Swarm and Evolutionary Computation xxx (xxxx) xxx–xxx

Contents lists available at ScienceDirect

Swarm and Evolutionary Computation

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

Binary Grey Wolf Optimizer for large scale unit commitment problem
⁎
Lokesh Kumar Panwara, , Srikanth Reddy Kb, Ashu Vermaa, B.K. Panigrahib, Rajesh Kumarc
a
Centre for Energy Studies, IIT Delhi, India
b
Department of Electrical Engineering, IIT Delhi, India
c
Department of Electrical Engineering, MNIT Jaipur, India

A R T I C L E I N F O A BS T RAC T

Keywords: The unit commitment problem belongs to the class of complex large scale, hard bound and constrained
Unit Commitment Problem optimization problem involving operational planning of power system generation assets. This paper presents a
Heuristics heuristic binary approach to solve unit commitment problem (UC). The proposed approach applies Binary Grey
Binary Grey Wolf Optimizer (BGWO) Wolf Optimizer (BGWO) to determine the commitment schedule of UC problem. The grey wolf optimizer
Constrained Optimization
belongs to the class of bio-inspired heuristic optimization approaches and mimics the hierarchical and hunting
Power system optimization
principles of grey wolves. The binarization of GWO is owing to the UC problem characteristic binary/discrete
search space. The binary string representation of BGWO is analogous to the commitment and de-committed
status of thermal units constrained by minimum up/down times. Two models of Binary Grey Wolf Optimizer are
presented to solve UC problem. The first approach includes upfront binarization of wolf update process towards
the global best solution (s) followed by crossover operation. While, the second approach estimates continuous
valued update of wolves towards global best solution(s) followed by sigmoid transformation. The Lambda-
Iteration method to solve the convex economic load dispatch (ELD) problem. The constraint handling is carried
out using the heuristic adjustment procedure. The BGWO models are experimented extensively using various
well known illustrations from literature. In addition, the numerical experiments are also carried out for different
circumstances of power system operation. The solution quality of BGWO are compared to existing classical as
well as heuristic approaches to solve UC problem. The simulation results demonstrate the superior performance
of BGWO in solving UC problem for small, medium and large scale systems successfully compared to other well
established heuristic and binary approaches.

1. Introduction
The unit commitment problem comprises the efficient utilization of
generation resources in power system operational planning. The UC
problem is a cost minimization problem which is often expressed as
optimization problem associated with various types of constraints with
respect to system as well as operation of generation units. The UC
problem is a complex optimization problem associated by many
constraints like load, reserve balance constraints, power generation
bounds, minimum up and down time constraints, ramp rate con-
straints etc. The complexity of UC problem is greatly affected by system
dimension and constant thrust to better the solution quality is
indispensable.
The earliest methods to solve the UC problem included classical
optimization methods like mixed integer linear programming (MILP)
[1], Dynamic programming (DP) [2], Priority list approach (PL) [3],
branch and bound approaches (BB) [4]. Some of other approaches
include dynamic programming with Lagrangian relaxation (DPLR) [5],
extended Lagrangian relaxation (ELR) [5], extended priority list (EPL)
[6], semidefinite Programming (SDP) [7] etc. The list of advantages of
classical methods lies in their simplest forms of representation and
application, fast convergence and integer solutions etc. However, suffer
from major drawbacks with poor solution quality (PL approach),
problems with system dimensionality (dynamic and linear program-
ming), exponentially increasing execution time with system dimension
(branch and bound) etc. The same resulted in origin of several nature/
bio inspired evolutionary and heuristic approaches.
The evolutionary and heuristic approaches are developed by
mimicking nature phenomenon. The same are adapted to solve UC
problem successfully. Some of the heuristic approaches include
genetic algorithm (GA) which functions on the principles of natural
selection and biological evolution of offspring of every generation
[8]. Whereas, approaches like particle swarm optimization (PSO)
[9], ant colony optimization (ACO) mimics the social behaviour and
coordination among the population [10]. In the similar lines of
inspiration from nature, many other optimization approaches like

⁎
Corresponding author.
E-mail address: lokesh.panwar28@gmail.com (L.K. Panwar).

http://dx.doi.org/10.1016/j.swevo.2017.08.002
Received 17 June 2016; Received in revised form 31 May 2017; Accepted 11 August 2017
2210-6502/ © 2017 Elsevier B.V. All rights reserved.

Please cite this article as: Panwar, L.K., Swarm and Evolutionary Computation (2017), http://dx.doi.org/10.1016/j.swevo.2017.08.002
L.K. Panwar et al. Swarm and Evolutionary Computation xxx (xxxx) xxx–xxx

Nomenclature TiMU Minimum up time of i th unit

Tih, off Consecutive hours of de-committed state of i th unit going
N Number of units into h th hour
H Total number of scheduling hours Tih, on Consecutive hours of committed state of i th unit going into
i Thermal unit index (i = 1,2,3…. N ) h th hour
h Scheduling hour index (h = 1,2,3…. H ) Pimin Minimum generation limit of i th unit
Fci Fuel cost function of i th unit Pimax Maximum generation limit of i th unit
γih Status bit (0 or 1) of i th unit for h th hour Pdh System load for h th hour
SUih Start-up cost of i th unit for h th hour Rsph Spinning reserve requirement for h th hour
Pih Scheduled power of i th unit for h th hour RiDR Ramp down rate of i th unit
SUihot Hot start-up cost of i th unit RiUR Ramp down rate of i th unit
SUicold Cold start-up cost of i th unit
TiMD Minimum down time of i th unit

evolutionary programming (EP) [11], simulated annealing (SA) 2. Problem formulation

[12], shuffled leaping frog approach (SFLA) [13], imperialistic
competition algorithm (ICA) [14], etc., are applied to solve UC
problem. Later, hybrid approaches are developed integrating the
expedient properties of classical and heuristic approaches to solve
the UC problem. Some of them are Lagrangian relaxation genetic
algorithm (LRGA) [15], Lagrangian relaxation particle swarm
optimization (LRPSO) [16], IPPDTM [17], hybrid harmony search
random search approach (HHSRSA) [18] are used to improve the
UC problem solution quality. Recently, the principles of quantum
computing viz. uncertainty, superposition and interference are
successfully applied to UC problem through evolutionary ap-
proaches. The applicability of quantum evolutionary approaches
[19] improved the exploration and exploitation of heuristic ap-
proaches at comparatively lower population size with respect to
other evolutionary approaches used to solve UC problem. The
harmony search algorithm with hybridization with random search
has been used to solve UC problem [30]. Also, the hybrid
approaches of nature inspired such as PSO-GWO are investigated
with application to UC problem [22]. The overview and insights of
Leading into the solution procedure, the formulation of objective
and constraints of the UC problems are explained before.
2.1. Objective function
The objective function of UC problem is modelled as a minimization
problem of total cost which constitutes of fuel cost, start-up and shut
down costs.
2.1.1. Fuel cost
All the committed thermal units incur fuel costs due to the
minimum power generation limits and the committed units are
dispatched economically so as to reduce the overall fuel cost yet
satisfying the system, thermal unit constraints. The fuel cost is
expressed as a quadratic equation given by,
2
Fci(Pih ) = ai + bi(Pih )+ci(Pih ) ∀ h ∈ H ; i ∈ N (1)
where ai , biandci are the fuel cost coefficients of i unit. th

some other recent nature inspired approaches for solving UC

problem can be found in literature review presented in [47]. 2.1.2. Start-up cost
Recently, Seyedali Mirjalili [21] proposed a meta-heuristic ap- The objective function also includes the start-up cost which is
proach named grey wolf optimizer (GWO) mimicking the specific incurred two the boiler temperature changes as a consequence of
hierarchical and hunting behaviour of grey wolves (Canis lupus). The commitment and de-commitment events. When returning to the
earlier models of GWO are economic load dispatch problem in power commitment status (γih = 1) from de-commitment state (γih = 0), the
system. Later, the GWO integrated with PSO is used to solve UC start cost depends on number of hours the unit is in de-committed
problem of different dimensions [22]. The GWO is also used to solve state. If the unit is in de-committed state for more than or equal to cold
many other industrial/research problems successfully [23]. Recently, start hours (Ticold ) after minimum off time, cold start-up (SUicold ) cost is
Binary Grey Wolf Optimizer is developed and successfully applied for associated with its commitment event. However, if the units is in de-
optimal feature selection purpose [24]. Motivated by the successful committed state after minimum off time but for less duration then cold
application of GWO & BGWO to industrial and research problems, this start hours, then the hot start-up cost SUihot is associated with the
paper presents a Binary Grey Wolf Optimizer application to solve commitment event. Therefore, the start-up cost applicable for i th
complex, non-linear and constrained UC problem. The presented thermal unit during h th hour is given by,
approach improves the solution quality of traditional GWO to solve
⎧
⎪ SU
hot
if TiMD ≤ Tioff ≤ TiMD + Ticold
UC problem efficiently. SUih = ⎨ i
∀ h ∈ H; i ∈ N
cold
The rest of the paper is organized as follows. The UC problem ⎪
⎩ SUi if Tioff ≥ TiMD + Ticold (2)
formulation and associated bounds, constraints are explained in
Section 2. The principles of real valued grey wolf optimizer and
Binary Grey Wolf Optimizer (BGWO) are presented in Section 3. A. Shutdown cost:In this paper, shut down costs are neglected which
Section 4 develops and describes the BGWO-UC approach. Section 5 are often modelled as constant values per de-commitment status of
presents the test system, parametric analysis and computational results the unit.
for different dimensions of the test system. Comparison of proposed
approach with the existing benchmarking algorithms in solving UC
problem is also presented in Section 5. The performance and statistical
significance of proposed BGWO models is also demonstrated in Section
5 using statistical tests. Finally, Section 6 concludes the paper with
contributions.

Fig. 1. Hierarchy of grey wolf pack [21].

2
L.K. Panwar et al. Swarm and Evolutionary Computation xxx (xxxx) xxx–xxx

Fig. 2. Representation of population structure of UC problem with BGWO.

Therefore, the total objective function is given by, F) Initial status: The status of generators in the ultimate/hour of
previous day/scheduling horizon before the beginning of current
TC = ∑∑ Fci(Pih )γih + SUih(1−γih )γih
scheduling horizon will reflect/affect parameters such as start-up
h i
costs, minimum up/down constraints etc.
∀ h ∈ H ; i ∈ N ; γih∈{0, 1} (3)
3. Grey Wolf optimizer (GWO)
2.2. Constraints
3.1. Overview of GWO
A) Generation limits: The actual generation of the units in committed
The grey wolf more often lives and hunts in pack with an average of
state should comply with the generation limits as given by,
5–12 wolves and follows certain hierarchy. The top of the hierarchy is
min h max
γihP ≤ P γih ≤γihP occupied by alpha wolf which are called as leaders of the pack ( Fig. 1).
i i i (4) This alpha wolf dictates the remaining wolves in tasks like hunting,
moving, sleeping place etc.
B) Load balance constraints: The system demand-supply balance The alpha may not be always the strongest of the pack physically
constraint given by, but intellectually it is considered as strongest of the pack. Often, the
remaining wolves show their obedience to decisions taken by alpha by
∑ Pihγih = Pdh;∀h ∈ H ; i ∈ N holding their tails down. The next level of hierarchy after alpha wolf is
i (5)
occupied by beta. The beta necessarily a subordinate of alpha and
reinforces alpha’s decisions among other lower level wolves. It helps
C) Spinning reserve: The hourly additional online capacity can be alpha in making decisions and is also the next candidate for alpha upon
summed up as follows. the expiration of latter. The last level of the hierarchy consists of omega
max wolves with no subordinates but to subordinate to other wolves and
∑ γihPi > Pdh + Rsph;∀h ∈ H ; i ∈ N
report them. These are least ranked in the system and are allowed to
i (6)
eat the remaining food that is left after all the other class of wolves are
finished. If any of the wolf in the pack does not belong to either alpha,
D) Minimum up/down time constraints: The time that should elapse beta or omega, then it is called as delta wolf.
between commitment and de-commitment events of units is The GWO employs particularly the hunting behaviour of the wolves
predefined based on the reliability and satisfactory performance which includes the following series of actions:
of particular unit.
⎧ 1, if 1 ≤ Tion MU • As the initial step, they track, chase and approach the prey.
⎪
γi = ⎨ 0,
, h −1 < Ti
• Followed by approaching the pray, wolves encircle it and harass it
•
h
if 1 ≤ Tioff
, h −1 < Ti
MD
Attacking the stationary prey after encircling it
⎪
⎩ 0 or 1, otherwise (7)
3.2. Continuous valued GWO
E) Ramp up/down rates: The ramp un/down rate constraints are
given by, The mathematical model of wolf behaviour involves ranking of
solution quality based on the hierarchy of the wolves. The first, second
h, min h h, max
γihP <P < γihP and third best solutions are designated asalpha(α ), beta ( β ) and delta
i i i (8)
(δ ) respectively. The wolf position is updated by using the distance
Where, from the updated position of prey (as decided by best three solutions of
α , β andδ ) of the wolves as follows.
Pih, min = max (Pimin, Pih −1 − RiDR )
→
⎯ →
⎯ ⎯→ →
⎯
Pih, max = min(Pimax , Pih −1 + RiUR ) (9) D = C . Xp (t ) − X (t ) (10)
⎯
→ ⎯→ →
⎯ → ⎯
X (t +1) = Xp(t ) − A . D (11)

3
L.K. Panwar et al. Swarm and Evolutionary Computation xxx (xxxx) xxx–xxx

Main Algorithm Heuristic Adjustment

Start
Take the wolf position
from main program
Initialization of the all
values

Initialize the all parameters of thermal

k=1 units

Initialize the position of ‘NP’ wolves in t=1

search space

Calculate the fitness of each position If t<=H No Return

using Economic Load Dispatch (ELD)

Yes
Find the values of alpha (α),

Heuristic Adjustment
beta (β) and delta (δ) wolf

Find the Up/Down time violation and

repair using Up/Down time repair
algorithm

k<=MaxIter No Stop

t=t+1
Check the reserve and load
constraint if violated then repair
Yes
using spinning reserve repair
algorithm
Update the position of wolfs
using equation (24/26)

Run Unit De-commitment algorithm

Generation of the Gaussian to remove the excessive reserve
spark available

Calculate the fitness of each position

using Economic Load Dispatch (ELD)

Find the values of alpha (α), beta

(β) and delta (δ) wolf

k=k+1

Fig. 3. Flowchart of UC problem solution using BGWO.

⎯
→ →⎯
Where the iteration number is denoted by t , A and C are coefficients, 3.3. Binary Grey Wolf Optimizer (BGWO)
⎯→ ⎯
→
Xp denotes the vector of prey position, and X represents the vector of
→⎯ ⎯
→ In the continuous valued grey wolf optimizer, the wolves update
grey wolf position around prey. The A and C can be estimated as
their position to a real value in the potential search space bounded by
follows:
problem constraints. However, for certain problems like UC, the
⎯
→ variables and search space is effectively limited to binary values [0,
A = 2→
a⎯ . r1⃗ − →
a⎯ (12) 1]. Therefore, mapping the real valued wolf position update to a binary
value is an essential procedure The same can be realized through
⎯
→
C = 2→
⎯r
2
different approaches at different phases of optimization process. The
two approaches employed in this study are explained as follows:

r1⃗ , →
⎯r ∈[0,1]
2 (13) A) Binarization- Model 1 (BGWO1)

4
L.K. Panwar et al. Swarm and Evolutionary Computation xxx (xxxx) xxx–xxx

Fig. 4. Algorithm for minimum up/down time constraint.

This model employs a crossover strategy which is stochastic in

nature. The wolves update their position according to the following
equation.
Fig. 5. Algorithm for reserve constraint satisfaction.
ϒi k+1 = Crossover (γ1, γ2, γ3), (14)
Where, γβD represents the vector of β wolf position in D dimension
Where, the ϒi k+1 = Crossover (γ1, γ2, γ3), represents a suitable cross-
over operation among γ1, γ2, γ3 which are binary values effected by and qβD denotes the binary step given by,
the movement of wolves in the direction of best three positions of ⎧ 1 if p D ≥ r
⎪

wolves namely alpha(α ), beta(β ) and delta(δ ) respectively. The update qβD = ⎨ β
⎩ 0 otherwise
⎪

process of binary vectors γ1, γ2, γ3 is determined by the following (19)

transformations. Where, r is uniformly distributed random number ∈[0, 1] and pβD
⎧ 1 if (γ D + q D ) ≥ 1 represents the continuous update step in D dimension estimated
γ1D = ⎨ α α
using sigmoid transformation as given by,
⎩0 otherwise (15)
1
Where, γαD
represents the vector of α wolf position in D dimension pβD =
1 + exp( − 10*(A1D DβD − 0.5) (20)
and qαD denotes the binary step given by,
Where, A1D , DβD
belongs to continuous valued grey wolf optimizer
⎧ 1 if p D ≥ r
qαD = ⎨ α and are estimated from (12), (10) for D dimension
⎩ 0 otherwise (16) The update operation of γ3D is also carried out in the similar
fashion as given by,
Where, r is uniformly distributed random number ∈[0, 1] and pαD
represents the continuous update step in D dimension estimated ⎧ 1 if (γ D + q D ) ≥ 1
γ3D = ⎨ δ δ
using sigmoid transformation as given by, ⎩0 otherwise (21)
1 γδD
pαD = where, represents the vector of δ wolf position in D dimension
1 + exp( − 10*(A1D DαD − 0.5) (17) and qδD denotes the binary step given by,
Where, A1D , DαD belongs to continuous valued grey wolf optimizer ⎧ 1 if p D ≥ r
and are estimated from (12), (10) for D dimension qδD = ⎨ δ
⎩ 0 otherwise (22)
Similarly, γ2D can be updated as follows,
⎧ 1 if (γ D + q D ) ≥ 1 where, r is uniformly distributed random number ∈[0, 1] and pδD
⎪

γ1D = ⎨ β β represents the continuous update step in D dimension estimated

⎩0
⎪

otherwise (18) using sigmoid transformation as given by,

5
L.K. Panwar et al. Swarm and Evolutionary Computation xxx (xxxx) xxx–xxx

2248500 600

2248000
500
2247500
400

Mean time (sec.)

2247000

Total cost ($)

2246500 300

2246000
200
2245500

Mean average cost 100

2245000
($)
2244500 0
5 10 15 20 25 30 35 40 45 50 55
Population size

Fig. 7. Variation of total cost and execution time with population size.

The second model employs a direct sigmoid transformation of

updated continuous valued function to arrive at binary values as given
by,
⎧ 1 if S{X (k + 1)}> r
γDk+1 = ⎨
⎩0 otherwise (25)
Where, r is the uniformly distributed random number over range [0, 1]
and S{X (k + 1)} is the sigmoid transformation function as given by,
1
S{X (k + 1)} = sigmoid (ϒ(k + 1)) = ⎧⎛ X + X + X ⎞ ⎫
−10⎨⎜ 1 2 3 ⎟ −0.5⎬
1+ e ⎩⎝ 3 ⎠ ⎭ (26)
where, X1, X2 , X3 are real valued position updates of wolf position with
respect to the positions of best three wolves (α , β , δ ) of the pack.
Fig. 6. Algorithm for de-commitment procedure.
4. BGWO Implementation to UC problem
1
pδD = The BGWO in this paper is used to find the optimal commitment
1 + exp( − 10*(A1D DδD − 0.5) (23)
schedule of thermal units. The power generation scheduling among the
where, A1D , DδD belongs to continuous valued grey wolf optimizer committed units through economic load dispatch is evaluated using
and are estimated from (12), (10) for D dimension. Lambda-iteration technique. The generalized representation of UC
Upon the estimation of γ1D, γ2D, γ3D the wolf position update problem is shown in Fig. 2.
operation through cross over operation is given by,
⎧γ D 1 A) Representation of Binary variables of UC Problem:The commit-
⎪1 if r < 3
⎪ ment status of i th generator of wth wolf during t th hour of k th iteration
γ D = ⎨ γ2D if 1 ≤ r ≤ 2 is represented byγit,,wk . Thus, the commitment of i th generator for t th
⎪ 3 3
is confirmed by assigning “1” and vice versa for de-commitment.
⎪ γ D otherwise
⎩3 (24) The BGWO is adapted to UC problem by assigning each wolf with a
where, γ1D, γ2D, γ3D represent the first, second and third best binary dimension of N-by-H matrix. Where, N and H represent the
values in dimension D, γ D denotes the output of dimension D, r is a number of units and number of scheduling hours respectively.
uniformly distributed random number over range [0.1]. Hence, the status bits of wth wolf in a population of NP range from
B) Binarization-Model 2 (BGWO2) γ1,1,wk to γNH,,wk for k th iteration. The commitment matrix of all units of

Table 1
Effect of population size on operational cost and mean computational time.

Population Best cost ($) Mean cost ($) Worst cost ($) Mean time Standard deviation (%)

5 2,247,472.885 2,248,158.293 2,248,966.053 47.62476968 0.017447527

10 2,246,561.534 2,247,465.041 2,247,922.831 95.44879995 0.019286349
15 2,246,305.835 2,246,998.242 2,247,706.541 144.3136118 0.014819079
20 2,246,155.426 2,246,648.633 2,247,119.906 191.1416483 0.012964035
25 2,246,034.335 2,246,451.574 2,247,217.468 239.9412435 0.015926715
30 2,245,795.973 2,246,274.076 2,246,816.112 283.8151838 0.011029745
35 2,245,462.311 2,246,126.048 2,247,113.193 334.2757163 0.019884921
40 2,245,366.781 2,246,017.991 2,246,896.702 384.0085725 0.019624868
45 2,245,195.54 2,245,938.252 2,246,883.871 430.2869865 0.02254337
50 2,245,188 2,245,911 2,246,872 479.0560583 0.022050963
55 2,245,179 2,245,901 2,246,861 527.486265 0.021294126

6
L.K. Panwar et al. Swarm and Evolutionary Computation xxx (xxxx) xxx–xxx

Table 2
Unit characteristics of 4 unit system.

Unit Pimax Pimin a b c TiMD TiMU SUihot SUicold Cs Initial status

(MW) (MW) ($/h) ($/MWh) ($/MW2h) (h) (h) ($) ($) (h) (h)

1 300 75 648.74 16.83 0.0021 5 4 500 1100 5 8

2 250 60 585.62 16.95 0.0042 5 3 170 400 5 8
3 80 25 213 20.74 0.0018 4 2 150 350 4 −5
4 60 20 252 23.6 0.0034 1 1 0 0.02 0 −6

Table 3
⎯→
⎯ ⎯→
⎯ ⎯→
⎯ ⎯→⎯
X2 = | Xβ − A2 . Dβ | (28)
Commitment and generation schedule of 4 unit system.

Commitment Schedule Generation schedule

⎯→
⎯ ⎯→
⎯ ⎯→
⎯ ⎯→⎯
X3 = | Xδ − A3 . Dδ |
Hour U1 U2 U3 U4 U1 U2 U3 U4
⎯⎯⎯→ ⎯→
⎯ ⎯→
⎯
where, Xα , Xβ , Xδ represents the first three best solutions,
1 1 1 0 0 300 149.9991 0 0
⎯→
⎯ ⎯→ ⎯ ⎯→ ⎯ ⎯→
⎯ ⎯→⎯ ⎯→⎯
2 1 1 1 0 300 204.9991 25 0 A1 , A2 , A3 are estimated using (12) Dα , Dβ , Dδ are estimated
3 1 1 1 0 300 250 49.9992 0 using (10). The binary position update of grey wolf is carried out
4 1 1 1 0 300 214.9991 25 0
using the real valued updates in (26) and is given by,
5 1 1 0 0 276.1901 123.8093 0 0
6 1 1 0 0 196.1901 83.8093 0 0 ⎧ ⎧ ⎯⎯⎯→ ⎯⎯⎯→ ⎯⎯⎯→ ⎫
7 1 1 0 0 202.8567 87.1426 0 0 ⎪ 1 if S ⎨ Xα + Xβ + Xδ ⎬> R
8 1 1 0 0 300 199.9991 0 0
γit, ,wk +1 =⎨ ⎩ 3 ⎭
⎪
⎩0 otherwise (29)

wth wolf over 24 scheduling horizon during k th iteration is denoted where, R is a uniformly distributed random number between 0 and
⎧ ⎯⎯⎯X→α + ⎯⎯⎯X→β + ⎯⎯⎯X→δ ⎫
byΥkw . The conventional numerical techniques can be used to solve 1, S ⎨ ⎬ is the sigmoid transformation of real valued wolf
ED problem as the fuel cost is approximately represented by a ⎩ 3 ⎭
quadratic equation. The detailed procedure for deciding ON/OFF position as given by,
schedule of units is explained in preceding sections. ⎯→
⎯ 1
S{ X (k + 1)} = ⎧ ⎛ ⎯⎯⎯⎯→ ⎯⎯⎯→ ⎯⎯⎯→ ⎞ ⎫
B) Binary grey wolf position initialization: The initialization ((k = 0) ) ⎪ Xα + Xβ + Xδ ⎪
−10⎨⎜⎜ ⎟ −0.5⎬
⎟
of wolves is assigned on random fashion to start with. A real valued 1+ e ⎩⎝
⎪ 3
⎠
⎪
⎭ (30)
random number Nrsr is generated within the range of sigmoid
function considered. The sigmoid transformation is applied to the
random number to arrive at probability/value in the range [0, 1]. D) Binary valued solution for UC schedule: The commitment and de-
Another random number Nrur is generator with is uniformly commitment status of units are represented by set of solutions ϒ kw
distributed between 0 and 1. Thereafter, the initial binary valued ϒ(k ) = {ϒ1k , ϒ2k ...... ϒ kw} of which each set consists of N-by-H matrix.
position of the grey wolves are updated as follows. The binary variables observed in BGWO can be mapped to UC
schedule ϒ kw as given by,
⎧ 0 S (N sr ) > N ur
γit,,1w = ⎨ ⎡ γ1, k γ 2, k … γ1,Hw, k ⎤
r r
⎩ 1 otherwise (27) ⎢ 1, w 1, w ⎥
⎢ 1, k 2, k … γ2,Hw, k ⎥
ϒ kw = ⎢ γ2, w γ2, w
… …⎥
For the next iteration, the positions of α , β , δ are assigned using
… …
first, second and third best fitness values obtained for initialized ⎢ 1, k 2, k ⎥
⎣⎢ γN , w γN , w … γNH,,wk ⎦⎥ (31)
wolf potions. The same can be evaluated by executing Lambda-
iteration procedure for the committed units for the first iteration.
C) Binary grey wolf position update: The position of wolf in potential E) Unit output continuous valued variables: The unit commitment
binary search space is updated based on the real valued update of schedules obtained in (31) above are supplied to Lambda-iteration
α , β , δ wolves. The real valued position update of α , β , δ is given as technique to obtain optimal generation schedules of committed
follows. thermal units. The vector of generation schedules is represented by
⎯→
⎯ ⎯⎯⎯→ ⎯→
⎯ ⎯→ ⎯
X1 = | Xα − A1 . Dα | P(k ) = [P1k , P2k.... Pwk ]. Where, each member of P(k ) is of dimension
N-by-H matrix and represents optimal generation schedules of
committed units as follows,

Table 4
Performance comparison of proposed BGWO for 4 unit test system.

Generation cost ($) Iteration time (s)

Method Best Mean Worst Best Mean Worst Change (%)

Improved Lagrangian Relaxation (ILR) [27] 75,231.9 NA NA – – – 1.726

B. SMP [26] 74,812 74,877 75,166 1.174
A.SMP [26] 74,812 74,877 75,166 1.174
Lagrangian Relaxation and PSO (LRPSO) [27] 74,808 NA NA – – – 1.169
Binary Differential Evolution (BDE) [28] 74,676 NA NA – – – 0.994
Genetic Algorithm (GA) [29] 74,675 NA NA – – – 0.993
Hybrid HS and Random Search Algorithm [30] 74,476 74,476 74,476 20.68 22.86 22.97 0.728
BGWO 73,933.1 73,933.1 73,933.1 3.455 – – 0

7
L.K. Panwar et al. Swarm and Evolutionary Computation xxx (xxxx) xxx–xxx

Fig. 8. Convergence characteristics of different test systems for BGWO2 & QEA. (a) 10 unit case (b) 20 unit case (c) 40 unit case (d) 60 unit case (e) 80 unit case (f) 100 unit case.

⎡ p 1, k p 2, k … p1,Hw, k ⎤
⎢ 1, w 1, w ⎥
⎢ 1, k F) The proposed approach considers number of iterations as termina-
2, k
Pwk = ⎢ p2, w p2, w … p2,Hw, k ⎥
… … ⎥
tion criteria. Therefore, the BGWO stops as soon as the iteration
… …
⎢ 1, k 2, k
⎥ reaches pre-defined maximum number of iterations. Fig. 3 shows
⎣⎢ pN , w pN , w … pNH,,wk ⎦⎥ (32) the flow chart of proposed BGWO approach for solving UC
problem.

8
L.K. Panwar et al. Swarm and Evolutionary Computation xxx (xxxx) xxx–xxx

Table 5 MATLAB 2010b environment on system with i5 INTEL processor,

Numerical results of BGWO approaches for different test systems. 16 GB RAM, Microsoft operating system. The simulation results for
proposed approach are presented and discussed for various test
Units Approach Cost ($)
systems. Later the comparison of proposed approach with other
Best Average Worst Std. benchmark approaches for solving UC problem is presented to
Deviation demonstrate the effectiveness of BGWO to solve UC problem.
Population size determination: The optimal population for carrying
10 QEA 563,936.3 563,980.04 564,017.6 18.97852
BGWO1 563,976.64 564,378.58 565,518.14 0.1067645 out the numerical experiments is evaluated using simulation studies
BGWO2 563,937.31 563,937.31 563,937.31 0 and observations of time/solution quality as a function of the popula-
20 QEA 1,123,404 1,123,455.9 1,123,505 36.9198 tion size. The numerical observation against the population variation
BGWO1 1,125,546.4 1,126,126.3 1,127,393.2 0.0557585 are summarized in Table 1. The simulations are performed for 40 unit
BGWO2 1,123,297 1,124,215 1,124,379.2 0.0016277
test case as a trade-off between small (10 unit) and large (100 unit)
40 QEA 2,245,520 2,245,675.4 2,245,933 164.8325
BGWO1 2,252,475.1 2,257,866.7 2,263,333 0.1536076 scale systems.
BGWO2 2,244,701 2,245,144.9 2,246,020.6 0.0135283 The total cost and execution time are inversely related to each other
60 QEA 3,364,972 3,366,000.2 3,367,247 927.0687 for variation in population (Fig. 7). The reduction of total cost with
BGWO1 3,368,934.4 3,375,220.5 3,384,305.8 0.144654
respect to population has reduced progressively. Whereas, the execu-
BGWO2 3,362,515.2 3,366,488.3 3,367,143.5 0.0192798
80 QEA 4,489,856 4,492,184.7 4,493,418 956.229 tion time witnessed a linear variation along with population. The
BGWO1 4,495,173.4 4,506,362.2 4,513,026.3 0.1094866 optimal population of 30 is selected based on the saturation of cost
BGWO2 4,483,381.1 4,486,675.6 4,488,568.4 0.0122751 reduction at higher population size yet resulting in higher execution
100 QEA 5,608,527 5,609,920.5 5,611,208 1042.705 time. Apart from population size, all the other parameters of the
BGWO1 5,628,975 5,637,659 5,643,899.3 0.0813199
algorithm are evaluated along the execution. The range of parameter a
BGWO2 5,604,145.9 5,607,031.1 5,607,722.7 0.0104419
in (12) is set to 0–2, which is adopted from literature upon the
observation of its optimality among all the ranges under study.
G) Constraint handling: The heuristic approach of handling con-
straints based on rule based mechanism is adopted in this work 5.1. Performance of BGWO for Test system 1
[19]. The possible violation of constraints like minimum up-down
times during initialization as well as update procedures has to be This test system consists of 4 thermal units, whose characteristics
handled appropriately. Therefore, the heuristic approach based are presented in Table 2. The schedule horizon is 8 h [26]. The mean
techniques are used to avoid the occurrence of infeasible solutions, execution time of the system with 30 wolf population is 3.455 s. The
thereby improving the solution quality. Apart from up/down time commitment and generation schedules of 4 unit system using BGWO
constraints, excessive reserve allocation may also prove costly. are produced in Table 3.
Therefore, the proposed approach used unit rule based de-commit- The solution quality and execution time of 4 unit UC problem using
ment algorithm to deal with reserve constraints. BGWO is compared with other existing approaches (Table 4). The same
demonstrates the superior performance of BGWO over existing
approaches. The maximum and minimum improvements in solution
4.2. Constraint repair
quality due to BGWO compared to the approaches is 1.725% and
0.728% respectively.
The heuristic approach of handling constraints based on rule based
mechanism is adopted in this work [19]. In the process of random
5.2. Performance of BGWO for Test system 2
initialization and update process of evolutionary algorithms like
BGWO, often the state transition of units may violate minimum-up
The total cost convergence characteristics of BGWO for different
down time constraints.
test systems are shown in Fig. 8. The comparison of convergence of
BGWO is made with existing QEA approach and the results shows
A) Minimum up/down constraints: The commitment and De-com-
considerable improvement in solution quality at comparable conver-
mitment event of unit must abide the up/down constraints of the
gence speed for lower systems. The numerical results for both the
same. This paper uses rule based heuristic adjustment process to
models are presented in Table 5.
tackle the up/down time violation as shown in Fig. 4.
The variation in best cost attainted in different rails of BGWO for
B) Spinning reserve and load satisfaction repair: The load and
various test systems is shown in Fig. 9. The variation is plotted along
spinning reserve satisfaction constraints are enforced to guarantee
with mean best cost and it can be observed that the deviation from
the reliability of the system. In case of these constraint violation,
mean value is lower in 10 and 100 unit case compared to all other test
the de-committed units are committed until the demand and
systems. The difference between highest and lowest best costs among
spinning reserve constraints are satisfied. The detailed procedure
different trails is shown in Fig. 10. It can be observed that, the
of load and reserve constraint repair strategy is explained in Fig. 5
difference increased as a function of test system size.
as follows.
C) De-commitment algorithm under excessive spinning reserve: The
5.2.1. Comparison of proposed approaches with various other
satisfaction of up/down time constraint and spinning reserve
approaches
constraint may end up in committing extra thermal units thereby
The comparison with other algorithms for various test systems from
resulting in residual spinning reserve which may increase the
10 to 100 units is presented in Table 4. The comparison of both the
operational cost unnecessarily. This can be avoided by implement-
proposed approaches with other approaches reveals the effectiveness of
ing De-commitment algorithm as explained in Fig. 6.
BGWO with respect to GA [8], EP [11], SA [12], SFLA [13], GSA [38],
GVNS [40], ICA [14], IPSO [36], QEA [20], HHSRSA [30], PSO-GWO
5. Numerical results and discussion [22] etc., for all test systems from 10 to 100 units. However, no
particular inferences can be drawn from comparison for computational
The proposed BGWO is modelled to solve UC problem with test times owing to the differences in computational facilities of different
systems of different dimensions starting from 10 to 100 units for 24 h approaches. The comparison of BGWO approaches with other ap-
scheduling horizon [25]. The simulation studies are performed in proaches for small scale systems of test system 2 is shown in Table 6.

9
L.K. Panwar et al. Swarm and Evolutionary Computation xxx (xxxx) xxx–xxx

Fig. 9. Variation of best cost among different trails (a) 10 unit case (b) 20 unit case (c) 40 unit case (d) 60 unit case (e) 80 unit case (f) 100 unit case.

10
L.K. Panwar et al. Swarm and Evolutionary Computation xxx (xxxx) xxx–xxx

0.07 comparison of BGWO against other existing algorithms is presented in

Table 11, which demonstrates the superior performance of BGWO to
0.06
solve UC problem under different operating conditions.
0.05
Difference (%)

5.3. Statistical analysis

0.04

0.03
0.02
0.01
0
10 20 40
Units
Fig. 10. Percentage difference between highest and lowest best cost.
The statistical analysis of mean and standard deviation of proposed
approach confirms the effectiveness of BGWO2 over BGWO1 and other
existing approaches.
The commitment and economic dispatch schedule for 10 unit and
24 h horizon using proposed BGWO approach is shown in Table 7. The
effect of system size on computational time is shown in Fig. 11. The
quadratic progression of execution time with system size can is
observed. The performance comparison of medium (40, 60 units)
and large scale (80, 100 units) systems is presented in Tables 8, 9
respectively.
The effectiveness of proposed BGWO approach for solving UC is
also evaluated for different operating conditions i.e. 5% spinning
reserve. The unit commitment schedule using BGWO under 5%
spinning reserve condition is given in Table 10. The performance
The statistical tests are often used to compare the performance and
demonstrate the statistical significance of approaches for a given
problem [48]. Statistical tests such as Friedman test, Wilcoxon test
etc., are used in recent studies to compare the effectiveness of different
approaches in solving UC problem [49]. The test ranks, test statistic
and p-value (frequentist inference) for null hypothesis rejection are
60 80 100 used as performance metrics/indices for comparing various ap-
proaches. This paper considers Friedman (aligned and non-aligned)
ranks, Wilcoxon test and Quade test for performance comparison of
proposed approaches with the existing binary approaches such as
BPSO, QEA, QBPSO, BFWA etc. The independent solution samples
required for performance comparison using these tests are obtained by
running respective algorithms for same number of independent trails
(30 trails each).
5.3.1. Friedman test
Friedman test provides a non-weighted and non-aligned ranking of
comparison for solutions/independent samples of various approaches.
The Friedman ranks (non-aligned) and p-values for BGWO and other
approaches are presented in Table 12. The individual test system
ranking and overall ranking shows the superiority of BGWO 2 over
BGWO 1 and other binary, quantum approaches. The p-values in all
three cases are well below 0.05 (significance level) demonstrates the
statistical significance of proposed approaches when compared to other
existing binary approaches. The increasing p-values from 10 unit test

Table 6
Comparison of BGWO with other approaches for small scale (10 and 20 unit) systems.

10 unit system 20 unit system

Approach Best ($) Average ($) Worst ($) Std. (%) Avg. Time (Sec) Best ($) Average ($) Worst ($) Std. (%) Avg. Time (Sec)

LR [8] 565,825 565,825 565,825 – – 1,130,660 1,130,660 1,130,660 – –

GA [8] 565,825 – 570,032 0.74 221 1,126,243 – 1,132,059 0.52 733
EP [11] 564,551 565,352 566,231 0.3 100 1,125,494 1,127,257 1,129,793 0.38 340
LRGA [15] 564,800 564,800 564,800 – 518 1,122,622 1147
SA [12] 565,828 565,988 566,260 0.08 3 1,126,251 1,127,955 1,129,112 0.25 17
ICGA [31] 566,404 566,404 566,404 – 7.4 1,127,244 – – – 22.4
MA [32] 566,686 566,787 567,022 0.06 61 1,128,192 1,128,213 1,128,403 0.02 113
IPPDTM [17] 563,977 563,977 563,977 – 0.52 – – – – –
GRASP [33] 565,825 565,825 565,825 – 17 565,825 – – – 571
LMBSI [34] 563,977 563,977 563,977 – 7.3 1,123,990 – – – 15.24
DPLR [5] 564,049 564,049 564,049 – 108 1,128,098 – – – 299
GACC [35] 563,977 564,791.5 565,606 0.0857 85 1,125,516 1,127,153 1,128,790 0.1004 225
LRPSO [16] 565,869 565,869 565,869 – 42 565,869 – – – 91
ELR [5] 563,977 563,977 563,977 – 4 1,130,660 – – – 16
EPL [6] 563,977 563,977 563,977 – – 1,124,369 – – – –
IPSO [36] 563,954 564,162 564,579 0.11 – 1,125,279 – 1,127,643 0.21 –
BFWA [25] 563,977 564,018 564,855 – 65.42 1,124,858 1,124,941 1,125,087 – 106.03
IBPSO [37] 563,977 564,155 565,312 0.0253 27 1,125,216 1,125,448 1,125,730 0.0155 55
SFLA [13] 564,769 564,769 564,769 – 1,123,261 – – –
ICA [14] 563,938 563,938 563,938 – 1,124,274 – – –
IQEA [19] 563,977 563,977 563977 0 15 1123,890 1,124,320 1,124,504 0.05 42
QEA [20] 563,938 563,969 564,672 0.13 19 1,123,607 1,124,689 1,125,715 0.19 28
SDP [7] 563,938 563,938 563,938 – 25.41 1,124,357 – – – 63.94
QBPSO [18] 563,977 563,977 563,977 0 18 1,123,297 1,123,981 1,124,294 0.09 50
HHSRSA [30] 563,937.68 563,965.3 563,995.3 – 16.83 1,124,889 1,124,912.8 1,124,951.5 – –
PSO-GWO [22] 565,210.2 – – – – – – – – –
RM [39] 563,977 – – – 1.15 1,123,990 – – – 2.14
GVNS [40] 563,938 – – – 0.23 1,123,297 – – – 2.46
hGADE/r1 [46] 563,938 564,044 564,283 0.06 24 1,123,386 1,124,436 1,125,045 0.15 51
hGADE/cur1 [46] 563,959 564,088 564,350 0.07 24 1,123,426 1,124,502 1,125,076 0.15 48
Enh-hGADE [46] 563,938 563,997 564,261 0.06 26 1,123,386 1,124,262 1,124,939 0.13 56
BGWO1 563,976.6 564,378.58 565,518.1 0.106 64.19 1,125,546.4 1,126126.3 1,127,393.2 0.055 80.47
BGWO2 563,937.3 563,937.3 563,937.3 0 66.15 1,123,297 1,124,215 1,124,379 0.0016 87.533

11
L.K. Panwar et al. Swarm and Evolutionary Computation xxx (xxxx) xxx–xxx

Table 7
Unit commitment schedule and reserve available for 10 unit system using BGWO2.

Hour Unit1 Unit2 Unit3 Unit4 Unit5 Unit6 Unit7 Unit8 Unit9 Unit10 Load Reserve

1 455 245 0 0 0 0 0 0 0 0 700 210

2 455 295 0 0 0 0 0 0 0 0 750 160
3 455 370 0 0 25 0 0 0 0 0 850 222
4 455 455 0 0 40 0 0 0 0 0 950 122
5 455 390 0 130 25 0 0 0 0 0 1000 202
6 455 360 130 130 25 0 0 0 0 0 1100 232
7 455 410 130 130 25 0 0 0 0 0 1150 182
8 455 455 130 130 30 0 0 0 0 0 1200 132
9 455 455 130 130 85 20 25 0 0 0 1300 197
10 455 455 130 130 162 33 25 10 0 0 1400 152
11 455 455 130 130 162 73 25 10 10 0 1450 157
12 455 455 130 130 162 80 25 43 10 10 1500 162
13 455 455 130 130 162 33 25 10 0 0 1400 152
14 455 455 130 130 85 20 25 0 0 0 1300 197
15 455 455 130 130 30 0 0 0 0 0 1200 132
16 455 310 130 130 25 0 0 0 0 0 1050 282
17 455 260 130 130 25 0 0 0 0 0 1000 332
18 455 360 130 130 25 0 0 0 0 0 1100 232
19 455 455 130 130 30 0 0 0 0 0 1200 132
20 455 455 130 130 162 33 25 10 0 0 1400 152
21 455 455 130 130 85 20 25 0 0 0 1300 197
22 455 455 0 0 145 20 25 0 0 0 1100 137
23 455 425 0 0 0 20 0 0 0 0 900 90
24 455 345 0 0 0 0 0 0 0 0 800 110

900 5.3.3. Quade test

800 Quade test can be used to for performance comparison among
different approaches, including the block differences that go unidenti-
700
fied in the Friedman as well as Friedman Aligned ranks test, where all
Execution time (Sec.)

600
500
400
300
200
100
0 aligned) and the superior performance of BGWO 2 over other
10 20 40
Untis
Fig. 11. Time complexity of BGWO approach across system dimension.
system to 100 unit test system signifies a higher statistical significance
of BGWO approaches for larger test systems compared to smaller test
systems.
5.3.2. Friedman aligned ranks test
The simple/non-aligned ranks of Friedman test provides a fair
intra-set comparison between various approaches. However, the same
fails in considering the relative performance of each algorithm with
respect to the average performance of all the approaches under
consideration. The Friedman Aligned ranks test provides the compar-
ison of any given approach against the average performance of all the
approaches under study. The Friedman aligned ranks and correspond-
ing statistical significance of BGWO against other approaches is
presented in Table 13. Similar to the Friedman ranks, BGWO 2
achieved superior performance over BGWO 1 as well as other
approaches. On the other hand, the p-values of Friedman aligned
ranks test are lower compared to their Friedman (non-aligned) test
counterparts. The same reveals a higher significance differences
between proposed and existing approaches at larger scale with respect
to the relative/mean performance.
the algorithms are given/assigned with same importance. The algo-
rithms/approaches are weighted in the Quade test based on block
difference and can help in establishing the statistical significance of a
particular algorithm within the group of algorithms considered for
study [48]. The p-values and ranks of different approaches estimated
using weighted ranks (using simple non-aligned ranks with exclusion of
block differences) method are presented in Table 14. The ranks
obtained using Quade test are comparable to the Friedman test (non-
60 80 100 approaches is also reflected in Quade ranks. The p-values of Quade
rank are considerably lower when compared to other two tests
(Friedman and Friedman aligned ranks test). The same reveals the
higher significant differences between BGWO and other approaches
when block differences are included through weighted ranks.
5.3.4. Wilcoxon pairwise compassion
The Friedman, Friedman aligned ranks and Quade test provide the
justification that proposed BGWO has significant differences with the
existing approaches. However, the comparison of differences in the
whole group of independent samples obtained for all the approaches
may mask any hidden similarities between proposed approach and the
existing approaches. Therefore, the pairwise comparison is performed
on different approaches across various test systems as presented in
Tables 15–17. In each comparison, the instances with no significance
differences (at a significance level of 0.05) are highlighted. It can be
observed that, except 10 unit systems (only BGWO 1), there exist
significance differences between the proposed BGWO variants and
other approaches. Also, the BGWO 2 which has shown superior
performance when compared to all other approaches is significantly
different (at 0.05 level of significance) with respect to each of the
algorithm/approach for all the test systems.
6. Conclusion and future scope
This paper presents two binary grey wolf (BGWO) models to solve
UC problem. The objective of UC problem is formulated as a cost

12
L.K. Panwar et al. Swarm and Evolutionary Computation xxx (xxxx) xxx–xxx

Table 8
Comparison of BGWO with other approaches for medium scale (40 and 60 unit) systems.

40 unit system 60 unit system

Approach Best ($) Average ($) Worst ($) Std. (%) Avg. Time (Sec) Best ($) Average ($) Worst ($) Std. (%) Avg. Time (Sec)

MA [32] 2,249,589 2,249,589 2,249,589 0 217 3,370,595 3,370,820 3,371,272 0 576

LR [8] 2,258,503 2,258,503 2,258,503 0 – 3,394,066 3,394,066 3,394,066 – –
GA [8] 2,251,911 – 2,259,706 0.35 2697 3,376,625 – 3,384,252 0.23 5840
EP [11] 2,249,093 2,252,612 2,256,085 0.31 1176 3,371,611 3,376,255 3,381,012 0.28 2267
LRGA [15] 2,242,178 2,242,178 2,242,178 0 2165 3,371,079 3,371,080 3,371,081 0 2414
SA [12] 2,250,063 2,252,125 2,254,539 0.2 88 – – – – –
ICGA [31] 2,254,123 2,254,123 2,254,123 0 58.3 3,378108 3,378,108 3,378,108 0 117.3
IPPDTM [17] 2,247,162 2,247,162 2,247,162 0 6.49 3,366874 3,366,874 3,366,874 0 11.39
GRASP [33] 2,259,340 2,259,340 2,259,340 0 1511 3,383,184 3,383,184 3,383,184 0 2638
LMBSI [34] 2,243,708 2,243,708 2,243,708 0 27 3,362,918 3,362,918 3,362,918 0 40
DPLR [5] 2,256,195 2,256,195 2,256,195 0 1200 3,384,293 3,384,293 3,384,293 0 3199
GACC [35] 2,249,715 2,253,270 2,256,824 0.08457 614 3,375,065 3,378,976 3,382,886 0 1085
LRPSO [16] 2,251,116 2,251,116 2,251,116 0 213 3,376,407 3,376,407 3,376,407 0 360
EPL [6] 2,246,508 2,246,508 2,246,508 0 – 3,366,210 3,366,210 3,366,210 0 –
IPSO [36] 2,248,163 – 2,252,117 0.18 – 3,370,979 – 3,379,125 0.24 –
BFWA [25] 2,248,228 2,248,572 2,248,645 – 238.02 3,367,445 3,367,828 3,367,974 0 422.29
IBPSO [37] 2,248,581 2,248,875 2,249,302 0.011561 110 3,367,865 3,368,278 3,368,779 0.009708 172
SFLA [13] 2,246,005 2,246,005 2,246,005 0 150 3,368,257 3,368,257 3,368,257 0 280
ICA [34] 2,247,078 2,247,078 2,247,078 0 151 3,371,722 3,371,722 3371722 0 366
IQEA [19] 2,245,151 2,246,026 2,246,701 0.07 132 3,365,003 3,365,667 3,366,223 0.04 273
QEA [20] 2,245,557 2,246,728 2,248,296 0.12 43 3,366,676 3,368,220 3,372,007 0.16 54
SDP [7] 2,243,328 – – – 157.73 3,363,031 – – – 260.7
HHSRSA [30] 2,248,508 2,248,653 2,248,757 – 179.666 – – – – –
hGADE/r1 [46] 2,243,724 2,245,582 2,247,130 0.15 137 3,363,470 3,365,587 3,368,196 0.14 277
hGADE/cur1 [46] 2,243,522 2,245,321 2,246,540 0.13 123 3,362,908 3,364,841 3,367,820 0.15 307
Enh-hGADE [46] 2,243,522 2,245,020 2,246,487 0.13 147 3,362,908 3,364,538 3,367,820 0.14 326
BGWO1 2,252,475 2,257,866 2,263,333 0.15360 169.24 3,368,934 3,375,221 3,384,306 0.14465 281.2
BGWO2 2,244,701 2,245,145 2,246,021 0.00345 153.5 3362515 3,366,488 3,367,144 0.00293 268.2

Table 9
Comparison of BGWO with other approaches for large scale (80 and 100 unit) systems.

80 unit system 100 unit system

Approach Best ($) Average ($) Worst ($) Std. (%) Avg. Time (Sec) Best ($) Average ($) Worst ($) Std. (%) Avg. Time (Sec)

MA [32] 4,494,214 4,494,378 4,494,439 0 664 5,616,314 5,616,699 5,616,900 0.01 1338
LR [8] 4,526,022 4,526,022 4,526,022 0 – 5,657,277 5,657,277 5,657,277 0 –
GA [8] 4,504,933 – 4,510,129 0.12 10,036 5,627,437 – 5,637,914 0.19 15,733
EP [11] 4,498,479 4,505,536 4,512,739 0.32 3584 5,623,885 5,633,800 5,639,148 0.27 6120
LRGA [15] 4,501,844 4,501,844 4,501,844 0 3383 5,613,127 5,613,127 5,613,127 0 4045
SA [12] 4,498,076 4,501,156 4,503,987 0.13 405 5,617,876 5,624,301 5,628,506 0.19 696
ICGA [5] 4,498,943 – – – 176 5,630,838 – – – 342.2
IPPDTM [17] 4,490,208 4,490,208 4,490,208 0 31.23 5,609,782 46.55
GRASP [33] 4,525,934 – – – 3308 – – – 4392
LMBSI [34] 4,483,593 4,483,593 4,483,593 0 54 5,602,844 5,602,844 5,602,844 0 73
DPLR [5] 4,512,391 4,512,391 4,512,391 0 8447 5,640,488 5,640,488 5,640,488 0 12,437
GACC [35] 4,505,614 4,516,731 4,527,847 0.0947 1975 5,626,514 5,636,522 5,646,529 0.1012 3547
LRPSO [16] 4,496,717 – – – 543 – – – 730
EPL [6] 4,489,322 4,489,322 4,489,322 0 – 5,608,440 5,608,440 5,608,440 0 –
IPSO [36] 4,495,032 – 4,508,943 0.31 – 5,619,284 – 5,628,506 0.24 –
BFWA [25] 4,491,284 4,492,550 4,493,036 – 676.53 5,610,954 5,612,422 5,612,790 – 1043.47
IBPSO [37] 4,491,083 4,491,681 4,492,686 0.0120 235 5,610,293 5,611,181 5,612,265 0.01554 295
SFLA [13] 4,503,928 4,503,928 4,503,928 0 690 5,624,526 5,624,526 5,624,526 0 1430
ICA [34] 4,497,919 4,497,919 4,497,919 0 994 5,617,913 5,617,913 5,617,913 0 1376
IQEA [19] 4,486,963 4,487,985 4,489,286 0.05 453 5,606,022 5,607,561 5,608,525 0.04 710
QEA [20] 4,488,470 4,490,128 4,492,839 0.1 66 5,609,550 5,611,797 5,613,220 0.07 80
SDP [7] 4,484,365 – – – 353 5,602,538 – – – 392
hGADE/r1 [46] 4,486,180 4,489,500 4,498,651 0.28 368 5,604,787 5,610,074 5,620,236 0.27 397
hGADE/cur1 [46] 4,485,160 4,487,968 4,494,013 0.2 343 5,606,075 5,610,336 5,620,839 0.26 451
Enh-hGADE [46] 4,485,160 4,487,293 4,489,114 0.09 404 5,604,787 5,607,487 5,612,131 0.13
BGWO1 4,495,173 4,506,362 4,513,026 0.1094 473.4 5,628,975 5,637,659 5,643,899 0.08131 836.54
BGWO2 4,483,381 4,486,676 4,488,568 0.0122 469.6 5,604,146 5,607,031 5,607,723 0.01044 822.23

13
L.K. Panwar et al. Swarm and Evolutionary Computation xxx (xxxx) xxx–xxx

Table 10
Commitment and generation scheduling under 5% spinning reserve.

Hour U1 U2 U3 U4 U5 U6 U7 U8 U9 U10

1 455 244.9992 0 0 0 0 0 0 0 0
2 455 294.9993 0 0 0 0 0 0 0 0
3 455 394.9992 0 0 0 0 0 0 0 0
4 455 364.9992 0 130 0 0 0 0 0 0
5 455 389.9991 0 130 25 0 0 0 0 0
6 455 455 0 130 59.999 0 0 0 0 0
7 455 409.9992 130 130 25 0 0 0 0 0
8 455 455 130 130 29.999 0 0 0 0 0
9 455 455 130 130 109.999 20 0 0 0 0
10 455 455 130 130 162 42.999 25 0 0 0
11 455 455 130 130 162 80 25 12.999 0 0
12 455 455 130 130 162 80 25 52.999 10 0
13 455 455 130 130 162 42.999 25 0 0 0
14 455 455 130 130 104.999 0 25 0 0 0
15 455 455 130 130 29.999 0 0 0 0 0
16 455 309.9991 130 130 25 0 0 0 0 0
17 455 259.9993 130 130 25 0 0 0 0 0
18 455 359.9991 130 130 25 0 0 0 0 0
19 455 455 130 130 29.999 0 0 0 0 0
20 455 455 130 130 162 0 25 42.999 0 0
21 455 455 130 130 104.999 0 25 0 0 0
22 455 455 0 0 162 0 27.9991 0 0 0
23 455 419.9992 0 0 25 0 0 0 0 0
24 455 344.9991 0 0 0 0 0 0 0 0

Table 11 natured problems in power systems such as profit based unit commit-
Performance comparison of BGWO for 5% spinning reserve condition. ment (PBUC). Other directions of possible extensions of the work can
include hybridization of BGWO with real coded heuristic approaches
Method Best cost ($) Average cost ($) Worst cost ($)
for allocating power dispatch of committed units in the UC problem.
BPSO [41] 565,804 566,992 567,251 Also, the proposed BGWO can be extended in future with real coded
GA [41] 570,781 574,280 576,791 hybrid approaches to solve combined economic and emission dispatch
APSO [42] 561,586 – – (CEED) problem in power system operation planning. Some other
BP [42] 565,450 – –
TSGB [43] 560,263.92 – –
interesting future works may be listed as follows.
IPSO [44] 558,114.8 – –
Hybrid PSO-SQP [45]
B.SMP [26]
568,032.3
558,844.76
–
558,937.24
–
559,155
• The BGWO approach proposed in this paper has been tested for
single objective function (cost minimization of UC problem).
Hybrid HS-Random Search 557,905.643 558,267.2 558,682
However, the same can be used to develop multi-objective problems
algorithm [30]
BGWO 556,840.099 558,026.296 558,358.3 involving UC framework [50]. The other objectives of UC problem
such as emission reduction and reliability maximization can form
multi-objective UC (MOUC) problem for future studies [51].
minimization optimization problem associated with bound, equality • Apart from binary natured problems such as UC and MOUC, the real
and inequality constraints. The first of the two Binarization models valued hybrid variants of GWO can be developed in future with
uses binary update upfront, followed by crossover operation. Whereas application to real time problems across various fields of engineer-
the second one employs real valued update of wolves followed by ing [52,53].
sigmoid transformation for binary update. The two models are tested • The extension of real valued and combinational variants of the
on various test systems operating under different conditions. The proposed BGWO and GWO frameworks for other multi-objective
simulation results comparison with other existing approaches demon- problems [54] may be considered for the future studies.
strates the superiority of BGWO approaches tin solving UC problem
efficiently. The proposed BGWO can be extended to solve other binary

Table 12
Performance comparison of various approaches for different test systems using Friedman test.

Approach 10 unit system 40 unit system 100 unit system Overall rank

Rank p-value Rank p-value Rank p-value

BPSO 3.666667 3.26E−12 3.866667 4.17E−14 4 6.08E−14 11.53333

QEA 5.733333 5.466667 5.533333 16.73333
QBPSO 5.133333 5.533333 5.466667 16.13333
BFWA 2.6 2.933333 2.133333 7.666667
BGWO 1 2.666667 2.2 2.866667 7.733333
BGWO 2 1.2 1 1 3.2

14
L.K. Panwar et al. Swarm and Evolutionary Computation xxx (xxxx) xxx–xxx

Table 13
Performance comparison of various approaches for different test systems using Friedman aligned ranks test.

Approach 10 unit system 40 unit system 100 unit system Overall rank

Rank p-value Rank p-value Rank p-value

BPSO 40 51.8667 53.1333 145

QEA 80.3333 74.7333 76.8667 231.9333
QBPSO 68.6 9.04E−13 76.2667 3.82E−16 74 6.54E−16 218.8667
BFWA 24.4 32 29.1333 85.5333
BGWO 1 41.3333 30.0667 31.8667 103.2667
BGWO 2 18.3333 8.0667 8 34.4

Table 14
Performance comparison of various approaches for different test systems using Quade test.

10 unit system 40 unit system 100 unit system Overall rank

Rank p-value Rank p-value Rank p-value

BPSO 3.75 3.883333 4 11.63333

QEA 5.9 5.4375 5.591667 16.92917
QBPSO 4.975 1.67E−15 5.5625 0 5.408333 0 15.94583
BFWA 2.641667 2.925 2.133333 7.7
BGWO 1 2.516667 2.191667 2.866667 7.575
BGWO 2 1.216667 1 1 3.216667

Table 15
Wilcoxon pairwise comparison for 10 unit system.

BPSO QEA QBPSO BFWA BGWO 1 BGWO 2

BPSO 1 6.10E−05 0.000122 6.10E−05 0.488708 6.10E−05

QEA 6.10E−05 1 0.005371 6.10E−05 0.000183 6.10E−05
QBPSO 0.000122 0.005371 1 6.10E−05 6.10E−05 6.10E−05
BFWA 6.10E−05 6.10E−05 6.10E−05 1 0.406067 6.10E−05
BGWO 1 0.488708 0.000183 6.10E−05 0.406067 1 0.003906
BGWO 2 6.10E−05 6.10E−05 6.10E−05 6.10E−05 0.003906 1

Table 16
Wilcoxon pairwise comparison for 40 unit system.

BPSO QEA QBPSO BFWA BGWO 1 BGWO 2

BPSO 1 6.10E−05 6.10E−05 0.000183 0.000183 6.10E−05

QEA 6.10E−05 1 0.910156 6.10E−05 6.10E−05 6.10E−05
QBPSO 6.10E−05 0.910156 1 6.10E−05 6.10E−05 6.10E−05
BFWA 0.000183 6.10E−05 6.10E−05 1 0.008362 6.10E−05
BGWO 1 0.000183 6.10E−05 6.10E−05 0.008362 1 6.10E−05
BGWO 2 6.10E−05 6.10E−05 6.10E−05 6.10E−05 6.10E−05 1

Table 17
Wilcoxon pairwise comparison for 100 unit system.

BPSO QEA QBPSO BFWA BGWO 1 BGWO 2

BPSO 1 6.10E−05 6.10E−05 6.10E−05 6.10E−05 6.10E−05

QEA 6.10E−05 1 0.719727 6.10E−05 6.10E−05 6.10E−05
QBPSO 6.10E−05 0.719727 1 6.10E−05 6.10E−05 6.10E−05
BFWA 6.10E−05 6.10E−05 6.10E−05 1 0.00116 6.10E−05
BGWO 1 6.10E−05 6.10E−05 6.10E−05 0.00116 1 6.10E−05
BGWO 2 6.10E−05 6.10E−05 6.10E−05 6.10E−05 6.10E−05 1

References
[1] B. Venkatesh, T. Jamtsho, H.B. Gooi, Unit commitment - a fuzzy mixed integer
linear programming solution, Gener. Transm. Distrib. IET 1 (5) (2007) 836–846.
[2] Chung-Ching Su, Yuan-Yih Hsu, Fuzzy dynamic programming: an application to
unit commitment, Power Syst. IEEE Trans. 6 (3) (1991) 1231–1237.
[3] T. Senjyu, K. Shimabukuro, K. Uezato, T. Funabashi, A fast technique for unit
commitment problem by extended priority list (in)Power Syst. IEEE Trans. on 18
(2) (2003) 882–888.
[4] Arthur I. Cohen, Miki Yoshimura, A Branch-and-Bound Algorithm for Unit
Commitment, in, IEEE Trans. Power Appar. Syst. 102 (2) (1983) 444–451.
[5] W. Ongsakul, N. Petcharaks, Unit commitment by enhanced adaptive Lagrangian
relaxation, IEEE Trans. Power Syst. 19 (2004) 620–628.
[6] D. Srinivasan, J. Chazelas A priority list based evolutionary algorithm to solve large
scale unit commitment problem. in: International conference on power system
technology Powercon 2004. Singapore. p. 21–24.
[7] R.A. Jabr, Rank-constrained semidefinite program for unit commitment, Int J.
Electr. Power Energy Syst. 47 (2013) 13–20.
[8] S.A. Kazarlis, A.G. Bakirtzis, V. Petridis, A genetic algorithm solution to the unit
commitment problem, IEEE Trans. Power Syst. 11 (1) (1996) 83–92.
[9] A.Y. Saber, T. Senjyu, A. Yona, T. Funabashi, Unit commitment computation by

15
L.K. Panwar et al.
fuzzy adaptive particle swarm optimisation (in)Gener., Transm. Distrib., IET 1 (3)
(2007) 456–465.
[10] K. Vaisakh, L.R. Srinivas, Evolving ant colony optimization based unit commit-
ment, Appl. Soft Comput. 11 (2) (2011) 2863–2870.
[11] K.A. Juste, H. Kita, E. Tanaka, J. Hasegawa, An evolutionary programming solution
to the unit commitment problem (in)Power Syst., IEEE Trans. on 14 (4) (1999)
1452–1459.
[12] D.N. Simopoulos, S.D. Kavatza, C.D. Vournas, Unit commitment by an enhanced
simulated annealing algorithm, IEEE Trans. Power Syst. 21 (1) (2006) 68–76.
[13] J. Ebrahimi, S. Hosseinian, G. Gharehpetian, Unit commitment problem solution
using shuffled frog leaping algorithm, IEEE Trans. Power Syst. 26 (2011) 573–581.
[14] M. Moghimi Hadji, B. Vahidi, A solution to the unit commitment problem using
imperialistic competition algorithm, IEEE Trans. Power Syst. 27 (2012) 117–124.
[15] C.P. Cheng, C.W. Liu, C.C. Liu, Unit commitment by Lagrangian relaxation and
genetic algorithm, IEEE Trans. Power Syst. 15 (2000) 707–714.
[16] H. Balci, J. Valenzuela, Scheduling electric power generations using particle swarm
optimization combined with the Lagrangian relaxation method, Int. J. Appl. Math.
Comput. Sci. 14 (3) (2004) 411–421.
[17] K. Chandram, N. Subrahmanyam, M. Sydulu, Unit Commitment by improved pre-
prepared power demand table and Muller method, Int J. Electr. Power Energy Syst.
33 (2011) 106–114.
[18] Vikram Kumar Kamboj, S.K. Bath, J.S. Dhillon, Implementation of hybrid harmony
search/random search algorithm for single area unit commitment problem, Int. J.
Electr. Power Energy Syst. 77 (2016) 228–249.
[19] Y.-W. Jeong, J.-B. Park, J.-R. Shin, K.Y. Lee, A thermal unit commitment approach
using an improved quantum evolutionary algorithm, Elect. Power Compon. Syst. 37
(7) (2009) 770–786.
[20] T.W. Lau, C.Y. Chung, K.P. Wong, T.S. Chung, S.L. Ho, Quantum-inspired
evolutionary algorithm approach for unit commitment (in)Power Syst., IEEE Trans.
on 24 (3) (2009) 1503–1512.
[21] Seyedali Mirjalili, Seyed Mohammad Mirjalili, Andrew Lewis, ”Grey Wolf
Optimizer,”, Adv. Eng. Softw. 69 (2014) 46–61.
[22] V.K. Kamboj A novel hybrid PSO–GWO approach for unit commitment problem
(2015) Neural Computing and Applications, 13 p. Article in Press.
[23] S. Mohanty, B. Subudhi, P.K. Ray, A new MPPT design using grey wolf optimization
technique for photovoltaic system Under partial shading conditions (in)IEEE
Trans. Sustain. Energy 7 (1) (2016) 181–188.
[24] E. Emary, Hossam M. Zawbaa, Aboul Ella Hassanien, Binary grey wolf optimization
approaches for feature selection, Neurocomputing 172 (2016) 371–381.
[25] Lokesh Kumar Panwar, Srikanth Reddy K, Rajesh Kumar, Binary Fireworks
algorithm based thermal unit commitment (Article 4)Int. J. Swarm Intell. Res.
(IJSIR) 6 (2) (2014) (Article 4).
[26] S. Khanmohammadi, M. Amiri, M. Tarafdar Haque, A new three-stage method for
solving unit commitment problem, Energy (2010) 3072–3080.
[27] P. Sriyanyong, Y.H. Song, Unit commitment using particle swarm optimization
combined with lagrange relaxation, IEEE Trans. (2005) 1–8.
[28] Yun-Won Jeong, Woo-Nam Lee, Hyun-Houng Kim, Jong-Bae Park, Joong-
Rin Shin, Thermal unit commitment using binary differential evolution, J. Electr.
Eng. Technol. 4 (3) (2009) 323–329.
[29] J. Valenzuela, A.E. Smith, a seeded memetic algorithm for large unit commitment
problems, J. Heuristics 8 (2002) 173–195.
[30] Vikram Kumar Kamboj, S.K. Bath, J.S. Dhillon, Implementation of hybrid harmony
search/random search algorithm for single area unit commitment problem, Int. J.
Electr. Power Energy Syst. 77 (2016) 228–249.
[31] I.G.S. Damousis, A.G. Bakirtziz, P.S. Dokopoulos, A solution to the unit commit-
ment problem using integer coded genetic algorithm, IEEE Trans. Power Syst. 19
(2004) 1165–1172.
[32] E.R. Vasudevan Saravanan, D.P. Kothari, Unit commitment problem solution using
invasive weed optimization algorithm, Int. J. Electr. Power Energy Syst. 55 (2014)
21–28.
[33] A. Viana, J. Sausa, M. Matos Using, GRASP to solve the unit commitment problem,
Ann. Oper. Res. 120 (1) (2003) 117–132.
[34] I. Silva Jr, S. Carneiro Jr, E.J. De Oliveira, J.L.R. Pereira, P.A.N. Garcia,
A.L.M. Marcato, A Lagrangian multiplier based sensitive index to determine the unit
16
Swarm and Evolutionary Computation xxx (xxxx) xxx–xxx
commitment of thermal units, Int J. Electr. Power Energy Syst. 30 (2008) 504–510.
[35] T. Senjyu, H. Yamashiro, K. Shimabukuro, K. Uezato, T. Funabashi, A unit
commitment problem by using genetic algorithm based on characteristic classifi-
cation, IEEE/Power Eng. Soc. Winter Meet. 1 (2002) 58–63.
[36] B. Zhao, C.X. Guo, B.R. Bai, Y.J. Cao, An improved particle swarm optimization
algorithm for unit commitment, Elect. Power Energy Syst. 28 (7) (2006) 482–490.
[37] Xiaohui Yuan, Hao Nie, Anjun Su, Liang Wang, Yanbin Yuan, An improved binary
particle swarm optimization for unit commitment problem, Expert Syst. Appl. 36
(2009) 8049–8055.
[38] K. Roy, Solution of unit commitment problem using gravitational search algorithm,
Int J. Electr. Power Energy Syst. 53 (2013) 85–94.
[39] R. Quan, J. Jian, Y. Mu, Tighter relaxation method for unit commitment based on
second-order cone programming and valid inequalities, Int J. Electr. Power Energy
Syst. 55 (2014) 82–90.
[40] Raca Todosijević, Marko Mladenović, Saïd Hanafi, Nenad Mladenović, Igor Crévits,
Adaptive general variable neighborhood search heuristics for solving the unit
commitment problem, Int. J. Electr. Power Energy Syst. 78 (2016) 873–883.
[41] Z.L. Gaing Discrete particle swarm optimization algorithm for unit commitment.
in: IEEE power engineering society general meeting, vol. 1. p. 7-13.
[42] V.S. Pappala, I. Erlich A new approach for solving the unit commitment problem by
adaptive particle swarm optimization. in: Power and energy society general
meeting-conversion and delivery of electrical energy in the 21st century. USA:
IEEE. p. 1–6.
[43] V.S. Pappala, I. Erlich A new approach for solving the unit commitment problem by
adaptive particle swarm optimization. in: Power and energy society general
meeting-conversion and delivery of electrical energy in the 21st century. USA:
IEEE. p. 1–6.
[44] A.S. Eldin, M.A.H. El-sayed, H.K.M. Youssef A two-stage genetic based technique
for the unit commitment optimization problem. in: Proceedings of the 12th
international middle east power system conference, MEPCO, Aswan. p. 25–30.
[45] W. Xiong, M.J. Li, Y.L. Cheng An improved particle swarm optimization algorithm
for unit commitment. in: Proceedings of the 2008 international conference on
intelligent computation technology and automation, vol. 01. p. 1–5.
[46] Anupam Trivedi, Dipti Srinivasan, Subhodip Biswas, Thomas Reindl, Hybridizing
genetic algorithm with differential evolution for solving the unit commitment
scheduling problem, Swarm Evolut. Comput. 23 (2015) 50–64.
[47] Rammohan Mallipeddi, Ponnuthurai Nagaratnam SuganthanUnit commitment–a
survey and comparison of conventional and nature inspired algorithms,”, Int. J.
Bio-Inspired Comput. 6 (2) (2014) 71–90.
[48] Joaquín Derrac, Salvador García, Daniel Molina, Francisco Herrera, A practical
tutorial on the use of nonparametric statistical tests as a methodology for
comparing evolutionary and swarm intelligence algorithms, Swarm Evolut.
Comput. 1 (1) (2011) 3–18.
[49] Anupam Trivedi, Dipti Srinivasan, Subhodip Biswas, Thomas Reindl, A genetic
algorithm – differential evolution based hybrid framework: case study on unit
commitment scheduling problem, Inf. Sci. 354 (2016) 275–300.
[50] A. Trivedi, D. Srinivasan, K. Pal, T. Reindl, "A multiobjective evolutionary
algorithm based on decomposition for unit commitment problem with significant
wind penetration," 2016 IEEE Congress on Evolutionary Computation (CEC),
Vancouver, BC, 2016, pp. 3939–3946.
[51] A. Trivedi, D. Srinivasan, K. Pal, T. Reindl, A MOEA/D with Non-uniform Weight
Vector Distribution Strategy for Solving the Unit Commitment Problem in
Uncertain Environment, in: M. Wagner, X. Li, T. Hendtlass (Eds.), , Artificial Life
and Computational Intelligence. ACALCI 2017. Lecture Notes in Computer Science
10142, Springer, Cham, 2017.
[52] S. Kundu, S. Das, A.V. Vasilakos, et al., A modified differential evolution-based
combined routing and sleep scheduling scheme for lifetime maximization of
wireless sensor networks, Soft Comput. 19 (2015) 637.
[53] S. Biswas, D. Bose, S. Das, S. Kundu, Decomposition-based evolutionary multi-
objective optimization approach to the design of concentric circular antenna arrays,
Progress. Electromagn. Res. B 52 (2013) 185–205.
[54] S. Biswas, S. Das, P.N. Suganthan, C.A.C. Coello, "Evolutionary multi-objective
optimization in dynamic environments: A set of novel benchmark functions," 2014
IEEE Congress on Evolutionary Computation (CEC), Beijing, 2014, pp. 3192–3199.