscillations. These analysis operates as guidance and helps to
drawbacks presented by conventional techniques. Particle improve swarm search ability and avoid premature
Swarm Optimization (PSO) and Genetic Algorithms (GA) convergence [17], [18]. Another work in [19] investigated the
have emerged as efficient and effective techniques for performance of PSO algorithm by using design experiments
handling protection coordination problems. However, tuning for decreasing simulations runs; nonetheless, it does not allow
particle swarm optimization and genetic algorithms control individual parametric sensitivity analyses thus, not much was
parameters to attain optimum overcurrent relay settings is a gained from the study [3], [19].
long-standing problem [6]. A comparative study conducted in Chui et al. [20] reduced Artificial Neural Network (ANN)
[7] demonstrated that GA algorithm fails to converge ineffectiveness by introducing radial basis function which is a
efficiently due to premature convergence caused by loss of training strategy designed to predict time series by adaptive
diversity which results in population deteriorating to local studying algorithm and response surface methodology.
optima. Nevertheless, radial basis function network involves
In order to improve converging speed and succeed determining the optimal core parameters and kernel function
premature convergence of genetic algorithms, the concept of width. Further study in [21] utilised ANN to identify faulty
altering crossover and mutation probability has been power line and location of the faulty from the source,
employed in [5], [8]. The significance of choosing crossover thereafter, a backtracking algorithm was adopted to coordinate
and mutation probability in such a way that GA behaviour is the primary and backup relay such that selectivity is
robust and efficient has been documented in research work [7] maximised [21]. The findings were compared with other
– [10], [11]. Rojas et al. [8] performed a basic genetic studies of ANNs presented in [22] and [23] which revealed it
algorithm sensitivity study where one genetic operator was performs better with hybrid algorithm. In [24], adaptive
varied at a time to analyse and study system behaviour, and it neuro-fuzzy inference systems (ANFIS) managed to
was discovered that mutation rate, crossover type, and outperform the ANNs due to its advantageous utilization of
population size have negligible effect on genetic algorithm hybrid model between neural network and fuzzy logic.
performance while crossover probability has substantial Consequently, this paper intends to present a parametric
impact [8]. In [11], author urged that mutation probability sensitivity analysis for particle swarm optimizer and genetic
must accelerates as population diversity increases [11]. algorithms, a comprehensive review of artificial intelligence
Another sensitivity study conducted in [12] and [13] showed models, and a comparative evaluation with respect to
combination of crossover and mutation probabilities yields convergence speed and fitness values. Table I highlights the
optimal global solution [12], [13]. Nevertheless, the best gaps in the existing research and the contributions of the
optimum solution is attained when integrating a low mutation proposed paper. The contributions of the paper are as follows:
probability with a high crossover rate. A review supplying a) A comprehensive review of evolutionary algorithms
dynamic method for selecting mutation and crossover with respect to parametric sensitivity analysis based
probabilities is presented in [9], and a further study in [10] on convergence speed and fitness function values.
reviewed the application of GA control parameters, pointing b) Analysis of overcurrent relay behaviour based on
out the pros and cons of genetic operators whereas in [5], a various literature to evaluate whether protection
review of control parameters was presented stating past, selectivity and operational speed is accomplished.
present, and future methods of assessing genetic algorithms. c) To review the application of artificial intelligence
In contrast, a study comparing PSO algorithm behaviour models in overcurrent protection coordination and
with other meta-heuristic methods shows that PSO algorithms the concept of adaptive control parameters.
manage to attain global minima with fewer number of Selectivity and speed studies are presented in section II. An
iterations [14]. However, PSO algorithm presented limitations overview of protection philosophy is presented in section III.
such as failing to fly pass local optima region hence, not Section IV provides algorithms that optimises overcurrent
succeeding to reach global optimal solutions. Therefore, relay coordination. Selectivity and speed analysis in terms of
research work in [14] and [15] presented alternative solutions convergence rate are presented in section V. Artificial
for improving PSO algorithm performance significantly. An intelligence models for optimizing overcurrent relays are
experimental study [15] established the introduction of weight provided in section VI. Lastly, recommendations for future
term into velocity update equation to enhance particles’ search studies as well as conclusive remarks, are presented in section
capability by stabilising local and global search. Swarm size VII and section VIII, respectively.
and inertia weight are extremely associated with algorithms’
2 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3387704

S.N. Langazane, A.K. Saha: A comparative review of current optimization algorithms for maximizing overcurrent relay selectivity and
speed

TABLE I. Summary of literature survey

Algorithm Current approach for tuning parameters Proposed approach

Particle swarm optimization In [15], authors suggested the use of larger Parametric sensitivity analysis is done for
inertia weight in the beginning, thereafter, self-adaptive control parameters such as
slowly decreasing to minimal value. self-adaptive mutation with dynamic non-
However, the time-varying based inertia linear inertia weight, evolutionary state-
weight variation may not lead to a global based acceleration coefficients, and
optimal solution. Ref. [19] investigated the adaptive velocity update to further improve
performance of PSO algorithm by using PSO algorithm performance.
design experiments for decreasing
simulations runs; nonetheless, it does not
allow individual parametric sensitivity
analyses thus, not much was gained from
the study.

Genetic algorithm Reference [12] and [13] adopted a This paper reviews current analysis of
combinational strategy that tunes both crossover probability, mutation probability
mutation and crossover probabilities and population size. Moreover, sensitivity
simultaneously to observe algorithm analysis to evaluate parameters that
behaviour. Similarly, authors in [8] varied significantly affects algorithms’
crossover and mutation probabilities at the performance is presented. Parametric
same time. The approaches only focused on analysis that changes on parameter at a time
crossover and mutation probabilities which while keeping others constant is proposed.
are not the only genetic operators that affect Analyses are performed based on
the behaviour of GA algorithm. Due to the convergence speed, selectivity, and fitness
variation of parameters simultaneously, it function.
was difficult to determine poor performing
parameters.

Artificial intelligence Authors in [24] utilised ANFIS structure As a result, this paper outlines the effect of
generating adaptive pickup current and time model architecture, quantity of If-Then
multiplier setting parameters based on fuzzy rules, data processing speed, membership
rules to provide input/output pairs. An function, and the anticipated output.
improvement in performance was achieved Comparative study is presented for AI
when ANFIS model used seven numbers of models with respect to overcurrent relay
Gaussian-type membership functions (MFs) coordination.
and more precise findings were obtained
using five numbers of triangular-type MFs
[24]. Nonetheless, the use of least-squares
estimation in training leads to complex
information extraction in ANFIS relays in
comparison to simple and straightforward
fuzzy logic overcurrent relays.

II. SELECTIVITY AND SPEED STUDIES as safety factor, circuit breaker interrupting time, maximum
The concept of relay selectivity measures system reliability fault currents, and load currents were taken into consideration
and effectiveness by ensuring continuous power supply even which resulted in protection miscoordination and longer fault
in instances when abnormal conditions occur in certain clearing time. According to [28] relay selectivity in a radial
protected zones. It plays an important role in segregating distribution network made up of one feeder and unidirectional
faulty section while leaving the healthy part intact and power flow can be attained by precise time grading of the
functional for continual supply of power to end-users. In a overcurrent relay, whereas other configurations such as ring
study attempting to investigate the factors affecting selectivity and mesh networks, the selectivity is not adequate for system
and operational speed of relays in [25], Sorrentino et al. protection; thus, directional overcurrent relay is favourable for
proved that it is not possible to achieve selectivity for mesh or such systems [28]. In [29], it was explicitly stated that majority
ring distribution network with more than two equivalent of distribution network adopts time grading method to achieve
power sources due to identical currents seen by the relay. selectivity. However, both techniques result in higher tripping
Author in [26] and [27] improved protection selectivity as well times which have detrimental effects on power systems
as speed of operation by using time grading margins of 0.2 and expensive equipment and safety of personnel [28], [29].
0.3 seconds, respectively. Distribution system constraints such Depending on the type of distribution network setup, heuristic,
2 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3387704

S.N. Langazane, A.K. Saha: A comparative review of current optimization algorithms for maximizing overcurrent relay selectivity and
speed

hybrid, and artificial intelligence techniques are preferred to signals into overcurrent relay [34]. The received signals are
maximise selectivity of relays and speed of operation since analysed and compared to the predetermined values.
these work in conjunction to optimise protection on the Distribution networks are typically safeguarded by
network. Furthermore, these optimization techniques mitigate overcurrent relays against extreme currents due to added
iterative trial and error as well as laborious process which is advantage of being cost-effective and thus preferred on
utilised in conventional methods [30]. Therefore, there is a distribution level in comparison to distance and differential
necessity to review optimization techniques for coordinating protection [35].
overcurrent relays such that maximum selectivity and At times, primary protective devices are bound to
optimised operational speed is obtained and to advance maloperation in the distribution network due to breakdown
converging speed and circumvent premature convergence of in circuit breaker trip mechanism, inadequate trip coil
some algorithms. voltage, or faulty protection relay [30]. Consequently,
backup protection provides a second line of defence to avoid
III. OVERVIEW OF PROTECTION PHILOSOPHY catastrophic damage to the scheme by detecting and clearing
In a power distribution system, the reliability of electrical abnormalities from the system. However, in order to enforce
protection system is as critical as selectivity, sensitivity, and sequential operation of primary and backup protection,
speed since these protection characteristic parameters are optimum protection coordination between relays is essential.
dependent such that an increase in one parameter leads to the Kudkelwar et al. [36] formulated overcurrent relay operating
other decreasing [3], [30], [31]. For this reason, protection time mathematically through the objective function as
quality was improved in [32] through the utilisation of analysis follows [13].
tools to evaluate all protection system characteristics. 𝑀𝑖𝑛(𝐶) = ∑𝑘𝑑=1 𝑊𝑜 × 𝑆𝐷𝐽 (1)
Similarly, authors in [31] studied energy distribution Where 𝐶 the objective function, 𝑘 the total number of
redundancy and reliability, sensitivity, speed, and selectivity relays installed, 𝑊𝑜 the weight factor, and 𝑆𝐷𝐽 is the
is not included in [31]. Nevertheless, the main attributes of functioning time of 𝐷𝑡ℎ main protection at its assigned zone
insignificant protection selectivity and speed of overcurrent 𝐽. The weight factor 𝑊𝑜 can be set at approximately one due
protection were not addressed. System disturbances triggered to its equivalent shorth line length and weight [36]. With the
by unanticipated faults, load transients, and maloperation of objective of maximising protection coordination between
electrical equipment require protection devices to respond upstream and downstream relay, standard inverse time
timely, speedily, and selectively [30]. Protective system characteristic curve is employed in [30]. According to IEC
comprises circuit breakers, relays, and other circuit 60255-151:2009 [33], inverse definite minimum time class
interrupters to isolate misoperational equipment. Circuit for overcurrent protection scheme adopt the standard inverse
breakers function to isolate abnormally operating part of the characteristic equation as follows:
distribution network when triggered by the overcurrent relay, 0.14
which senses, localises a fault and issues a trip command to 𝑇𝑜𝑝 = ( 0.02 ) × 𝑇𝑀𝑆 (2)
𝑃𝑆𝑀 −1
the circuit breaker for promptly discrimination of faulty Where 𝑇𝑜𝑝 the relay operating time, 𝑇𝑀𝑆 the time
section [30], [33]. Protection system must satisfy the multiplier setting, and 𝑃𝑆𝑀 the plug setting multiplier. It was
predominate objectives as follows [30], [33], [34]: stated in [33] that 𝑇𝑜𝑝 values are dependent on utmost fault
• To minimise duration of the fault. current level, IDMT curve type, and the operating time of
• To protect customers' apparatus and continual power downstream relays whereas 𝑃𝑆𝑀 contains the ratio of
supply. maximum fault current 𝐼𝐹 to pick-up current 𝐼𝑃 [33], [38].
• Reduction of overhaul outages to slightest section of The overcurrent relay issue trip command when 𝐼𝐹 ≥ 𝐼𝑃 but
the system. retain normal operating state if 𝐼𝐹 < 𝐼𝑃 [37] under 𝑇𝑀𝑆
• Improvement of system performance, stability, and values ranging between 0.01 < 𝑇𝑀𝑆 < 1.0 seconds [38].
reliability. When fault current 𝐼𝐹 is equal to pickup current 𝐼𝑃 ,
• Deisolation of defective power lines, transformers, overcurrent relay prompts the operation of the circuit breaker
and other apparatus timeously. to isolate the faulty section [37]. According to [34] 𝐼𝑃
settings for phase-to-phase faults can be selected between
A. OVERCURRENT PROTECTION 50% and 200% in steps of 25% whereas authors in [38]
Overcurrent relay provides protection against excessive suggested 10% to 70% in steps of 10% is sufficient for earth
current magnitudes in distribution networks due to leakage faults. Therefore, 𝐼𝑃 is the product of current
malfunctioning of the system [30], [34]. These extreme transformer secondary current and current setting. Current
current magnitudes can be used to signify the existence of setting can be defined as the adjustment of tappings on the
faulty conditions and help activates the operation of relay coil to obtain the desired relay pickup current. The
protection devices accordingly, which vary in system more current setting the relay has, the greater current the
complexity and design requirements [30], [33], [34]. relay needs to send the trip command [37], [38]. References
Auxiliary devices such as current transformers and voltage [31] and [32] presented two common approaches to compute
transformers act as protection and metering components that the pickup current. The first method stipulates that the pickup
measure current and voltage levels which serves as input current is twice the maximum load current, or it must be one-
2 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3387704

S.N. Langazane, A.K. Saha: A comparative review of current optimization algorithms for maximizing overcurrent relay selectivity and
speed

third of the minimum fault current at the nearest busbar [31]. metaheuristic algorithms such as ant colony optimization
Second method proposes that pickup current must be [42] and firefly algorithm [43], studies have proven PSO
selected between 125% of the maximum load current and 2/3 robustness and efficiency in attaining optimal solutions [42]
times of minimum fault current [32]. – [44]. Nonetheless, PSO at times experiences a condition
It is challenging to execute non-linear standard inverse that leads it to converge prematurely to local solution due to
characteristics like (1) directly due to it exponential incapability of particles to fly from local minima. Poor
expression in the order of 0.02. It is imperative to indicate setting and selection of PSO control parameters may
that former researchers and scholars have prevented this contribute significantly to unsatisfactory performance and
inverse definite minimum time characteristic in their work premature convergence [30]. Additionally, PSO algorithm
[39], [40]. Due to this constraint, Amin et al. [41] proposed may be inefficient as it needs higher function evaluations to
the exploitation of artificial intelligence-based optimization obtain optimum solutions in hyperdimensional problems
techniques, particularly neural networks by using universal [45]. Consequently, new PSO variants are developed to
function approximation capabilities. Reference [30] and [41] address algorithms’ premature convergence and improve
utilised evolutionary algorithm and artificial intelligence converging speed.
models, respectively, to accomplish optimal protection Table II depicts some state-of-the-art PSO algorithm
coordination and assure precise sequential operation of variants. An earlier study modified PSO algorithm to find
upstream and downstream relays which is achieved by solutions to multi-objective [34] and discrete [47]
coordination time interval (CTI) [30], [31]. Under normal optimization problems. A number of PSO algorithm
system conditions, the backup protection is not active until variations have been introduced either by incorporating
CTI exceeds the predetermined value [30]. Once the CTI is theories of PSO algorithm with other meta-heuristic
exceeded, the backup relay must operate within coordination techniques or by developing novel mechanisms. In [48], PSO
constraints as formulated in the following equation [30], was coupled with differential evolution (DE) for constraint
[31]: handling by using feasibility principles to obtain best
𝑇𝑏𝑎𝑐𝑘𝑢𝑝 − 𝑇𝑚𝑎𝑖𝑛 ≥ 𝐶𝑇𝐼 (3) personal solution of individual particles in the swarm.
Where 𝑇𝑏𝑎𝑐𝑘𝑢𝑝 the backup protection operating time, Furthermore, a study in [15] introduced the weight
𝑇𝑚𝑎𝑖𝑛 the time of operation for main protection. The backup expression into velocity update equation to improve search
relay 𝑇𝑏𝑎𝑐𝑘𝑢𝑝 must function within the coordination capability of particles by stabilising search abilities.
constraints of with coordination time interval set between 0.1 Reference [16] categorised velocity update equation into
seconds and 0.2 seconds for microprocessor-based relay, three segments to control the effects of particles’ past
while electromagnetic relay utilises 0.3 seconds to 0.4 velocity to prevailing speed. In a study aiming to learn the
seconds [31], [32]. performance of PSO algorithm particles [17], it was
proposed that acceleration coefficients in velocity update
IV. ALGORITHMS TO OPTIMISE OVERCURRENT RELAY equation must not exceed 4. A further analysis in [18]
COORDINATION demonstrated that when acceleration coefficients exceed 4,
Scholars have been exploring different algorithms to find the particle tends to experience higher oscillations. The
solutions to complex and multi-dimensional problems, and literature reviewed serves as a guide and assists in enhancing
taking into consideration factors such as problem swarm search abilities and avoiding converging prematurely
complexity, computational strength, time availability, and [17], [18]. Another work in [19] examined PSO algorithm
understanding of function behaviour, an optimization behaviour using design experiments, this method decreased
algorithm may be chosen. Nature is commonly utilised as an the number of runs; nonetheless, it does not allow analyses
inspiration in establishing and solving sophisticated of individual parameter, thus not mush was discovered [19].
optimization techniques such as hybrid algorithms which are Constraint handling technique problems were highlighted
utilised to enhance converging behaviour of meta-heuristics. in [49] as crucial in effective performance of PSO, it was
In this section, algorithms for optimizing overcurrent relays clearly stated that constraint handling approaches provide
and their developments are discussed. essential information with regards to solution feasibility
[49]. Therefore, PSO algorithm modifications are necessary
A. PARTICLE SWARM OPTIMIZATION to handle constrained optimization problems. Ref. [6]
To evaluate distinctive optimization techniques that will proposed adaptive constraint handling approach that prefers
yield superior performance throughout optimization any feasible solution over unfeasible ones, and among two
problems, a statistical approach demanding unfeasibly great feasible solutions, the one with best objective function values
number of simulations may be required, which is not is favoured. It was observed in [50] that by choosing feasible
practical; however, greater certainty can be set on techniques solutions only, the algorithm position itself to favour feasible
that constantly achieve improved results than others. Particle solutions and the particles moves towards feasible regions
swarm optimization has the capability to utilise its memory with optimal solution.
for updating particles' location which other nature-inspired
algorithms lack this feature. In contrast with other

2 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3387704

S.N. Langazane, A.K. Saha: A comparative review of current optimization algorithms for maximizing overcurrent relay selectivity and
speed

TABLE I. Some state-of-the-art PSO variants

PSO Variant Description

CVI – PSO [51] In CVI-PSO, constraints are handled by attracting problem solutions
to the feasible search regions through interval arithmetic that
normalises overall violations and the subsequent objective function is
executed by means of basic lexicographic approach.
PSO – DE [48] In an attempt to minimise premature convergence in the original PSO,
differential evolution (DE) was introduced to handle constraints
through the utilisation of feasibility principles to best individual
solutions of particles in the region.
APSO [52] An accelerated PSO was introduced to accelerate algorithm
convergence to transverse into global best position by velocity vector.
PSOGSA [53] PSOGSA was developed to integrate optimization intelligence of
gravitational search algorithm (GSA) in local search with the robust
capability PSO in social learning global best solution mathematically.
DPSO [54] By incorporating deterministic (DPSO), a significant reduction in
CPU time cost was achieved by eradicating stochasticity in PSO. In
[55] and [56], the technique has been effective in determining single
and multi-objective functions, respectively.
FPSOGSA [57] A concept of integrating fractional calculus within the mathematical
model of canonical PSO algorithm and coupled with traditional GSA
to further improve optimization characteristics such as convergence
rate.
PSO – TVAC [58] Acceleration coefficients are varied with respect to time such that
cognitive coefficient is lowered while social coefficient is accelerated
as the search advances. Re-initialisation approach is enforced to
particles’ velocity in such a manner that static performance is
prevented.
MAPSO [30] A modified PSO with self-adaptive inertia weight, and acceleration
coefficients based on feedback control parameters generated by the
fitness of individual particles.
ALC – PSO [59] Inspired by aging leaders and challengers, the swarm leader is
assigned with a propagating age and a life expectancy, and particles
are motivated to protest each other for leadership when swarm leader
has aged. Therefore, diversity is facilitated, and algorithms' efficiency
is enhanced.
PSO – DFCM [60] In PSO-DFCM, premature convergence was prevented, and
algorithms' performance significantly improved by introducing a
damping factor and collaborative mechanism into particle swarm
optimizer. Local exploitation and global exploration of particles are
regulated by damping factor, whereas the cooperative mechanism
updates particles' velocity through the utilisation of global and local
best swarm learning systems.

The literature highlights PSO algorithm premature (𝑘)

particle velocity 𝑣𝑖 changes with current velocity and
convergence and other factors contributing to unsatisfactory (𝑘)
position from the 𝑝𝑏𝑒𝑠𝑡𝑖 solution and 𝑔𝑏𝑒𝑠𝑡 (𝑘) solution.
behaviour of the algorithm [42] – [45], other authors [48] – (𝑘) (𝑘)
[60] introduce variants to optimise the performance of PSO. Therefore, the 𝑣𝑖 and 𝑠𝑖 updated in accordance with the
However, setting particle swarm optimizer parameters to subsequent equation [61]
yield optimal overcurrent relay settings is a long-standing (𝑘+1) (𝑘) (𝑘) (𝑘) (𝑘)
problem. The subsequent paragraph presents an overview of 𝑣𝑖 = 𝑣𝑖 + 𝑐1 𝑟𝑎𝑛𝑑1 (𝑝𝑏𝑒𝑠𝑡𝑖 − 𝑠𝑖 ) +
(𝑘) (𝑘)
PSO algorithm background and its sensitivity parameters. 𝑐2 𝑟𝑎𝑛𝑑2 (𝑔𝑏𝑒𝑠𝑡 (𝑘) − 𝑠𝑖 ); 𝑖 = 1 𝑡𝑜 𝑁 (4)
In PSO, individual particle initially transverse through the (𝑘+1) (𝑘) (𝑘+1)
multidimensional search space at spontaneous velocity and 𝑠𝑖 = 𝑠𝑖 + 𝑣𝑖 ; 𝑖 = 1 𝑡𝑜 𝑁 (5)
its prevailing position in the 𝑖-th dimension is indicated by (𝑘) (𝑘)
(𝑘) Where 𝑁 the swarm size, and 𝑟𝑎𝑛𝑑1 are 𝑟𝑎𝑛𝑑2
𝑠𝑖 where 𝑘 is the iteration number, and 𝑖 the individual numbers produced randomly every 𝑘 iteration ranging
particle. Each particle learns its best position and its own between 0 and 1. Acceleration coefficients 𝑐1 and 𝑐2 , named
(𝑘)
experience indicated by 𝑝𝑏𝑒𝑠𝑡𝑖 , and the whole swarm as cognitive and social components, separately, are positive
experience is expressed by 𝑔𝑏𝑒𝑠𝑡 (𝑘) . At each iteration, the constants [61].

2 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3387704

S.N. Langazane, A.K. Saha: A comparative review of current optimization algorithms for maximizing overcurrent relay selectivity and
speed

FIGURE 1. Particle velocity and position update in a two-dimensional search space [6]

Figure 1 depicts how particles' velocity, cognitive • Inertia weight: Due to limitations presented by
component, and global component affect particles' search 𝑣𝑚𝑎𝑥 , authors in [15] suggest incorporating weight
capabilities for best optimal solutions. In [26], the particles’ term into velocity update equation to improve
velocity update equation is grouped into three segments, the particles’ exploring capability by balancing the
first part of the equation denotes particles’ momentum which 𝑝𝑏𝑒𝑠𝑡𝑖
(𝑘)
and 𝑔𝑏𝑒𝑠𝑡 (𝑘) search [15]. The inertia
includes the influence of the past velocity on current weight 𝑤 is the scaling factor associated with the
velocity, second segments represent cognitive constant velocity iteration during the last time step and helps
which indicates particles’ pull velocity toward its individual to enhance PSO algorithm converging rate. In
𝑝𝑏𝑒𝑠𝑡 while the last segment demonstrates 𝑔𝑏𝑒𝑠𝑡 or social accordance with the alteration proposed in [56],
collaboration between particles, as depicted in Fig. 1. inertia weight is included into (4) as follows:
Subsequent calculation of particles’ new velocity and (𝑘+1) (𝑘) (𝑘) (𝑘) (𝑘)
(𝑘) 𝑣𝑖 = 𝑤𝑣𝑖 + 𝑐1 𝑟𝑎𝑛𝑑1 (𝑝𝑏𝑒𝑠𝑡𝑖 − 𝑠𝑖 ) +
position, 𝑝𝑏𝑒𝑠𝑡𝑖 and 𝑔𝑏𝑒𝑠𝑡 (𝑘) are updated according to (𝑘) (𝑘)
the following equations: 𝑐2 𝑟𝑎𝑛𝑑2 (𝑔𝑏𝑒𝑠𝑡 (𝑘) − 𝑠𝑖 ); 𝑖 = 1 𝑡𝑜 𝑁 (8)
(𝑘) (𝑘) (𝑘) A higher inertia weight value promotes
(𝑘) 𝑠 , 𝑖𝑓 𝑓(𝑠𝑖 ) < 𝑓(𝑝𝑏𝑒𝑠𝑡𝑖 )
𝑝𝑏𝑒𝑠𝑡𝑖 = { 𝑖 (𝑘) (𝑘) (𝑘)
(6) exploration, whereas a lower value facilitates
𝑝𝑏𝑒𝑠𝑡𝑖 , 𝑖𝑓 𝑓(𝑠𝑖 ) ≥ 𝑓(𝑝𝑏𝑒𝑠𝑡𝑖 ) exploitation which maximise local search ability of
PSO algorithm. An earlier study presented in [15]
(𝑘) (𝑘)
𝑠𝑖 , 𝑖𝑓 𝑓(𝑠𝑖 ) < 𝑓(𝑔𝑏𝑒𝑠𝑡 (𝑘) ) demonstrated significant improvement in PSO
𝑔𝑏𝑒𝑠𝑡 (𝑘) = { (𝑘)
(7) algorithm performance with inertia weight set
𝑔𝑏𝑒𝑠𝑡 (𝑘) , 𝑖𝑓 𝑓(𝑠𝑖 ) ≥ 𝑓(𝑔𝑏𝑒𝑠𝑡 (𝑘) )
between 0.9 to 1.2. Presently, researchers have
Where 𝑓 is the fitness function of PSO algorithm. favoured the use of linearly descending inertia
Normally, the velocity of the particle is constant to the range weight which was original executed in [62], the 𝑤
[−𝑣𝑚𝑎𝑥 , 𝑣𝑚𝑎𝑥 ] to reduce the likelihood of particles value was regulated between 0.9 to 0.4 of which
traversing out of feasible search area [61]. It has been yields better results. The subsequent weighting
observed that setting higher 𝑣𝑚𝑎𝑥 causes particles to function is employed in linearly decreasing inertia
transverse past optima solution, whereas lower 𝑣𝑚𝑎𝑥 values weight [62]:
𝑤 −𝑤
minimise particles’ exploitation abilities thus, particles get 𝑤 = 𝑤𝑚𝑎𝑥 − 𝑚𝑎𝑥 𝑚𝑖𝑛 × 𝑖𝑡𝑒𝑟 (9)
𝑖𝑡𝑒𝑟𝑚𝑎𝑥
trapped in local minima solution [47], [61]. Where 𝑤𝑚𝑎𝑥 the maximum inertia weight, 𝑤𝑚𝑖𝑛
the minimum inertia weight, 𝑖𝑡𝑒𝑟 is the current
1) PSO CONTROL PARAMETERS iteration, and 𝑖𝑡𝑒𝑟𝑚𝑎𝑥 the maximum iterations.
The PSO algorithm discussed in the former section has Reference [61] conducted a comprehensive
control parameters that govern algorithms' performance and sensitivity analysis by evaluating inertia weight
have an influence on the whole search abilities of the values between 0.8 − 1.2, and it was discovered
technique. These parameters include inertia weight, swarm that bigger values facilitate global search whereas,
size, acceleration coefficients, number of iterations, and small values encourage local search [15], [19].
velocity clamping-limit, are described as:
2 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3387704

S.N. Langazane, A.K. Saha: A comparative review of current optimization algorithms for maximizing overcurrent relay selectivity and
speed

• Acceleration coefficients: The two positive Tuning control parameters to solve optimization problems
constants, 𝑐1 and 𝑐2 related to velocity of traversing still remains an issue. The aforementioned paragraphs
particles into optimum regions and its individual discuss setting control parameters based on theoretical and
position, these components regulate search abilities mathematical assumptions, which at time leads to premature
and period taken to accomplish feasible solutions by convergence of the algorithm. Harrison et al. [68] utilised
individual particles. Authors in [61] set both 𝑐1 and function analysis of variance to compute original PSO
𝑐2 to 2 and observed substantial advancement in control parameters i.e., inertia weight 𝑤, cognitive 𝑐1 and
terms of algorithms’ behaviour whereas, social 𝑐2 parameters. It was seen that 𝑤 provides the greatest
altering 𝑐1 and 𝑐2 values resulted in particles moving sensitivity, thus the most crucial parameter to be tuned as it
to infeasible regions [61]. For higher 𝑐1 and 𝑐2 affects the probability of particles’ momentum in reaching
parameters, particle transverse pass feasible regions global optima. The optimization problem complexity to be
and at lower values, particles stagnate local solution solved serve as a guide to select reasonable parameter values
due to getting trapped in the local regions before and by testing the proposed control parameters on a common
moving towards optimum solutions [61]. benchmark function, as done in [6]. Reference [6] conducted
Consequently, 𝑐1 and 𝑐2 have been set to 2 since the control parameter sensitivity analysis and introduced the
beginning of PSO algorithm [15], [62]. unique adapting control parameters to circumvent algorithm
• Number of iterations: A study conducted in [63] converging prematurely. As extensively discussed, some
showed that bigger maximum number of iterations scholars tend to utilise the published values in the literature
𝑖𝑡𝑒𝑟𝑚𝑎𝑥 results in longer computational effort, and while others prefer the use of fine-tuned static values.
it was observed that the chosen 𝑖𝑡𝑒𝑟𝑚𝑎𝑥 value Nonetheless, these methods were found inefficient and
directly affects the likelihood of PSO algorithm ineffective due to control parameters being time dependent.
succeeding to global solutions [63]. Furthermore, Consequently, new novel self-adapting and dynamic PSO
poor selection of 𝑖𝑡𝑒𝑟𝑚𝑎𝑥 values lead to algorithm approaches are developed and currently favoured to solve
converging prematurely. Smaller 𝑖𝑡𝑒𝑟𝑚𝑎𝑥 values modern optimization problems.
lessen algorithms’ probability of reaching global
solution whereas, for larger 𝑖𝑡𝑒𝑟𝑚𝑎𝑥 , convergence B. DEVELOPMENTS ON PSO CONTROL PARAMETERS
speed improves at the expense of computing time Numerous variances of PSO algorithm based on self-
[63], [64]. adapting strategies have been developed using various
• Swarm size: Typically, swarm size, 𝑁, selection is methods as the feedback parameter to keep record of
based on the optimization problem application and algorithms’ condition and make adjustments in accordance
complexity. It performs an imperative part in the with transient states. The algorithm advancements have been
behaviour of PSO and affects diversity of the categorized in the subsequent subsection.
population as it controls number of particles
1) INERTIA WEIGHT ADAPTATION STRATEGIES
traversing toward optimum space [56], [60].
Authors in [65] specified that swarm size, 𝑁, A novel self-adaptive mutation with dynamic non-linear
inertia weight was proposed in [69] to improve searching
selected between 5 and 10 particles is precise
momentum by using the average particle spacing approach.
estimate; however, utilising swarm size with the
The average particle spacing 𝑆(𝑡) describes dispersion
range 10 and 50 particles is commonly favoured to
degree of the swarm among individual particles as follows
solve optimization complex problems [65]. For
[69]:
larger population size, particles learn to traverse
1
additional search space and algorithm improves 𝑆(𝑡) = ∑𝑁 𝐷 𝑘 −𝑘 2
𝑖=1 √∑𝑑 (𝑥𝑖𝑑 − 𝑥𝑖𝑑 ) (10)
𝑁𝐻
performance but at computational efforts [60], [65].
• Velocity clamping limit: Ref. [66] explicitly Where 𝐻 is the diagonal maximum length of the search
stipulates that [−𝑣𝑚𝑎𝑥 , 𝑣𝑚𝑎𝑥 ] bounds particles’ area, 𝑁 and 𝐷 are swarm size and space dimension,
𝑘 −𝑘
velocity with 𝑣𝑚𝑎𝑥 = 𝛿 × (𝑥𝑚𝑎𝑥 − 𝑥min ), where separately. 𝑥𝑖𝑑 and 𝑥𝑖𝑑 are coordinate average value of the
𝑥𝑚𝑎𝑥 and 𝑥𝑚𝑖𝑛 are design variables (minimum and 𝑖-th particle in 𝑘-th iteration, and it was reported that smaller
maximum), and 𝛿 the clamping constant set at average spacing results in more concentrated swarm and
[0.1 − 1.0]. The 𝑣𝑚𝑎𝑥 value regulates the poorer species diversity [69]. For altering the nonlinear
smoothness of the constraining changes in velocity inertia weight 𝑤(𝑡) as swarm diversity transverse, a new
[67]. Higher 𝑣𝑚𝑎𝑥 value facilitates global nonlinear dynamic approach based on average particle
exploration; nevertheless, very high 𝑣𝑚𝑎𝑥 value spacing was adopted as defined:
leads to particles flying past global optima which 𝑤(𝑡) = 1/(1 + 𝑒 −10(𝑆(𝑡)−0.5) ) (11)
results in premature convergence [66]. On the
contrary, smaller 𝑣𝑚𝑎𝑥 value promotes local search; Another effective group fitness variance strategy
however, a very small 𝑣𝑚𝑎𝑥 results in particles employed in [70] was also utilised in [69] to accomplish a
demanding more iterations to obtain global self-adaptive adjustment of global and local search ability.
optimum solution [66], [67].
2 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3387704

S.N. Langazane, A.K. Saha: A comparative review of current optimization algorithms for maximizing overcurrent relay selectivity and
speed

TABLE II. Impact of different acceleration coefficient values on PSO performance

Acceleration coefficient value Impact on PSO performance

𝒄𝟏 = 𝒄𝟐 = 𝟎 [71] Particles transverse with constant speed towards the search space as its
velocity is independent from the effect of 𝑝𝑏𝑒𝑠𝑡 and 𝑔𝑏𝑒𝑠𝑡. The position
update is entirely dependent on the inertia weight and swarm is attracted to
the 𝑔𝑏𝑒𝑠𝑡 position.
𝒄𝟏 > 𝟎 and 𝒄𝟐 = 𝟎 [72] Only the cognitive component influences particles' velocity such that the
particle travels toward global search space based on individual experience.
𝒄𝟏 = 𝟎 and 𝒄𝟐 > 𝟎 [72] The cognitive component does not influence particle momentum, whereas
social component directly facilitates exploitative behaviour and displays
quicker convergence as particles rely on the experience of the greatest
particle in the whole swarm.
𝒄𝟏 = 𝒄𝟐 > 𝟎 [71] Particles travels towards median of 𝑝𝑏𝑒𝑠𝑡 and 𝑔𝑏𝑒𝑠𝑡 solution.

𝒄𝟏 > 𝒄𝟐 [71] Cognitive acceleration that facilitates exploration behaviour.

𝒄𝟏 < 𝒄𝟐 [71] Social acceleration promotes exploitation search.

Nagra et al. [73] presented outstanding advancement in the 2) ACCELERATION COEFFICIENT

performance of PSO algorithm when the self-adaptive inertia Since the beginning of PSO algorithm, the cognitive and
weight with gradient-based local search was utilised to social parameters have been set positive and constant [36],
balance exploitation and exploration capabilities [73]. A [44]. As depicted in Table III, researchers observed the
ranked-based inertia weight system was proposed in 74] and impact of various acceleration coefficient values on
[75] to improve converging speed by means allocating ranks algorithms' behaviour by changing either cognitive or social
to particle based on its fitness value. However, the strategy component and in some instances, both parameters were
yields more efficient results with the increasing iterations varied. Shirazi et al. [71] experimented acceleration
which may increment computational time. Reference [75] coefficient by means of dynamic change such that when 𝑐1
proposed cosine inertia weight approach in the form of increases, 𝑐2 decreases and vice versa. Similarly, Maharana
variable-period cosine function ranging between [0, 𝜋] to et al. [72] tested 𝑐1 > 0 while setting 𝑐2 to zero and
meet the multi-stage inertia weight requirements. At the conversely. However, both approaches are time-dependent
range [0, π/6], the cosine function preserves a larger value only and have proven to be inefficient at times. An approach
(≤ √3/2) and traverse slowly whereas the cosine function that considers particles' evolutionary state was proposed in
declines rapidly at [π/6, 5π/6]. In the range [5π/6, π], the [16], it allows evolutionary state to command acceleration
coefficient into converging towards the best optimum
cosine function keeps a smaller value (≤ −√3/2) and
solution in the hyperdimensional space. Particles are forced
changes slowly. Due to inconsistence of original cosine
into performing in a predefined manner to enhance searching
function, an iterative term 𝐼(𝑡) was introduced into the
abilities of the swarm by the acceleration coefficient
cosine function to modify the period and 𝑤𝑐𝑜𝑠 (𝑡) was equations that follows [6]:
rescaled in the range [𝑤𝑓𝑖𝑛 , 𝑤𝑖𝑛𝑖 ] and denoted by the (𝑐𝑚𝑎𝑥 −𝑐𝑚𝑖𝑛 )×(𝑖𝑡𝑒𝑟𝑚𝑎𝑥 −𝑘)
following equation [75]: 𝑖𝑡𝑒𝑟𝑚𝑎𝑥
(𝑤𝑖𝑛𝑖 +𝑤𝑓𝑖𝑛 ) (𝑤𝑖𝑛𝑖 −𝑤𝑓𝑖𝑛 )
𝑤𝑐𝑜𝑠 (𝑡) = + × 𝑐𝑜𝑠 (
𝐼(𝑡)𝜋
) (12) +𝑐𝑚𝑖𝑛 𝑖𝑓 0 ≤ 𝐸𝑆𝑖𝑘 ≥ 0.5
2 2 𝑡𝑚𝑎𝑥 𝑐1𝑘 = (𝑐𝑚𝑎𝑥 −𝑐𝑚𝑖𝑛 )×(𝑖𝑡𝑒𝑟𝑚𝑎𝑥 −𝑘)
(15)
𝐼(𝑡 + 1) = 𝐼(𝑡) + 𝑎, 𝐼(1) = 0 (13) 𝑐𝑚𝑎𝑥 − 𝑖𝑓
𝑖𝑡𝑒𝑟𝑚𝑎𝑥
𝐼(𝑡)≤𝑡𝑚𝑎𝑥
𝑎1 , { 0.5 ≤ 𝐸𝑆𝑖𝑘 ≥ 1.0
6
𝑡𝑚𝑎𝑥 5𝑡𝑚𝑎𝑥 (𝑐𝑚𝑎𝑥 −𝑐𝑚𝑖𝑛 )×(𝑖𝑡𝑒𝑟𝑚𝑎𝑥 −𝑘)
𝑎 = 𝑎2 , < 𝐼(𝑡) ≤ (14) 𝑐𝑚𝑎𝑥 −
6 6 𝑖𝑡𝑒𝑟𝑚𝑎𝑥
𝑡𝑚𝑎𝑥
𝑖𝑓 0 ≤ 𝐸𝑆𝑖𝑘 ≥ 0.5
{ 𝑎3 , 6 < 𝐼(𝑡) ≤ 𝑡𝑚𝑎𝑥 𝑐2𝑘 = (𝑐𝑚𝑎𝑥 −𝑐𝑚𝑖𝑛 )×(𝑖𝑡𝑒𝑟𝑚𝑎𝑥 −𝑘)
(16)
Where 𝑎 the constant for adjusting 𝑤𝑐𝑜𝑠 (𝑡), the values of 𝑎1 ,
𝑖𝑡𝑒𝑟𝑚𝑎𝑥
𝑎2 , and 𝑎3 control the length of each stage in 𝑤𝑐𝑜𝑠 (𝑡). The
requirements of cosine inertia weight 𝑤𝑐𝑜𝑠 (𝑡) limits the { +𝑐𝑚𝑖𝑛 𝑖𝑓 0.5 ≤ 𝐸𝑆𝑖𝑘 ≥ 1.0
phase (𝐼(t)𝜋/𝑡𝑚𝑎𝑥 ) in the range [0, 𝜋] and (𝐼(t)𝜋/𝑡𝑚𝑎𝑥 ) 0 ≤ 𝐸𝑆𝑖𝑘 ≥ 0.5 refers to low evolutionary state and global
increases from 0 to 𝜋 while 𝑡 moves from 0 to 𝑡𝑚𝑎𝑥 . search is promoted by permitting more global search at the
1 2 1
Consequently, ( + + = 1) must be satisfied beginning and towards the end, local search is facilitated.
6𝑎1 3𝑎2 6𝑎3
[75]. It should be noted that cosine inertia weight approach Larger evolutionary state value 0.5 ≤ 𝐸𝑆𝑖𝑘 ≥ 1.0 promotes
demonstrated significant improvement, but more exploitive search in early iteration stages, allowing particles
experiments need to be performed on practical engineering to traverse towards global solution and developing gradually,
fields. further exploration is facilitated at later stages of the
iterations with two constants 𝑐𝑚𝑖𝑛 and 𝑐𝑚𝑎𝑥 set at 0 and 4.0,
2 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3387704

S.N. Langazane, A.K. Saha: A comparative review of current optimization algorithms for maximizing overcurrent relay selectivity and
speed

respectively [6]. The self-adapting acceleration coefficient sequential constructive crossover operator that
was validated on benchmark engineering constrained builds offspring from selective parents based on
problems which proven to be effective, it managed to parents' structure qualities using better edges
promote convergence speed and overcome premature approach. However, the technique presented the
convergence [27]. drawback of algorithm not utilizing any local search
method for enhancing quality due to the small
C. GENETIC ALGORITHM
population size [80]. A cycle crossover that counts
Contrary to PSO, Genetic Algorithm (GA) search functions' bits circularly from parents and their current
solution space by adopting survival of the fittest strategy, position was proposed in [81], but the produced
whereas PSO algorithm inspiration emerges from social offspring had identical characteristics as parents.
behaviour of animal and birds [3], [18]. In GA, mutated Simple or single-point crossover produces a
chromosomes improve algorithms' solution by maintaining a random number 𝑟 from a recurring allocation and
certain probability of population diversity with a great generates two new individuals (𝑥𝑖′ and 𝑦𝑖′ )
percentage of genetic parameter selection to support new corresponding to the equations that follows [78].
population [30]. Roulette wheel selection approach assigns 𝑥 𝑖𝑓 𝑖 < 𝑟
selection likelihood to individual chromosomes based on 𝑥𝑖′ = { 𝑖 (19)
𝑦𝑖 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
their fitness function values [76]. The numbers are generated 𝑦 𝑖𝑓 𝑖 < 𝑟
randomly to correlates cumulative probability for computing 𝑦𝑖′ = { 𝑖 (20)
𝑥𝑖 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
new population selection [15], [76]. This approach presented Recombination presents a new neighbourhood
drawbacks such as prematurely converging to local optima for supplemental implementation within the
due to the supremacy of individual chromosomes that hyperdimension, which are indicated by either
steadily succeed in the competitiveness and are selected as parent assembly [76], [77]. Consequently, the
parents. The probability 𝑃𝑖 (𝑡 + 1) for individual probability of attaining better performing offspring
chromosome 𝑖 is stated in the equation below, where 𝑓𝑖 (𝑡) is substantially risen. In [64], it was seen that higher
the chromosome 𝑖 fitness function, and 𝑛 denotes population crossover rate leads to introduction of new
size [76]. population quickly, whereas very high
𝑓 (𝑡)
𝑃𝑖 (𝑡 + 1) = 1 𝑛 𝑖 (17) recombination likelihood results to structure
∑𝑗=0 𝑓𝑖 (𝑡)
𝑛 discarding rapidly before selection produces
Due to limitations encountered on roulette wheel approach, improvements [64]. Smaller recombination
variations such as ranking method, scaling technique, and probability leads to stagnating search due to
tournament selection were developed to permit negativity insufficient exploration [64].
and minimization on genetic algorithms [76], [77]. In the • Mutation: Establishes heterogeneousness into the
ranking-based selection method, the probability of individual population by extending the search space for
chromosome 𝑃𝑖 is allocated based on the success of genetic algorithm to assess and mitigate faster
individual solution 𝑖 when all the solutions are mapped based convergence prior to whole search area exploration
on their fitness values to permit minimisation. Chromosomes [76], [77]. An incremental in mutation probability
constituting of larger fitness values have a greater likelihood leads to population discovering beyond current
of succeeding to the subsequent generation. The randomly search region of variable area which may results in
generated number in a range [0,1] contributes to impairment of population by changing surviving
reproduction of a next population 𝑛𝑘𝑒𝑒𝑝 of optimal solutions. feasible solutions. Therefore, smaller mutation
Probability of individual chromosome 𝑃𝑖 can be determined probability advocated [79]. Uniform mutation
as follows [76], [77]. chooses one variable 𝑗 randomly and equate it into
𝑛𝑘𝑒𝑒𝑝 −𝑖+1 a constant number 𝑈(𝑎𝑖 , 𝑏𝑖 ) where 𝑎𝑖 and 𝑏𝑖 are
𝑃𝑖 = 𝑛𝑘𝑒𝑒𝑝 (18)
∑𝑖=1 𝑖 lower and upper bound, separately [79].
Similarly, to PSO algorithm, genetic algorithm consists of 𝑈(𝑎𝑖 , 𝑏𝑖 ) 𝑖𝑓 𝑖 < 𝑗
𝑥𝑖′ = { (21)
control parameters that require tuning such that premature 𝑥𝑖 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
convergence is avoided, and algorithms' performance is • Population size: As stated in [77], the group of
maximised. Eberhart et al. [78] noticed a property that chromosomes known as population affects GA
affected GA algorithm performance that was crossover or algorithm behaviour. It was explicitly asserted that
recombination. Michalewics [79], on the hand, studied small population size results in algorithm
genetic operators and found that mutation have significant performing poorly due to inadequate trail size for
impact on the algorithm convergence. Population size also exploring hyperplane [77]. Contrarily, larger
plays a role in behaviour of the algorithm and computational population size prevents algorithm from converging
effort [77]. prematurely by permitting more particles to occupy
• Crossover: Recombination draws two search space but at computational expense [77],
chromosomes from the reproduced population sets [78]. According to [83], any values between 10-50
and employs crossover. Ahmed [80] proposed
2 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3387704

S.N. Langazane, A.K. Saha: A comparative review of current optimization algorithms for maximizing overcurrent relay selectivity and
speed

is a precise choice however, in other study [82], any with self-adaptive mutation scheme [88], GA algorithm
value between 25-250 produces efficient solutions. overcomes premature convergence when coupled with
overlapping and subpopulation convergence. Overlapping
1) DEVELOPMENTS ON GA CONTROL PARAMETERS
In 2019, Hassanat et al. [84] undertook a comprehensive evaluated between two subpopulations which are determined
study to review GA parameter selection and proposed the by the comparison of finest solutions distance of the
change of crossover and mutation rate dynamically. The subpopulation, and if the search radius of the best individuals
study adopted a deterministic method to decrease crossover of subpopulation is less than the threshold value, then
individuals associated with subpopulation are removed.
probability from 100% to 0% and increase mutation
Subsequent overlapping search, the overall subpopulations
probability linearly from 0% to 100%, and vice versa.
experience convergence process to evaluate converging
However, the approach lacks diversity, operates on smaller
robustness.
population size only, and requires larger number of
mutations. On the other hand, Akter et al. [85] suggested a Another adaptive mechanism to set mutation probability
new crossover operator made up of two crossover points dynamically is proposed in [89], it controls the utilisation of
selection and new offspring reproduction by comparing cost population entropy toward the end of epoch (consecutive
between two parents; the approach needs to be validated generations number) by computing the variation of the
through benchmark problems and compared with other current entropy from the preceding 𝑘 + 1 epoch 𝐻𝑘−1
adaptive as well as self-adaptive methods. A further study in denoted as ∆𝐻𝑘 = 𝐻𝑘 − 𝐻𝑘−1 . A comparison between the
[86] detailed improvements in genetic algorithms and change ∆Hk and the one evaluated in the previous epoch
proposed adaptive GA by modifying important genetic ∆𝐻𝑘−1 = 𝐻𝑘−1 − 𝐻𝑘−2 is performed in such a way that when
operators, that is, crossover and mutation probabilities. The the change in entropic decreases at least by factor 𝜀, loss of
adaptive crossover probability continuously adjusts the diversity is signalized which then triggers the mutation
probability with respect to fitness function value of probability by means of including a constant factor (𝛼).
individuals in the population, it adapts such that for Otherwise, the mutation probability 𝑃𝑚 is reduced by the
individuals made up of smallest and largest fitness functions, subtraction of the constant factor 𝛼 value, 𝑃𝑚 value prevents
the crossover operates with specific probability to overshooting by keeping the value on the interval of
accommodate changes. The crossover probability 𝑃𝑐 is [𝑃𝑚𝐿𝐵 , 𝑃𝑚𝑈𝑃 ]; where the lower bound is set at 0.001 and the
adjusted according to the following equation: upper bound set at 0.1 [89]. In [90], mutation probability was
𝑓 −𝑓
𝑘1 𝑚𝑎𝑥 𝑐 , 𝑖𝑓 𝑓𝑐 ≠ 𝑓𝑚𝑎𝑥 , 𝑓𝑚𝑖𝑛 modified based on stochastic Manhattan learning algorithm,
𝑓𝑚𝑎𝑥 −𝑓𝑚𝑖𝑛
𝑃𝑐 = { 𝑘 , (22) whereas in [91] 𝑃𝑚 was changed by means of fitness
2 𝑖𝑓 𝑓 = 𝑓 𝑐 𝑚𝑖𝑛
frequency distribution. Lastly, ref. [92] utilised the entropy
𝑘3 , 𝑖𝑓 𝑓𝑐 = 𝑓𝑚𝑎𝑥 value for modification of mutation operator instead mutation
Where 𝑃𝑐 the crossover probability, 𝑓𝑐 consists of higher probability values, as reviewed in the aforementioned
fitness in the first two parents of crossover operation, 𝑓𝑚𝑎𝑥 subsection.
and 𝑓𝑚𝑖𝑛 are maximum and minimum fitness, respectively.
𝑘1 , 𝑘2 , 𝑘3 are constants ranging between 0 - 1 and 𝑘2 > 𝑘3 . V. SELECTIVITY AND SPEED ANALYSIS
In another study [87], it was empirically depicted that by In an attempt to discover control parameters that influence
generating mutation probability based on chromosome rank overcurrent relay selectivity and operational speed, a
in the population, quicker convergence was obtained. sensitivity analysis of GA and PSO algorithms was
Besides population size, mutation probability plays a performed in [30]. With a swarm size ranging between 10 to
significant role in general algorithms’ performance. 500 particles, inertia weight set at 0.9, acceleration
The motivation for implementing a rank-based adaptive coefficient set at 2 and maximum velocity set at 50, it was
mutation was to overcome insufficient genetic information observed that an increase in swarm size caused PSO
in the initial population and loss of such information during algorithm performance to be more efficient but at
optimization process. It assigns the fittest chromosome a computational time expense [30]. Authors in [61] utilised
rank 𝑁 in a population of 𝑁 individuals which ranges swarm size set between 20 and 160 particle and noticed that
between [1, 𝑁] depending on their fitness function values. swarm size have minimal effects on the performance of
Mutation probability 𝑃𝑚 adapts the rank of chromosomes 𝑟 particle swarm optimizer. Another study in [65] proposed
as follows: that choosing swarm size must be done based on variables
𝑟−1
𝑃𝑚 = 𝑝𝑚𝑎𝑥 (1 − ) (23) number. Nevertheless, the response of overcurrent relay with
𝑁−1
The best chromosome has zero mutation probability and respect to smaller swarm size was unsatisfactory, the relays
the poorer consists of the maximum probability 𝑝𝑚𝑎𝑥 , took long to operates which resulted in violation of
meaning 𝑃𝑚 distributes linearly between 0 and 𝑝𝑚𝑎𝑥 . protection principle to discriminate faults speedily. A
Nevertheless, if one or more chromosomes obtain identical comparative study conducted in [93] showed that GA
fitness, ranks are allocated randomly to them, and the converges quicker than PSO algorithm meaning the relay
mutation probability remains unaffected by the asymmetry was more selective and operated speedily when configured
of fitness distribution [87]. By applying multi-population with GA algorithm; however, the contributary factors to PSO
algorithm poor convergence were not clearly indicated, as
2 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3387704

S.N. Langazane, A.K. Saha: A comparative review of current optimization algorithms for maximizing overcurrent relay selectivity and
speed

Ref. [30] suspected inertia weight. It was seen that larger seconds which indicated slower response of overcurrent
inertia weight 0.8 – 1.2 was unsuccessful in optimizing relay. In a subsequent study on the linearly increasing
protection coordination hence, selectivity and operational crossover probabilities [0.3, 0.9], Bikirli and Kut [82]
speed were neglected [30] which opposes the study reported that a higher crossover value (0.9) leads to dominant
conducted in [61] that found inertia weight in this range individuals with best fitness function values getting lost in
facilitates global search. Poor selectivity and speed are the hyperplane space. Further experiments conducted in [30]
evident that PSO converged prematurely due to larger inertia utilised a uniform single-point crossover value of 0.3 while
weight and failed to fulfill its goal of reducing iteration varying mutation probability from 0.02 – 0.3 to observe the
number and improving particles' exploration and behaviour of GA algorithm with respect to overcurrent
exploitation abilities. Furthermore, a decreasing inertia selectivity and speed of operation. A behaviour similarly to
weight 0.9 – 0.4 permits particles to transition from increasing the crossover probability was observed,
exploration mode to exploitative mode to produce global increasing mutation rate improved algorithm fitness and aids
solutions, as claimed by [94] and [95]. to avoid premature convergence by generating distinct
Parveen [96] proposed a new hybrid optimization chromosomes. It introduced diversity and preservation
technique particle swarm optimization gravitational search which are predominately the primary purpose of mutation,
algorithm to coordinate relays, both GA and PSO algorithms and the relays were more selective and operated promptly
were utilised for comparison purposes which showed the when required. Another genetic operator that influences the
latter yields infeasible results. The hybrid algorithm showed behaviour GA algorithm is population size as alluded in
its superiority by obtaining more sensitive time multiplier aforementioned section. A sensitivity analysis performed in
setting values which signifies speed relay operation and [30] set population size at 10 – 500 particles, it was proven
protection selectivity [96]. However, control parameters through plots that larger population size results in GA
impact in algorithms’ performance and the relay selectivity algorithm performing robustly and efficient at computational
as well as operational speed was not considered [96]. A efforts expense, which agrees with ref. [82]. Overcurrent
detailed sensitivity analysis is presented in [6] considers the relays operated speedily when population size was set at 500
effect of velocity clamping-limit on the PSO algorithm and when the population was set at 10, the relays took longer
behaviour, it studied distinct clamping constants [0.0, 1.0], to operate with CTI higher than the stipulated value which
taking into consideration local search enhancer, velocity violated one of protection philosophies to discriminate fault
clamping-limit decreasing technique, active penalty scheme, promptly [30].
and reinitialization methods. The attained results revealed Both PSO and GA algorithms parameters were analysed
that the clamping-limits have least effect on the algorithms' and reviewed in terms of their selection influence on the
behaviour which might be due to the problems' nature as the behaviour of overcurrent relay selectivity and speed. From
velocity is normally set to range dynamically [6]. Contrary analysis of PSO algorithm, it can be seen that parameters
to Ref. [6], Barrera et al. [97] experimented different such as swarm size, inertia weight, acceleration coefficient,
polynomial functions that lessen maximum velocity and and number of iterations influences the performance of the
developed a parabola function to positions particles near a algorithm. Therefore, selection of the operating parameters
localized search was proven successfully. The number of plays an essential role in the response of overcurrent relays.
iteration selection is problem-nature dependent, and Sensitivity analysis of GA algorithm revealed that crossover
complexity dependent as larger values increase rate, mutation probability and population size possess a
computational time, while smaller value reduces the direct influence on the behaviour of genetic algorithms.
probability of obtaining global solution [30]. It was reported Slower convergence and higher fitness function was
that protection relay remained selective throughout the experienced when larger crossover, mutation, and population
variation of iterations and the operational speed was size were chosen whereas smaller values yielded faster
minimised [30], the overall overcurrent relay response convergence and optimised fitness function.
showed that larger iterations fail to enhance PSO Selectivity and speed analysis can be substantiated by the
performance due to algorithms’ ability to regulate search performance comparison of GA with PSO algorithm. A
period and not particles movement in search area [30]. comparative study conducted in [98] highlighted that GA
The genetic operators, i.e., crossover and mutation algorithm presents slower convergence behaviour than PSO
probability displayed significant impact on the performance algorithm. Consequently, the authors introduced a hybrid
of GA algorithm [30], the fitness value increased optimization algorithm by combining PSO and GA by
proportionately to crossover and mutation probability. The studying the results of one algorithm as an input for the other.
findings demonstrated that an increase in crossover and An observation of interest in [30] showed that PSO
mutation probability resulted in overcurrent relays response convergence speed was slightly faster than GA; however,
time was increased and the CTI was exceeded on other GA algorithm fitness curves were smoother due to fewer
relays. With crossover probability set at 30%, the yield time alterations and efficient exploitation at the beginning of the
multiplier setting was 2.30 seconds which was optimal value search.
and selectivity was maximised [30]. A crossover probability
of 80% obtained optimum time multiplier value of 4.98

2 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3387704

S.N. Langazane, A.K. Saha: A comparative review of current optimization algorithms for maximizing overcurrent relay selectivity and
speed

TABLE III. Selected state-of-the-art algorithm for the comparative study

Algorithm Description
Artificial Bee Colony (ABC) [100] ABC adopts the concept of self-organization and labour division inspired by
behaviour of a colony of bees. The constraints are handled through employed
bees, onlooker bees, and scouts.
Class Topper Optimization (CTP) [101] CTP imitates from student learning behaviour to improve performance. The
algorithms' premature convergence was circumvented by introducing chaotic
search class topper optimization which uses a logic map in the initial stage and
varying sigmoid functions towards updated stage to maintain diversity in the
search space.
Adaptive Fuzzy Directional Bat Algorithm Calculates optimum overcurrent relay settings in any distribution network
(AFDBA) [102] topology without requiring initial tuning of parameters. It relies on the evaluation
of bat algorithm and fuzzy inference system to determine essential parameters on
a real-time basis. Due to AFDBA dependence on fuzzy inference system, a
higher computational time is incurred which leads to a slower relay response.
Cascade Forward Neural Network (CFNN) CFNN modelling includes cascade neural networks by associating neurons
[103] number and learning rates in the testing as well as learning processes. Its merit is
the ability to maximise the weight at iteration phase to accomplish appropriate
error value based on the set goal. Nonetheless, high infiltration of distributed
generation affects the sampling rate reliability.
ANN Levenbergs Marquardt Algorithm Ref. [102] explored the ability of ANN to predict overcurrent relay
[102] miscoordination time interval between main and backup relay operating time by
assigning the output of the ANN to curve fitting technique. Quicker relay
operating time and less miscoordination time were attained which proves the
robustness of the algorithm.

The experimental work [99] proved that PSO yields global However, a further hybridization [111] OBL with GSA
best solution with at least 100 particles which agrees with the (OBL-GSA) outperformed traditional methods and
sensitivity study in [30]. Furthermore, Beielstein et al. [19] successfully solved overcurrent relay problem. A modified
investigated the behaviour of PSO with GA algorithm and form of class topper optimization with the concept of OBL
found that PSO yields best fitness in fewer iterations and and Fractional Order (FO) was introduced in [28] to enhance
overall performed efficient and effective. exploitation and exploration capabilities, which
Although PSO algorithm converging prematurely and demonstrated it superiority through optimum selectivity and
reliance on the preliminary control parameter settings optimised operational speed of overcurrent relays.
problems were addressed in earlier sections, the algorithm Nevertheless, an incremental in computational effort and
performance can be further improved through hybridization complexity was observed, another observation of interest
with other algorithms. Eberhart and Kennedy [16], [104] was the failure of the algorithm to coordinate small
conducted the first experiment to train Artificial Neural distribution network [28]. Some of the favoured algorithms
Network (ANN) using PSO algorithm, which successfully are tabulated in Table IV with their merits/demerits.
feedforward multilayer perceptron ANN to arrange Fisher's A comprehensive review and analysis of optimization
Iris dataset. Authors in [105] utilised an ANN with 60 input, techniques with respect to relay selectivity and speed
12 hidden nodes, and 2 input nodes to train a feedforward revealed the following:
multilayer perceptron, which resulted in successful • The tuning of GA genetic operators, i.e., population
application of PSO to train ANN with minimal error and size, crossover rate, and mutation probability have
maximised performance. ANNs are used to solve protection the substantial effect on the algorithms'
problems in [106] and a nonlinear signal transformation- performance and overall response of overcurrent
based ANN was suggested to optimize differential protection relays. Parametric sensitivity revealed that varying
scheme. A significant percentage error was encountered in one parameter while keeping the other constant
[107] when the ANN was integrated with the discrete helps to identify operators responsible for the model
wavelet transform algorithm, the error was a result of the failure.
quantity of transient data utilised to train ANN algorithm. • Similarly, PSO algorithm sensitivity analysis
Baran et al. [108] solved overcurrent coordination revealed that inertia weight and swarm size have
problem through ANNs for smart distribution systems, the greater effect on the algorithms’ performance,
work was further modified in [109] to maximise the whereas number of iterations presents an
effectiveness of the artificial intelligence protection strategy. insignificant impact on convergence speed and
In [110], Opposition Based Learning (OBL) is hybridized to fitness function.
enhance cuckoo optimization techniques known as CH- • A comparative review between PSO and GA
EOBCCS, the algorithm updates the population by means of algorithm showed that PSO managed to perform
OBL strategy with chaotic type of cuckoo search technique. efficiently and effectively, the maximised
2 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3387704

S.N. Langazane, A.K. Saha: A comparative review of current optimization algorithms for maximizing overcurrent relay selectivity and
speed

FIGURE 2. Typical ANN architecture for determining optimised TMS and Fault location
selectivity and optimised speed were attained. Although the second order need to be exploited for instructions [114].
optimal PSO performance was obtained, the algorithm is Consequently, Levenberg Marquardt algorithm aids in
sensitive and depends on initial settings. tuning of backpropagation neural network of smaller and
medium sized configurations as it improves convergence rate
VI. ARTIFICIAL INTELLIGENCE MODELS FOR [114]. In another work [115], Levenberg Marquardt
OPTIMIZING OVERCURRENT RELAY algorithm was utilised to teach multilayer perceptron and
In this section, the literature of preceding case studies are optimized structure was obtained which agrees with work in
analysed for appropriate selection of artificial intelligence [116]. Karupiah et al. [117], further adopted the utilization
models based on data accessibility and the possibility of of Levenberg Marquardt ANNs for relay coordination
solving optimization problem. problem and recommended a new efficient relay operational
speed with capabilities of forecasting the likelihood of
A. ARTIFICIAL NEURAL NETWORK protection miscoordination in the distribution network;
In ANNs, transfer functions determine the connection however, the proposed solution was not validated
between the input and output nodes in a system by experimentally. A typical architecture of ANN, as can be
incorporating a non-linearity level and exploiting ANNs seen in Fig. 2, comprises input neurons (𝑥1 … . 𝑥𝑛 ) with
merits of insensitivity to noise data, which leads to excellent given respective weights (𝑊1 … . 𝑊𝑛 ) and the addition of bias
capability for generalization [112]. Reference [20] trained (b) to all inputs [112]. Activation function (𝐹) determines
ANN by using multi-layer feedforward back propagation the association between the weighted input nodes and neural
neural network to decrease the predictable percentage error network output by incorporating non-linearity level required
between virtual and target parameters through searching for in majority of ANN applications [112].
the greatest compilation of relative weights [20]. The A further study in [21] utilised ANN to identify faulty
mechanism tends to be extensively favoured as it presents power lines and location of the fault from the source,
effective function and has the capability of learning in the thereafter, a backtracking method was adopted to provide
user absence. However, in certain circumstances, it results in coordination between primary and secondary relay such that
increased computational time due to an intrinsic failure to selectivity is maximised [21]. The findings were associated
comprehend the output produced by ANN algorithm [20], with other studies of ANNs studied in [22] and [23] which
[112]. Chui et al. [113] reduced this optimization revealed it performs better with the hybrid algorithm. Ref.
ineffectiveness by introducing radial basis function which is [118] hybridised feedforward neural network with support
a training strategy designed to predict time series by adaptive vector machine to detect faults and estimate fault location,
studying algorithm and response surface methodology [113]. respectively. Improved findings were demonstrated, and the
Nevertheless, radial basis function network involves protection settings managed to adapt to system changes,
determining the optimal core parameters and kernel function unlike the traditional ANN that stagnates to changes. A study
width [113]. [119] attempting to estimate protection miscoordination time
Backpropagation learning trains the weight of ANNs of relay operations utilised ANN, which managed to obtain
based on error probability attained in previous iterations, but minimal solution for medium size of radial system; however,
to accomplish faster and efficient fine-tuning, algorithms of miscoordination was produced.

2 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3387704

S.N. Langazane, A.K. Saha: A comparative review of current optimization algorithms for maximizing overcurrent relay selectivity and
speed

FIGURE 3. Flow diagram depicting adaptive relaying using ANFIS

A. ADAPTIVE NEURO-FUZZY INFERENCE SYSTEM improvement in performance was achieved when ANFIS
To achieve optimal protection settings performance, ref. [24] model used seven numbers of Gaussian-type MFs and more
developed adaptive neural-fuzzy inference system (ANFIS) precise findings were obtained using five numbers of
structure generating adaptive 𝐼𝑝 and 𝑇𝑀𝑆 parameters. It triangular-type MFs [24]. It was observed that ANFIS relays
produces suitable input/output mapping and membership yield more precise results under varying load currents and
functions (MFs) based on fuzzy If-Then rules to provide TMS values. Nevertheless, the use of least-squares
input/output pairs [24], as can be seen in Fig. 3. The ANFIS estimation in training leads to complex information
technique was implemented on the normal inverse extraction in ANFIS relays in comparison to simple and
overcurrent relay characteristics, and further experiments are straightforward fuzzy logic overcurrent relays.
required to prove algorithms' effectiveness as well as its
In [120], an ANFIS architecture was included into the
superiority in comparison with other competitive algorithms
development of protective relay model and the structure of
[24]. In ANFIS, every node 𝑖 is equivalent with node
individual relay was associated with the system state.
function 𝑂𝑖1 = 𝜇𝐴𝑖 (𝑥), where 𝑥 the input to node 𝑖 and 𝐴𝑖
Consequently, the protective relay demonstrated better
denotes the linguistic label related to node function. 𝜇𝐴𝑖 (𝑥) performance due to less complicated selection of network
maintains interval 0 ≤ 𝜇𝐴𝑖 (𝑥) ≤ 1 irrespective of bell and the amount of topological state was extremely slightest.
function (24) or Gaussian function (25). Nevertheless, the approach utilised a single inverse relay
1 characteristics and other characteristic curves were not taken
𝜇𝐴𝑖 (𝑥) = 𝑥−𝑐𝑖 2
(24)
1+[( ) ]𝑏𝑖 into consideration [120]. Further study [23] suggested fine
𝑎𝑖
tuning of membership-functions due to highly non-linear
𝑥−𝑐𝑖 2 mapping and self-adaptive nature of tuning MFs of ANFIS.
𝜇𝐴𝑖 (𝑥) = 𝑒𝑥𝑝 [− ( ) ] (25)
𝑎𝑖 The following was drawn from different literature reviewed:
The model architecture, quantity of If-Then rules, data
Where 𝑎𝑖 , 𝑏𝑖 , and 𝑐𝑖 are parameter sets that adjust
processing speed, and the anticipated output are dependent
accordingly with varying bell-shaped function thus
on the quantity of inputs for the model.
displaying different forms of MFs on 𝐴𝑖 label. A circle node
Membership function plays an essential role in model
𝑤𝑖 = 𝜇𝐴𝑖 (𝑥)𝜇𝐵𝑖 (𝑦) signifies the firing strength of a
performance; however, the absence of a standard technique
particular rule and square node 𝑖 with a function 𝑂𝑖4 = for selection of suitable membership function results in
𝑤𝑖 𝑖 𝑥 + 𝑞𝑖 𝑦 + 𝑟𝑖 ) where 𝑤𝑖 the output of previous layer
̅̅̅(𝑝 stagnant performance.
and 𝑝𝑖 , 𝑞𝑖 , and 𝑟𝑖 are consequent parameter set. An
2 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3387704

S.N. Langazane, A.K. Saha: A comparative review of current optimization algorithms for maximizing overcurrent relay selectivity and
speed

Adaptive neuro-fuzzy inference systems managed to models’ robustness and effectiveness in optimizing
outperform the artificial neural network due to its overcurrent relay coordination. Therefore, it is
advantageous utilization of hybrid model between neural advocated to employ more than two various
network and fuzzy logic [24]. artificial intelligence models in a case study for
more accurate results.
VII. RECOMMENDATIONS FOR FUTURE STUDIES • Future studies to entail the effects of renewable
Optimising overcurrent relay selectivity and speed by means energy sources on overcurrent relay coordination
of optimization techniques has demonstrated greater problems. There is a lack of research papers
progress in recent research and implementation; however, focusing on renewable energy association with
the challenge of algorithms' converging prematurely still overcurrent relay selectivity and sensitivity studies.
remains. The poor parametric selection and absence of
models to serve as a guide for choosing critical parameters VIII. CONCLUSION
contribute significantly to unsatisfactory performance. Some The objective of this paper was to review optimization
of the recommendations for future studies to tackle these techniques that maximize overcurrent relay selectivity and
challenges are as follows: operational speed. Due to the disadvantages of traditional
• The problems of protection coordination in modern methods, evolutionary and artificial intelligence techniques
distribution networks are tackled adequately by are favoured to solve overcurrent relay coordination
multiple meta-heuristic techniques, particularly problems in distribution networks. However, literature
evolutionary algorithms with major improvements revealed that tuning control parameters to yield optimum
such as adaptive inertia weight, crossover, overcurrent relay settings is a long-standing problem. The
mutation, and variation acceleration coefficients. comparative study performed in this paper found that control
Nonetheless, it is worth determining a standard parameters definitely influence algorithms' behaviour which
range for coefficients such that divergence risks are ultimately affects overcurrent relay discrimination time. The
eliminated. Furthermore, the hybridization of varying of one parameter at a time while keeping others
metaheuristic methods with other algorithms constant was useful in identifying parameters contributing to
improves selectivity and speed but the likelihood of poor protection settings.
premature convergence remnants which leads to A theoretical review conducted presented numerous
stagnant behaviour of the algorithm. To mitigate variations of PSO algorithm based on self-adapting strategies
stagnating performance, the selection of control by means of various methods as the feedback parameter to
parameters must be standardised based on problem keep track of algorithms' condition and make adjustments in
complexity and a comprehensive sensitivity accordance with transient states. Inertia weight adaptation
analysis needs to be conducted. strategies was presented to accomplish self-adaptive tuning
• In artificial intelligence models, the selection of of global search capabilities, but the approach lacks practical
input variables plays an essential role in experiments to prove its effectiveness. Similarly,
determining the models’ performance and precision acceleration coefficient was experimented by means of
of outcomes. Consequently, it is commended that dynamic change which is time-dependent and proven to be
more research interest into discovering optimum inefficient at times.
input variable combination. Similarly, the filtration There are developments in terms of tuning GA control
of data during pre-processing phase needs to be parameters, adaptive crossover probability continuously
reliable and more accurate for subsequent adjusted likelihoods with respect to fitness function value of
utilisation as input. individuals in the population. Insufficient genetic
• Another observation of interest, artificial information in the initial population and loss of such
intelligence model parameters, that is, number of information during optimization process was mitigated by
hidden layers, membership functions, epsilon in implementing a rank-based adaptive mutation. The
support vector machine, and number of fuzzy rules comparison of PSO and GA algorithm depicts that genetic
affects the models’ performance and the obtained algorithm converges slower than particle swarm optimizer.
results. However, these parameters lack optimal This means particle swarm optimizer maximised overcurrent
values and are varied in different case studies which relay settings and achieved optimum protection coordination
have drawbacks such as longer computation time. It in distribution networks. Although optimal PSO
is advisable to perform a sensitivity study for performance was obtained, the algorithm is sensitive and
parametric tuning using respective algorithms and depends on initial settings.
eliminate the utilisation of a trial-and-error Artificial intelligence models have emerged as faster, most
approach. accurate, and more efficient solutions for protection
• Majority of literature reviewed [24], [117], [118] coordination problems. Adaptive neuro-fuzzy inference
and etc, adopt a maximum of two artificial systems managed to outperform the artificial neural network
intelligence models and performs a comparative due to its advantageous utilization of hybrid model between
study between two models which fails to prove neural network and fuzzy logic. Nonetheless, ANN and

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3387704

S.N. Langazane, A.K. Saha: A comparative review of current optimization algorithms for maximizing overcurrent relay selectivity and
speed

ANFIS performed efficiently in comparison to both [19] T. Beielstein, K. E. Parsopoulos, and M. N. Vrahatis, “Tuning PSO
parameters through sensitivity analysis,” in Proc. Int. Conf. Evol.
evolutionary algorithms, that is, particle swarm optimization
Algorithm, Dortmund, Germany, Jan. 2002, pp. 1 - 6.
and genetic algorithms. Challenges currently faced in the [20] S. Daniel, V. Kvasnicka, and J. Pospichal, “Introduction to multi-layer
research domain and recommendations were detailed with a feed-forward neural networks,” Chem. and intel. Lab. Syst. vol. 39,
main focus on optimization techniques for maximising no.1, pp. 43-62, 1997.
[21] H. Zayandehroodi, A. Mohamed, H. Shareef, and M. Farhoodnea, “A
overcurrent relay selectivity and speed.
novel neural network and backtracking based protection coordination
scheme for distribution system with distributed generation,” Inter. J.
of Electr. Pow. & Ener. Syst., vol. 43, no.1, pp.868-879, 2012.
REFERENCES [22] N. Rezaei, and M.R. Haghifam, “Protection scheme for a distribution
[1] N. I. Nkhasi, “Real-time modelling and simulation of distribution system with distributed generation using neural networks,” Inter. J. of
system protection with and without renewable distributed generation,” Electr. P. & Ener. Syst., vol. 30, no.4, pp.235-241, 2008.
M.S. thesis, Dept. Elect., Electron., Comput. Eng., Univ. KwaZulu- [23] S.A.M. Javadian, M.R. Haghifam, and N. Rezaei, “A fault location
Natal, Durban, South Africa, 2017. and protection scheme for distribution systems in presence of dg using
[2] M. H. Hussain, S. R. A. Rahim, and I. Musirin, “Optimal overcurrent MLP neural networks,” In IEEE Pow. & Ener. Soc. Gen.l
relay coordination: A review,” Proc. Eng., vol. 53, no. 1, pp. 332 - 336, Meet. Tehran, Iran, Jul. 2009, pp. 1-8.
2013. [24] M. Daryani, A. Jamaliyan, K. Esmaeili, and U. Omer, “A hybrid
[3] A. R. Haron, A. Mahamed, and H. Shareef, “Coordination of approach for microgrid protection system based on neural network and
overcurrent, directional and differential relays for the protection of fuzzy logic,” In IEEE PES Innovative Smart Grid Techn. Europe
microgrid system,” in Proc. Int. Conf. Electr. Eng. Informat., vol. 11, (ISGT-Europe), pp. 1-5, 2019.
2013, pp. 366 - 373. [25] E. Sorrentino and N. Gupta, “Summary of useful concepts about the
[4] F. B. Bottura, W. M. S. Bernardes, M. Oleskovicz, and E. N. Asada, coordination of directional overcurrent protections,” CSEE J. of Pow.
“Setting directional overcurrent protection parameters using hybrid and Ener. Syst. vol. 5, no. 3, pp. 382-390, 2019.
GA optimizer,” Electr. Power Syst. Res., vol. 143, pp. 400 - 408, Feb. [26] E. Sorrentino and M. Navas, “Two improvements related to
2017. overcurrent functions for bus protection in distribution
[5] S. Katoch, S. S. Chauhan, and V. Kumar, “A review on genetic systems,” IEEE Trans. on Pow. Deliv. vol. 30, no. 3, pp.1634-1635,
algorithm: Past, present, and future,” Multimedia Tools Appl., vol. 80, 2015.
no. 5, pp. 8091 - 8126, 2021. [27] L. Cheng-Hung and C. Chen, “Using genetic algorithm for overcurrent
[6] M. Isiet and G. Mohamed, “Sensitivity analysis of control parameters relay coordination in industrial power system,” In Inter. Conf. on
in particle swarm optimization,” J. Comput. Sci., vol. 41, pp. 1 - 13, Intell. Syst. Applic. to Pow. Syst., Nov. 2007, pp. 1-5.
Mar. 2020. [28] P.K. Choudhary and D.K. Das, “Optimal coordination of over-current
[7] J. Moirangthem, K. R. Krishnanand, S. S. Dash, and R. Ramaswami, relay in a power distribution network using opposition-based learning
“Adaptive differential evolution algorithm for solving non-linear fractional order class topper optimization (OBL-FOCTO)
coordination problem of directional overcurrent relays,” IET Gener. algorithm,” Applied. Soft Computing, vol. 113, p.107916, 2021.
Transmiss. Distrib., vol. 7, no. 4, pp. 329 -336, 2013. [29] T. Keil and J. Jager, “Advanced coordination method for overcurrent
[8] I. Rojas, H. Pomares, J. J. Gonzalez, R. Merelo, P. Castillo, and G. protection relays using nonstandard tripping characteristics,” IEEE
Romero, “Statistical analysis of the main parameters involved in the trans. on pow. Deliv., vol. 23, no.1, pp.52-57, 2007.
design of a genetic algorithm,” IEEE Trans. Syst., Man, Cybern. C, [30] S.N. Langazane, and A.K. Saha, “Effects of particle swarm
Appl. Rev., vol. 32, no. 1, pp. 31 - 37, 2002. optimization and genetic algorithm control parameters on overcurrent
[9] A. Hassanat, K. Almohammadi, and E. Abunawas, “Choosing relay selectivity and speed,” IEEE Access, vol. 10, pp. 4550-4567,
mutation and crossover ratios for genetic algorithms_A review with a 2022.
new dynamic approach,” Information, vol. 10, no. 12, pp. 390 - 399, [31] E. O. Schweitzer, D. Whitehead, H. J. A. Ferrer, D. A. Tziouvaras, D.
2019. A. Costello, and D. S. Escobedo, “Line protection: Redundancy,
[10] C. K. H. Lee, “A review of applications of genetic algorithms in reliability, and affordability,” in Proc. 37th Annu. Conf. Western
operations management,” Eng. Appl. Artif. Intell., vol. 76, pp. 1 - 12, Protective Relay Conf., Spokane, WA, USA, Apr. 2011, pp. 26-45.
2018. [32] H. J. Altuve, K. Zimmerman, and D. Tziouvaras, “Maximizing line
[11] P. Charbonneau, “An introduction to genetic algorithms for numerical protection reliability, speed, and security,” in Proc. 42nd Annu.
optimization,” Nat. Center Atmos. Res., Boulder, CO, USA, Tech. Western Protective Relay Conf., Spokane, WA, USA, Oct. 2015, pp.
Rep. 450, Mar. 2002, pp. 4 - 7. 128.
[12] M. I. G. Arenas, P. Á. C. Valdivieso, A. M. M. Garcia, J. J. M. [33] B. Ram and N. D. Vishwakarma, Power System Protection and
Guervos, J. L. J. Laredo, and P. García-Sánchez, “Statistical analysis Switchgear. New Delhi, India: McGraw-Hill, 2011.
of parameter setting in real-coded evolutionary algorithms,” in Proc. [34] J. M. Gers and E. J. Holmes, Protection of Electricity Distribution
Int. Conf. Parallel Problem-Solving Nature, Berlin, Germany, Sep. Networks. London, U.K.: The Institution of Electrical Engineers,
2010, pp. 452 - 461. 1998.
[13] L. Núñez-Letamendia, “Fitting the control parameters of a genetic [35] T. Keil and J. Jager, “Advanced coordination method for overcurrent
algorithm: An application to technical trading systems design,” Eur. protection relays using nonstandard tripping characteristics,” IEEE
J. Oper. Res., vol. 179, no. 3, pp. 847 - 868, 2007. Trans. Power Del., vol. 23, no. 1, pp. 5257, Jan. 2008.
[14] X. Hu, R. C. Eberhart, and Y. Shi, “Engineering optimization with [36] S. Kudkelwar, D. Sarkar, “Online implementation of time
particle swarm,” in Proc. IEEE Swarm Intell. Symp., Indianapolis, IN, augmentation of over current relay coordination using water cycle
USA, Apr. 2003, pp. 53 – 57. algorithm,” SN Appl. Sci.1, vol.12, pp. 16-28, 2019.
[15] Y. Shi and R. Eberhart, “A modified particle swarm optimizer,” in [37] M. Rojnic, R. Prenc, H. Bulat, and D. Frankovic, “A comprehensive
Proc. IEEE Int. Conf. Evol. Comput., May. 1998, pp. 69 - 73. assessment of fundamental overcurrent relay operation optimization
[16] J. Kennedy and R. C. Eberhart, “Particle swarm optimization,” in function and its constraints,” Energies, vol.15, no.4, pp. 1-20, 2022.
Proc. IEEE Int. Conf. Neural Netw., vol. 6, Nov./Dec. 1995, pp. 1942 [38] H.H. Zeineldin, “Planning the coordination of directional overcurrent
- 1948. relays for distribution systems considering DG,” IEEE Trans. Smart
[17] J. Kennedy, “The behavior of particles,” in Proc. Int. Conf. Evol. Grid, vol. 7, pp. 1642–1649, 2016.
Program., San Diego, CA, USA, Mar. 1998, pp. 581 - 589. [39] V. Maheshwari, B. Das, A.K. Saxena, “FPGA-based digital
[18] M. Juneja and S. Tiwari, “Reduced order modeling of triple link overcurrent relay with concurrent sense-process- communicate
inverted pendulum using particle swarm optimization algorithm,” Int. cycles,” Intr.J. Electr. Power Energy Syst. vol. 55, pp. 66–73, 2014.
J. Adv. Electron. Electr. Eng., vol. 3, no. 3, pp. 77 - 82, 2014.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3387704

S.N. Langazane, A.K. Saha: A comparative review of current optimization algorithms for maximizing overcurrent relay selectivity and
speed

[40] C.R. Chen, C.H. Lee, C.J. Chang, “Optimal overcurrent relay [60] M. He, M. Liu, R. Wang, X. Jiang, B. Liu, H. Zhou, “Particle swarm
coordination in power distribution system using a new approach,” Int. optimization with damping factor and cooperative mechanism,” Appl.
J. Electr. Power Energy Syst., vol. 45, pp. 217–222, 2013. Soft Comput. vol. 76, pp. 45–52, 2019.
[41] H. Amin, “Piecewise linear approximation applied to the nonlinear [61] Y. Shi and R. C. Eberhart, “Empirical study of particle swarm
function of a neural network,” IEEE Proc. Circuits., Devices Syst. vol. optimization,” in Proc. Congr. Evol. Comput., Washington, DC, USA,
144, no.6, pp. 313–337, 1997. Jun. 1999, pp. 1945 - 1950.
[42] M.Y. Shih, C.A.C. Salazar, and A.C. Enríquez, “Adaptive directional [62] R. C. Eberhart and Y. Shi, “Comparing inertia weights and
overcurrent relay coordination using ant colony optimisation,” IET constriction factors in particle swarm optimization,” in Proc. Congr.
Gen., Trans. & Distr., vol. 9, no. 14, pp. 2040-2049, 2015. Evol. Comput., La Jolla, CA, USA, Jun. 2000, pp. 84 - 88.
[43] R. Benabid, M. Zellagui, A. Chaghi, and M. Boudour, “Application of [63] J. Xin, G. Chen, and Y. Hai, “A particle swarm optimizer with multi-
firefly algorithm for optimal directional overcurrent relays stage linearly decreasing inertia weight,” in Proc. Int. Joint Conf.
coordination in the presence of IFCL,” Inter. J. of Intell. Sys. and Comput. Sci. Optim., vol. 1, Xian, China, Apr. 2009, pp. 505508.
Appl., vol. 6, no. 2, pp. 44-53, 2014. [64] A. P. Engelbrecht, Computational Intelligence: An Introduction, 2nd
[44] X. Hu, R.C. Eberhart, and Y. Shi, “Engineering optimization with ed. Hoboken, NJ, USA: Wiley, 2007.
particle swarm,” in Proc. of the IEEE Swarm Intell. Symposium, [65] J. S. Arora, Introduction to Optimum Design, 4th ed. New York, NY,
Indianapolis, IN, USA, May. 2003, pp. 53–57. USA: Academic, 2017.
[45] J.J. Liang, A.K. Qin, S. Baskar, and P.N. Suganthan, “Comprehensive [66] A. Banks, J. Vincent, and C. Anyakoha, “A Review of Particle Swarm
learning particle swarm optimizer for global optimization of Optimization.Part I: Background and Development,” Nat. Comput.
multimodal functions,” IEEE Trans. Evol. Comput. Vol. 10 no.3, pp. vol. 6 no. 4, pp. 467–484, 2007.
281–295, 2006. [67] F. Van den Bergh, and A. Engelbrecht, “A study of particle swarm
[46] X. Hu, and R. Eberhart, “Multi-objective optimization using dynamic optimization particle trajectories,” Inf. Sci. vol. 176, no. 8, pp. 937–
neighborhood particle swarm optimization,” in Proc. of the Congr. on 971, 2006.
Evol. Comput., Honolulu, HI, USA, Jun. 2002, pp. 1667–1681. [68] K.R. Harrison, B.M. Ombuki-Berman, and A.P. Engelbrecht, “An
[47] J. Kennedy, R.C. Eberhart, “A discrete binary version of the particle analysis of control parameter importance in the particle swarm
swarm algorithm,” in IEEE Inter. Conf. on Syst., Man, and optimization algorithm,” in: International Conference on Swarm
Cybernetics, Orlando, FL, USA, 1997, pp. 4104–4109. Intelligence, Springer, 2019, pp. 93–105.
[48] T.H. Agung, T.A. Nugroho, M. Abdillah, and N.I. Pertiwi, [69] H. Liang, and F. Kang, “Adaptive mutation particle swarm algorithm
“Coordination of Directional Overcurrent Relay Using PSO-DE with dynamic non- linear changed inertia weight”, Optik, vol. 127
Algorithm,” In Inter. Conf. on Tech. and Poli. in Ener. and Electri. no.19, pp. 8036–8042, 2016.
Pow. (ICT-PEP), India, Sept. 2021, pp. 412-417. [70] C. Ze, D. Mengnan, Y. Tiankai, and H. Lijie, “Extraction of solar cell
[49] C.A. Coello, “Theoretical and numerical constraint-handling model parameters based on self-adaptive chaos particle swarm
techniques used with evolutionary algorithms: a survey of the state of optimization algorithm,” Trans. China Electrotech. Soc. Vol. 29, no.
the art,” Comput. Methods Appl. Mech. Engrg., vol. 191, no.11, pp. 9, pp. 245–252, 2014.
1245–1287, 2002. [71] M.Z. Shirazi, T. Pamulapati, R. Mallipeddi, and K.C. Veluvolu,
[50] E. Mezura-Montes, C.A. Coello, “An empirical study about the “Particle swarm optimization with ensemble of inertia weight
usefulness of evolution strategies to solve constrained optimization strategies,” in Inter. Conf. on Swarm Intel. Springer, 2017, pp. 140–
problems,” Int. J. Gen. Syst. vol. 37, no.4, pp. 443–473, 2008. 147.
[51] I. Mazhoud, K. Hadj-Hamou, J. Bigeon, and P. Joyeux, “Particle [72] H.S. Maharana, and S.K. Dash, “Comparative optimization analysis
swarm optimization for solving engineering problems: a new of ramp rate constriction factor based PSO and electro magnetism
constraint-handling mechanism.” Engi. Appl. of Artif. Intell., vol. 26, based PSO for economic load dispatch in electric power system,” in
no. 4, pp.1263-1273, 2013. Inter. Conf. on Applied Machine Learning (ICAML), USA, Mar. 2019,
[52] P.A. Shah, A.S. Nanoty, and V.N. Rajput, “Comparative analysis of pp. 63–68.
different optimization methods for optimal coordination of directional [73] A.A. Nagra, F. Han, F., Q.H. Ling, M. Abubaker, F. Ahmad, S. Mehta,
overcurrent relays,” in Inter. Conf. on Electri., Electro., and Optimi. and A.T. Apasiba, “Hybrid self-inertia weight adaptive particle swarm
Techn. (ICEEOT), Mar. 2016, pp. 3927-3931. optimisation with local search using C4.5 decision tree classifier for
[53] S. Mirjalili, G.G. Wang, L.D.S. Coelho, “Binary optimization using feature selection problems,” Connection Science, vol. 32, no.1, pp.16-
hybrid particle swarm optimization and gravitational search 36, 2020.
algorithm,” Neural Comput. Appl. vol. 25, no.6, pp. 1423–1435, 2014. [74] A. Iftikhar, “Rank-Based Self-Adaptive Inertia Weight Scheme to
[54] E.F. Campana, G. Liuzzi, S. Lucidi, D. Peri, V. Piccialli, and A. Pinto, Enhance the Performance of Novel Binary Particle Swarm
“New global optimization methods for ship design problems”, Opt. Optimization,” in Inter. Scient. Conf. on Inform. Techn. and Data
Eng. vol. 10, pp. 533–555, 2009. Related Research, Singidunum University, Sept. 2021, pp. 63-69.
[55] A. Serani, C. Leotardi, U. Iemma, E.F. Campana, G. Fasano, and M. [75] J. Zhang, J. Sheng, J. Lu, L. and Shen, “UCPSO: A uniform initialized
Diez, “Parameter selection in synchronous and asynchronous particle swarm optimization algorithm with cosine inertia
deterministic particle swarm optimization for ship hydrodynamics weight,” Compu.l Intelli. and Neuro., China, Mar. 2021, pp. 40 -50.
problems,” Appl. Soft Comput. vol. 49, pp. 313–334, 2016. [76] J. H. Holland, “Adaptation in Natural and Articial Systems: An
[56] R. Pellegrini, A. Serani, C. Leotardi, U. Iemma, E.F. Campana, and Introductory Analysis with Applications to Biology, Control, and
M. Diez, “Formulation and parameter selection of multi-objective Articial Intelligent,” Cambridge, MA, USA: MIT Press, 1992.
deterministic particle swarm for simulation-based optimization,” [77] R. L. Hauptand and S. E. Haupt, Practical Genetic Algorithms.
Appl. Soft Comput. vol. 58, pp. 714–731, 2017. Hoboken, NJ, USA: Wiley, 2004.
[57] Y. Muhammad, M.A.Z. Raja, M.A.A. Shah, S.E. Awan, F. Ullah, N.I. [78] X. Hu and R. Eberhart, “Solving constrained nonlinear optimization
Chaudhary, K.M. Cheema, A.H. Milyani, and C.M. Shu, “Optimal problems with particle swarm optimization,” in Proc. 6th World
coordination of directional overcurrent relays using hybrid fractional Multiconf. Syst., Cybern. Inform., Orlando, FL, USA, Jul. 2002, pp.
computing with gravitational search strategy,” Energy Reports, vol. 7, 203206.
pp.7504-7519, 2021. [79] Z. Michalewicz, “Genetic Algorithms C Data Structures D Evolution
[58] K.T. Chaturvedi, T. Pandit, and L. Srivastava, “Particle swarm Programs (AI Series)”, New York, NY, USA: Springer-Verlag, 1994.
optimization with time varying acceleration coefficients for non- [80] H. Z. Ahmed, “Genetic algorithm for the traveling salesman problem
convex economic power dispatch,” Inter. J. of Electr. Pow. & Ener. using sequential constructive crossover operator,” in Proc. of the Inter.
Syst., vol. 31, no. 6, pp. 249–257, 2009. J. of Biometrics & Bioinformatics (IJBB), vol. 3, p. 96, 2010.
[59] W.N. Chen, J. Zhang, Y. Lin, N. Chen, Z.H. Zhan, H.S.H. Chung, Y. [81] K. P. Ghadle and Y. M. Muley, “Travelling salesman problem with
Li, and Y.H. Shi, “Particle swarm optimization with an aging leader {MATLAB} programming,” Inter. J. of Advances in Applied
and challengers,” IEEE Trans. Evol. Comput. vol. 17, no.2, pp. 241– Mathematics and Mechanics, vol. 2, no. 3, pp. 258–266, 2015.
258, 2013.
VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3387704

S.N. Langazane, A.K. Saha: A comparative review of current optimization algorithms for maximizing overcurrent relay selectivity and
speed

[82] G. Bakrl, D. Birant, and A. Kut, “An incremental genetic algorithm [103] M. Alam, “Adaptive protection coordination scheme using numerical
for classication and sensitivity analysis of its parameters,” Expert Syst. directional overcurrent relays,” IEEE Trans. on Industr. Inform., vol.
Appl., vol. 38, no. 3, pp. 26092620, 2011. 15, no. 1, pp. 64-73, 2018.
[83] F. G. Lobo, C. F. Lima, and Z. Michalewicz, “Parameter Setting in [104] R. Eberhart, J. Kennedy, “A new optimizer using particle swarm
Evo-lutionary Algorithms,” Berlin, Germany: Springer, vol. 54. Mar. theory, in: MHS’95,” Proc. of the Sixth Inter. Symposium on Micro
2007, pp. 118. Machine and Human Science, IEEE, Oct. 1995, pp. 39–43.
[84] A. Hassanat, K. Almohammadi, E.A. Alkafaween, E. Abunawas, A. [105] R.C. Eberhart, X. Hu, “Human tremor analysis using particle swarm
Hammouri, and V.S. Prasath, “Choosing mutation and crossover ratios optimization,” in Proc. of the Congress on Evolutionary Computation
for genetic algorithms - a review with a new dynamic (CEC), Washington, WA, USA, Aug. 2002; vol.3, pp. 1927–1930.
approach,” Information, vol. 10, no. 12, p.390 – 399, 2019. [106] M. Mishra, R.R. Panigrahi, and P.K. Rout, “A combined mathematical
[85] S. Akter, N. Nahar, M. ShahadatHossain, and K. Andersson, “A new morphology and extreme learning machine techniques-based
crossover technique to improve genetic algorithm and its application approach to micro-grid protection,” Ain Shams Eng. J., vol.10, 307–
to TSP,” in Intern. Conf. on Electr., Compu. and Commu. Eng. 318, 2019.
(ECCE), Japan, Feb. 2019, pp. 1-6. [107] J.J. Yu, Y. Hou, A.Y.S Lam, and V.O.K. Li, “Intelligent Fault
[86] S.C. Chu, J.C.W. Lin, and J. Li, “Genetic and Evolutionary Detection Scheme for Microgrids with Wavelet-based Deep Neural
Computing”. Springer, Singapore, 2022. Networks,” IEEE Trans. Smart Grid, vol. 3053, 1694–1703, 2017.
[87] A. Basak, “A rank based adaptive mutation in genetic algorithm” [108] M.E. Baran, H. Hooshyar, Z. Shen, A. Huang, “Accommodating high
arXiv preprint arXiv:2104.08842, 2021. PV penetration on distribution feeders,” IEEE Trans. Smart Grid, vol.
[88] I. Korejo, S. Yang, K. Brohi, and Z.U.A. Khuhro, “Multi-population 3, 1039–1046, 2012.
methods with adaptive mutation for multi-modal optimization [109] M. Bakkar, S. Bogarra, F. Córcoles, A. Aboelhassan, S. Wang, and J.
problems,” 2013. Iglesias, “Artificial Intelligence-Based Protection for Smart
[89] N. Stark, G.F. Minetti, and C. Salto, “A new strategy for adapting the Grids,” Energies, vol.15, no.13, p.4933, 2022.
mutation probability in genetic algorithms,” in XVIII Congreso [110] N.A. Alawad, and B.H. Abed-alguni, “Discrete island-based cuckoo
Argentino de Ciencias de la Computación., Argentina, Mar. 2012. search with highly disruptive polynomial mutation and opposition-
[90] D. Thierens, “Adaptive mutation rate control schemes in genetic based learning strategy for scheduling of workflow applications in
algorithms,” in Proc. of the 2002 Congress on Evolutionary cloud environments,” Arab. J. Sci. Eng. vol. 46, no. 4, pp. 3213–3233,
Computation. CEC'02 (Cat. No. 02TH8600), vol. 1, pp. 980-985, 2021.
2002. [111] J. Radosavljević, and M. Jevtić, “Hybrid GSA-SQP algorithm for
[91] S. Blum, R. Puisa, J. Riedel, and M. Wintermantel, “Adaptive optimal coordination of directional overcurrent relays,” IET Gener.
mutation strategies for evolutionary algorithms,” in Annual Transm. Distrib. vol.10, no. 8, pp. 1928–1937, 2016.
Conference: EVEN at Weimarer Optimierungsund [112] F.G. Roba, D.S. Arya, and N.K. Goel, “Streamflow forecasting using
Stochastiktage, vol. 2, pp. 501 -511, 2001. Artificial Neural Network,” Water Energy Int. vol. 57, pp. 30–37,
[92] J. Wang Zai, “Design of mutation operator based on information 2000.
entropy,” in Second Inter. Conf. on Intell. Comp. Tech. and [113] C.C. Chiu, D.F. Cook, J.J. Pignatiello, and A.D. Whittaker, “Design
Automation, vol. 1, pp. 264-266. 2009. of a radial basis function neural network with a radius modification
[93] S.D. Saldarriaga-Zuluaga, J.M. López-Lezama, and N. Muñoz- algorithm using response surface methodology,” J. Intell. Manuf.
Galeano, “Hybrid harmony search algorithm applied to the optimal vol.8, pp. 117–124, 1997.
coordination of overcurrent relays in distribution networks with [114] N.M. Nawi, A. Khan, and M.Z. Rehman, “A New Levenberg
distributed generation,” Applied Sciences, vol. 11, no. 19, p.9207, Marquardt based Back Propagation Algorithm Trained with Cuckoo
2021. Search,” Procedia Technol. vol.11, pp. 18–23, 2013.
[94] R. Poli, “Analysis of the publications on the applications of particle [115] H. Yonaba, F. Anctil, and V. Fortin, “Comparing sigmoid transfer
swarm optimisation,” J. Artif. Evol. Appl., vol. 8, Feb. 2008, pp. 110. functions for neural network multistep ahead streamflow forecasting,”
[95] M. Arumugam, M. Rao, and A. Chandramohan, “A new and improved J. Hydrol. Eng. vol. 15, pp. 275–283, 2010.
version of particle swarm optimization algorithm with globallocal best [116] Y. Tan, and A. Van Cauwenberghe, “Neural-network-based d-
parameters,” Knowl. Inf. Syst., vol. 16, no. 3, pp. 324350, 2008. stepahead predictors for nonlinear systems with time delay,” Eng.
[96] R. Perveen, “Hybrid Optimal Coordination of Overcurrent Relay and Appl. Artif. Intell. vol. 12, pp. 21–35, 1999.
Distance Relay for Offshore Wind Farm,” in 2nd Global Conference [117] S. Karupiah, M. Hussain, I. Musirin I, S.R.A Rahim, “Prediction of
for Advancement in Technology (GCAT), Oct. 2021, pp. 1-6. overcurrent relay miscoordination time using urtificial neural
[97] J. Barrera, O. Álvarez-Bajo, J.J. Flores, C.A. Coello Coello, “Limiting network,” Indonesian J. of Electr. Eng. and Comp. Sc., vol. 14, no.1,
the velocity inthe particle swarm optimization algorithm,” Comput. Y pp. 319-326, 2019.
Sist. vol. 20 no. 4, pp. 635–645, 2016. [118] H. Lin, K. Sun, Z.H. Tan, C. Liu, J.M. Guerrero, and J.C. Vasquez,
[98] B. Yang, Y. Chen, and Z. Zhao, “A hybrid evolutionary algorithm by “Adaptive protection combined with machine learning for
combination of PSO and GA for unconstrained and constrained microgrids,” IET gen., transmit. & distri. vol. 13, no.6, pp.770-779,
optimization problems,” in IEEE Inter. Conf. on Control and 2019.
Automation, April. 2007, pp. 166–170. [119] H. Saberi, T. and Amraee, “Coordination of directional over‐current
[99] S. T. P. Srinivas and S. K. Swarup, “Application of improved invasive relays in active distribution networks using generalised benders
weed optimization technique for optimally setting directional decomposition” IET Gen., Transmit. & Distri., vol. 11, no. 16,
overcurrent relays in power systems,” Appl. Soft Comput., vol. 79, Jun. pp.4078-4086, 2017.
2019 pp. 113. [120] K.S.M.H. Ibrahim, Y.F. Huang, A.N. Ahmed, C.H. Koo, and A. El-
[100] B. Akay, and D. Karaboga, “Artificial bee colony algorithm for large- Shafie, A, “A review of the hybrid artificial intelligence and
scale problems and engineering design optimization,” J. Intell. Manuf. optimization modelling of hydrological streamflow
vol .23 no. 4, pp. 1001–1014, 2012. forecasting,” Alexandria Engineering Journal, vol. 61, no.1, pp.279-
[101] D. Acharya and D.K. Das, “An efficient optimizer for optimal 303, 2022.
overcurrent relay coordination in power distribution system,” Expert
Systems with Applications, vol.199, p.116858, 2022.
[102] F.C. Sampaio, F.L. Tofoli, L.S. Melo, G.C. Barroso, R.F. Sampaio,
and R.P.S Leão, “Adaptive fuzzy directional bat algorithm for the
optimal coordination of protection systems based on directional
overcurrent relays,” Electric Power Systems Research, vol. 211,
pp.108619, 2022.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3387704

S.N. Langazane, A.K. Saha: A comparative review of current optimization algorithms for maximizing overcurrent relay selectivity and
speed

AKSHAY KUMAR SAHA (Senior Member,

IEEE) is currently a Professor and Academic
Leader Research and Higher Degrees at the School
SETHEMBISO NONJABULO LANGAZANE
of Engineering, University of KwaZulu-Natal,
was born in Margate, KwaZulu-Natal, South
Durban, South Africa. He is also a registered
Africa. She received the MSc and BSc. degree in
Professional Engineer with the Engineering
Electrical Engineering from the University of
Council of South Africa and a Fellow of the South
KwaZulu-Natal, where she is currently pursuing
African Institute of Electrical Engineers, South
PhD degree in Electrical Engineering. Her research
African Academy of Engineering. He is also a
interests includes power systems protection and
Senior Member of SAIMC, Member of Academy of Science of South Africa
artificial intelligence.
(ASSAf) and an Individual Member Cigre. His research works are related
to advances of power systems in various areas including engineering
education. He has published more than 60 articles in top-tier international
journals and over 100 international conference papers in relevant areas. He
was awarded the Best Lecturer Electrical Engineering in 2013-2014, 2016-
2019 by the School of Engineering, Research Excellence Award in the years
2015-2022 and Top-30 Publishing Research Award for 2020-2022 by the
University of KwaZulu-Natal. He is also acting as an Editorial Board
Member for a number of top-tier international journals.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4