Вы находитесь на странице: 1из 10

Ecient real coded genetic algorithm to solve

the non-convex hydrothermal scheduling problem


Sushil Kumar
*
, R. Naresh
Electrical Engineering Department, National Institute of Technology, Hamirpur (HP) 177 005, India
Received 23 June 2006; received in revised form 29 May 2007; accepted 3 June 2007
Abstract
A simple and ecient optimisation procedure based on real coded genetic algorithm is proposed for the solution of short-term hydro-
thermal scheduling problem with continuous and non-smooth/non-convex cost function. The constraints like load-generation balance,
unit generation limits, reservoir ow balance, reservoir physical limitations and reservoir coupling are also considered. The eectiveness
of the proposed algorithm is demonstrated on a multichain-cascaded hydrothermal system that uses non-linear hydro generation func-
tion, includes water travel times between the linked reservoirs, and considers the valve point loading eect in thermal units. The proposed
algorithm is equipped with an eective constraint-handling technique, which eliminates the need for penalty parameters. A simple strat-
egy based on allowing infeasible solutions to remain in the population is used to maintain diversity. The same problem is also solved
using binary coded genetic algorithm. The features of both algorithms are same except the crossover and mutation operators. In real
coded genetic algorithm, simulated binary crossover and polynomial mutation are used against the single point crossover and bit-ipping
mutation in binary coded genetic algorithm. The comparison of the two genetic algorithms reveals that real coded genetic algorithm is
more ecient in terms of thermal cost minimisation for a short-term hydrothermal scheduling problem with continuous search space.
2007 Elsevier Ltd. All rights reserved.
Keywords: Short-term hydrothermal scheduling; Constraint-handling; Valve point loading eect; Real coded genetic algorithm; Simulated binary cros-
sover; Polynomial mutation
1. Introduction
Short-term hydrothermal scheduling is a crucial task in
the economic operation of a power system. A good gener-
ation schedule reduces the production cost, increases the
system reliability, and maximises the energy capability of
reservoirs. By utilizing the limited water resource, the main
objective of short-term hydrothermal scheduling is to deter-
mine the optimal amount of generated powers for hydro
and thermal units such that the total fuel cost of thermal
units can be minimised over a short time horizon subject
to various equality and inequality constraints. The equality
constraints may include generation-demand balance,
dynamic balance of reservoir storage for cascaded hydro
plants, boundary conditions for the reservoir storage, etc.
Among the inequality constraints are the generation limits
on the hydro and thermal generators, bounds on the reser-
voir storage levels and turbine discharge rates, etc. Further,
the valve point loading eect (VPLE) of thermal units com-
plicates the non-linearity in the problem [1,2].
All the above requirements as well as electric and
hydraulic coupling create a multidimensional, non-linear,
non-convex, large-scaled, and combinatorial problem.
Therefore, short-term hydrothermal scheduling problem
poses a signicant challenge to the scheduler, as it possesses
a large parametric space to choose from, a possibility of
large infeasible and non-uniform areas, and the presence
of numerous local optima. To solve this problem, either
analytical gradients or numerical researches have been
largely resorted to in the past [24]. These techniques lead
0142-0615/$ - see front matter 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ijepes.2007.06.001
*
Corresponding author. Tel.: +91 9418177999.
E-mail addresses: skhpindian@gmail.com (S. Kumar), rnareshnith@
gmail.com (R. Naresh).
www.elsevier.com/locate/ijepes
Available online at www.sciencedirect.com
Electrical Power and Energy Systems 29 (2007) 738747
only to locally optimal solutions. The problem is sometimes
simplied to satisfy the optimisation technique, resulting in
a solution that is optimal for only an approximate problem.
With the emergence of articial and computational intel-
ligence technology, attention has been gradually shifted to
applications of such technology-based approaches to han-
dle the complexity involved in real world problems.
Recently, researchers have shown keen interest in solving
generation scheduling problems using articial intelligent
techniques such as articial neural networks (ANN), simu-
lated annealing, tabu search, genetic algorithm (GA), evo-
lutionary programming (EP) and have got successful results
to certain level. A review of these techniques can be found
in [36]. While the ANN [79] and other techniques suer
from stability problem and require gradient information
which often leads to local optima, the techniques based
on GA [1014] and EP [1517] are more suitable for han-
dling such complex problems.
In applications [1014], binary coded genetic algorithm
(BCGA) has been reported to give near optimal results for
generation scheduling problem whereas the performance
of real coded genetic algorithm (RCGA) for solving this
problem needs to be investigated. GAs dier from conven-
tional search methods as they (i) search from a population
of solutions (and not a single solution at a time); (ii) use
probabilistic transition rules to propagate towards global
optimum; (iii) require no gradient information; (iv) implic-
itly handle any constraints in the optimisation problem;
(v) have a built-in global search mechanism and; (vi) do
not care about the discontinuities in the variables and
therefore easily adapt to the irregular search space of the
optimisation problem.
In the present work, a comparison of the computational
eciency of BCGA and RCGA is presented while solving
the non-convex and continuous hydrothermal scheduling
problem with a new constraint-handling technique. The
following section presents a detailed review of the mathe-
matical model of short-term scheduling of hydrothermal
power systems. Section 3 summarises the principal aspects
of GA with a brief comparison of RCGA and BCGA. The
optimisation scheme of GA for solving the hydrothermal
generation scheduling problem is elaborated in Section 4.
Simulation and results of the hydro generation scheduling
problem are presented in Section 5 followed by the con-
cluding remarks.
2. Problem description
The short-term hydrothermal scheduling problem can
be stated as to nd out water release from each reservoir
and through each power house, and the corresponding
thermal unit generations over all the planning time inter-
vals so as to minimise the total cost of fuel for thermal gen-
eration while satisfying diverse hydraulic, thermal and load
balance constraints. Typically, the total planning period is
one day or one-week and each planning time interval is one
hour.
2.1. Notation
a
k
, b
k
, c
k
cost coecients of kth thermal unit
C
1j
C
6j
generation coecients of jth hydro plant
j index for reservoir or hydro plant
k index for thermal unit
Ng total number of thermal plants
Nh total number of hydro plants
Nu
j
number of upstream plants for jth reservoir
Pd
t
load demand in tth time interval
Pg
t
k
power generation of kth thermal unit at t
Ph
t
j
power generation of jth hydro plant at t
Pl
t
total transmission losses at time t
q
t
j
water discharge rate of jth plant at t
q
k
, r
k
valve-point coecients of kth thermal unit
t index for time interval
T total planning (time) horizon
s
m
j
water transport delay from mth reservoir to jth res-
ervoir
U
j
vector of indices of the plants upstream to the jth
reservoir
x
t
j
storage volume of jth reservoir at t
x
begin
j
storage of jth reservoir at the beginning of the
planning horizon
x
end
j
storage of jth reservoir at the end of the planning
horizon
y
t
j
side or river inow to jth reservoir at t
z
t
j
inow from upstream plant to jth reservoir at t
2.2. Hydropower generation characteristics
Hydropower generation is a function of net head and
turbine discharge. Generally, constant water head is
assumed in short-term scheduling. However, this assump-
tion is true only in case of large capacity reservoirs. Head
variation cannot be ignored if there is a strong relationship
between inow and capacity. Since the net head is a func-
tion of stored water, hydropower generation Ph can be
written in terms of turbine discharge rate q and storage
x, and a frequently used expression [8,10] is
Ph
t
j
C
1j
x
t
j

2
C
2j
q
t
j

2
C
3j
x
t
j
q
t
j
C
4j
x
t
j
C
5j
q
t
j
C
6j
;
1
where j = 1, 2, . . . , Nh and t = 1, 2, . . . , T. Here Nh is the to-
tal number of hydro plants and T is the total planning
(time) horizon. C
1
C
6
are the generation coecients of hy-
dro plants.
2.3. Objective function and constraints
The objective function is taken as to minimise the sum-
mation of the fuel cost for all the thermal units over the
complete planning period. The unit fuel cost is generally
assumed a quadratic function of unit thermal power Pg.
S. Kumar, R. Naresh / Electrical Power and Energy Systems 29 (2007) 738747 739
However, due to the wire drawing eects occurring during
the opening of each steam admission valve in a turbine, a
rippling eect is produced on the thermal-unit inputout-
put curve and the fuel cost is no more a quadratic function
[1]. Mathematically, the non-linear and non-smooth fuel
cost function F of the thermal plants is expressed as
F min
X
T
t1
X
Ng
k1
a
k
Pg
t
k

2
b
k
Pg
t
k
c
k
jq
k
sinr
k
Pg min
k
Pg
t
k
j; 2
where a, b, c are the cost coecients and q, r are the valve-
point coecients of thermal units. Ng is the total number
of thermal plants.
This model is subject to the following constraints:
1. System active load balance:
Total generated power is equal to the total demand Pd
including losses Pl in each time interval
X
Ng
k1
Pg
t
k

X
Nh
j1
Ph
t
j
Pd
t
Pl
t
3
2. Thermal and Hydro power limits:
Thermal and hydro units can generate power between
specied upper and lower limits
Pg min
k
6 Pg
t
k
6 Pg max
k
; k 1; 2; . . . ; Ng 4
Ph min
j
6 Ph
t
j
6 Ph max
j
; j 1; 2; . . . ; Nh 5
3. Reservoir ow balance (continuity equation):
In this constraint, water transportation delay s between
reservoirs is considered. The ow balance equation
relates the storage during previous interval with the stor-
age, net inowy plus z, and net outow during the cur-
rent time interval
x
t
j
x
t1
j
y
t
j
q
t
j
z
t
j
; where z
t
j

X
Nuj
m1
q
ts
m
j
Ujm
6
Here, Nu gives the number of upstream plants for any
reservoir and U gives indices of the plants upstream to
any reservoir.
4. Physical constraints:
Due to physical limitations, following minimum and
maximum limits may exist for jth reservoir or hydro
unit:
Reservoir storage volume x min
j
6 x
t
j
6 x max
j
Water discharge rate q min
j
6 q
t
j
6 q max
j
7
5. Coupling constraint:
Terminal reservoir volumes x
begin
and x
end
are previ-
ously set by the midterm scheduling process. This equal-
ity constraint implies that the total quantity of available
water should be used
x
0
j
x
begin
j
and x
T
j
x
end
j
8
3. An overview of genetic algorithms
GA is a general purpose search procedure that uses
principles inspired by natural genetic populations to
evolve solutions to problems [1820]. The basic idea is
to maintain a population of chromosomes representing
candidate solutions to the concrete problem, which
evolves over time through a process of competition and
controlled variation. Each chromosome in the population
has an associated tness to determine which chromosomes
are used to form new ones in the competition process,
which is called selection. The new ones are created using
genetic operators such as crossover and mutation. A
major reason for the success of GA is its ability to exploit
the information accumulated about an initially unknown
search space in order to bias subsequent searches into use-
ful subspaces. This is their key feature, particularly in
large, complex, and poorly understood search spaces,
where classical search tools are inappropriate, oering a
valid approach to problems requiring ecient and eective
search techniques. The basic procedure of GA consists of
the following steps:
Step 1. Set the generation counter n = 0. Initialize ran-
domly the genes of each chromosome in the pop-
ulation gen(0). Evaluate the tness of each
chromosome.
Step 2. Generate gen(n + 1) from gen(n) as follows:
(i) Select chromosomes from gen(n) based on their
tness.
(ii) Recombine them using genetic operators.
(iii) Evaluate tness of chromosomes in current
population.
Step 3. If the terminal condition is satised, stop and
return the best-t chromosome. Otherwise set
n = n + 1 and go to step 2.
3.1. Real-coded versus binary-coded chromosomes
GAs good properties do not stem from the use of bit
strings for chromosomes. The binary representation meets
certain diculties when dealing with continuous search
space. One diculty is the Hamming clis associated with
certain strings (such as 01111 and 10000) from which a
transition to a neighbouring solution (in real space)
requires the alteration of many bits. The other diculty
is the inability to achieve any arbitrary precision in the
optimal solution. On the other hand, the real number rep-
resentation seems particularly natural when optimisation
problems with variables in continuous search spaces are
tackled. Since real-parameters are used directly (without
any bit-coding), solving real-parameter optimisation prob-
lems is a step easier when compared to the BCGA. Unlike
in the BCGA, decision variables can be directly used to
compute the tness values. Since the selection operator
works with the tness value, any selection operator used
with BCGA can also be used in RCGA. However, the
740 S. Kumar, R. Naresh / Electrical Power and Energy Systems 29 (2007) 738747
diculty arises with the search operators. In the BCGA,
exchanging portions of two nite-length parent strings is
easier to implement and visualize the crossover. Simply
ipping a bit to perform mutation is also convenient and
resembles a natural mutation event. In RCGA, the main
challenge is how to use a pair of real-parameter decision
variable vectors to create a new pair of ospring vectors
or how to perturb a decision variable vector to a mutated
vector in a meaningful manner. A detailed discussion on
RCGA can be found in [20,21].
4. GA based solution approach for hydrothermal scheduling
The general scheme of the genetic algorithm for short-
term hydrothermal scheduling is summarised as follows:
4.1. Solution representation and population initialization
The water discharge rates of various reservoirs in vari-
ous intervals are used as decision variables in this work.
An initial parent population of N
p
chromosomes is created
randomly using uniform distribution, in which the decision
variables are chosen from respective feasible ranges. The
trial chromosome designating the rth individual of the pop-
ulation is given as below:
Q
r
q
1
; q
2
; . . . ; q
j
; . . . ; q
Nh
where
q
j
q
1
j
; q
2
j
; . . . ; q
t
j
; . . . ; q
T
j
9
Thus, Q
r
is a vector encoding Nh T independent vari-
ables. However, in order to meet out the nal reservoir sto-
rages in (8), a dependent interval d 2 [1T] is chosen
arbitrarily and Q
r
is modied as below:
Q
r
q
1
; q
2
; . . . ; q
j
; . . . ; q
Nh
where
q
j
q
1
j
; q
2
j
; . . . ; q
d1
j
; q
d1
j
; . . . ; q
t
j
; . . . ; q
T
j
10
Clearly, Q
r
is now a vector encoding Nh (T 1) indepen-
dent variables. The discharge in the dth (dependent) inter-
val is then calculated by
q
d
j
x
T
j
x
0
j

X
T
t1
t6d
q
t
j

X
T
t1
y
t
j

X
Nuj
m1
X
T
t1
q
ts
m
j

Ujm
11
The chromosome representation of (10) along with (11)
inherently avoids the GA to be misled into generating solu-
tions with no water in the reservoirs at the end of the plan-
ning horizon. For BCGA, each gene q
t
j
in (10) is encoded as
a xed-length binary substring with suitable and reason-
able resolution. For RCGA, each gene is represented by
the decimal value of q
t
j
.
4.2. Constraint-handling and tness evaluation
The short-term hydrothermal scheduling problem for-
mulated so far can be expressed in a general way as a
non-linear programming problem of the following kind:
Minimise f Q; the objective function
subject to g
l
Q 6 0; l 1; 2; . . . ; L
12
In (12), there are L lesser-than-equal-to type inequality
constraints g(Q) where Q is given by (10). The equality
constraints are not shown, as they are all satised. Gener-
ally, (12) is handled by using a penalty function pf where
infeasible solutions are penalised depending on the amount
of constraint violation:
pf Q; R f Q R
X
l
hg
l
Qi
2
" #
13
where h i denotes the absolute value of the operand, if the
operand is positive and returns a zero value otherwise; R
is the penalty parameter and xing a correct penalty
parameter is an essential element of a successful simulation.
However, xing the value of R is always problem specic
and may not work for all sort of problems.
In this work, we have used a constraint-handling
method [20] based on penalty function approach that does
not require any penalty parameter:
pf Q
f Q; if Q is feasible;
f
max

P
L
l1
hg
l
Qi; otherwise:
8
<
:
14
The parameter f
max
is the objective function value of the
worst feasible solution in the current population. The pen-
alty function of a particular solution is then converted to
its corresponding tness value as below:
fit pf
max
pf Q 15
where pf
max
is the highest among all the N
p
penalty func-
tions in the current population. It is evident from (15) that
a better solution would have higher tness. This constraint-
handling technique works only with population based ap-
proaches such as GA.
It may be noted that if the discharge rate of a hydro-
plant for the dependent interval goes outside permissible
region then the corresponding solution is infeasible and
has a higher value of pf. Such a solution has lower tness
and poor chance of survival.
4.3. Competition and selection
A selection mechanism is implemented to determine the
number of copies of each parent chromosome that can par-
ticipate in the crossover process. The tournament selection
is well suited for the proposed constraint-handling scheme.
During tournament selection, N
p
sets of S randomly picked
chromosomes are formed where S is the tournament size. A
competition takes place among the chromosomes within
each set to select the best one. During this competition pro-
cess, a comparison mechanism is used in which feasible
solutions are preferred over infeasible ones, infeasible solu-
tions with smaller violations are preferred over solutions
with larger violations, and feasible solutions with lower
objective function values are preferred over other feasible
S. Kumar, R. Naresh / Electrical Power and Energy Systems 29 (2007) 738747 741
solutions. Hence, there will be selective pressure towards
the feasible region. The N
p
chromosomes in the survivor
set are new parents for the next generation.
4.4. Creation of ospring population
Equal number of ospring Q
0
r
, r = 1, 2, . . . , N
p
, are gen-
erated by subjecting the parents to crossover, mutation
and elite-preserving operators. These ospring may be sub-
jected to a few other operators for better search as
described below.
4.4.1. Crossover operator
The crossover operator is mainly responsible for the glo-
bal search property of the GA. For the genetic search to be
eective, every possible point in the search space must be
reachable from the initial population through crossover
only. For BCGA, the single point crossover [18] has been
used with a crossover rate pcross. For RCGA, the Simu-
lated Binary Crossover (SBX) [20] has been used in the
present work as explained in appendix.
4.4.2. Mutation operator
In order to prevent the permanent loss of any particular
gene value, the mutation operator is used. However, muta-
tion is typically a secondary operator and cannot be relied
upon for reaching the entire search space. For BCGA, the
bit-ipping mutation has been used with a variable muta-
tion rate, which decreases gradually from a specied max-
imum value mut
max
to a minimum value mut
min
with the
increase in the number of generations [19] as given below:
mutgen mut
max
1 gen=gen
max
mut
min
where gen
max
is the maximum number of generations in the
GA. For RCGA, the polynomial mutation [20] has been
used as explained in the appendix.
4.4.3. Elite-preserving operator
An elitist strategy retains intact a copy of the best solu-
tion in successive generations. The elite-preserving opera-
tor can be implemented simply by directly copying the
best e% chromosomes from the current population to the
next generation. The rest (100 e)% of the new population
is created by the crossover applied on the entire current
population (including the chosen e% elite members).
4.4.4. Replacement operator
In order to maintain pressure towards global optimum,
a chromosome among the children is replaced with a better
chromosome among the parents to form the population for
the next generation using the previously mentioned com-
parison mechanism. Replacement is done with a probabil-
ity p
rep
.
4.4.5. Diversity mechanism
Diversity is maintained to prevent premature conver-
gence. A diversity mechanism [22] is applied on the chro-
mosomes that do not undergo replacement by allowing
the best infeasible solution (from either the parents or the
ospring population with 50% probability each) to be cop-
ied in the population for the next generation with a prob-
ability p
div
. The best infeasible solution is the one having
the best value of the objective function and with the lowest
amount of constraint violation.
4.5. Stopping rule
The processes of generating new chromosomes and
selecting those with better function values are continued
until the given stopping conditions are satised. The pro-
cess can be stopped after a xed number of generations,
or when any signicant improvement in the solution ceases
to occur. In this paper, GA is run for a xed number of
generations.
4.6. Steps of GA for hydrothermal scheduling problem
4.6.1. Main procedure
1. Scan input data and topology of the interconnected
hydro system.
2. Randomly select a dependent interval from the plan-
ning horizon. The discharge rates q
t
j
are represented
by (10) as a Nh (T 1) dimensional vector Q in
which the discharge rates of each reservoir in the
dependent interval are not included.
3. Set the generation counter to zero, n = 0.
4. Using uniform distribution, randomly generate initial
population gen(0) of trial vectors Q
r
for r =
1, 2. . . , N
p
from a feasible range in each dimension.
5. Evaluate tness of each individual of current
population.
6. Increase the generation counter by one, n = n + 1.
7. Apply GA to produce gen(n + 1): Preserve the elite
chromosomes. Select N
p
parents from population of
trial vectors using tournament selection. Recombine
them using crossover and mutation operators to pro-
duce N
p
child vectors.
8. Evaluate tness of each chromosome of current popu-
lation and store the solution corresponding to the best-
t chromosome.
9. Apply the replacement operator and diversity mecha-
nism to the current population.
10. Check for the convergence criterion: If current genera-
tion number n is equal to the maximum generations,
stop and print the results corresponding to the best-
t vector of the population. Otherwise, go to step 6.
4.6.2. Subroutine for tness evaluation
1. Calculate using (11) the discharge rates of each
reservoir in the dependent interval for the entire
population.
742 S. Kumar, R. Naresh / Electrical Power and Energy Systems 29 (2007) 738747
2. Calculate storage (6) of the reservoirs using current val-
ues of water discharge rates for the entire population.
3. Using discharge rates and storage, determine hydro
plant generations (1) for the entire population.
4. Calculate thermal power generation using the load bal-
ance equation (3) for the entire population.
5. Evaluate the inequality constraints (4), (5), and (7)
using current values of thermal powers, hydro powers,
discharge rates, and storage for the entire population.
6. Calculate the objective function values (2) for entire
population.
7. Using the objective function and inequality constraint
values, calculate the tness (15) of each individual.
4.7. Test system
In this work, the unit commitment problem has been
assumed to be already solved and only the dispatch prob-
lem is considered. In addition, the cost functions of the
number of committed thermal units are represented by a
composite fuel cost function [2]. The performance of the
proposed GA is veried on a test system that comprises
an equivalent thermal unit and a multichain-cascade of sev-
eral hydro units with no transmission losses. The planning
horizon is 24 h with 1 h intervals. The hydraulic subsystem
is characterised by the following:
(i) A multichain-cascade ow network, with all the
plants on one stream.
(ii) River transport delay between successive reservoirs.
(iii) Variable head hydro plants.
(iv) Variable natural inow rates into each reservoir.
(v) Variable load demand over scheduling period.
This hydraulic test network models most of the com-
plexities encountered in practical hydro networks. To make
the thermal model more realistic, the cost function of the
equivalent thermal unit includes VPLE as shown in (16).
The rest of the data of the test system considered here is
same as in [10].
f Pg 5000 19:2Pg 0:002Pg
2
j2500 sin0:0075Pg min Pgj 16
5. Simulation and results
This section employs two cases to illustrate the eective-
ness of RCGA against BCGA with respect to the quality of
the solution obtained. The rst case solves the hydrother-
mal scheduling problem for above test system without con-
sidering the VPLE whereas the second case includes the
VPLE. All of the programs were implemented in MAT-
LAB 7 on a PC (Pentium-IV, 512 MB, 3.0 GHz).
The GA parameters were tuned on convergence test
with respect to some nearby optimal value by trial and
error procedure. For the sake of comparison of BCGA
and RCGA, the common GA parameters were kept
same:
COMMON: N
p
50; e 20; S 6; p
rep
0:2;
and p
div
0:03:
BCGA: pcross 1; resolution 0:01; mut
max
0:009;
mut
min
0:0001
RCGA: p
c
0:5; g
c
0:1; g
mmin
100
The programs were run for dierent values of gen
max
start-
ing from 500 until the results cease to show a gradual and
large improvement in the minimum cost. The results for ten
values of gen
max
from 500 to 5000 are presented here. The
discussion of results for GA applies to both the BCGA as
well as RCGA unless otherwise stated explicitly.
5.1. Simulation and results without considering the valve-
point loading eect
To investigate the eect of initial solutions (population
of trial vectors), the programs were executed with 20 dier-
ent randomly generated initial solutions. The statistical
results for the case without considering the VPLE are
reported in Table 1.
It is evident from Table 1 that RCGA has outperformed
BCGA in terms of cost minimisation and execution time.
The BCGA suers from premature convergence problem.
The minimum generation cost obtained with BCGA is
$926922.71 whereas RCGA gives a much better cost in just
one-fth of the gen
max
as used by BCGA. The minimum
value of cost obtained with RCGA is $925932.75. Conver-
gence characteristics for the two trial runs giving minimum
cost for BCGA and RCGA are compared in Fig. 1.
It may be noted that for the same case solved in [10]
using BCGA, the minimum cost is $926707 with certain
end reservoir volume violations. In the present results,
there is no constraint violation and RCGA gives a much
better cost. Moreover, in comparison to [10], the popula-
tion size is just half and the value of gen
max
is much smaller
in the present work.
During the execution of the GA code, it has been
observed that in the beginning, i.e. for gen(0), even the
Table 1a
Statistical test results of 20 runs for BCGA
gen
max
Cost (in $) Avg. time (in s)
Minimum Average Maximum
500 928361.63 929789.21 931869.42 6.47
1000 927135.14 928578.16 930606.58 12.87
1500 927476.89 928306.39 929206.82 19.48
2000 927473.84 928053.04 928736.78 25.81
2500 927207.37 927895.15 928723.67 32.24
3000 927084.98 927924.39 928475.93 39.29
3500 926985.81 927913.46 929642.95 45.19
4000 926970.03 927698.36 928577.03 51.49
4500 927004.15 927877.97 929309.39 58.10
5000 926922.71 927815.35 929451.09 64.51
S. Kumar, R. Naresh / Electrical Power and Energy Systems 29 (2007) 738747 743
best-t solution chromosome was in the infeasible region
with a large constraint violation and a large thermal gener-
ation cost. However, with the progress of the code execu-
tion, water discharge rate solution trajectory quickly
moved into the feasible region decreasing the generation
cost at fast rate. This is because the modelling of the solu-
tion chromosomes as depicted from Section 4.1 has already
eliminated the end reservoir volume constraint. Therefore,
the end reservoir volume requirements are always satised
and the algorithm has to satisfy only the inequality con-
straints on discharge rate, storage volume, hydro genera-
tion, and thermal generation. Hence, most of the time the
GA searches in the feasible region. It may be noted that
in [10], all the constraints including end reservoir volume
had been handled with the help of penalty multipliers;
hence, whereas all other constraints were satised, the
end reservoir volumes were violated. The notable feature
of the present work is that need for penalty multipliers to
handle the constraints has been eliminated. However, this
constraint-handling technique can be applied only with
population-based approaches.
The interval wise hydro discharges, reservoir storage
volumes, and power generations are shown in Figs. 24,
respectively. These plots correspond to the results of the
trial giving minimum cost by RCGA.
It can be seen from Fig. 2 that some of the releases in
initial time intervals are scheduled at their minimum per-
missible values. This is because load demand for rst 5
6 h is relatively low and therefore less hydro-energy is
required to partially meet this load demand. It is evident
that optimal solution here lies at the boundary of the fea-
sible region.
Looking at Fig. 3, it is observed that there is a sharp
fall in the storage of 4th reservoir during hours 14. This
occurs mainly due to delay in water travel from 3rd
Table 1b
Statistical test results of 20 runs for RCGA
gen
max
Cost (in $) Avg. time (in s)
Minimum Average Maximum
500 927159.71 927703.55 928919.26 6.08
1000 926359.53 926846.80 927504.38 12.01
1500 926195.22 926549.33 927658.23 17.92
2000 926060.12 926298.64 926972.26 23.78
2500 925977.53 926212.50 927394.17 29.45
3000 925985.55 926194.14 927374.97 35.01
3500 925967.58 926266.29 926811.97 40.60
4000 925937.24 926163.29 926995.89 46.29
4500 925932.75 926138.82 926896.81 51.67
5000 925940.03 926120.26 926538.81 57.52
0 1000 2000 3000 4000 5000
9.25
9.3
9.35
9.4
9.45
9.5
x10
5
Number of Generation
C
o
s
t
,

$
RCGA
BCGA
Fig. 1. Convergence with quadratic cost curve ignoring VPLE.
1 6 12 18 24
5
10
15
20
25
30
Number of Interval
H
y
d
r
o

D
i
s
c
h
a
r
g
e
Reservoir 1
Reservoir 2
Reservoir 3
Reservoir4
Fig. 2. Hydro plant discharges (10
4
m
3
).
Number of Interval
0 1 6 12 18 24
60
80
100
120
140
160
180
R
e
s
e
r
v
o
i
r

S
t
o
r
a
g
e

V
o
l
u
m
e
s
Reservoir 1
Reservoir 2
Reservoir 3
Reservoir 4
Fig. 3. Hydro reservoir storage volumes (10
4
m
3
).
1 6 12 18 24
0
500
1000
1500
2000
2500
Number of Interval
D
e
m
a
n
d

a
n
d

G
e
n
e
r
a
t
i
o
n
,

M
W
Load Demand
Thermal Generation
Hydro Generation
Fig. 4. Load demand and the corresponding hydro and thermal
generation.
744 S. Kumar, R. Naresh / Electrical Power and Energy Systems 29 (2007) 738747
reservoir and no side inow to 4th reservoir. It is also noted
that the storage level for 4th reservoir did reach its upper
bound value during hours 1520.
5.2. Simulation and results considering the valve-point
loading eect
The eect of initial solutions on both algorithms was
investigated with 20 dierent randomly generated initial
solutions. The statistical results for this case considering
the VPLE are reported in Table 2.
It is evident from Table 2 that RCGA has again outper-
formed BCGA in terms of cost minimisation and execution
time and the BCGA suers from premature convergence
problem. The minimum generation cost obtained with
BCGA is $952618.00 whereas RCGA gives a much better
cost in just one-fth of the gen
max
as used by BCGA.
The minimum value of cost obtained with RCGA is
$951451.35. Once again, it was observed that most of the
time the proposed GA searches in the feasible region and
Table 2a
Statistical test results of 20 runs for BCGA
gen
max
Cost (in $) Avg. time (in s)
Minimum Average Maximum
500 954006.62 955782.07 958019.25 6.57
1000 953784.57 955091.28 957606.70 13.08
1500 953156.09 954639.82 956453.54 19.62
2000 953090.92 954413.82 955638.67 26.15
2500 952963.44 953986.22 955196.83 32.58
3000 953125.99 954083.71 955117.09 39.07
3500 952831.02 954063.04 955661.00 45.62
4000 952958.14 954117.87 955470.26 52.70
4500 952997.79 953898.30 954901.81 59.48
5000 952618.00 953609.15 954285.03 66.30
Table 2b
Statistical test results of 20 runs for RCGA
gen
max
Cost (in $) Avg. time (in s)
Minimum Average Maximum
500 953978.65 955731.90 958412.04 6.14
1000 952245.47 953298.42 954657.31 12.14
1500 951820.16 952722.30 953857.22 18.04
2000 951810.25 952502.34 953035.02 23.77
2500 951734.46 952502.87 953685.06 29.38
3000 951842.27 952543.47 953901.94 35.08
3500 951675.51 952098.66 952854.52 40.55
4000 951626.23 952051.01 952815.93 46.04
4500 951451.35 952142.51 953800.96 51.74
5000 951559.24 952009.32 952932.96 57.32
0 1000 2000 3000 4000 5000
0.95
1
1.05
x 10
6
Number of Generation
C
o
s
t
,

$
RCGA
BCGA
Fig. 5. Convergence with valve point loading eect.
1 6 12 18 24
5
10
15
20
25
30
Number of Interval
H
y
d
r
o

D
i
s
c
h
a
r
g
e
Reservoir 1
Reservoir 2
Reservoir 3
Reservoir 4
Fig. 6. Hydro plant discharges (10
4
m
3
) with VPLE.
0 1 6 12 18 24
60
80
100
120
140
160
180
Number of Interval
R
e
s
e
r
v
o
i
r

S
t
o
r
a
g
e

V
o
l
u
m
e
s
Reservoir 1
Reservoir 2
Reservoir 3
Reservoir 4
Fig. 7. Hydro reservoir storage volumes (10
4
m
3
) with VPLE.
1 6 12 18 24
0
500
1000
1500
2000
2500
Number of Interval
D
e
m
a
n
d

a
n
d

G
e
n
e
r
a
t
i
o
n
,

M
W
LoadDemand
Thermal Generation
Hydro Generation
Fig. 8. Load demand and the corresponding hydro and thermal
generation.
S. Kumar, R. Naresh / Electrical Power and Energy Systems 29 (2007) 738747 745
all the solutions reported are feasible solutions. Conver-
gence characteristics for the two trial runs giving minimum
cost for BCGA and RCGA are compared in Fig. 5.
The interval wise reservoir hydro discharges, storage
volumes, and power generations are shown in Figs. 68
respectively. These plots correspond to the results of the
trial giving minimum cost by RCGA. The eect of valve
point loading can be easily visualised through the poor
smoothness of these plots as compared to Figs. 24.
6. Conclusions
Eective GA with both binary and real coding has been
designed and implemented successfully for solving the
hydrothermal generation scheduling problem. In the pres-
ent work, the complexity in the form of non-linearity and
non-convexity involved in the scheduling problem has been
handled in a very simple and eective manner. The beauty
of the proposed algorithms is that it comes into the feasible
region after a few generations and then searches for better
optima in the feasible region. Hence, there is no constraint
violation in the results presented here. The results obtained
by RCGA are better than the results by BCGA. The pres-
ent study reveals that for handling a non-linear and non-
convex hydrothermal generation scheduling problem with
continuous search space, the real coded GA has various
merits such as: simple concept; easy implementation; better
eciency than binary coded GA; and requirement of a
small population size.
Acknowledgement
This work was supported by Council of Scientic &
Industrial Research, New Delhi, India under Grant
22(0372)04/EMR-II.
Appendix A. Simulated binary crossover (SBX) operator
Consider decision parameters q
t
j1
and q
t
j2
belonging to
parent solutions Q
1
and Q
2
, respectively, and q
t0
j1
and q
t0
j2
belonging to child solutions Q
0
1
and Q0
2
, respectively. The
procedure of computing q
t0
j1
and q
t0
j2
from q
t
j1
and q
t
j2
is as
follows:
1. Create a random number u between 0 and 1.
2. Find a parameter c using a polynomial probability dis-
tribution as follows:
c
ua
1=g
c
1
; if u 6 1=a;
1=2 ua
1=g
c
1
; otherwise;
(
where a 2 b
g
c
1
and b is calculated as follows:
b 1
2
q
t
j2
q
t
j1
minq
t
j1
q
min j
; q
max j
q
t
j2
:
Here, the parameter g
c
is the distribution index for SBX
and can take any non-negative value. A small value of
g
c
allows the creation of children solutions far away
from parents and a large value restricts only near-par-
ent solutions to be created as children solutions.
3. The children solutions are then calculated as follows:
q
t
0
j1
0:5q
t
j1
q
t
j2
cjq
t
j2
q
t
j1
j;
q
t
0
j2
0:5q
t
j1
q
t
j2
cjq
t
j2
q
t
j1
j:
It is assumed here that q
t
j1
6 q
t
j2
. Each variable q
t
j
is
chosen with a probability p
c
and the above SBX oper-
ator is applied variable-by-variable.
Appendix B. Polynomial mutation operator
A polynomial probability distribution is used to create a
decision variable q
t0
j
(for the child solution Q
0
, say) in the
vicinity of a decision variable q
t
j
(in the parent solution
Q, say) under the mutation operator. The following proce-
dure is used:
1. Create a random number u between 0 and 1.
2. Calculate the parameter d as follows:
d
2u12u1/
g
m
1

1
g
m
1
1; if u60:5;
121u2u0:51/
g
m
1

1
g
m
1
; otherwise;
(
where /
minq
t
j
q
min j
;q
max j
q
t
j

q
max j
q
minj

.
The parameter g
m
is the distribution index for muta-
tion and takes any non-negative value.
3. Calculate the mutated child as follows:
q
t
0
j
q
t
j
dq
max j
q
min j
:
The perturbance in the solutions can be adjusted by
varying g
m
and p
m
with generations as given below:
g
m
g
mmin
gen;
ap
m
1=N gen=gen
max
1 1=N;
where g
mmin
is the user dened minimum value for g
m
;
p
m
is the probability of mutation; and N = Nh
(T 1)is the maximum number of decision variables
in solution vector Q.
References
[1] Walters DC, Sheble GB. Genetic algorithm solution of economic
dispatch with valve point loading. IEEE Trans Power Systems 1993;8
(3):132532.
[2] Wood AJ, Wollenberg BF. Power generation, operation and
control. New York: John Wiley and Sons; 1996.
[3] Bansal RC. Optimization methods for electric power systems: an
overview. Int J Emerging Electric Power Systems 2005;2 (1). article
1021, Berkeley Electronic Press (bepress).
[4] Yamin HY. Review on methods of generation scheduling in electric
power systems. Electric Power Systems Res 2004;69 (23):22748.
[5] Kumar S, Naresh R, Sharma AK. An overview of articial
intelligence techniques for short term hydrothermal scheduling, In:
Proc. PEPEM05, IEE sponsored Nat. Conf. on Power Engineering
Practices and Energy Management, Patiala, India, January 2930,
2005.
746 S. Kumar, R. Naresh / Electrical Power and Energy Systems 29 (2007) 738747
[6] Miranda V, Srinivasan D, Proenca LM. Evolutionary computation in
power systems, In: Proc. 12th Power Systems Computation Conf.,
Dresden, Germany, August 1923, 1996, p. 2540.
[7] Naresh R, Sharma J. Hydro system scheduling using ANN approach.
IEEE Trans Power Systems 2000;15 (1):38895.
[8] Naresh R, Sharma J. Two-phase neural network based solution
technique for short term hydrothermal scheduling. IEE Proc Gen
Trans Distri 1999;146 (6):65763.
[9] Liang RH, Hsu YY. Short-term hydro-scheduling using hopeld
neural network. IEE Proc Gen Trans Distri 1996;143 (3):26975.
[10] Orero SO, Irving MR. A genetic algorithm modeling framework and
solution technique for short-term optimal hydrothermal scheduling.
IEEE Trans Power Systems 1998;13 (2):50118.
[11] Zoumas CE, Bakirtzis AG, Theocharis JB, Petridis V. A genetic
algorithm solution approach to the hydrothermal coordination
problem. IEEE Trans Power Systems 2004;19 (3):135664.
[12] Gil E, Bustos J, Rudnick H. Short-term hydrothermal generation
scheduling model using a genetic algorithm. IEEE Trans Power
Systems 2003;18 (4):125664.
[13] Chang HC, Chen PH. Hydrothermal generation scheduling package: a
genetic basedapproach. IEEProc GenTrans Distri 1998;145(4):4517.
[14] Chen Po-Hung, Chang Hong-Chan. Genetic aided scheduling of
hydraulically coupled plants in hydro-thermal coordination. IEEE
Trans Power Systems 1996;11 (2):97581.
[15] Yang PC, Yang HT, Huang CL. Scheduling short-term hydrothermal
generation using evolutionary programming techniques. IEE Proc
Gen Trans Distri 1996;143 (4):3716.
[16] Sinha N, Chakrabarti R, Chattopadhyay PK. Fast evolutionary
programming techniques for short-term hydrothermal scheduling.
IEEE Trans Power Systems 2003;18 (1):21420.
[17] Cau TDH, Kaye RJ. Using constructive evolutionary programming
to optimize multi-storage electrical power supply systems operation.
Proc CEC 03, The 2003 Congress on Evolutionary Computation
2003;2:13104. December 812.
[18] Goldberg DE. Genetic algorithms in search optimization and
machine learning. Addison-Wesley Pub. Co.; 1989.
[19] Pham DT, Karaboga D. Intelligent optimization techniques: genetic
algorithms, tabu search, simulated annealing and neural net-
works. Springer-Verlag London Ltd.; 2000.
[20] Deb K. Multi-Objective Optimization using Evolutionary Algo-
rithms. John Wiley & Sons, Ltd.; 2002.
[21] Herrera F, Lozano M, Verdegay JL. Tackling real-coded genetic
algorithms: operators and tools for behavioral analysis, Articial
Intelligence Review 12, Kluwer Academic Publishers, 1998, p. 265
319.
[22] Mezura-Montes E, Coello CAC. A simple multimembered evolution
strategy to solve constrained optimization problems. IEEE Trans
Evolution Comput 2005;9 (1):117.
S. Kumar, R. Naresh / Electrical Power and Energy Systems 29 (2007) 738747 747

Вам также может понравиться