1400 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 22, NO.

3, JULY 2007

A Modified Particle Swarm Optimizer for the

Coordination of Directional Overcurrent Relays
Mohamed M. Mansour, Member, IEEE, Said F. Mekhamer, and
Nehad El-Sherif El-Kharbawe, Student Member, IEEE

Abstract—The coordination of directional overcurrent relays power networks using digital computers. Traditionally, to solve
(DOCR) is treated in this paper using particle swarm optimization such problem the trail and error approach was used, but it suf-
(PSO), a recently proposed optimizer that utilizes the swarm fered a slow rate of convergence, due to the large number of
behavior in searching for an optimum. PSO gained a lot of in-
terest for its simplicity, robustness, and easy implementation. The iterations needed to reach a suitable relay setting. In a trial to
problem of setting DOCR is a highly constrained optimization minimize the number of iterations needed for the coordination
problem that has been stated and solved as a linear programming process, a technique to break all the loops at the so-called “break
(LP) problem. To deal with such constraints a modification to the points” and locate the starting relays at these points (where the
standard PSO algorithm is introduced. Three case studies are coordination process starts) is suggested. Topological methods,
presented, and the results are compared to those of LP technique
to demonstrate the effectiveness of the proposed methodology. including graph theory and functional dependency, are used to
determine the break points [2]–[4]. It is important to mention
Index Terms—Optimal coordination, particle swarm optimiza- that the solution found using topological methods is the best of
tion, power system protection, swarm intelligence.
the alternative settings considered, but not optimal in any strict
sense. In other words the time dial settings of the relays are rel-
atively high. In the year 1988, the coordination of DOCR in the
I. INTRODUCTION
frame of the optimization theory was reported [5]. The values of
the time dial setting (TDS) have been calculated using LP (sim-
URING the operation of a modern interconnected plex method) for a given values of the pick-up currents .
D power system, abnormal conditions (faults, overload,
over-voltage, under-frequency, etc.) can frequently occur.
Recently, the interest in applying artificial intelligence (AI) in
optimization has grown rapidly. Genetic algorithm [6] and evo-
Such conditions cause interruption of the supply, and may lutionary algorithm [7] have been used in the literature to find
damage the equipments connected to the system, arising the an optimal setting of the protective relays.
importance of designing a reliable protective system. In order In 1995, a new evolutionary computation (EC) technique was
to achieve such reliability, a back-up protective scheme is proposed by Kennedy and Eberhart [8], which they called par-
provided to act as the second line of defense in case of any ticle swarm optimizer (PSO). It was reported as an optimization
failure in the primary protection (the first line of defense). To technique as claimed by its co-founder. Actually, PSO is a new
insure reliability of the protective system, the back-up scheme AI technique that can be considered as a member of the wide
shouldn’t come in to action unless the primary (main) fails to category of swam intelligence [9]. It was used to solve a wide
take the appropriate action. In other words, it should operate variety of optimization problems such as ANN training [10], and
after a certain time delay known as coordination time interval function minimization [11].
(CTI), giving the chance for the primary protection to operate. PSO has been recently adopted by a lot of researchers due to
The forementioned situation leads to the formulation of the its superiority to other evolutionary algorithms (EA) regarding
well-known protective relay setting coordination, that consists its memory, and computational time requirements as it relies on
of the selection of a suitable setting of each relay such that very simple mathematical operations, also it requires very few
lines of computer code to implement [12].
their fundamental protective function is met under the desirable
qualities of protective relaying, namely sensitivity, selectivity, A common feature between PSO and most EA is that it is ini-
reliability, and speed [1]. tialized with a random population, but unlike all EA it doesn’t
In the past four decades, several efforts has been devoted to rely on the famous Darwinian natural selection “survival of
the fittest”, but mainly depends on “constructive co-operation”
the automation of the coordination process of directional over-
current relays (DOCR) and distance relays in interconnected among individuals (agents). Another important difference
between PSO and EA is the ability of PSO to keep track of the
position, and the change in position (velocity) of each particle
Manuscript received June 8, 2005; revised March 27, 2006. Paper no.
TPWRD-00342–2005.
(agent), while EA can only keep information regarding the
M. M. Mansour and S. F. Mekhamer are with the Department of Electric position of the members of the population.
Power and Machines, Ain Shams University, Cairo 11517, Egypt (e-mail: This paper presents the solution of the coordination problem
mmsmansour@ieee.org; sfmekhamer@yahoo.com). of DOCR using a modified PSO approach. The development and
N. E. El-Kharbawe is with the University of Saskatchewan, Saskatoon, SK
S7N 5A9, Canada (e-mail: nee917@mail.usask.ca). implementation of the proposed algorithm on three case studies
Digital Object Identifier 10.1109/TPWRD.2007.899259 are given.
0885-8977/$25.00 © 2007 IEEE
MANSOUR et al.: MODIFIED PARTICLE SWARM OPTIMIZER 1401

II. FORMULATION OF THE COORDINATION PROBLEM where

DOCR coordination problem can be stated as a parametric (6)
optimization problem, where the objective function to be mini-
mized is the sum of the operating times of the relays connected
to the system, subject to the following constraints [5].
1) Coordination Criteria: It is a common practice that every
primary protection has its own backup one for guaranteeing (7)
a dependable protective system. The two protective schemes
should be coordinated together, i.e., a predefined coordination where and are constants
time interval (CTI) collapses before the backup scheme comes depending on the type of the relay simulated.
into action. This CTI depends upon the type of the relays 3) Bounds on Relay Settings: The essence of the directional
(electromechanical or microprocessor based), speed of the cir- overcurrent relay coordination study is the calculation of its
cuit breakers, and other system parameters. Typically, the CTI TDS and . It worth mentioning that DOCR allow for contin-
used for electromechanical relays is 0.3 to 0.4 s, while a CTI in uous time dial setting, but rather discrete pickup current setting.
the order of 0.1 to 0.2 s is used in the case of microprocessor Formulating the above constraints gives:
based protective relays. The above situation can be described
(8)
by
(9)
(1)
This problem can be formulated mathematically as [5]
where
(10)
operating time of the backup relay;
operating time of the primary relay. where is the operating time of the primary relay for
In some particular cases, it is important to consider the dy- a fault subject to
namic changes in the network topology that occurs during the (11)
transient conditions. Such transient configurations take place
(12)
when only one relay of the protective zone operates, while the
other one is still inoperative [5], [13]. The transient situation can (13)
be described mathematically by where
(2) coordination criteria used represented by (1);
set of feasible settings, i.e., TDS and ;
where the superscript (’) indicates transient quantities.
2) Relay Operational Characteristics: Typically, the inverse relay characteristics represented by (4) and (5).
time overcurrent relay (OCR) consists of two elements, an in- Once more, in order to determine the operating time of the
stantaneous unit, and a time dependent unit. The time dependent relay for a given fault there are two variables to be set (TDS
unit has two values to be set, the pickup current value , and and ). To solve such problem, one variable is optimized as-
the time dial setting (TDS). The pickup value is the minimum suming that the other one is predefined. Accordingly, there are
current value for which the relay operates, and the time dial set- two methods to solve the coordination problem of directional
ting defines the operating time of the device for each current over current relays.
value. Normally, the characteristics of the OCR are given as a 4) Finding TDS for a Given : In this case, the pickup cur-
curve of versus , where (multiple of pickup current) is rent values of the relays are known previously. For a predefined
the ratio of the relay current, , to the pickup current value, . , (4) is reduced to
(3) (14)
where
Usually, the manufacturers provide Current-Time curves to
describe the operation of the electromechanical relays. These (15)
curves should be represented in an appropriate way to facilitate
its use with computer programs, cause they cannot be stored
and the problem is reduced to a LP problem.
in the computer. Instead, data-fitted formulas are used to repre-
It is worthy to mention that even if (5) is used to represent
sent these curves [14]. Due to the rapid increase in the use of
the characteristics of the relays, the coordination problem can
digital computers in protective relays, standard formulas mod- still be stated as a LP one. In this case, the problem is solved in
eling the inverse OCR characteristics are proposed, e.g., IEC
terms of the variables PTDS, then the corresponding TDS can
and ANSI/IEEE standards [15]. In this work, the following two
be calculated by finding the roots of the polynomial defined by
formulas are used to approximately represent the inverse OCR
(6) using the optimum values of PTDS calculated.
characteristics [14]
5) Finding for a Given TDS: In this case, using (4) con-
(4) verts the problem to a nonlinear optimization one, whose vari-
ables are the pickup currents of the relays. If the relays are rep-
(5) resented by the characteristic equation indicated by (5), then the
1402 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 22, NO. 3, JULY 2007

problem can still be considered as a LP one if it is solved in terms

of ’s. The values obtained for ’s in conjunction with the
relay current would be used to compute the ’s using (7).
In this paper, PSO is used to solve a LP problem of finding
TDS of the relays for a given ’s. The coordination problem
is addressed as a linear optimization one just for the sake of
comparison with previously published results. It should be clear
that PSO is capable of addressing both linear and nonlinear op-
timization problems.

III. PARTICLE SWARM OPTIMIZER

PSO is a population-based optimization technique that is
originally inspired by the sociological behavior associated with
bird flocking and fish schooling [16]. One of the main advan-
tages of PSO is that it needs no gradient information derived
from the objective function. The aforementioned feature is a
common property of all EA (including PSO) allowing them to
be used on functions where the gradient is either unavailable
(due to the discontinuity of most of real functions), or compu- Fig. 1. Three components of the velocity update equation.
tationally expensive to obtain.
The main idea of the PSO algorithm is to maintain a popu- , where [16]. After that, each par-
lation of particles (agents), referred to as “swarm,” where each ticle is allowed to update its position using its current velocity to
particle represents a potential solution to the objective function explore the problem search space for a better solution as follows:
under consideration. Each particle in the swarm can memorize
its current position that is determined by evaluation of the ob- (17)
jective function, velocity, and the best position visited during
It is a common practice in PSO literatures to choose a unity time
its flying tour in the problem search space referred to as “per-
step , accordingly is set to one through out this work.
sonal best position” (pbest). The personal best position, is the
The personal best position is updated after the th iteration ac-
one that yields the highest fitness value for that particle. For a
cording to
minimization task, the position having a smaller function value
is regarded to as having a higher fitness. Also the best position (18)
visited by all the particle is memorized, i.e., the best position
among all pbest positions referred to as “global best position” where is the fitness function.
(gbest). The particles of the swarm are assumed to travel the Referring to (16), the velocity update equation has three
problem search space in a discrete rather than continuous time terms; the first term represents the particle’s memory of its cur-
steps. At each time step (iteration) the velocity of each particle rent velocity (change in position) in the different dimensions of
is modified using its current velocity and its distance from pbest the search space, the second term is associated with “cognition”
and gbest according to since it only takes into account the particle’s own experience,
while the third one represents the “social interaction” between
the particles. These three components are shown in Fig. 1 [17].
Each agent updates each location according to the interaction
(16) of the above three components, are illustrated in Fig. 2.
A better way to understand the mechanism of the stochastic
where search done by PSO is to think of each iteration not a process
of replacing the previous population with a new one (death and
th velocity component at iteration ; birth), but rather as a process of adaptation [16].
rand() random number between 0 and 1; Attempting to increase the rate of convergence of the stan-
current position in the th dimension; dard PSO algorithm to a global optimum, the inertia weight has
been introduced in the velocity update equation [18]. The inertia
acceleration coefficients;
weight is a scaling factor associated with the velocity during
personal best position in the th dimension; the previous time step. According to this modification proposed,
global best position in the th dimension; (16) is modified to
time step.
Usually, the value of the velocity is clamped to the range
to reduce the possibility that the particle might (19)
fly out of the search space. If the space is defined by the bounds
, then the value of is typically set so that where is the inertia weight.
MANSOUR et al.: MODIFIED PARTICLE SWARM OPTIMIZER 1403

A. Repair Algorithm
After position updating of each particle, a position check is
carried out to make sure that none of the particles have flied out
of the search space bounds or violated the constraints, i.e., all
the generated solutions are feasible. If a violation is detected, a
repair algorithm is used to force the violating particle to return
to the feasible region as follows:
When a violating particle is detected, the velocity of this par-
ticle at the kth iteration is reset to zero. Thus, the velocity of the
particle at the iteration is only influenced by the best posi-
tion found so far for that particle and the current best position in
the swarm. The position of the particle is updated just as before
using this new velocity vector. In most cases this new velocity
Fig. 2. Position update operation of agents. vector will point back to a feasible region of the search space.
The result is to have the violated particle move back towards the
feasible region in the next iteration.
The inertia weight governs how much of the previous ve-
However, in some cases, resetting the current velocity of the
locity should be retained from the previous time step. In this
particle is just not enough to force it back to the feasible region.
work a linearly decreasing inertia weight is used [16]. The in-
In this situation, the velocity of the particle is assumed to have
ertia weight is set to decrease linearly from 0.9 to 0.4 during the
only the gbest component, i.e., the trust is put in the social com-
course of a simulation. This setting allows the particles to ex-
ponent of the velocity, again the position of the particle is up-
plore a large area at the start of the simulation (when the inertia
dated using this new velocity vector. This assumption is justified
weight is large), and to refine the search later by using a small
by the fact that the essence, or at least an important aspect, of
inertia weight. In addition, damping the oscillations of the par-
intelligence can be phrased by “the mind is social”. According
ticles around gbest is another advantage gained by using a de-
to this viewpoint, individuals learn primarily from the success
creasing inertia weight. These oscillations are recorded when
of their neighbors, we compare ourselves to others and we imi-
a large constant inertial weight is used. Accordingly, damping
tate the behaviors of others who have succeeded with things we
such oscillations assists the particles of the swarm to converge
care about.
to the global optimal solution. This way, the inertia weight can
Unfortunately, in highly constrained optimization problems,
be likened to the temperature parameter encountered in simu-
some particles still fly out of the search space, i.e., neither of
lated annealing [16].
the previously mentioned modifications succeeds in forcing the
It is important to realize that the velocity term models the rate
particle back to the feasible region. In this case the particle is
of change in the position of the particle. Therefore, the changes
moved back to its pbest position found so far, and allowed to
induced by the velocity update equation represent acceleration,
explore the search space again with its velocity set to zero in
which explains the name of acceleration coefficients for the con-
the next iteration in order to fine-tune the search in the neigh-
stants and . The acceleration coefficients can be thought of
borhood of pbest. A pseudocode that concisely summarizes the
as a balance between exploration (searching for a good solu-
repair algorithm is as follows.
tion) and exploitation (taking advantage of someone else’s suc-
cess). Too little exploration and the particles will all converge on
if Violation is Detected for the First Time then
the first good solution encountered, while too little exploitation
and the particles will never converge, i.e., they will just keep
searching. There is another way of looking at this rather than
behaviors (exploration and exploitation). What must be prop-
erly balanced is individuality and sociality, i.e., traits that influ-
ence behavior. Ideally, individuals prefer being individualistic
yet they still like to know what others have achieved so that they elseif violation is detected for the second time then
can learn from.

IV. MODIFIED PSO ALGORITHM

In most published literature, the standard PSO algorithm is elseif violation is detected for the third time then
used for unconstrained optimization tasks. PSO in its standard
form is not capable of dealing with the coordination of DOCR,
which is a constrained optimization problem. In this paper, two
modifications are introduced to the original PSO algorithm. The
repair algorithm that gives the PSO algorithm the capability endif
of tackling the coordination constraints imposed on the relays,
while searching for an optimal setting. In addition, another tech- The repair algorithm described above is not driven from a set
nique for initializing PSO, rather than the random initialization, of equations and has no theoretical proof. In fact, it is a kind of
is proposed. Both modifications are discussed in the following heuristic algorithm that we tried to improve by computer simu-
sections. lation, till it has reached its final form.
1404 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 22, NO. 3, JULY 2007

TABLE I
PSO PARAMETERS USED DURING THE SIMULATION

V. IMPLEMENTATION RELATED ISSUES

Before presenting the results of the simulated case studies,
some issues related to the earlier stages of implementing the
proposed algorithm are discussed. For instance, selecting the
values of the parameters of PSO used during simulation, fault
current simulation, and the objective function to be optimized.
In what follows, every one of these issues is discussed in detail.

A. Implementation of the Proposed Algorithm

A computer program is developed to implement the proposed
PSO algorithm using Matlab, and executed on a Pentium III
1.3-GHz PC with 256-MB RAM. The program is equipped with
graphical output to demonstrate the progress of the algorithm in
finding an optimal solution.
One of the main advantages of PSO is that it has a few param-
eters to adjust with no probabilistic ones unlike other EA. Inertia
weight and acceleration coefficients are the only parameters to
be adjusted. For the value of inertia weight , it is assumed to
decrease linearly during the course of the simulation from 0.9
to 0.4 according to

(20)

where
Fig. 3. Flowchart of the modified PSO algorithm.
iter current iteration;
minimum (initial) inertia weight;
B. Initialization of PSO w maximum (final) inertia weight;
maximum allowable number of iterations.
Recalling that the PSO algorithm is initialized with a random
The PSO parameters used during the simulation are summa-
population in order to start its search, a problem arises when
rized in Table I. Table I shows that the population size used
dealing with large-scale optimization problems, for instance,
during the simulation is five particles, which is a very little
when coordinating DOCR with distance relays or when the con-
swarm size compared to the previously published PSO litera-
straints on the search space of the problem are in conflict, for ture. In addition, the values of and are equally set to 2. In
instance, when coordinating ground relays. In these cases it is what follows, these two points are discussed.
practically impossible to find a random feasible solution to start According to [21], the optimal number of particles per swarm
the algorithm with, so another technique should be used to ini- depends on the function itself. There is no certain method to de-
tialize the algorithm. This obstacle can be handled by using the cide a suitable number of agents. Yet parametric studies have
idea of Phase I of the simplex method [19], where an auxiliary found that a population size of about 30 is optimal for many
problem having an optimal solution feasible to the problem in problems [22]. A smaller population size of around 10 to 20
hand is solved. The solution obtained for the auxiliary problem particles have been effective for engineering problems using the
is then used to initialize the PSO algorithm to solve the orig- UCLA-PSO proposed in [17]. On the contrary, our modified
inal problem. The auxiliary problem is formulated by adding a PSO requires only 5 agents, and yet still converging to a sat-
new dimension to the original problem, a nonnegative variable isfactory result.
known as the artificial variable [20]. A flow chart of the modi- The equal values of and used in this work are the values
fied PSO is shown in Fig. 3. originally suggested by Kennedy and Eberhart [8]. Although
MANSOUR et al.: MODIFIED PARTICLE SWARM OPTIMIZER 1405

more recent empirical studies of the standard PSO algorithm the TDS variables, as long as they are positive real numbers.
have recommended other values than the standard ones used Accordingly, (21), is reduced to:
here, testing such values during the simulation proved that
no significant improvement in the results could be obtained.
Accordingly, an equal value of the acceleration coefficients is (22)
adopted in a trial to balance the exploration and exploitation
capabilities of the algorithm.
This objective function is subject to the following constraints:
B. Relay Current Simulation (23)
The determination of the relay current during fault intervals is (24)
very important for proper relay setting. Usually the fault current
where
calculation for relay setting is based on 3- and 1- close-in
faults for phase and ground relays respectively. In this work, operating time of the backup relay as a function
commercial packages are used for the calculation of the fault of ;
currents. During the simulation of the system shown in Fig. 8, operating time of the primary (main) relay as a
the two infinite busses are simulated using three single-phase function of ;
ac voltage sources. The pre-fault (load) current is neglected and minimum allowable TDS of the th relay;
the fault inception time is at the start of the simulation, i.e., at maximum allowable TDS of the th relay.
.
D. Using a More Precise Characteristics
C. Objective Function Used During the Simulation
Equation (4) represents the relay characteristics approxi-
There are different forms of objective functions that can be mately. In some cases, it is required to model the relay with
used in the formulation of the coordination problem. In this a more precise formula. In this case, formulas like (5) can be
paper, the minimized objective function is the sum of the op- used. Doing so a problem arises due to the nonlinearity of
erating times of the primary relays as given by (10). The objec- the characteristics. Referring to (5) and (6), it is clear that the
tive function is usually selected to achieve a certain performance operating time of the relay is a nonlinear function of TDS.
index. The form given by (10) is selected to achieve the objective Accordingly, the coordination problem is no more a LP one.
of minimizing the stress on the relays [5]. This can be justified This problem can be manipulated simply by recognizing that
by realizing that the energy dissipated by an equipment as heat the operating time of the relay can be considered as a linear
is approximately proportional to time, assuming constant cur- function of PTDS as long as the value of is known.
rent magnitude. Therefore, the coordination problem can still be formulated
There is no favor in minimizing the operating times of the as a LP problem. Consequently, the problem is solved first in
relays when operating as primary relays over minimizing their terms of PTDS by minimizing the summation of PTDS, and
operating times when operating as backup relays. The reason stating all the constraints in terms of PTDS instead of the TDS
for this is that for a given relay, if its primary operating time is of the relays. Another important thing that should be beard in
minimized, then its operating time when acting as a backup relay mind is the bounds on the values of PTDS. This range can be
is also minimized. In other words, the operating times of the easily determined by substituting the minimum and maximum
primary relays and those of the backup relays are not in conflict values of TDS in (6). Using these results, the values of TDS can
when considered as separate objectives, and therefore a reduc- be calculated by finding the roots of (6), using any numerical
tion of one leads necessarily to the reduction of the other [5]. method, for instance, Gauss-Seidel or Newton-Raphson. In
As mentioned before, the coordination problem is solved in this work, Maltab is used to find such roots. Obviously, for
terms of TDS, given that values of all the relays are pre- each PTDS, Matlab finds three roots; the required TDS is the
defined. In order to solve the coordination problem in terms of root that lies within the bounds, i.e., the one lies between the
TDS, the relay characteristics given by (4) or (5) should be used. minimum and maximum TDS allowed.
For a predefined values of all relays, the objective function
given by (10) turns to VI. SIMULATION RESULTS AND DISCUSSION
In order to validate the applicability of the proposed PSO al-
(21) gorithm, the algorithm is used to coordinate the DOCR of three
different systems adopted from previous literature. The simula-
tion results are compared to those of LP obtained using Matlab
where
optimization toolbox, to demonstrate the effectiveness of the
number of the relays to be set; proposed PSO.
coefficient of the ith relay given by (15). A. System-1: 3-Bus System
It should be noted that due to the particular characteristics The proposed PSO is applied to the 3-bus system shown in
of the mathematical formulation of the optimization problem, Fig. 4. This system has been previously adopted to illustrate the
the solution is independent on the values of the coefficients of usage of the LP in calculating the setting of DOCR [5], and is
1406 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 22, NO. 3, JULY 2007

TABLE II
SIMULATION RESULTS OF THE 3-BUS SYSTEM

Fig. 4. System-1: 3-bus system.

It is important to mention that the authors of [5] have not re-
port the time taken by the Simplex method to find the optimal
setting. In addition, it is believed that the above convergence
time can be further improved by using the concept of auxiliary
Phase I to initialize the algorithm, instead of the random initial-
ization performed in the above case study.
Fig. 5 shows the very fast convergence rate of the proposed
PSO to reach the optimal solution, just about 200 iterations,
while the algorithm stagnates for the rest of the 1000 iterations
simulated. It worth mentioning that the time stated above for
the proposed PSO is the time taken by the algorithm to reach
the optimal solution excluding the stagnation time, i.e., the time
taken to perform just the needed 200 iterations. The reason for
the use of larger number of iterations than the actually required
to reach the optimal solution is to insure a proper inertia weight
at each iteration.
Referring to (20), it is clear that the rate of decrease of
depends on the maximum allowable number of itera-
tions ( . Consequently, caution should be taken when
Fig. 5. Convergence of the proposed PSO to the optimal solution. choosing . If is too large, PSO may stagnate
waiting for the decrease in inertia weight to begin exploration
of the search space. Likewise too few iterations could result
adopted here for the sake of comparison. All relays are identical, in exploitation of local minimum (or maximum) before the
having inverse time characteristic that can be approximated by swarm has had time to adequately explore the solution space
(4), where . The TDS ranges and find the global minimum (or maximum) [17]. As a result a
are from 0.1 to 1.1. CTI of 0.2 s is adopted, and three phase compromise is necessary for choosing a reasonable .
faults at the middle of the transmission lines are considered One way to overcome the use of more iterations than actually
as the relevant faults. Midpoint faults are considered here for the required is to monitor the progress of the algorithm in finding the
sake of comparison with the results reported in [5]. optimal solution. If stagnation is detected, then the simulation
During the formulation of the coordination constraints, the is stopped before reaching . The stagnation of the algo-
transient changes in the network topology are taken into ac- rithm can be judged according to a user-predefined criterion, for
count. For a full system data [5] should be referenced. The re- instance, if the algorithm fails to find any better solution than the
sults of the proposed method compared to those of the Simplex one already found for a certain number of iterations, then stag-
method, reported in [5], and the results obtained by using the nation has occurred.
Linear Interior Point Solver (LIPSOL) algorithm, implemented
by the Matlab optimization toolbox function ‘linprog’ are given B. System-2: 6-Bus System
in Table II, where refers to the objective function. The In this case, our proposed methodology is used to find the
convergence of the proposed PSO to the optimal solution is il- optimum TDS of the 6-system shown in Fig. 6. The 3- faults
lustrated in Fig. 5. are applied at the near-end of each relay (close in faults). The
From the results obtained, it is clear that the proposed PSO primary/backup (P/B) relay pairs and the close in fault currents
algorithm successfully converges to the same optimal relay set- are given in Table III. The CT ratios and the pickup taps of the
tings reached by the classical LP methods (Simplex method and relays are given in Table IV.
LIPSOL algorithm). Regarding the convergence time, it is quite The relays used in the network are the Westinghouse Co-9
obvious from Table II that the time taken by the proposed PSO that can be modeled by (5). The coefficients of (6) and (7) re-
to reach the optimal setting (0.5129 s) is very close to that of quired to calculate the operating times of the relays using (5)
Matlab LIPSOL algorithm (0.4370 s). are given in Table V and Table VI respectively. The TDS are
MANSOUR et al.: MODIFIED PARTICLE SWARM OPTIMIZER 1407

TABLE IV
CT RATIOS AND THE PICKUP TABS FOR THE 6-BUS SYSTEM

Fig. 6. System-2: 6-bus system.

TABLE III
P/B PAIRS AND THE CLOSE IN FAULT CURRENTS FOR THE 6-BUS SYSTEM
TABLE V
COEFFICIENTS OF (6)

TABLE VI
COEFFICIENTS OF (7)

TABLE VII
PTDS VALUES OF THE 6-BUS SYSTEM

assumed to vary between a minimum value of 0.5 and max-

imum value of 11. A CTI of 0.2 s is adopted, while the transient
changes in the network topology are not considered.
The PTDS values obtained from the simulation are given in
Table VII, while The TDS values obtained from solving (6) after
substituting for PTDS are shown in Table VIII. The best values
of the objective function found during the simulation are shown
in Fig. 7.
This case study is so important in demonstrating the effec-
tiveness of the proposed PSO. The algorithm manages to find a TABLE VIII
feasible setting of the relays, while the LIPSOL algorithm fails TDS VALUES OF THE 6-BUS SYSTEM
to converge to any solution at all. The reason for this failure is
caused by effect of the residuals of the function that are neither
growing nor shrinking.
It is clear that the time taken by the algorithm to converge
is approximately four times that of the previous case study
(2.1708 s). This can be interpreted, simply, by the fact that the
dimensionality of this system has increased, 14 DOCR relays to
be set instead of 6 and 31 coordination constraints to be fulfilled
instead of 12. In other words, the algorithm takes longer time It should be clear that PSO performs a stochastic search
to complete a single iteration. process exploring the solution space for the optimal solution.
1408 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 22, NO. 3, JULY 2007

TABLE IX
PTDS VALUES OF THE 8-BUS SYSTEM FOR BOTH PHASE AND GROUND RELAYS

Fig. 7. Best value of the objective function found during the course of
simulation.

TABLE X
TDS VALUES OF THE 8-BUS SYSTEM FOR BOTH PHASE AND GROUND RELAYS

system consists of 2-generators, 2 (Y-Y) earthed power trans-

formers and 7 transmission lines. The system data are given in
[23].
The relays used are the Westinghouse Co-9 having TDS
varying from 0.5 to 11, and a CTI of 0.3 s is used. This system
is equipped with both phase and ground relays. The setting of
Fig. 8. System-3: 8-bus system. the phase relays is based on close-in faults, while the
setting of the ground relays is based on close-in faults. In
this system the transient configuration is neglected.
Accordingly, a nonlinear increase in the convergence time of The PTDS values for both phase and relays are shown in
the algorithm is expected, for instance, if the dimensionality of Table IX, and the TDS values are given in Table X. The output
the problem is doubled, it is not a must that the convergence of the developed program for both phase and ground relays is
time of PSO be doubled too. shown in Figs. 9 and 10 respectively.
From Table IX, it clear that with the increased dimensionality
C. System-3: 8-Bus System of the problem solved, the algorithm converges to a near-optimal
solution. In spite of the very close results obtained, the conver-
This time the proposed methodology is applied to set the gence time of the proposed PSO is very large compared to that
phase and ground relays of the system shown in Fig. 8. This of the Matlab LIPSOL algorithm.
MANSOUR et al.: MODIFIED PARTICLE SWARM OPTIMIZER 1409

number of iterations. In spite of the fact that the constraints in

the case of ground relays model are more restricted than those
of the phase relays model, as can be seen clearly in [25]. Again
this can be interpreted by the stochastic nature of PSO.

VII. CONCLUSION
Optimal coordination of DOCR is presented in this work
using PSO, which is a new optimization technique that still
unpopular in the power engineering community. PSO has been
used before to solve many engineering problems in various
fields. We believe that PSO will gain more popularity in the
upcoming years for its robustness and easiness.
As seen from the simulation results, the proposed PSO suc-
ceeds to converge to the same optimal setting found by the sim-
plex method, as in the first case study. Moreover PSO is capable
Fig. 9. Output of the developed program for setting the phase relays. of finding a feasible setting while the LIPSOL algorithm fails to,
as in the second case study. Regarding the computational speed,
PSO is able to reach a very good result in an acceptable time.
In spite of the fact that most of previously published PSO lit-
erature recommended the use of a population size of about 10 to
20 agents, it has been found during simulation that using only
5 agents is enough to reach a satisfactory result. Consequently,
PSO proposed outperforms the standard PSO in terms of com-
putational speed and memory requirements, as it requires only
25% of the agent recommended by most of the previously pub-
lished PSO literature (5 instead of 20).
There is no attempt made in this work to adjust the parameters
of PSO to the problem in hand. We believe that by doing so, less
execution time and better rate of convergence can be achieved.
Tuning PSO will be a subject of our future work. Another
aspect that will come in our concern is the coordination problem
of a mixed protective scheme, i.e., having other types of relays
along with DOCR. Distance and breaker failure relays are the
Fig. 10. Output of the developed program for setting the ground relays. candidates.

REFERENCES
The execution time of the program is influenced by several [1] P. M. Anderson, Power System Protection. New York: McGraw-Hill,
factors, among which is the syntax of the program. Conse- 1999.
[2] V. V. B. Rao and K. S. Rao, “Computer aided coordination of direc-
quently, rewriting the program can improve the execution time tional relays: Determination of break points,” IEEE Trans. Power Del.,
tremendously. Another important factor that also influences vol. 3, no. 2, pp. 545–548, Apr. 1988.
the execution time remarkably is the programming language [3] V. C. Prasad, K. S. P. Rao, and A. S. Rao, “Coordination of directional
relays without generating all circuits,” IEEE Trans. Power Del., vol. 6,
used to implement the program. As mentioned, Matlab is used no. 2, pp. 584–590, Apr. 1991.
as the platform for implementing the proposed algorithm. It [4] L. Jenkines, H. Khincha, S. Shivakumar, and P. Dash, “An applica-
is believed that the solution time obtained from a Matlab im- tion of functional dependencies to the topological analysis of protec-
tion schemes,” IEEE Trans. Power Del., vol. 7, no. 1, pp. 77–83, Jan.
plementation of the proposed PSO will be much slower than a 1992.
C++ implementation of the same algorithm. The reason behind [5] A. J. Urdaneta, R. Nadira, and L. G. Perez, “Optimal coordination of
this is that Matlab runs in an interpretive mode, i.e., it compiles directional overcurrent relays in interconnected power systems,” IEEE
Trans. Power Del., vol. 3, no. 3, pp. 903–911, Jul. 1988.
and then executes each command one at a time [24]. [6] C. W. So, K. K. Li, K. T. Lai, and K. Y. Fung, “Application of genetic
Moreover, the time taken by the algorithm to reach the above algorithm for overcurrent relay coordination,” in Proc. IEE Develop-
ground relay settings is smaller than that of the phase relay set- ments in Power System Protection Conf., 1997, pp. 66–69.
[7] C. W. So and K. K. Li, “Time coordination method for power system
tings. The reason for this is that the phase relays are initialized protection by evolutionary algorithm,” IEEE Trans. Ind. Appl., vol. 36,
using the random generator function described before. It has no. 5, pp. 1235–1240, Sep./Oct. 2000.
been found that using the Phase I concept to initialize the al- [8] J. Kennedy and R. Eberhart, “Particle swarm optimization,” in Proc.
IEEE Neural Networks Conf., Piscataway, NJ, 1995, pp. 1942–1948.
gorithm can reach the same results but in a shorter time (about [9] P. Tarasewich and P. R. McMullen, “Swarm intelligence: Power in
4 s instead of 6.5 s). numbers,” Commun. ACM pp. 62–67, Aug. 2002 [Online]. Available:
The convergence rate of the algorithm in the case of ground http://www.ccs.neu.edu/home/tarase/TaraseMcMullSwarm.pdf.
[10] R. Eberhart and X. Hu, “Human tremor analysis using particle swarm
relays is better than that of phase relays as depicted in Figs. 9 and optimization,” in Proc. Evolutionary Computation Congr., 1999, vol.
10. In other words, the algorithm reaches the solution in a fewer 3.
1410 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 22, NO. 3, JULY 2007

[11] J. B. Park, K. S. Lee, J. R. Shin, and K. Y. Lee, “A particle swarm Mohamed M. Mansour (M’82) was born in Cairo,
optimization for economic dispatch with nonsmooth cost functions,” Egypt, on October 7, 1952. He received the B.Sc.
IEEE Trans. Power Syst., vol. 20, no. 1, pp. 34–42, Feb. 2005. and M.Sc. degrees in electrical engineering from Ain
[12] J. Kennedy and R. Eberhart, “A new optimizer using particle swarm Shams University, Cairo, in 1975 and 1980, respec-
theory,” in Proc. Symp. Micro Machine and Human Science, Piscat- tively, and the Ph.D. degree in electrical engineering
away, NJ, 1995, pp. 39–43. from the University of Manitoba, Winnipeg, MB,
[13] A. J. Urdenta, L. G. Perez, and H. Resterbo, “Optimal coordination of Canada, in 1984.
directional overcurrent relays considering dynamic changes in the net- He is currently a Professor in the Department of
work topology,” IEEE Trans. Power Del., vol. 12, no. 4, pp. 1458–1464, Electric Power and Machines, Ain Shams University.
Oct. 1997.
[14] IEEE Committee Report, “Computer representation of overcurrent
relay characteristics,” IEEE Trans. Power Del., vol. 4, no. 3, pp.
1659–1667, Jul. 1989.
[15] “Numerical Time Overcurrent Protection and Thermal Overload Re-
lays with Auto-reclosure Option SIPROTEC 7SJ600 V3.2,” Siemens Said F. Mekhamer was born in Egypt in 1964. He
Instruction Manual [Online]. Available: http://siemens.siprotec.de/ received the B.Sc. and M.Sc. degrees in electrical
download_neu/devices/7SJ600/Manual/7SJ6 00_Manual_V3.2.pdf. engineering from Ain Shams University, Cairo,
[16] F. V. D. Bergh, “An analysis of particle swarm optimizer” Ph.D. disser- Egypt, and the Ph.D. degree in electrical engineering
tation, Univ. Pretoria, Pretoria, South Africa, 2001 [Online]. Available: from Ain Shams University with joint supervision
http://www.cs.up.ac.za/cs/fvdbergh/publications.php. from Dalhousie University, Halifax, NS, Canada, in
[17] J. Robinson and Y. Rahmat-Samii, “Particle swarm optimization in 2002.
electromagnetics,” IEEE Trans. Antennas Propag., vol. 52, no. 2, pp. He is currently an Assistant Professor in the
397–407, Feb. 2004. Department of Electric Power and Machines, Ain
[18] Y. Shi and R. Eberhart, “A modified particle swarm optimizer,” in Proc. Shams University. His research interests include
IEEE Int. Conf. on Evolutionary Computation, 1998, pp. 69–73. power system analysis, power system protection,
[19] R. J. Vanderbei, Linear Programming: Foundations and Exten- and applications of AI in power systems.
sions [Online]. Available: http://www.princeton.edu/~rvdb/LPbook/
index.html. ch. 2
[20] J. W. Chinneck, Practical Optimization: A Gentle Introduction [On-
line]. Available: http://www.sce.carleton.ca/faculty/chinneck/po.html. Nehad El-Sherif El-Kharbawe (S’07) was born in
ch. 5 Cairo, Egypt, on October 17, 1978. He received the
[21] F. V. D. Bergh and A. P. Engelbrecht, “Effects of swarm size on coop- B.Sc. and M.Sc. degrees in electrical engineering
erative particle swarm optimizers,” in Proc. Genetic and Evolutionary from Ain Shams University, Cairo, in 2001 and
Computation Conf., 2001, pp. 892–899. 2005, respectively. He is currently pursuing the
[22] A. Carlisle and G. Doizier, “An off-the-shelf PSO,” in Proc. Particle Ph.D. degree at the University of Saskatchewan,
Swarm Optimization Workshop, West Lafayette, IN, 2001, School Eng. Saskatoon, SK, Canada.
Technol, Purdue Univ..
[23] L. G. Pérez and A. J. Urdaneta, “Optimal coordination of directional
overcurrent relays considering definite time backup relays,” IEEE
Trans. Power Del., vol. 14, no. 4, pp. 1276–1284, Oct. 1999.
[24] M. S. Mirotznik, “Enhancing Matlab-to-C++ programming,” IEEE
Spectr., vol. 36, no. 2, p. 64, Feb. 1999.
[25] N. El-Sherif, “Intelligent optimization techniques for protective relays
coordiantion,” M.Sc. dissertation, Dept. Elect. Power Mach., Univ. Ain
Shams, Cairo, Egypt, 2005.