Applied Soft Computing 29 (2015) 310–327

Contents lists available at ScienceDirect

Applied Soft Computing

journal homepage: www.elsevier.com/locate/asoc

A novel hybrid gravitational search and pattern search algorithm for

load frequency control of nonlinear power system
Rabindra Kumar Sahu ∗ , Sidhartha Panda, Saroj Padhan
Department of Electrical Engineering, Veer Surendra Sai University of Technology (VSSUT), Burla 768018, Odisha, India

a r t i c l e i n f o a b s t r a c t

Article history: In this paper, a hybrid gravitational search algorithm (GSA) and pattern search (PS) technique is proposed
Received 29 July 2013 for load frequency control (LFC) of multi-area power system. Initially, various conventional error criteri-
Received in revised form 8 December 2014 ons are considered, the PI controller parameters for a two-area power system are optimized employing
Accepted 5 January 2015
GSA and the effect of objective function on system performance is analyzed. Then GSA control parame-
Available online 19 January 2015
ters are tuned by carrying out multiple runs of algorithm for each control parameter variation. After that
PS is employed to fine tune the best solution provided by GSA. Further, modifications in the objective
Keywords:
function and controller structure are introduced and the controller parameters are optimized employ-
Load frequency control (LFC)
PID controller
ing the proposed hybrid GSA and PS (hGSA-PS) approach. The superiority of the proposed approach is
Gravitational search algorithm (GSA) demonstrated by comparing the results with some recently published modern heuristic optimization
Pattern search (PS) techniques such as firefly algorithm (FA), differential evolution (DE), bacteria foraging optimization algo-
Governor dead-band non-linearity rithm (BFOA), particle swarm optimization (PSO), hybrid BFOA-PSO, NSGA-II and genetic algorithm (GA)
Generation rate constraint (GRC) for the same interconnected power system. Additionally, sensitivity analysis is performed by varying the
system parameters and operating load conditions from their nominal values. Also, the proposed approach
is extended to two-area reheat thermal power system by considering the physical constraints such as
reheat turbine, generation rate constraint (GRC) and governor dead band (GDB) nonlinearity. Finally, to
demonstrate the ability of the proposed algorithm to cope with nonlinear and unequal interconnected
areas with different controller coefficients, the study is extended to a nonlinear three unequal area power
system and the controller parameters of each area are optimized using proposed hGSA-PS technique.
© 2015 Elsevier B.V. All rights reserved.

1. Introduction
The main objective of a power system utility is to maintain
continuous supply of power with an acceptable quality to all the
consumers in the system. The system will be in equilibrium, when
there is a balance between the power demand and the power gen-
erated. There are two basic control mechanisms used to achieve
power balance, reactive power balance (acceptable voltage profile)
and real power balance (acceptable frequency values). The former
is called the automatic voltage regulator (AVR) and the latter is
called the automatic load frequency control (LFC). For multiarea
power systems, which normally consist of interconnected control
area, in this regard LFC is an important aspect to keep the system
frequency and the interconnected area tie-line power as close as
∗ Corresponding author. Tel.: +91 9438251162.
E-mail addresses: rksahu123@gmail.com (R.K. Sahu),
panda sidhartha@rediffmail.com, prof.dr.spanda@gmail.com (S. Panda),
callsaroj201@rediffmail.com (S. Padhan).
possible to the intended values [1]. The mechanical input power
to the generators is used to control the system as it is affected by
the output electrical power demand and to maintain the power
exchange between the areas as planned. LFC monitors the system
frequency and tie-line flows, calculates the net change in the gener-
ation required according to the change in demand and changes the
set position of the generators within the area so as to keep the time
average of the ACE (area control error) at a low value. ACE is gen-
erally treated as controlled output of LFC. As the ACE is adjusted
to zero by the LFC, both frequency and tie-line power errors will
become zero [2].
Several control strategies for LFC of power systems have been
proposed in order to maintain the system frequency and tie-line
power flow at their scheduled values during normal and dis-
turbed conditions. In [3], a critical literature review on the LFC of
power systems has been presented. It is observed that considerable
research work is going on to propose better LFC systems based on
modern control theory [4], neural network [5], fuzzy system theory
[6], reinforcement learning [7] and ANFIS approach [8]. But, these
advanced approaches are complicated and need familiarity of users

http://dx.doi.org/10.1016/j.asoc.2015.01.020
1568-4946/© 2015 Elsevier B.V. All rights reserved.
 R.K. Sahu et al. / Applied Soft Computing 29 (2015) 310–327
to these techniques thus reducing their applicability. Alternatively,
a classical proportional integral derivative (PID) controller and its
variant remain an engineer’s preferred choice due to its structural
simplicity, reliability and the favourable ratio between perform-
ances and cost. Additionally, it also offers simplified dynamic
modelling, lower user-skill requirements and minimal develop-
ment effort, which are major issues of in engineering practice.
In recent times, new artificial intelligence-based approaches have
been proposed to optimize the PI/PID controller parameters for
LFC system. In [9], several classical controllers structures such as
Integral (I), proportional integral (PI), integral derivative (ID), PID
and integral double derivative (IDD) have been applied and their
performance has been compared for LFC system. Nanda et al. [10]
have demonstrated that bacterial foraging optimization algorithm
(BFOA) optimized controller provides better performance than
GA-based controllers and conventional controllers for an intercon-
nected power system. Ali and Abd-Elazim [11] have employed a
BFOA to optimize the PI controller parameters and shown its supe-
riority over GA in a two area non-reheat thermal system. Shabani
et al. [12] employed an imperialist competitive algorithm (ICA) to
optimize the PID controller parameters in a multiarea multi unit
power system. In [13], a modified objective function using integral
of time multiplied by absolute value of error (ITAE), damping ratio
of dominant eigenvalues and settling time is proposed where the
PI controller parameters are optimized employed differential evo-
lution (DE) algorithm and the results are compared with BFOA and
GA optimized ITAE-based PI controller to show its superiority.
Recently, new hybrid/modified optimization techniques have
been proposed in the literature. A self-organizing migration algo-
rithm (SOMA) has been proposed in [14] for online tuning of an
actively compensated Langmuir probe system. The results are com-
pared with simulated annealing (SA) and differential evolution (DE)
algorithm to show the superiority of SOMA. In [15], support vector
regression (SVR) model with chaotic GA (SVRCGA) has been pro-
posed to forecast the tourism demands where the superiority of
SVRCGA model is demonstrated by comparing the results of other
approaches reported in the literature. In [16], an enhanced cultural
algorithms by diversified social networks has been employed to
resolve complex mechanical design optimization problems and is
compared to other well known algorithms from literature to illus-
trate its efficiency. An improved algorithm based on particle swarm
optimization (PSO) by trying to increase the ability of local search
around optimum with focusing on best found peak in each environ-
ment is proposed for optimization in dynamic environments [17].
The results of the proposed approach are evaluated on moving peak
benchmarks and are compared with results of several state of the
art algorithms to show its advantage. A cat swarm optimization
(CSO) algorithm is applied in [18] to determine the best optimal
impulse response coefficients of FIR filters, while trying to meet the
respective ideal frequency response characteristics. The results of
the CSO-based approach have been compared to those of other well
known optimization methods such as real coded genetic algorithm
(RCGA), standard PSO and DE to verify its superiority.
It obvious from literature survey that the performance of
the power system not only depends on the artificial tech-
niques employed, but also on the controller structure and chosen
objective function. Hence, proposing and implementing new
high-performance heuristic optimization algorithms to real world
problems are always welcome. Gravitational search algorithm
(GSA) is a newly developed heuristic optimization method based on
the law of gravity and mass interactions [19]. It has been reported
in the literature that GSA is more efficient in terms of CPU time and
offers higher precision with more consistent results [20]. However,
the success of GSA in solving a specific problem crucially depends
on appropriately choosing its control parameter values namely
gravitational constant G0 , population size NP, number of iteration T
311
and constant ˛. It is highly desirable to determine appropriate con-
trol parameter values which are often problem-dependent. The key
to achieving high performance for any meta-heuristic algorithm is
to maintain a good balance between exploitation and exploration
during search. GSA being a global optimizing method is designed
to explore the search space and most likely gives an optimal/near-
optimal solution if used alone. On the other hand, local optimizing
methods like pattern search (PS) are designed to exploit a local
area, but they are usually not good at exploring wide area and
hence not applied alone for global optimization [21,22]. Due to
their respective strength and weakness, there is motivation for the
hybridization of GSA and PS.
In a PID controller, the derivative mode improves stability of
the system and increases speed of the controller response but it
produces unreasonable size control inputs to the plant. Also, any
noise in the control input signal will result in large plant input sig-
nals which often lead to complications in practical applications.
The practical solution to these problems is to put a first filter on
the derivative term and tune its pole so that the chattering due to
the noise does not occur since it attenuates high frequency noise.
In view of the above, an attempt has been made in the present
paper for the optimal design of hybrid GSA and PS (hGSA-PS)-based
PI/PIDF controller for LFC in a multi-area interconnected power
system.
The aim of the present work is
(i) to study the effect of objective function of the system perfor-
mance
(ii) to tune the control parameters of GSA
(iii) to demonstrate the advantages of proposed hGSA-PS tech-
nique over other techniques such as GSA, FA, PSO, hBFOA-PSO,
NSGA-II, DE, BFOA and GA for the similar problem
(iv) to show advantages of using a modified controller structure
and objective function to further increase the performance of
the power system
(v) to study the effect of the physical constraints such as time
delay, generation rate constraints and governor dead band
nonlinearity on the system performance and design the con-
trollers for this conditions.
2. Materials and methods
2.1. Modelling of power system for LFC studies
Load frequency control (LFC) provides the control only during
normal changes in load which are small and slow. So the nonlin-
ear equations which describe the dynamic behaviour of the system
can be linearized around an operating point during these small load
changes and a linear incremental model can be used for the anal-
ysis, thus making the analysis simpler. The linear model of LFC
for an interconnected power system is presented in this section.
Each area of the power system consists of speed governing sys-
tem, hydraulic valve actuator (governor), turbine, generator and
load as shown in Fig. 1. To simplify the frequency-domain analy-
ses, transfer functions are used to model each component of the
area [2].
The speed governing system has two inputs reference power
setting Pref and frequency F and one output, i.e. governor output
command PG (s) given by
1
PG (s) = Pref (s) − F(s) (1)
R
The governor valve is represented by transfer function as
PV (s) 1
GG (s) = = (2)
PG (s) 1 + sTg
312 R.K. Sahu et al. / Applied Soft Computing 29 (2015) 310–327

Fig. 1. Block diagram representation of one-area of interconnected power system.

The transfer function of turbine is given by
PT (s) 1
GT (s) = = (3)
PV (s) 1 + sTt
The generator and load is represented by the transfer function:
KP
GP (s) =
1 + sTP
where KP = 1/D and TP = 2H/FD.
Each area has three inputs and two outputs. The inputs are the
controller input Pref (also denoted as u), load disturbance PD and
tie-line power error Ptie . The outputs are the generator frequency
F and area control error (ACE) given by equation (5).
ACE = BF + Ptie
where B is the frequency bias parameter.
The generator load system has two inputs PT (s) and PD (s)
with one out put F(s) given by
F(s) = GP (s)[PT (s) − PD (s)]
The generator always adjusts its output so as to meet the power
demand. In normal steady state, the turbine power PT keeps balance
with the total power demand PD resulting in zero acceleration and a
constant speed or frequency. Perturbations PT and PD in these
powers will upset the above balance and if the difference power
PT − PD is positive, then the generator rotor unit will accelerate
and vice versa. The turbine power increment PT depends on the
valve power increment PV which in turn depends on the governor
output command PG .
(4) 2.2. System under study
The system under investigation consists of two-area intercon-
nected power system of non-reheat thermal plant as shown in
Fig. 2. Each area has a rating of 2000 MW with a nominal load of
1000 MW. The system is widely used in literature is for the design
and analysis of automatic load frequency control of interconnected
(5) areas [11,13]. In Fig. 2, B1 and B2 are the frequency bias parameters;
ACE1 and ACE2 are area control errors; u1 and u2 are the control
outputs from the controller; R1 and R2 are the governor speed reg-
ulation parameters in pu Hz; Tg1 and Tg2 are the speed governor
time constants in s; PG1 and PG2 are the governor output com-
mand (pu); Tt1 and Tt2 are the turbine time constant in sec; PT1
(6)
and PT2 are the change in turbine output powers; PD1 and PD2
are the load demand changes; PTie is the incremental change in
tie-line power (pu); KPS1 and KPS2 are the power system gains; TPS1
and TPS2 are the power system time constant in s; T12 is the syn-
chronizing coefficient and F1 and F2 are the system frequency
deviations in Hz. The relevant parameters are given in Appendix.

Fig. 2. Two-area thermal power system.

 R.K. Sahu et al. / Applied Soft Computing 29 (2015) 310–327 313

3. The proposed approach 3.2. Objective function

3.1. Controller structure While designing a controller, the objective function is first
defined based on the desired specifications and constraints. The
Nonlinear control systems usually pose substantially higher design of objective function to tune controller parameters is gen-
data, design, implementation, and maintenance demands than erally based on a performance index that considers the entire closed
linear control systems. Power electronics and power system loop response. Some of the realistic control specifications for load
applications employ various forms of linear controllers such frequency control (LFC) are [2]:
as proportional-integral (PI) and proportional-integral-derivative
(PID) controllers, to achieve control objectives mainly specified by (i) The frequency error should return to zero following a load
the required transient and steady-state requirements. The classi- change.
cal PID controller and its variants are preferred because of their (ii) The integral of frequency error should be minimum.
structural simplicity, reliability, and the favourable ratio between (iii) The control loop must be characterized by a sufficient degree
performance and cost. Beyond these benefits, these controllers also of stability.
offer simplified dynamic modelling, lower user-skill requirements, (iv) Under normal operating conditions, each area should carry
and minimal development effort, which are issues of substantial its own load and the power exchange between control
importance to engineering practice. A control without derivative areas following a load perturbation should maintained at its
(D) mode is used when: fast response of the system is not required, prescheduled value as quickly as possible.
large disturbances and noises are present during operation of the
process, and there are large transport delays in the system. PID To determine the optimum values of controller parameters, con-
controllers are used when stability and fast response are required. ventional objective functions are considered at the first instance.
Derivative mode improves stability of the system and enables The objective functions considered are IAE, ISE, ITSE and ITAE as
increase in proportional gain and decrease in integral gain which given below:
in turn increases speed of the controller response. However, when 
tsim
     
the input signal has sharp corners, the derivative term will produce J1 = ISE = F1  + F2  + Ptie  2 dt (10)
unreasonable size control inputs to the plant. Also, any noise in the
control input signal will result in large plant input signals. These 0

tsim
reasons often lead to complications in practical applications. The      
practical solution to the these problems is to put a first filter on the J2 = IAE = F1  + F2  + Ptie  dt (11)
derivative term and tune its pole so that the chattering due to the
0
noise does not occur since it attenuates high frequency noise [26].

tsim
In view of the above, a filter is used for the derivative term in the      
present paper. J3 = ITSE = F1  + F2  + Ptie  2 t dt (12)
The structure of PID controller with derivative filter is shown 0
in Fig. 3 where KP , KI , and KD are the proportional, integral, and 
tsim
     
derivative gains, respectively, and N is the derivative filter coeffi-
J4 = ITAE = F1  + F2  + Ptie  · t dt (13)
cient. When used as PI controller, the derivative path along with
the filter is removed from Fig. 3. The error inputs to the controllers 0
are the respective area control errors (ACE) given by In the above equations, F1 and F2 are the system frequency
deviations, Ptie is the incremental change in tie-line power, and
e1 (t) = ACE1 = B1 F1 + Ptie (7) tsim is the time range of simulation.
The problem constraints are the PI/PIDF controller parameter
e2 (t) = ACE2 = B2 F2 − Ptie (8) bounds. Therefore, the design problem can be formulated as the
following optimization problem:
The control inputs of the power system u1 and u2 are the outputs
of the controllers. The transfer function of the controller is given by Minimize J (14)
 1  Ns  Subject to
TFPID = KP + KI + KD (9) For PI controller:
s s+N
KP min ≤ KP ≤ KP max , KI min ≤ KI ≤ KI max (15)

For PIDF controller:

KP min ≤ KP ≤ KP max , KI min ≤ KI ≤ KI max ,

KD min ≤ KD ≤ KD max , Nmin ≤ N ≤ Nmax (16)

where J is the objective function (J1 , J2 , J3 , and J4 ) and KPID min and
KPID max , are the minimum and maximum value of the PI/PID control
parameters, respectively. As reported in the literature [10,13,28],
the minimum and maximum values of PID controller parameters
are chosen as −2.0 and 2.0, respectively. The range for filter coeffi-
cient N is selected as 1 and 300 [26–28].

4. Gravitational search algorithm

Gravitational search algorithm (GSA) is one of the newest

Fig. 3. Structure of PID controller with derivative filter. heuristic algorithms inspired by the Newtonian laws of gravity and
314 R.K. Sahu et al. / Applied Soft Computing 29 (2015) 310–327

motion [20]. In GSA, agents are considered as objects and their per- The acceleration of the agent i at the time t and in the direction
formance is measured by their masses. All these objects attract each dth, adi (t) is given by the law of the motion as
other by the force of gravity and this force causes a global move-
ment of all objects towards the objects with a heavier mass. Hence Fid (t)
masses co-operate using a direct form of communication through adi (t) = (21)
Mii (t)
gravitational force. The heavy masses which correspond to good
solution move more slowly than lighter ones, this guarantees the where Mii (t) is the inertia mass of ith agent.
exploitation step of the algorithm. The velocity of an agent is updated depending on the cur-
In GSA, each mass (agent) has four specifications: position, rent velocity and acceleration. The velocity and position are
inertial mass, active gravitational mass, and passive gravitational updated as
mass. The position of the mass corresponds to a solution of
the problem and its gravitational and inertia masses are deter- vdi (t + 1) = randi ∗ vdi (t) + adi (t) (22)
mined using a fitness function. In other words, each mass presents
a solution and the algorithm is navigated by properly adjust- xid (t + 1) = xid (t) + vdi (t + 1) (23)
ing the gravitational and inertia masses. By lapse of time, it
is expected that masses be attracted by the heavier mass. This where randi is a uniform random variable in the interval (0, 1). The
mass will present an optimum solution in the search space. The random number is used to give a randomized characteristic to the
GSA could be considered as an isolated system of masses. It is search process.
like a small artificial world of masses obeying the Newtonian The gravitational constant G is initialized at the beginning. To
laws of gravitation and motion. Masses obey the following laws control the search accuracy, it is reduced with time and expressed
[19,20,23,24]. as function of the initial value (G0 ) and time t as

G(t) = G0 e−˛t/T (24)

4.1. Law of gravity
where ˛ is a constant and T is the number of iteration.
Each particle attracts every other particle and the gravita-
The masses (gravitational and inertia) are evaluated by the fit-
tional force between the two particle is directly proportional
ness function. Efficient agents are characterized by heavier masses.
to the product of their masses and inversely proportional to
Assuming the equal gravitational and inertia mass, the values of
the distance between them R. It has been reported in literature
masse are calculated using the map of fitness. The gravitational
that R provides better results than R2 in all experiment cases
and inertial masses are updated as
[19].
Mai = Mpi = Mii = Mi , i = 1, 2, . . ., n (25)
4.2. Law of motion
fiti (t) − worst(t)
mi (t) = (26)
The current velocity of any mass equals the sum of the fraction best(t) − worst(t)
of its previous velocity and the variation in the velocity. Variation
in the velocity or acceleration of any mass is equal to the force acted mi (t)
on the system divided by mass of inertia.
Mi (t) = N (27)
j=1
mj (t)
For a system with n agents (masses), the ith position of an agent
Xi is defined by where fiti (t) represents the fitness value of the agent i at time t and
best(t) is defined for a minimization problem as
Xi = (xi1 , . . ., xid , . . ., xin ) for i = 1, 2, . . ., n (17)
Best(t) = min fitj (t) (28)
where xid represents the position of ith agent in the dth dimen- j∈{1,...,n}
sion.
At a specific time t, the force acting on mass i from mass j is Worst(t) = max fitj (t) (29)
j∈(1,...,n)
defined as
Mpi (t) ∗ Maj (t) To achieve a good compromise between exploration and
Fijd (t) = G(t) (xjd (t) − xid (t)) (18) exploitation, the number of agents is reduced with lapse of Eq. (20)
Rij (t)+ ∈
and therefore a set of agents with bigger mass are used for applying
where Maj is the active gravitational mass related to agent j, Mpi is their force to the other.
the passive gravitational mass related to agent i, G(t) is the grav- The performance of GSA is improved by controlling exploration
itational constant at time t, ∈ is small constant, and Rij (t) is the and exploitation. To avoid trapping in a local optimum, GSA must
Euclidian distance between two agents i and j given by use the exploration at beginning. By lapse of iterations, exploration


must fade out and exploitation must fade in. In GSA only the Kbest
Rij (t) =
Xi (t), Xj (t)
(19) (which is a function of time, with the initial value K0 at the begin-
2
ning and decreasing with time) agents attract the others. At the
The stochastic characteristic in GSA algorithm is incorporated by beginning, all agents apply the force, and as time passes Kbest is
assuming that the total forces that act on agent i in a dimension d decreased linearly and at the end there is just one agent applying
be a randomly weight sum of dth components of the forces exerted force to the others. Therefore, Eq. (20) is modified as
from other agents as


n Fid (t) = randj Fijd (t) (30)
Fid (t) = randj Fijd (t) (20) j∈Kbest,j =
/ i

j=1,j =
/ i
where Kbest is the set of first K agents with the best fitness value
where randj is a random number in the interval [0,1]. and biggest mass k.
 R.K. Sahu et al. / Applied Soft Computing 29 (2015) 310–327
The different steps of the GSA are the followings:
i. Identify the search space of parameters to be searched.
ii. Initialize the variables.
iii. Evaluate the fitness of each agent.
iv. Update G(t), best(t), worst(t), and Mi (t) for i = 1,2, . . ., n.
v. Calculate the total force in various directions.
vi. Calculate the acceleration and velocity.
vii. Update the position of the agents.
viii. Repeat steps (iii) to (vii) until the stop criteria is reached.
ix. End
GSA is characterized as a simple concept which is easy to
implement and computationally efficient. In order to improve
exploration and exploitation capabilities, GSA has a flexible and
balanced mechanism. More precise search is achieved by assum-
ing a higher inertia mass which causes a slower motion of agents
in the search space. Faster convergence is obtained by consider-
ing a higher gravitational mass which causes a higher attraction
of agents. GSA is a memory-less algorithm but works power-
fully like the other memory-based algorithms. The nature inspired
population-based techniques have proved themselves to be effec-
tive solutions to optimization problems control parameters and
objective function are involved in these optimization techniques,
and appropriate selection of these is a key point for success. It has
been reported that GSA tends to find the global optimum faster
than other algorithms and has a higher convergence rate for uni-
modal high-dimensional functions. The performance of GSA for
multi-modal functions is comparable to other algorithms [19].
5. Pattern search algorithm
The pattern search (PS) optimization technique is a deriva-
tive free evolutionary algorithm suitable to solve a variety of
optimization problems that lie outside the scope of the standard
optimization methods. It is simple in concept, easy to implement,
and computationally efficient. It possesses a flexible and well-
balanced operator to enhance and adapt the global search and fine
tune local search [25]. The PS algorithm computes a sequence of
points that may or may not approaches to the optimal point. The
algorithm starts with a set of points called mesh, around the ini-
tial points. The initial points or current points are provided by the
GSA technique. The mesh is created by adding the current point to
a scalar multiple of a set of vectors called a pattern. If a point in
the mesh is having better objective function value, it becomes the
current point at the next iteration [22].
The PS method consists of a sequence of polls or iterations xk ,
k ∈ N. A number of trial steps ski with i = 1, 2, . . ., p are added to
the iteration xk to obtain a number of trial points xki = xk + ski at
each iteration k. The objective function J to be optimized is eval-
uated on these trial points through a series of exploratory moves,
which describe a method in which the trial points are evaluated,
and the obtained values are compared with J(xk ). Then, the trial
step sk∗ corresponding to minimum value of J(xk + ski ) − J(xk ) < 0
is selected to produce the subsequent estimation of the patterns
polls xk+1 = xk + sk∗ . The trial steps ski are produced by a step length
parameter k ∈ R+ n which is also updated through time depending
on the value of xk+1 . The evolution of k establishes the conver-
gence properties of the algorithm. These elements are explained in
more details below.
5.1. The patterns
The aim of a pattern is to provide a direction to produce the new
trial points. A pattern is defined by two elements: a basis matrix B
315
and a generating matrix Ck . These matrixes are used to find out the
possible directions of the trial steps which will shape the next pat-
tern. The basis matrix can be any non-singular matrix B ∈ Rn×n . The
generating matrix is a matrix Ck ∈ Zn×p , where p > 2n and essentially
contains the directions in the space Rn on which the trial points are
to be produced. The structure of matrix Ck is given by
. .
Ck = [Mk .. − Mk ..Lk ] (31)
Generally, the matrix Ck is composed of a set of search direc-
tions defined as the columns of a matrix Mk ∈ Zn×n along with its
opposites, −Mk and a matrix Lk ∈ Zn×(p−2n) . Lk ensures that the orig-
inal iterate xk is also a possible trial point. The size of Mk and Lk are
determined in function on the size of the domain of J, i.e. n.
A pattern is then defined by the columns of the matrix Pk = BCk .
The length of the direction vectors is adjusted through the step size
parameter k . The final trial step is formed by the columns of the
following equation:
ski = k Bcki (32)
p
where cki denotes a column of Ck = [ck1 . . .ck ]. k provides as a step
length parameter as it acts to adjust the length of ski . At iteration k,
it defines a sequence of trial point xk + ski with i = 1, 2, . . ., p as any
point of the form J(xki ) = J(xk + ski ), where xk is the current iterate.
5.2. The exploratory moves
PS methods progress by calculating the value of the objective
function J on each trial point before computing xk+1 . The choice of
exploratory moves must ensure two things to allow the broadest
possible choice of exploratory moves.
The direction of any step sk accepted at iteration k is defined by
the pattern Pk and its length is determined by k . Mathematically,
this can be expressed as
sk = k Pk ≡ k BCk (33)
A legitimate exploratory moves algorithm would be one that
by some means guesses which of the steps defined by k Pk will
produce a less value and then calculates the function at only one
such step. At worst case, a legitimate exploratory moves algorithm
is one that evaluates all sk modifications and returns the step that
produced the least function value. This step trial is then used to gen-
erate the new estimate xk+1 . This can be expressed mathematically
as
If 0 ≤ min{f (xk + ski ), ski ∈ k [Mk, − Mk ]} < f (xk ), then xk+1
= xk + sk∗ (34)
This process repeated until the minimum is reached.
A PS method has the following operation sequence. First, the
objective function is evaluated in each trial points xk + ski , and the
value sk∗ associated with the minimum of k = f (xk + ski ) − f (xk ) is
chosen subject to k < 0. Otherwise, xk+1 = xk . Finally, the PS pro-
ceeds to update the values of Ck and k to produce a new series
of exploratory moves. The aim of the updating for k is to force
ki = f (xk + ski ) − f (xk ) < 0. A poll with k < 0 is successful; other-
wise, the poll is unsuccessful.
6. Results and discussions
6.1. Implementation of hGSA-PS algorithm
The model of the system under study shown in Fig. 2 is devel-
oped in MATLAB/SIMULINK environment and GSA program is
316 R.K. Sahu et al. / Applied Soft Computing 29 (2015) 310–327

Table 1
Tuned controller parameters and performance index for each objective function.

Objective function Controller parameters TS (s) Peak overshoot 

Proportional gain (KP ) Integral gain (KI ) F1 F2 Ptie F1 F2 Ptie

J1 : ISE −0.0120 0.8641 21.30 21.30 15.10 0.0594 0.0877 0.0152 0.0626
J2 : IAE −0.0117 0.7668 18.00 18.00 12.90 0.0450 0.0710 0.0119 0.0733
J3 : ITSE −0.1228 0.7849 16.20 16.20 11.10 0.0499 0.0838 0.0127 0.0882
J4 : ITAE −0.1701 0.6492 12.00 11.90 8.90 0.0390 0.0590 0.0083 0.1144

written (in .mfile). Initially, similar PI controllers are considered for
each area. The developed model is simulated in a separate program
(by .m file using initial population/controller parameters) consid-
ering a 10% step load change in area-1. The objective function is
calculated in the .m file and used in the optimization algorithm. At
the first instance, the following parameters are chosen for the appli-
cation of GSA: population size NP = 30; maximum iteration T = 500;
gravitational constants G0 = 30 and ˛ = 10; K0 = total number of
agents and decreases linearly to 1 with time [19]. Optimization is
terminated by the prespecified number of generations. Simulations
were conducted on an Intel, core 2 Duo CPU of 2.4 GHz and 2 GB MB
RAM computer in the MATLAB 7.10.0.499 (R2010a) environment.
The optimization was repeated 50 times and the best final solution
among the 50 runs is chosen as final controller parameters. The best
final solutions obtained in the 50 runs for each objective function
are shown in Table 1. To investigate the effect of objective function
on the dynamic performance of the system, settling times (2% of
final value) and peak overshoots in frequency and tie-line power
deviations along with minimum damping ratios are also provided
in Table 1. It can be seen from Table 1 that best system perfor-
mance is obtained in terms of maximum value of damping ratio of
system modes and minimum settling times and peak overshoots
in frequency and tie-line power deviations when ITAE is used as
objective function.
The success of GSA is heavily dependent on setting of control
parameters, namely constant ˛, initial gravitational constant G0 ,
population size NP, and number of iterations T. While applying
GSA, these control parameters should be carefully chosen for the
successful implementation of the algorithm. A series of experi-
ments were conducted to properly tune the GSA control parameters
in order to optimize the PI parameters employing ITAE objective
function. Table 2 shows the GSA outcomes as a result of varying
its control parameters. To quantify the results, 50 independent
runs were executed for each parameter variation. It is clear from
the results shown in Table 2 that the best settings for constant
˛, gravitational constant G0 , population size NP, and number of
iterations T are ˛ = 20, G0 = 100, NP = 20, and T = 100, respectively.
Note that increasing the population size NP beyond 20 and iter-
ations T beyond 100 will improve the average, maximum, and
standard deviation values slightly (with same minimum value) at
the expense of increasing the computation time significantly.
Pattern search (PS) is then employed to fine tune the best solu-
tion provided by GSA. The final values of controller parameters
obtained by GSA algorithm are taken as the initial points for PS
algorithm. The PS is executed with a mesh size of 1, mesh expan-
sion factor of 2, and mesh contraction factor of 0.5. The maximum
number of objective function evaluations and generations are set to
10 each. The flow chart of proposed hybrid pattern search (PS) and
gravitational search (GSA) approach is shown in Fig. 4. The results of
proposed hGSA-PS algorithm over 50 independent runs are shown
in Table 3. Also better results are observed in terms of average, max-
imum, and standard deviation values with the proposed hGSA-PS
technique compared to GSA as evident from Tables 2 and 3. The
performance of proposed hGSA-PS technique is also compared with
individual GSA technique and other recently published techniques
as shown in Table 4. It should be noted here that conventional [11],
GA [11], BFOA [11], DE [13], FA [27], PSO [28], hybrid BFOA-PSO
[28], and NSGA-II [26] values correspond to same power system,
controller structure (PI), and objective function employed (ITAE)
for proper comparison of techniques.
It is clear from Table 4 that with PI structured con-
troller and ITAE objective function, minimum ITAE value is
obtained with hGSA-PS (ITAE = 609.2 × 10−3 ) compared to ITAE
values with GSA (ITAE = 665.9 × 10−3 ), FA (ITAE = 869.5 × 10−3 ),

Table 2
Simulation results over 50 independent runs of tuning GSA parameters.

Parameter Min (×10−3 ) Ave (×10−3 ) Max (×10−3 ) St.Dev (×10−3 ) Other parameters

˛ = 10 665.9897 898.0626 1012.3918 126.5678

˛ = 15 665.9897 784.1139 1000.1651 126.7459
˛ = 20 665.9897 729.1873 918.2768 80.2699 NP = 20, T = 50, G0 = 100
˛ = 25 674.9712 900.0713 975.4091 122.8549
˛ = 30 687.0195 916.4638 1012.3918 137.5834

G0 = 30 665.9945 804.6012 1012.3918 118.4529

G0 = 70 665.9988 851.2434 1022.5763 128.2661
G0 = 100 665.9897 729.1873 918.2768 80.2699 NP = 20, T = 50, ˛ = 20
G0 = 130 668.0531 779.0509 1022.5763 105.5190
G0 = 150 667.8854 799.1217 1003.5452 198.8937

NP = 10 688.1325 823.1840 1031.5944 198.8937

NP = 15 670.6229 801.9416 942.6240 125.0767
NP = 20 665.9897 729.1873 918.2768 80.2699 T = 50, ˛ = 20, G0 = 100
NP = 25 665.9897 728.0129 914.6560 70.2121
NP = 30 665.9897 727.9858 913.5081 67.9424

T = 30 665.0960 878.1403 974.2163 104.0270

T = 50 665.9897 729.1873 918.2768 80.2699
NP = 20, G0 = 100, ˛ = 20
T = 100 665.9897 677.4539 873.9539 57.1679
T = 200 665.9897 676.4037 873.9539 57.0019
 R.K. Sahu et al. / Applied Soft Computing 29 (2015) 310–327 317

Inilitize A

Specify the GSA parameters

Specify the PS paremeters
Iter. =1 Read the initial values from GSA

Evalute the fitness of each agent by Iter. = 0

time-domain simulation

Yes
Update G, best and worst Is termination conditions Stop
reached?
Iter. = Iter. +1 Iter. = Iter. +1
Calculate the mass and acceleration for No
each agent
Set the starting point

Update the velocity and position of

agents using Eq. (22) and (23)
Construct Pattern vectors and create mesh
points
Iter. = Iter.+1

Stoping
Evalute the mesh points
No criteria reached?

Yes Yes No
Expand the Contract the
Is the poll successful?
mesh size mesh size
A
Fig. 4. Flow chart of proposed hGSA-PS optimization approach.

Table 3
Simulation results over 50 independent runs of hybrid GSA-PS.

Min. (×10−3 ) Ave. (×10−3 ) Max. (×10−3 ) St.Dev. (×10−3 ) Algorithm parameters

609.2095 598.9347 720.9421 40.5834 Tuning GSA parameters:

Population size NP = 20, Generations T = 100, G0 = 100, ˛ = 20
PS parameters:
Mesh size = 1, Mesh expansion factor = 2
Mesh contraction factor = 0.5
Maximum number of objective function evaluations =10
Maximum number of Iterations = 10

DE (ITAE = 991.1 × 10−3 ), hybrid BFOA-PSO (ITAE = 1186.5 × 10−3 ),
PSO (ITAE = 1214.2 × 10−3 ), NSGA-II (ITAE = 1676.4 × 10−3 ), BFOA
(ITAE = 1827.0 × 10−3 ), GA (ITAE = 2747.4 × 10−3 ), and conventional
(ITAE = 3756.8 × 10−3 ) techniques. Hence it can be concluded that
for the similar controller structure (PI) and same objective function
(ITAE), hGSA-PS outperforms GSA, FA, DE, hybrid BFOA-PSO, PSO,
NSGA-II, BFOA, GA, and conventional techniques.
6.2. Modified objective function and controller structure

All the above analyses are performed for the system under
Table 4
Tuned controller parameter and error with ITAE objective function. study by applying a 10% step load increase in area-1 and optimi-
zing the PI controller parameters employing an ITAE criterion. It
Technique Tuned controller parameter ITAE (×10−3 )
is worthwhile to mention that in this initial study less effective PI
KP KI controllers are considered for a better illustration of capabilities of
Conventional PI [11] −0.3317 0.4741 3756.8 proposed hGSA-PS technique. To further improve the performance
GA PI [11] −0.2346 0.2662 2747.4 of the system, modifications in the objective function and con-
BFOA PI [11] −0.4207 0.2795 1827.0 troller structure are introduced [13]. A modified objective function
DE PI [13] −0.2146 0.4335 991.1 given by Eq. (31) is used and proportional integral (PI) and propor-
FA PI [27] −0.3267 0.4296 869.5
PSO tuned PI [28] −0.3597 0.4756 1214.2
tional integral derivative controller with derivative filter (PIDF) are
hBFOA-PSO tuned PI [28] −0.3317 0.4741 1186.5 considered.
NSGA-II [26] −0.428 0.2967 1676.4 The modified objective function J5 tries to minimize the ITAE
GSA PI −0.1880 0.6179 665.9 error, maximizes the minimum damping ratios of dominant eigen-
hGSA-PS PI −0.2034 0.5621 609.2
values, and minimizes the settling times of F1 , F2 , and Ptie as
318 R.K. Sahu et al. / Applied Soft Computing 29 (2015) 310–327

Table 5
Performance index with modified objective function.

Technique Tuned controller parameter Settling time Ts (s)  ITAE

KP KI KD N F1 F2 Ptie

DE: PI [13] −0.4233 0.2879 – – 5.38 6.95 6.21 0.2361 1.6766

hGSA-PS: PI −0.4313 0.2986 – – 4.79 6.75 6.04 0.2384 1.5189
hGSA-PS: PIDF 1.3842 1.9897 0.6573 225.1048 1.85 3.17 2.81 0.4468 0.1319

given by Eq. (35):

12 g

tsim 
tsim h h
      g
J5 = ω1 · F1  + F2  + Ptie  t dt 10 g
f
0 0 f

Settling Time(S)
8 e e h
1 d
b c
+ ω2 · n + ω3 (ST) (35)
min( i=1
(1 − i ))
where F1 and F2 are the system frequency deviations, Ptie is
the incremental change in tie-line power, tsim is the time range
of simulation,  i is the damping ratio, n is the total number of
the dominant eigen values, ST is the sum of the settling times of
frequency and tie-line power deviations, respectively, and ω1 to
ω3 are weighting factors. Inclusion of appropriate weighting fac-
tors to the right-hand individual terms helps to make each term
competitive during the optimization process. Wrong choice of the
weighting factors leads to incompatible numerical values of each
term involved in the definition of fitness function which gives mis-
leading result. The weights are so chosen that numerical value of
all the terms in the right hand side of Eq. (35) lie in the same range.
Repetitive trial runs of the optimizing algorithms are executed with
both PI and PIDF controller to select the weights. To make each term
competitive during the optimization process the following weights
are chosen: PI controller: ω1 = 1.0, ω2 = 0.3, and ω3 = 0.08 and PIDF
controller: ω1 = 1.0, ω2 = 0.05, and ω3 = 0.02.
A 10% step load change in area-1 is considered at t = 0 s and the
proposed hGSA-PS algorithm is applied 50 times. The final PI/PIDF
controller parameters corresponding to the minimum objective
function values along with the system performance in terms of
settling time, minimum damping ratio, and ITAE value are shown
in Table 5. For comparison, the system performance with DE opti-
mized PI [13] is also provided in Table 5. It is evident from Table 5
that for similar power system and controller structure employing
same objective function, better system performance is obtained
with proposed hGSA-PS optimized PI/PIDF controller compared
to DE optimized PI controller [3]. The ITAE value is reduced by
9.4 and 92.13% with proposed hGSA-PS optimized PI and PIDF
controllers, respectively, compared to DE optimized PI controller.
0.05
-0.05
ΔF1(Hz)
-0.1
-0.15
-0.2
-0.25
0 1 2 3 4 5
Time (S)
Fig. 6. Change in frequency of area-1 for 10% change in area-1.
b c d e f
6
c d
b
4
a
a
2 a
0
ΔF1 ΔF1 ΔPtie
Fig. 5. Comparison of settling time: (a) hGSA-PS PIDF: J5; (b) hGSA-PS PI: J5; (c) DE
PI: J5 ; (d) BFOA PI: J4; (e) FA PI: J4; (f) DE PI: J4; (g) GSA PI: J4; (h) GA PI: J4.
Also, the settling time for F1 is improved by 10.96 and 65.61% for
proposed hGSA-PS optimized PI and PIDF controllers, respectively,
compared to the results given in [13]. The improvements in settling
time for F2 are 2.87 and 54.38%, respectively, with proposed
PI and PIDF controllers. For the tie-line power deviations, Ptie
the improvements with PI/PIDF controllers are 2.73 and 54.75%,
respectively, compared to the DE optimized PI controller. It should
be noted here that DE values correspond to same power system
and objective function (J5 ) for a proper comparison. For a better
illustration of improvements, the settling times of frequency and
tie-line deviations for each case are shown in Fig. 5 from which it
is clear that minimum settling times are achieved with proposed
hGSA-PS optimized PIDF controller optimized employing modified
objective function J5 .
To study the dynamic performance of the proposed controllers,
a step increase in demand of 10% is applied at t = 0 s in area-1
and the system dynamic responses are shown in Figs. 6–8. For
comparison, the simulation results with FA optimized PI controller
(corresponding to minimum ITAE value reported in literature
and shown in Table 5) using ITAE objective function [27] and DE
Modified Obj. fun.: DE tunned PI [13]
Modified Obj. fun.: hGSA-PS tunned PI
Modified Obj. fun.: hGSA-PS tunned PIDF
ITAE Obj. fun.: FA tunned PI [27]
6 7 8 9 10
 R.K. Sahu et al. / Applied Soft Computing 29 (2015) 310–327
-0.05
Δ F2 (Hz)
-0.1
-0.15
-0.2
0 1 2 3
Fig. 7. Change in frequency of area-2 for 10% change in area-1.
optimized PI controller using modified objective function [13] for
the same power system are also shown in Figs. 6–8. Critical anal-
ysis of the dynamic responses clearly reveals that better dynamic
performance is obtained with proposed hGSA-PS optimized PI
controller compared to GA, BFOA, and DE optimized PI controllers
and significant improvement is observed with PIDF controller
optimized using proposed hGSA-PS technique employing modified
objective function (J5 ).
6.3. Sensitivity analysis
Sensitivity analysis is carried out to study the robustness the
system to wide changes in the operating conditions and system
parameters [13,26–29]. Taking one at a time, the operating load
condition and time constants of speed governor, turbine, tie-line
power are changed from their nominal values (given in appendix)
in the range of +50% to −50% in steps of 25%. In all the cases, PIDF
controller optimized using proposed hGSA-PS technique employ-
ing modified objective function J5 is considered due to its superior
performance.
The optimum values of controller parameters, at changed load-
ing conditions and changed system parameters (for a step increase
in demand of 10% at t = 0 s in area-1) are provided in Table 6. The cor-
responding performance indexes (ITAE values, settling times and
minimum damping ratios) with the above varied system conditions
are also given in Table 6. As an example, the frequency deviation
response of area-1 with varied loading conditions is shown in Fig. 9.
-0.02
Δ Ptie (P.U.)
-0.04
-0.06
-0.08
0 1 2 3
Fig. 8. Change in tie-line power for 10% change in area-1.
319
Modified Obj. fun.: DE tunned PI [13]
Modified Obj. fun.: hGSA-PS tunned PI
Modified Obj. fun.: hGSA-PS tunned PIDF
ITAE Obj. fun.: FA tunned PI [27]
4 5 6 7 8 9 10
Time (S)
It can be observed from Table 6 and Fig. 9 that the effect of the vari-
ation of operating loading conditions and system time constants on
the system performance is negligible. So it can be concluded that
the proposed control strategy provides a robust control.
6.4. Inclusion of physical constraints
To get an accurate insight of the LFC topic, it is essential to
include the important inherent requirement and the basic phys-
ical constraints. In order to demonstrate the ability of the proposed
algorithm to cope with nonlinearity, the study is extended by
including the physical constraints such as reheat turbine, gover-
nor dead band (GDB) nonlinearity and generation rate constraints
(GRC) [27,30]. As most of the thermal plants are of reheat type,
reheat turbines are considered in the power system model. The
speed governor dead band has a great effect on the dynamic per-
formance of electric energy system. In thermal power plants, power
generation can change only at a specified maximum rate known as
GRC. Governor dead band (GDB) is defined as the total amount of a
continued speed change within which there is no change in valve
position. The effect of the GBD is to increase the apparent steady-
state speed regulation. The speed-governor dead band has makes
the system oscillatory. In the present study, a GRC of 3%/min and a
GDB of 0.06% [30] are considered. A describing function approach
is used to incorporate the governor dead band nonlinearity. In
this case, the modified objective function given by Eq. (35) is used
due to its superior performance. The same procedure as explained
Modified Obj. fun.: DE tunned PI [13]
Modified Obj. fun.: hGSA-PS tunned PI
Modified Obj. fun.: hGSA-PS tunned PIDF
ITAE Obj. fun.: FA tunned PI [27]
4 5 6 7 8 9 10
Time (S)
320 R.K. Sahu et al. / Applied Soft Computing 29 (2015) 310–327

Table 6
Sensitivity analysis of two-area system without physical constraints.

Parameter variation % Change Tuned controller parameter Settling time Ts (s)  ITAE

KP KI KD N F1 F2 Ptie

Nominal 0 1.3842 1.9897 0.6573 225.1048 1.85 3.17 2.81 0.4468 0.1319

+50 1.3814 1.9887 0.6564 228.3574 1.86 3.21 2.84 0.4494 0.1324
+25 1.3803 1.9896 0.6576 226.8468 1.85 3.20 2.84 0.4490 0.1321
Loading condition
−25 1.3873 1.9897 0.6587 225.2587 1.85 3.20 2.84 0.4459 0.1318
−50 1.3816 1.9876 0.6587 226.0523 1.85 3.19 2.84 0.4456 0.1317

+50 1.4362 1.9880 0.6800 227.9429 1.73 3.25 2.87 0.2907 0.1293
+25 1.3804 1.9895 0.6574 227.8123 1.77 3.19 2.83 0.3572 0.1289
Tg
−25 1.3807 1.9898 0.6552 229.5481 1.91 3.20 2.85 0.5565 0.1346
−50 1.3818 1.9893 0.6505 226.2032 1.96 3.22 2.86 0.6321 0.1373

+50 1.5701 1.9374 0.8299 225.6409 1.83 3.34 2.98 0.3547 0.1344
+25 1.5004 1.9972 0.7299 227.1837 1.80 3.26 2.88 0.3808 0.1282
Tt
−25 1.3816 1.9989 0.5145 227.7261 1.92 3.36 2.86 0.4744 0.1363
−50 1.3801 1.9983 0.5029 225.1471 2.12 3.42 2.91 0.5112 0.1470

+50 1.3904 1.9987 0.6589 228.1939 2.17 2.94 2.66 0.3034 0.1288
+25 1.3904 1.9986 0.6580 229.8030 1.95 3.05 2.73 0.3650 0.1297
T12
−25 1.3922 1.9988 0.6517 229.7053 3.46 3.46 3.01 0.5518 0.1325
−50 1.2236 1.9971 0.6500 227.7779 4.19 3.73 3.22 0.7068 0.1292

-0.02
Δ F1 (Hz)

-0.04

-0.06
+25% of nominal load
- 25% of nominal load
-0.08 +50% of nominal load
- 50% of nominal load

0 1 2 3 4 5 6 7 8 9 10
Time (S)
Fig. 9. Change in frequency of area-1 for change in nominal load.

earlier is followed to optimize the PI/PIDF controller parameters
and the results are provided in Table 7. The performance of pro-
posed hGSA-PS tuned PIDF controller optimized using the modified
objective function is compared with GSA and hGSA-PS optimized
PI controllers for ITAE and modified objective functions as shown
in Table 7. It is clear from Table 7 that with PI structured con-
troller and ITAE objective function, minimum ITAE value is obtained
with hGSA-PS (ITAE = 52.7864) compared to ITAE values with GSA
(ITAE = 59.9091) technique. Hence it can be concluded that for the
similar controller structure (PI) and same objective function (ITAE),
hGSA-PS outperforms GSA technique. The corresponding perfor-
mance index in terms of minimum damping ratio and settling
times in frequency and tie-line power deviations are also provided
in Table 7. Further, it can be seen from Table 7 that better sys-
tem performance is obtained with proposed hGSA-PS optimized
PI controller employing the modified objective function compared
to the conventional ITAE objective function. It is also evident from
Table 7 that system performance is greatly improved with proposed
hGSA-PS optimized PIDF controller compared to proposed hGSA-PS
optimized PI controller. The ITAE value is reduced by 26.09% with
proposed hGSA-PS optimized PIDF controller compared to hGSA-PS
optimized PI controller. Also, the settling time for F1 is improved
by 36.86% for proposed hGSA-PS optimized PIDF controllers com-
pared to hGSA-PS optimized PI controller. The improvements in
settling time for F2 is 33.96% for proposed hGSA-PS optimized
PIDF controllers compared to hGSA-PS optimized PI controller. For

Table 7
Tuned controller parameter and performance index using different technique/controller/objective functions for two-area system with physical constraints.

Technique/controller Objective function Tuned controller parameter Settling time Ts (s)  ITAE

KP KI KD N F1 F2 Ptie

GSA: PI ITAE −0.0546 0.1051 – – 30.77 30.69 50.36 0.0997 59.9091

hGSA-PS: PI ITAE −0.0828 0.1025 – – 30.71 30.07 49.67 0.1025 52.7864
hGSA-PS: PI Modified −0.3248 0.1403 – – 28.20 29.48 41.23 0.1244 53.4282
hGSAPS: PIDF Modified 0.8538 0.0774 1.9941 54.7980 17.86 19.47 38.89 0.4651 39.4879
 R.K. Sahu et al. / Applied Soft Computing 29 (2015) 310–327 321

0.4

0.2

ΔF1(Hz)
0

-0.2
Modified Obj. function : hGSA-PS: PIDF
-0.4 Modified Obj. function: hGSA-PS: PI
ITAE Obj. function: hGSA-PS: PI
-0.6 ITAE Obj. function : GSA: PI

0 5 10 15 20 25 30 35 40
Time (S)
Fig. 10. Change in frequency of area-1 for 10% change in area-1.

0.4

0.2
Δ F2 (Hz)

-0.2
Modified Obj. function : hGSA-PS: PIDF
-0.4 Modified Obj. function : hGSA-PS: PI
ITAE Obj. funct ion : hGSA-PS: PI
-0.6 ITAE Obj. function: GSA: PI

0 5 10 15 20 25 30 35 40
Time (S)
Fig. 11. Change in frequency of area-2 for 10% change in area-1.

the tie-line power deviations Ptie the improvements with PIDF
controllers is 5.67% compared to the hGSA-PS optimized PI con-
troller.
To study the dynamic performance of the proposed hGSA-PS
optimization technique, a step increase in demand of 10% is applied
at t = 0 s in area-1 and the system dynamic responses are shown
in Figs. 10–12. Critical analysis of the dynamic responses clearly
reveals that better dynamic performance is observed with hGSA-PS
tuned PIDF controller using proposed modified objective func-
tion. Sensitivity analysis is carried out as before and the results
are provided in Table 8. The corresponding performance indexes
(ITAE values, settling times and minimum damping ratios) with
the above varied system conditions are also given in Table 8. As an
example, the frequency deviation response of area-1 with varied
loading conditions is shown in Fig. 13. Comparison of Tables 6 and 8
and Figs. 9 and 13 clearly reveals that the system performance

0.02

-0.02
Δ Ptie (P.U.)

-0.04

-0.06
Modified Obj. function : hGSA-PS: PIDF
Modified Obj. function : hGSA-PS: PI
-0.08 ITAE Obj. function : hGSA-PS: PI
ITAE Obj. function: GSA: PI
-0.1
0 10 20 30 40 50 60 70
Time (S)
Fig. 12. Change in tie-line power for 10% change in area-1.
322 R.K. Sahu et al. / Applied Soft Computing 29 (2015) 310–327

0.2

Δ F1 (Hz)
-0.2
Nominal loading
-0.4 +25% of nominal loading
-25% of nominal loading
+50% of nominal loading
-0.6 - 50% of nominal loading

0 5 10 15 20 25 30 35 40
Time (S)
Fig. 13. Frequency deviation of area-1 for change in nominal load.

degrades when physical constraints are included in the system
model.
6.5. Extension to nonlinear three unequal-area power system
To demonstrate the ability of the proposed algorithm to cope
with unequal areas with different controller parameters, the study
is extended to a three-area system [10,31]. The relevant param-
eters are given in Appendix. Owing to the growing complexity
of power systems in deregulated environment, communication
delays become a significant challenge in the LFC analysis. Time
delays can degrade a system’s performance and even cause sys-
tem instability. In view of the above, a time delay of 50 ms [32]
is considered in the present study in addition to reheat turbine,
GRC and GDB. The MATLAB/SIMULINK-based model of three-area
hydro-thermal system under study is shown in Fig. 14. The PI/PIDF
controller parameters are obtained for each area by applying the
proposed hGSA-PS technique employing the modified objective
function (J5 ). To make each term competitive during the opti-
mization process, the following weights are chosen: PI controller:
ω1 = 1.0, ω2 = 0.01, and ω3 = 0.05 and PIDF controller: ω1 = 1.0,
ω2 = 0.1, and ω3 = 0.01. A 1% step increase in load demand in area-1
is applied at t = 0.0 s and the optimum values of PI/PIDF controller
parameters are shown in Table 9. The corresponding performance
index in terms of ITAE value, minimum damping ratio, and sett-
ling times in frequency and tie-line power deviations is shown in
Table 10 from which it is evident that best system performance is
obtained with proposed hGSA-PS optimized PIDF controller. The
dynamic performance of the system with proposed hGSA-PS opti-
mized PI/PIDF controller employing the modified objective function

Fig. 14. MATLAB/SIMULINK model of three-area unequal and nonlinear power system.
 R.K. Sahu et al. / Applied Soft Computing 29 (2015) 310–327 323

Table 8
Sensitivity analysis of two-area system with physical constraints.

Parameter variation % Change Tuned controller parameter Settling time Ts (s)  ITAE

KP KI KD N F1 F2 Ptie

Nominal 0 0.8538 0.0774 1.9941 54.7980 17.86 19.47 38.89 0.4651 39.4879

+50 0.8626 0.0718 1.9907 53.8862 16.44 15.19 41.21 0.4641 30.4745
+25 0.8556 0.0746 1.9964 53.1229 15.24 14.71 40.26 0.4615 34.0896
Loading condition
−25 0.8544 0.0792 1.9986 132.9548 19.71 21.13 36.98 0.5303 44.6948
−50 0.8529 0.0801 1.9955 194.2783 21.29 21.50 35.87 0.5448 50.4185

+50 0.8533 0.0789 1.9957 54.6498 20.51 21.41 37. 36 0.3586 46.9176
+25 0.8524 0.0804 1.9974 53.1851 19.43 20.36 36.68 0. 4015 43.4225
Tg
−25 0.8500 0.0759 1.9904 52.2137 16.30 15.36 40.61 0.5445 36.2472
−50 0.8528 0.0716 1.9979 52.1021 17.72 16.29 41.11 0.6754 33.9986

+50 0.8517 0.0775 1.9917 54.0526 17.94 19.50 38.64 0.5110 39.6111
+25 0.8508 0.0766 1.9964 54.3016 17.91 19.59 39.46 0.4976 39.4111
Tt
−25 0.8502 0.0775 1.9971 52.2364 17.67 19.30 38.84 0.4157 39.4194
−50 0.8507 0.0780 1.9988 52.7138 17.44 19.10 38.51 0.3678 39.3853

+50 0.8537 0.0789 1.9944 54.3057 20.67 21.20 40.68 0.4329 45.6643
+25 0.8611 0.0760 1.9981 133.4022 19.28 20.16 41.32 0.4879 40.8882
T12
−25 0.8663 0.0703 1.9919 153.7261 30.82 15.22 42.34 0.5682 32.3384
−50 0.8639 0.0703 1.9914 144.4516 17.86 15.34 40.51 0.5871 32.3134

Table 9 J5 is shown in Fig. 15 for the above disturbance. It is clear from

Controller parameters with modified objective function for three area power system.
Fig. 15a–f that proposed hGSA-PS optimized PIDF controller out-
hGSAPS: PI hGSAPS: PIDF performs the PI controller in terms of settling times and overshoots
KP1 0.1535 0.1297 in frequency and tie-line power deviations. To study the effect of
KP2 0.9489 0.8660 location of disturbance on the system performance, a simultaneous
KP3 −1.2751 −1.2828 step increase in load of 1% in area-1 and area-3 are considered
KI1 −0.2950 −0.6539 and the system dynamic response is shown in Fig. 16. It is clear
KI2 −1.6073 −1.2491
from Fig. 16a–f that when double disturbance is applied the system
KI3 −0.7871 −0.7123
KD1 − −0.3846 becomes highly oscillatory with PI controller whose parameters
KD2 − 0.1943 were obtained considering single disturbance only. However, the
KD3 − −1.8050 performance of the system with proposed hGSA-PS optimized PIDF
N1 − 170.2923
controller is hardly affected by type of disturbance (single or dou-
N2 − 89.6953
N3 − 80.0355 ble) and the PIDF controller parameters which were obtained for
single disturbance are effective for double disturbance also. Hence,
it can be concluded that the designed PIDF controllers are robust
and perform satisfactorily irrespective of location of load disturb-
ance.

Table 10
Performance index with hGSAPS optimized PI/PIDF controllers for three-area power system.

Technique Settling time Ts (s)  ITAE

F1 F2 F3 Ptie-12 Ptie-13 Ptie-23

hGSAPS: PI 14.66 13.65 13.67 14.14 11.23 12.09 0.0028 4.6435

hGSAPS: PIDF 8.26 8.16 9.37 11.84 7.47 9.40 0.1563 0.6508

Table 11
Tuned parameters with varied conditions (loading, Tg , Tt ) for three-area power system.

Loading condition Tg Tt

+50% +25% −25% −50% +50% +25% −25% −50% +50% +25% −25% −50%

KP1 0.1236 0.1208 0.1272 0.1229 0.1299 0.1231 0.1222 0.1310 0.1222 0.1293 0.1238 0.1266
KP2 −0.8662 −0.8616 −0.8602 −0.8671 −0.8603 −0.8606 −0.8603 −0.8711 −0.8687 −0.8610 −0.8652 −0.8686
KP3 −1.2881 −1.2832 −1.2857 −1.2854 −1.2842 −1.2822 −1.2844 −1.2937 −1.2821 −1.2860 −1.2889 −1.2827
KI1 −0.6502 −0.6530 −0.6505 −0.6519 −0.6549 −0.6508 −0.6594 −0.6679 −0.6584 −0.6523 −0.6541 −0.6590
KI2 −1.2408 −1.2401 −1.2442 −1.2469 −1.2458 −1.2495 −1.2426 −1.2558 −1.2486 −1.2403 −1.2460 −1.2421
KI3 −0.7197 −0.7154 −0.7147 −0.7105 −0.7108 −0.7102 −0.7157 −0.7174 −0.7152 −0.7158 −0.7109 −0.7196
KD1 −0.3865 −0.3810 −0.3802 −0.3818 −0.3866 −0.3811 −0.3836 −0.3841 −0.3848 −0.3884 −0.3834 −0.3841
KD2 0.1923 0.1915 0.1907 0.1905 0.1905 0.1901 0.1903 0.1917 0.1989 0.1956 0.1914 0.1923
KD3 −1.8081 −1.8126 −1.8185 −1.8177 −1.8111 −1.8043 −1.8100 −1.8154 −1.8013 −1.8168 −1.8049 −1.8164
N1 170.6101 174.2966 172.6707 174.1990 173.5601 170.0571 174.1351 164.1135 172.5473 171.0248 171.8949 171.8022
N2 86.3422 89.8711 86.8340 85.5908 87.4396 88.2122 86.2949 88.8827 88.1040 88.1063 86.3516 87.5076
N3 79.2892 80.8542 79.8197 80.0521 81.0880 80.5850 78.2294 83.5178 81.6678 78.8701 79.0780 81.7384
324 R.K. Sahu et al. / Applied Soft Computing 29 (2015) 310–327

x 10

(a) 5
0.01
(b)
0.005
0

ΔF2(Hz)
ΔF1(Hz)

-0.005
-5
-0.01
hGSA-PS: PI hGSA-PS: PI
-0.015 hGSA-PS: PIDF -10 hGSA-PS: PIDF

0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Time(S) Time(S)

x 10 x 10
5
5
(d)
(c)

ΔPtie 1-2 (P.U.)

0
0
ΔF3(Hz)

-5
-5

-10
-10
hGSA-PS: PI hGSA-PS: PI
hGSA-PS: PIDF hGSA-PS: PIDF
-15
-15
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Time(S) Time(S)

x 10 x 10
4 6
(e) (f)
3
4
ΔPtie 1-3 (P.U.)

ΔPtie 2-3 (P.U.)

2
1

0 0

-1 hGSA-PS: PI hGSA-PS: PI
hGSA-PS: PIDF -2
hGSA-PS: PIDF
-2
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Time(S) Time(S)

Fig. 15. System response for 1% step increase in load demand in area-1: (a) frequency deviation of area-1; (b) frequency deviation of area-2; (c) frequency deviation of area-3;
(d) tie-line power deviation between area-1 and area-2; (e) tie-line power deviation between area-1 and area-3; (f) tie-line power deviation between area-2 and area-3.

Finally, sensitivity analysis is done to study the robustness the settling times and minimum damping ratios) with varied condi-
system to wide changes in the operating conditions and system tions are shown in Table 14. It can be noticed from Table 14 that
parameters as before. In all the cases, PIDF controllers are consid- when physical constraints are introduced, the variations in per-
ered and the controller parameters are optimized using proposed formance index are more prominent. As an example, the frequency
hGSA-PS technique employing modified objective function J5 . The deviation response of area-1 with varied loading condition is shown
tuned controller parameters under normal and varied conditions in Fig. 17. It can be observed from Fig. 17 that the effect of the vari-
are shown in Tables 11–13. The performance index (ITAE values, ation of operating loading conditions on the system responses is

Table 12
Tuned parameters with varied conditions (Kr , Tr , H) for three-area power system.

Kr Tr H

+50% +25% −25% −50% +50% +25% −25% −50% +50% +25% −25% −50%

KP1 0.1300 0.1250 0.1290 0.1241 0.1202 0.1286 0.1292 0.1231 0.1222 0.1224 0.1287 0.1219
KP2 −0.8655 −0.8665 −0.8658 −0.8671 −0.8605 −0.8628 −0.8608 −0.8606 −0.8608 −0.8607 −0.8641 −0.8629
KP3 −1.2852 −1.2831 −1.2858 −1.2814 −1.2814 −1.2862 −1.2848 −1.2822 −1.2893 −1.2830 −1.2866 −1.2846
KI1 −0.6533 −0.6514 −0.6585 −0.6587 −0.6589 −0.6578 −0.6513 −0.6508 −0.6560 −0.6581 −0.6578 −0.6535
KI2 −1.2443 −1.2448 −1.2403 −1.2408 −1.2447 −1.2495 −1.2425 −1.2495 −1.2438 −1.2408 −1.2425 −1.2432
KI3 −0.7149 −0.7136 −0.7189 −0.7146 −0.7156 −0.7192 −0.7188 −0.7102 −0.7166 −0.7135 −0.7155 −0.7146
KD1 −0.3807 −0.3879 −0.3841 −0.3803 −0.3849 −0.3838 −0.3820 −0.3811 −0.3879 −0.3813 −0.3823 −0.3824
KD2 0.1989 0.1978 0.1904 0.1975 0.1907 0.1916 0.1912 0.1901 0.1933 0.1916 0.1901 0.1903
KD3 −1.8013 −1.8134 −1.8149 −1.8140 −1.8180 −1.8159 −1.8109 −1.8043 −1.8139 −1.8012 −1.8153 −1.8132
N1 172.1809 170.6675 170.7741 171.0726 171.4428 170.5691 171.5731 170.0571 171.0191 173.5092 170.1092 170.7940
N2 89.1331 85.1078 85.7195 88.3995 86.3452 85.7941 86.9102 88.2122 89.7936 85.4324 86.9655 89.0133
N3 79.9727 80.7992 81.0298 80.7865 80.9710 79.7791 81.9577 80.5850 81.5592 81.0839 79.2627 80.0429
 R.K. Sahu et al. / Applied Soft Computing 29 (2015) 310–327 325

0.02 0.02
(a) (b)
0.01 0.01

0
ΔF1(Hz)

ΔF2(Hz)
-0.01
-0.01
-0.02

-0.02
hGSA-PS: PI -0.03 hGSA-PS: PI
hGSA-PS: PIDF hGSA-PS: PIDF
-0.03 -0.04
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Time(S) Time(S)

0.02 0.01
(c) (d)
0.01 0.005

ΔPtie 1-2 (P.U.)

0 0
ΔF3(Hz)

-0.005
-0.01

-0.01
-0.02
hGSA-PS: PI -0.015 hGSA-PS: PI
-0.03 hGSA-PS: PIDF hGSA-PS: PIDF
-0.02
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Time(S) Time(S)

-3
x 10
6 0.015
(e) (f)
4
0.01
ΔPtie 1-3 (P.U.)

ΔPtie 2-3 (P.U.)

2
0.005

0
0
-2
hGSA-PS: PI -0.005
hGSA-PS: PI
-4 hGSA-PS: PIDF hGSA-PS: PIDF
0 10 20 30 40 50 60 70 80 90 100 -0.01
0 10 20 30 40 50 60 70 80 90 100
Time(S) Time(S)

Fig. 16. System response for simultaneous step increase in load of 1% in area-1 and area-3: (a) frequency deviation of area-1; (b) frequency deviation of area-2; (c) frequency
deviation of area-3; (d) tie-line power deviation between area-1 and area-2; (e) tie-line power deviation between area-1 and area-3; (f) tie-line power deviation between
area-2 and area-3.

Table 13
Tuned parameters with varied tie-line constant for three-area power system.

Tie-line

+50% +25% −25% −50%

KP1 0.1203 0.1297 0.1263 0.1211

KP2 −0.8683 −0.8656 −0.8656 −0.8630
KP3 −1.2876 −1.2877 −1.2854 −1.2838
KI1 −0.6572 −0.6568 −0.6539 −0.6547
KI2 −1.2482 −1.2421 −1.2442 −1.2422
KI3 −0.7176 −0.7183 −0.7114 −0.7132
KD1 −0.3812 −0.3887 −0.3829 −0.3823
KD2 0.1938 0.1904 0.1901 0.1902
KD3 −1.8006 −1.8123 −1.8187 −1.8003
N1 170.0065 171.3664 173.0463 173.9192
N2 85.0350 88.7335 89.4592 85.0932
N3 78.8389 80.4394 80.1041 82.5136
326 R.K. Sahu et al. / Applied Soft Computing 29 (2015) 310–327

Table 14
Performance index under varied conditions for three-area power system.

Parameter variation % Change Settling time Ts (s)  ITAE

F1 F2 F3 Ptie-12 Ptie-13 Ptie-23

Nominal 0 8.26 8.16 9.37 11.84 7.47 9.40 0.1563 0.6508

+50 8.63 8.74 8.81 11.95 2.78 8.08 0.1093 0.6256

+25 8.80 8.92 8.98 12.21 2.85 8.26 0.1087 0.6375
Loading condition
−25 8.70 8.80 8.86 12.11 2.81 8.19 0.1054 0.6981
−50 8.80 8.87 8.95 12.22 3.21 8.26 0.1027 0.7682

+50 8.14 8.89 8.37 12.02 3.29 8.21 0.1100 0.6583

+25 8.57 8.72 8.75 11.88 2.81 8.08 0.1083 0.6522
Tg
−25 8.74 8.82 8.90 12.17 2.80 8.20 0.1074 0.6488
−50 8.18 8.22 8.33 11.39 2.60 7.65 0.1063 0.6310

+50 8.03 8.50 8.35 11.64 3.02 7.88 0.1379 0.6593

+25 8.05 8.67 8.31 11.78 2.90 7.96 0.1472 0.6585
Tt
−25 8.63 9.64 8.78 12.10 3.07 7.98 0.0724 0.6476
−50 9.38 10.36 9.47 12.24 3.65 8.74 0.0564 0.6447

+50 11.10 12.53 11.60 9.19 4.69 7.99 0.0050 0.5365

+25 8.98 10.06 9.07 10.95 3.31 8.17 0.0482 0.3604
Kr
−25 15.95 15.81 15.97 12.75 2.93 9.10 0.1485 0.9254
−50 24.13 24.10 24.07 19.61 10.21 16.31 0.1173 1.5233

+50 9.01 9.75 9.20 14.35 2.78 8.11 0.1071 0.7108

+25 8.86 8.89 9.03 13.17 2.80 8.13 0.1061 0.6806
Tr
−25 16.68 16.53 16.74 10.50 3.17 7.90 0.1063 0.6576
−50 13.18 13.26 13.00 8.75 3.28 6.97 0.1058 0.5799

+50 8.38 8.36 8.51 11.88 5.20 7.87 0.1184 0.6774

+25 8.25 8.25 8.43 12.06 3.61 7.66 0.1183 0.6694
H
−25 8.19 8.95 8.39 11.82 2.51 7.62 0.0825 0.6338
−50 8.97 9.85 9.14 11.78 3.14 8.42 0.0403 0.6238

+50 8.09 8.35 8.27 11.75 2.88 7.53 0.1494 0.6301

+25 8.15 8.53 8.38 11.86 2.60 7.76 0.1467 0.6347
Tie-line
−25 9.74 9.67 9.88 12.19 3.73 9.02 0.0745 0.6613
−50 10.58 11.47 12.03 12.35 7.92 10.64 0.0555 0.8299

-3
x 10
2

-2
Δ F1 (Hz)

-4

-6

-8 +50% of nominal load

+25% of nominal load
-10 -25% of nominal load
-12 -50% of nominal load

0 5 10 15 20 25 30 35 40 45 50
Time (S)
Fig. 17. Frequency deviation of area-1 for 1% change in area-1 with different loadings.

negligible. Hence it can be concluded that the proposed approach
provides a robust control under wide changes in the system
loading or system parameters.
7. Conclusion
An attempt has been made for the first time to apply a hybrid
gravitational algorithm and pattern search (hGSA-PS) technique
to optimize PI and PIDF controller parameters for load frequency
control (LFC) of a multi-area interconnected power system. The
proposed technique takes the advantage of global search capabil-
ity of GSA and local search capability of PS. Initially, a tow-area
interconnected power system is considered and the parameters
of PI controllers are optimized employing various integral based
conventional objective functions using GSA. The study reveals that
better system response in terms of settling times, maximum over-
shoots, and minimum damping ratio is obtained with ITAE objective
function. Then, investigations are carried out to select the controller
parameters of GSA. It is observed that for the system under study,
the best GSA parameters are: constant ˛ = 20, gravitational constant
G0 = 100, population size NP = 20, and generation T = 100. To fine
tune the search technique, pattern search (PS) is also employed,
where the initial starting point for PS is provided by GSA. Mod-
ified objective function and controller structure are then used
 R.K. Sahu et al. / Applied Soft Computing 29 (2015) 310–327
and controller parameters are obtained using the proposed hybrid
approach. Simulation results are compared with FA. The superior-
ity of the proposed approach is demonstrated by comparing the
results with some recently published modern heuristic optimiza-
tion techniques such as FA, DE, hybrid BFOA-PSO, PSO, NSGA-II,
BFOA, and GA techniques for the same interconnected power sys-
tem. Additionally, sensitivity analysis is performed by varying the
system parameters and operating load conditions from their nom-
inal values. Also, physical constraints such as reheat turbine, GRC,
and GBD are included in the system model and the superiority
of proposed hGSA-PS technique is over GSA technique is demon-
strated. Further, the ability of the proposed control strategy to cope
with nonlinear and unequal interconnected areas with different
controller coefficients is demonstrated by extending the study to
a nonlinear three unequal area power system. It is observed that
the proposed approach ensures satisfactory system performance
in presence of physical constraints.
Appendix A.
Two area system [11,13,26–28]:
PR = 2000 MW (rating), PL = 1000 MW (nominal loading);
f = 60 Hz, B1 , B2 = 0.045 pu/Hz; R1 = R2 = 2.4 Hz/pu; Tg1 = Tg2 = 0.08 s;
Tt1 = Tt1 = 0.3 s; KPS1 = KPS1 = 120 Hz/pu; TPS1 = TPS2 = 20 s;
T12 = 0.545 pu; a12 = −1, Kr1 = Kr2 = 0.5, Tr1 = Tr2 = 10.
Three are unequal and nonlinear power system [10,31]:
f = 60 Hz; B1 = 0.3483, B2 = 0.3827, B3 = 0.3692 pu Hz;
D1 = D3 = 0.015, D2 = 0.016 pu Hz; 2H1 = 0.1667, 2H2 = 0.2017,
2H3 = 0.1247 pu s; R1 = 3.0, R2 = 2.73, R3 = 2.82 Hz/pu;
Tg1 = 0.08, Tg2 = 0.06, Tg3 = 0.07 s; Tt1 = 0.4, Tt2 = 0.44, Tt3 = 0.3 s;
Kr1 = Kr2 = Kr3 = 0.5 p, Tr1 = Tr2 = Tr3 = 10 s, T12 = 0.2, T23 = 0.12,
T31 = 0.25 pu/Hz, PR1 = 2000, PR2 = 4000, PR3 = 8000 MW.
References
[1] P. Kundur, Power System Stability and Control, 8th ed., Tata McGraw-Hill,
reprint, 2009.
[2] O.I. Elgerd, Electric Energy Systems Theory. An Introduction, Tata McGraw-Hill,
New Delhi, 1983.
[3] P. Ibraheem, D.P. Kothari, Recent philosophies of automatic generation control
strategies in power systems, IEEE Trans. Power Syst. 20 (1) (2005) 346–357.
[4] R.R. Shoults, J.A. Jativa Ibarra, Multi area adaptive LFC developed for a compre-
hensive AGC simulation, IEEE Trans. Power Syst. 8 (2) (1993) 541–547.
[5] D.K. Chaturvedi, P.S. Satsangi, P.K. Kalra, Load frequency control: a general-
ized neural network approach, Int. J. Electr. Power Energy Syst. 21 (6) (1999)
405–415.
[6] S.P. Ghosal, Optimization of PID gains by particle swarm optimization in fuzzy
based automatic generation control, Electr. Power Syst. Res. 72 (3) (2004)
203–212.
[7] T.P.I Ahamed, P.S.N. Rao, P.S. Sastry, A reinforcement learning approach to
automatic generation control, Electr. Power Syst. Res. 63 (2002) 9–26.
[8] S.R. Khuntia, S. Panda, Simulation study for automatic generation control of a
multi-area power system by ANFIS approach, Appl. Soft Comput. 12 (1) (2012)
333–341.
[9] L.C. Saikia, J. Nanda, S. Mishra, Performance comparison of several classical
controllers in AGC for multi-area interconnected thermal system, Int. J. Electr.
Power Energy Syst. 33 (2011) 394–401.
327
[10] J. Nanda, S. Mishra, L.C. Saikia, Maiden application of bacterial foraging based
optimization technique in multi area automatic generation control, IEEE Trans.
Power Syst. 24 (2) (2009) 602–609.
[11] E.S. Ali, S.M. Abd-Elazim, Bacteria foraging optimization algorithm based load
frequency controller for interconnected power system, Electr. Power Energy
Syst. 33 (2011) 633–638.
[12] H. Shabani, B. Vahidi, M. Ebrahimpour, A robust PID controller based on impe-
rialist competitive algorithm for load-frequency control of power systems, ISA
Trans. 52 (2012) 88–95.
[13] U.K. Rout, R.K. Sahu, S. Panda, Design and analysis of differential evolution algo-
rithm based automatic generation control for interconnected power system,
Ain Shams Eng. J. 4 (3) (2013) 409–421.
[14] L. Nolle, I. Zelinka, A.A. Hopgood, A. Goodyear, Comparison of an self-
organizing migration algorithm with simulated annealing and differential
evolution for automated waveform tuning, Adv. Eng. Software 36 (10) (2005)
645–653.
[15] W.C. Hong, Y. Dong, L.Y. Chen, S.Y. Wei, SVR with hybrid chaotic genetic
algorithms for tourism demand forecasting, Appl. Soft Comput. 11 (2) (2011)
1881–1890.
[16] M.Z. Ali, K. Alkhatib, Y. Tashtoush, Cultural algorithms: emerging social struc-
tures for the solution of complex optimization problems, Int. J. Artif. Intell. Eng.
11 (13) (2013) 20–42.
[17] D. Yazdani, B. Nasiri, R. Azizi, A. Sepas-Moghaddam, M.R. Meybodi, Optimiza-
tion in dynamic environments utilizing a novel method based on particle
swarm optimization, Int. J. Artif. Intell. Eng. 11 (13) (2013) 170–192.
[18] S.K. Saha, S.P. Ghoshal, R. Kar, D. Mandal, Cat swarm optimization algo-
rithm for optimal linear phase FIR filter design, ISA Trans. 52 (6) (2013)
781–794.
[19] E. Rashedi, H. Nezamabadi-Pour, S. Saryazdi, GSA: a gravitational search algo-
rithm, Inf. Sci. 179 (2009) 2232–2248.
[20] E. Rashedi, H. Nezamabadi-pour, S. Saryazdi, Filter modeling using gravitational
search algorithm, Eng. Appl. Artif. Intell. 24 (2011) 117–122.
[21] Y. Bao, Z. Hu, T. Xiong, A PSO and pattern search based mimetic algo-
rithm for SVMs parameters optimization, Neurocomputing 117 (2013)
98–106.
[22] A.K. Al-Othman, N.A. Ahmed, M.E. AlSharidah, A. Hanan, A hybrid real coded
genetic algorithm—pattern search approach for selective harmonic elimination
of PWM AC/AC voltage controller, Int. J. Electr. Power Energy Syst. 44 (2013)
123–133.
[23] L. Chaoshun, J. Zhou, Parameters identification of hydraulic turbine governing
system using improved gravitational search algorithm, Energy Conv. Manage.
52 (2011) 374–381.
[24] R.E. Precup, R.C. David, E.M. Petriu, M.B. Radac, S. Preitl, J. Fodor, Evolution-
ary optimization-based tuning of low-cost fuzzy controllers for servo systems,
Knowl.-Based Syst. 38 (2013) 74–84.
[25] E.D. Dolan, R.M. Lewis, V. Torczon, On the local convergence of pattern search,
SIAM J. Optim. 14 (2003) 567–583.
[26] S. Panda, N.K. Yegireddy, Automatic generation control of multi-area power
system using multi-objective non-dominated sorting genetic algorithm-II, Int.
J. Electr. Power Energy Syst. 53 (2013) 54–63.
[27] S. Padhan, R.K. Sahu, S. Panda, Application of firefly algorithm for load frequency
control of multi area interconnected power system, Electr. Power Compon. Syst.
42 (13) (2014) 1–12.
[28] S. Panda, B. Mohanty, P.K. Hota, Hybrid BFOA–PSO algorithm for automatic
generation control of linear and nonlinear interconnected power systems, Appl.
Soft Comput. 13 (2013) 4718–4730.
[29] W. Tan, Unified tuning of PID load frequency controller for power systems via
IMC, IEEE Trans. Power Syst. 25 (2010) 341–350.
[30] R.K. Sahu, S. Panda, U.K. Rout, DE optimized parallel 2-DOF PID controller for
load frequency control of power system with governor dead-band nonlinearity,
Int. J. Electr. Power Energy Syst. 49 (2013) 19–33.
[31] H. Golpira, H. Bevrani, H. Golpira, Application of GA optimization for automatic
generation control design in an interconnected power system, Energy Conv.
Manage. 52 (2011) 2247–2255.
[32] S. Panda, N.K. Yegireddy, S. Mahapatra, Hybrid BFOA-PSO approach for coordi-
nated design of PSS and SSSC-based controller considering time delays, Int. J.
Electr. Power Energy Syst. 49 (2013) 221–233.