Abraham George et. al.

/ International Journal of Engineering Science and Technology

This paper presents two novel algorithms for multi-objective, short-term hydrothermal scheduling. The former, a
hybrid algorithm, is the offspring of union between genetic algorithm and traditional Newton-Raphson method and
the latter involves mainly heuristic searches with genetic algorithm. In the hybrid algorithm, a population of weight
vectors is genetically generated, for each weight vector the objective function values are computed for the
optimization interval by N-R method and the overall satisfaction/fitness of the solution is found using fuzzy
techniques. The population of weight vectors is modified and fitness values are computed for members of the
modified population. The cycle continues till the highest fitness value obtained, attains near saturation. The major
steps in both the algorithm are the same, but in heuristic search algorithm the objective function values for the
optimization interval are computed by a sub-process which employs GA and Fuzzy logic. Special crossover
techniques are used in all genetic searches to improve the efficiency of the algorithm. Another special feature of this
paper is the introduction of coal-constrained thermal plants.

Keywords: Genetic algorithm, fuzzy logic, hydrothermal scheduling, search space reduction
technique

1 Introduction
In the present setup of large power systems with hydro and thermal stations, the integrated operation of them is
inevitable with due consideration to economic and environmental aspects. The available energy sources are to be
effectively scheduled for efficient operation of a power system keeping economy in view. As emission levels are
touching very high levels and also due to the public outcry for environmental protection, environmental objectives
have been added to economy objective resulting in a multi-objective optimization problem. Thus the goal of multi-
objective hydrothermal scheduling is not only to minimize the system operating cost, but also to minimize the
quantity of gaseous pollutants emitted by thermal units. The researchers who explored the complex field of
hydrothermal scheduling and its constituent parts lately, were focusing on heuristic search methods rather than
traditional methods. A few of the recent explorations are briefed below.
A heuristic search technique based on binary successive approximation using stochastic models was proposed by
Dhillon et al. [2007]. They considered five objectives; economy, emission of SO
2
, emission of NO
x
, emission of
CO
2
and variance of power. Normally the input system data is believed to be deterministic, but practically it is found
to vary depending on the uncertainties and hence it is worthwhile to assume the system data to be uncertain for a
ISSN: 0975-5462 7021
Abraham George et. al. / International Journal of Engineering Science and Technology
Vol. 2(12), 2010, 7021-7034
more realistic approach to the problem. The new algorithm was tested on a small system and promising results were
obtained.
An algorithm based on evolutionary particle swarm optimization for solving the pumped-storage scheduling
problem was proposed by Chen[2008]. The proposed approach combined a basic particle swarm optimization with
binary encoding/decoding techniques as well as a mutation operation. The binary encoding/decoding techniques
were adopted to model the discrete characteristics of a plant. The mutation operation was applied to accelerate
convergence and to escape local optima.
A short-term hydro scheduling study was carried out by Catalao et al. [2009] on a system with seven head
dependent cascaded reservoirs with the goal to maximize the value of total hydroelectric power generation
throughout the time horizon considered, satisfying all physical and operational constraints, and consequently to
maximize the profit of the hydroelectric utility from selling energy into the market.
Economic-environmental dispatch using Marginal Rate of Substitution Decision Approach was proposed by Chen
and Wang [2009]. The approach incorporated CO
2
emission constraints in the solution of unit commitment problem.
An efficient particle swarm optimization algorithm was used for generating fuel cost-emission trade off function to
determine the final optimal compromise decision for unit commitment-dispatch under controlled emission
allowance.
A co-evolutionary algorithm based on Lagrangian method was proposed by Liang et al. [2009] for hydrothermal
generation scheduling in which a genetic algorithm was successfully incorporated into the Lagrangian method. The
genetic algorithm searched for the optimum using multiple-path technique and had the ability to deal with
continuous and discrete variables.

Lots more literature is available about metaheuristics and optimization which are worth considering for solving
several complex multi-objective optimization problems.

2 Multi-objective optimization

2.1 Approaches

There are two general approaches for multi-objective optimization [Abdulla Konak et al., (2006)]. One is to
combine the individual objective functions to a composite function and the other is to consider one objective as
prime and the remaining as constraints. Weighted sum method can be used to combine objective functions in the
first approach, but it is often found difficult to accurately select the weights to be assigned to objectives even for one
quite familiar with the problem domain. The major difficulty with the second approach is in establishing the
constraining values of objectives other than the prime one. In both the approaches, optimization would result in a
single solution. But, as decision-makers often prefer a set of solutions that can be considered for trade-offs, the first
approach is widely preferred. With the first approach, the optimization problem could be solved for a set of weight
vectors and out of the Pareto-optimal solutions obtained; the one having the highest overall satisfaction would be the
most preferred solution.

2.2 Problem formulation

2.2.1 Notations

ob number of objectives
n total number of thermal units
n
1
number of coal-constrained thermal units
h number of hydro units
a
1i
, b
1i
, c
1i
cost coefficients of i
th
thermal unit
a
2i
, b
2i
, c
2i
NO
x
emission coefficients of i
th
thermal unit
a
3i
, b
3i
, c
3i
SO
2
emission coefficients of i
th
thermal unit
a
4i
, b
4i
, c
4i
CO
2
emission coefficients of i
th
thermal unit
ISSN: 0975-5462 7022
Abraham George et. al. / International Journal of Engineering Science and Technology
Vol. 2(12), 2010, 7021-7034
w
i
weight assigned to i
th
objective
P
D k
Power Demand during k
th
sub-interval
P
Lk
Real power Loss during k
th
sub-interval
P
ik
Power output of i
th
unit during k
th
sub-interval
P
i
max
, P
i
min

maximum/minimum power output of i
th
unit
I total number of sub-intervals
d
n+i,k
discharge of i
th
hydro unit or (n+i)
th
unit during k
th
sub-interval
Va
n+i
allocated volume of water for i
th
hydro unit or (n+i)
th
unit
, Lagrange multipliers

k
Lagrange multiplier for sub-interval k

n+i
,
n+i
,
n+i
discharge coefficients of

i
th
hydro unit or (n+i)
th
unit
Q
Ti
Total coal allocation for i
th
coal-constrained thermal unit

1i
,
1i
,
1i
coal discharge coefficients of

i
th
coal constrained thermal unit
H Hessian
J Jacobian
z vector of mismatches
R symbol used for currency

(Duration of a sub-interval is taken as 1 hour)

2.2.2 Problem Objectives

The objective function is minimization of fuel cost and emission of pollutants NO
x
, SO
2
and CO
2
.


ob
m m
m=1
Minimise w F

2

1 1
( ) (1)
I n
m mi mi ik mi
ik
k i
where F = a P b P c +
= =
+

2.2.3 Constraints
(1)Power Balance constraint
1
0 ,
n h
Dk Lk ik
i
P P P k =1,2,..., I
+
=
+ =

(2) Generation limit constraint

min max
i i i P P P i =1,2,.....,n+h s s


(3) W
eight constraint

1
( )
ob
m m
m
w 1 w is zero / +ive
=
=


(4) Hydro constraint

2

1
3
, /
I
ik i ik i ik i ik i
k
d Va i = n+1,....,n+h where d P P m h o |
=
= = + +

(5) Coal constraint
2
, 1 1 1
1
/
I
ik Ti 1 ik i ik i ik i
k
Q Q i =1,2,...,n where Q P P kg h o |
=
= = + +

ISSN: 0975-5462 7023
Abraham George et. al. / International Journal of Engineering Science and Technology
Vol. 2(12), 2010, 7021-7034
2.2.4 Optimality conditions

The approach used in this algorithm for solution of the above problem is to find the conditions for optimality by
application of calculus. By method of Lagrange multipliers, Eq. (1), augmented by constraints, for a particular
weight combination can be expressed as
1 1

1 1 1 1 1 1 1 1 1
( ) + (2)
ob I n h I n h n h I n n
m m k Dk Lk ik i ik i i i ik i Ti
m k i k i n i n k i i
L w F P P P d Va Q Q
+ + +
= = = = = + = + = = =
= + + +


Differentiating Eq. (2) with respect to decision variables results in optimality conditions, numbering six, which are
given below.
1
( / ) ( / 1) ( / ) 0, (3)
ob
m m ik k Lk ik i ik ik
m
w F P P P Q P i =1,..n1, k =1,..., I
=
c c + c c + c c =

4
1
( / ) ( / 1) 0 (4) m m ik k Lk ik 1
m
w F P P P i = n +1,... n, k =1,..., I
=
c c + c c =

k ( / )+ ( / - )=0 (5) i ik ik Lk ik d P P P 1 i = n+1,..n+h, k =1,2,...I c c c c

1
0 (6)
n h
Dk Lk ik
i
P P P k =1,....I
+
=
+ =

1
0 (7)
I
ik i
k
d Va i = n+1,..,n+h
=
=

1
1
0 1, 2,... (8)
I
ik Ti
k
Q Q i = n
=
=

2 The hybrid algorithm
Fig.1 shows multi-objective, short-term hydrothermal scheduling algorithm. The major steps in this algorithm are
explained in sub-sections 3.1 - 3.6.
ISSN: 0975-5462 7024
Abraham George et. al. / International Journal of Engineering Science and Technology
Vol. 2(12), 2010, 7021-7034

3.1 Genetic generation of weight vectors
A widely used technique to solve the multi-objective hydrothermal scheduling problem is to generate a set of Pareto-
optimal solutions using user-fed weight vectors and identify the most satisfactory solution using Fuzzy techniques.
Auto-generating weight vectors is another option. But, in either case, a number of weight vectors are to be tried to
get a solution of high fitness value. Genetic search for a high fitness weight vector [George et al., (2010)] is highly
recommended as genetic algorithms are found to be efficient global optimum locaters.
A population of L individuals is generated genetically. An individual has a length of l bits, is split into ob sub-
strings of length l/ob bits each and hence a combination of ob weights or a weight vector is derived from a string.
Two sample stings are shown in Fig.2.

ISSN: 0975-5462 7025
Abraham George et. al. / International Journal of Engineering Science and Technology
Vol. 2(12), 2010, 7021-7034
3.2 Sub-process A
Optimality conditions described by Eq. (3)-(8) is reduced to Newton-Raphson equations and expressed in the form
[ ] [ z]=[J] (9) H A


Sub-process A involves solving Eq. (9) iteratively to obtain the objective function values for the sub-intervals and
hence for the optimization interval.


3.3 Determining fitness of a weight vector
While generating L weight vectors genetically, the first ob weight vectors should be pre-defined, of the format
shown in Fig.3, essential for determining the extremes of objective function values.


max min
, i , , i F F i =1,2,... ob , maximum and minimum values of objective functions, can be obtained once
computations for the first ob weight vectors are done. Membership functions of objective function values and
fitness of a weight vector are computed using Eq. (10) and Eq. (11)

max max min
) / ( ) ( ( ) (10) i i i i i m F F F F F for i =1,2,...,ob =
1
( ) / (11)
ob
i
i
fitness m F ob
=
=

3.4 Check for Pareto optimality

Solution corresponding to each weight vector is checked for dominancy. Dominated solutions are discarded and the
weight vector corresponding to each dominated solution is replaced with the weight vector corresponding to a non-
dominated solution, having the closest fitness value. An external Pareto set is not used in this algorithm as
dominated solutions are found rare in several case studies.


3.5 Convergence criterion

In many simulation trial runs made, it has been observed that the highest fitness value saturates after a definite
number of generations. Alternately, if the improvement in highest fitness value over a definite number of
generations is marginal, the process can be terminated.


3.6 Modification of weight vectors

After computing fitness of each element in the population, it is sorted in descending order of fitness and further
divided into two halves of L/2 individuals each. First string from the upper half and a random one from lower half
ISSN: 0975-5462 7026
Abraham George et. al. / International Journal of Engineering Science and Technology
Vol. 2(12), 2010, 7021-7034
are selected for mating. Two random numbers between 1 and l/ (2*ob), are generated, which give the crossover site
as well as the number of bits to be selected/replaced for upper half /lower half strings.
A typical crossover is shown in Fig.4. Crossover steps are to be repeated (ob-1) times for the remaining sub-strings.
Second mating pair consists of second string from the upper half and a random one from lower half, other than the
one selected already. The crossover site is restricted almost to the left half of a sub-string and only the sub-stings of
the lower half individual are modified. The upper half of the population is untouched and hence high fitness
individuals are retained. A low probability mutation is performed on the lower half strings to get the new generation
individuals. Mutation rate can be made higher for genetic diversity. Many low fitness individuals in the lower half
become stronger after genetic operations and a few very efficient ones will be pushed to the upper half when the
population is sorted next.

3.7 Strength of hybrid algorithm

The major difficulty encountered while solving the multi-objective optimization problem by Weighting method is
that a number of weight vectors are to be tried to get a solution of high fitness value. This difficulty is overcome by
genetic search for a high fitness weight vector. In many trial simulations made, a 16-member population yielded
high fitness solutions in five to six generations i.e. a maximum of 96 weight vectors were tried to get the desired
result, but in the traditional case this number will be definitely more. Moreover, determining objective function
values for the optimization interval by traditional N-R method has many plus points such as: sure to converge for
most of the systems, high solution speed and good accuracy. Thus the hybrid algorithm is a powerful tool to solve
multi-objective, short-term hydrothermal scheduling problems.

4. Heuristic search algorithm
The flow diagram in Fig.1 applies also to the Heuristic search algorithm introduced here. The sub-process A in
this case is shown in Fig. 5.



ISSN: 0975-5462 7027
Abraham George et. al. / International Journal of Engineering Science and Technology
Vol. 2(12), 2010, 7021-7034



4.1 Optimal power allocations for a sub-interval

Optimality conditions (3), (4) and (5) are simplified to
k k 1 1
1,
1 1
=( - 2 ) / ( 2 2 2 ) (12)
ob ob
n h
ik m mi ij jk i i m mi k ii i i 1
j j i
m m
P w b B P w a B i =1,2,...n | o
+
+
= =
= =
+


k k
1,
1 1
=( - 2 ) / ( 2 2 ) (13)
ob ob
n h
ik m mi ij jk m mi k ii 1
j j i
m m
P w b B P w a B i = n +1,.....,n
+
= =
= =
+


k k
1,
=( - 2 ) / (2 2 ) (14)
n h
ik i i ij jk i i k ii
j j i
P B P B i = n+1,n+2,...n+h | o
+
= =
+

Eq. (12) (14) gives power output of coal-constrained thermal units, other thermal units and hydro units
respectively, during each sub-interval. Losses in each sub-interval can be determined using standard formula.

Initial values of Lagrange multipliers and , values of and power allocations for each sub-interval can be
computed assuming losses equal to zero. With a genetic population of N individuals a search is carried out to
locate the optimum during each sub-interval. A new technique, search space reduction technique [George,
ISSN: 0975-5462 7028
Abraham George et. al. / International Journal of Engineering Science and Technology
Vol. 2(12), 2010, 7021-7034
(2009) ] is incorporated in determining optimum , which reduces the search range in each generation. As the initial
value of is known and its final value is in the vicinity of initial value, search for is a search for local optimum.
Search space reduction technique forces the individuals to move towards the optimum at a faster rate comparing
with a normal genetic search, thus locating the optimum in minimum trials.

4.1.1 Search space reduction technique

This technique reduces the search range in each generation such that optimum is located faster. This technique is
used in locating optimum value of Lagrange multiplier , for satisfying power balance constraint in each
optimization sub-interval. Initial is determined assuming losses equal to zero. Initial search range can be fixed as
0.5 2.0 allocating sufficient space below and above the initially determined . The variation of from its initial
value is marginal as only transmission losses are incorporated in the final solution. To prove the effectiveness of
search space reduction technique, the following test has been conducted.

A system having six thermal units had been tested for 161 weight combinations to obtain the best compromise
solution. In a particular trial run, out of 161 executions with normal genetic search, solutions in the cases of 4 weight
vectors converged in 1-5 generations; 12 in 6-10 generations and the remaining 145 took more than 10 generations
to converge. With search space reduction technique incorporated, even with a convergence tolerance of 0.0001,
these numbers became 102, 47 and 12 respectively and the time of execution reduced by 397 seconds. With
convergence tolerance less than 0.1 in a normal genetic search, the number of generations for many weight
combinations exceeded 500. Results of execution are shown in Table 1.

Table 1 Effectiveness of search space reduction technique


4.2 Determining fuzzy acceleration factors/multipliers

At the end of optimization interval, water withdrawals and coal consumption are computed. If constraints 4, 5 of (1)
are not satisfied then the associated Lagrange multipliers are to be modified for the next iteration.
j
and
j
are
modified as shown below.


1 ( ) / (15) i i i i i i i i i where k V Va Va + A A =

2 ( ) / (16) i i i i i i i Ti Ti where k Q Q Q + A A =
The program execution is very sensitive to choice of k. A wrong selection results in retardation/divergence. A set of
suitable values can be suggested only by one who is very familiar with the problem domain and the selected values
remain constant throughout the execution. If k

is made proportional to per unit mismatches in water
withdrawals/coal consumption at the end of an iteration, the solution is found converging at a faster rate and an easy
way to get the proportional k

is using fuzzy logic. Hence k
1i,
(i=n+1, ,n+ h) and k
2i,
(i=1,2,,n
1
) are modified at
the end of each iteration and they have became variables rather than constants. In many cases best results have been
obtained when output surface for the FIS has an approximate shape as shown in Fig. 6. Fuzzy acceleration factors
were successfully applied in load flow studies by Chakravorty et al., [2009].

ISSN: 0975-5462 7029
Abraham George et. al. / International Journal of Engineering Science and Technology
Vol. 2(12), 2010, 7021-7034
A system having two thermal and two hydro units was tested for 10 user-defined weight vectors to prove the
efficiency of fuzzy multipliers k
1
and k
2
. Number of iterations for satisfying hydro and coal constraints reduced
between 20-40% of their original values with substantial reduction in computation time.



4.3 Strength of heuristic search algorithm

The algorithm simplifies the problem to a considerable extend. The computational complexity associated with N-R
method in determining objective function values for the optimization interval is totally eliminated in Sub-process
A. Search space reduction technique accelerates determining sub-interval power allocations and fuzzy acceleration
factors help speedy convergence. An overall improvement in execution speed could be achieved due to the above
modifications. Still the algorithm works much slower than the hybrid algorithm due to numerous genetic
manipulations to be performed. This cannot be considered as a major drawback as superfast computers with multiple
processors are on the rise.

5. Results and discussions.

5.1 Test system

A system having two thermal and two hydro units is considered for testing the novel algorithms introduced. Table 2
& Table 3 list the features of hydro and thermal units. Table 4 lists loss coefficients.

Table 2 Features of hydro units.






ISSN: 0975-5462 7030
Abraham George et. al. / International Journal of Engineering Science and Technology
Vol. 2(12), 2010, 7021-7034




Table 4 Loss coefficients



5.2 Test results: hybrid algorithm

When the system described in 5.1 is tested for an optimization interval of 72 hours, the highest fitness value
obtained in most of the trial runs reached near saturation in about five generations. The final Pareto front always
contained a set of weight combinations with fitness values lying within a narrow range. Weight combinations in the
last Pareto optimal front in two different trial runs are given in Table 5. Weight combinations vary over a wider
range whereas fitness values lie in a closer range providing an opportunity for one to select a weight combination
depending on the priority to be accorded to objectives.


ISSN: 0975-5462 7031
Abraham George et. al. / International Journal of Engineering Science and Technology
Vol. 2(12), 2010, 7021-7034
Table 5 High fitness weight combinations in the last
Pareto optimal front



Fig. 7 shows the cost- total emission characteristic corresponding to the final Pareto optimal front obtained in Trial
run 1 shown in Table 5. Total emission is the sum of individual emission of all the three pollutants during the
optimization interval.

Fig. 8 shows the system demand, total thermal generation and total hydro generation during the optimization interval
corresponding to the Pareto optimal solution with highest fitness in Trial run 1as shown in Table 5.

Seven scheduling intervals from 24 hours to 168 hours, in steps of 24 hours, were considered to obtain the
optimization interval-execution time characteristic shown in Fig. 9. For optimization intervals 24-hour and 168-
hour, the first and the last considered, the execution time were 5.594s and 1807.586s respectively.


Fig. 7 Cost- total emission characteristic


ISSN: 0975-5462 7032
Abraham George et. al. / International Journal of Engineering Science and Technology
Vol. 2(12), 2010, 7021-7034


Fig. 8 Total hydro and thermal generations




Fig. 9 Scheduling interval - execution time characteristic


5.3 Test results: comparison of hybrid and heuristic search algorithms

Both the algorithms introduced are the same except for the step to compute objective function values for the
optimization interval. Algorihm-1 uses N-R method and Algorithm-2 uses Sub-process A as shown in Fig. 5. Both
the algorithms give identical results; the only difference is the time factor. Table 6 shows the difference in time of
execution with scheduling interval and number of generations limed respectively to 24 and 5.








ISSN: 0975-5462 7033
Abraham George et. al. / International Journal of Engineering Science and Technology
Vol. 2(12), 2010, 7021-7034
6. Conclusion

A brief comparison of the two new algorithms is done here. The traditional N-R method with genetic generation and
modification of weight vectors is an excellent tool for short-term multi objective hydrothermal scheduling with the
only drawback of numerous, elaborate mathematical operations involved. The process of short-term multi objective
hydrothermal scheduling is made extremely simple in heuristic search algorithm but with the major drawback of
high time consumption. The time consumption is substantially reduced by introducing search space reduction
technique and fuzzy acceleration factors. With both the methods in front, one has to compromise between problem
complexity and speed of execution.

References

[1] Abdulla Konak; David W. Coit; Alice E. Smith, (2006): Multi-objective Optimization using genetic algorithms: A tutorial,ew
Reliability Engineering and System Safety, Elsevier 91, pp 992-1007.
[2] Abraham George (2009): Emission constrained thermal dispatch and hydrothermal scheduling based on genetic algorithm: search space
reduction technique, 18
th
Annual Symposium, IEEE Bangalore Section.
[3] Abraham George; Channa Reddy M; Sivaramakrishnan A.Y (2010): Short term Hydrothermal Scheduling based on Multi-objective
Genetic Algorithm , International Journal of Electrical Engineering, ISSN 0974-2158 Volume 3, No. 1, pp. 13-26.
[4] Catalao J.P.S et al., (2009): Scheduling of Head Sensitive Cascaded Hydro Systems : A Nonlinear Approach, IEEE Transactions on
Power System, Vol. 24, No. 1 .
[5] David E. Goldberg (2004): Genetic Algorithms, Pearson Education , New Delhi.
[6] Jarnail S. Dhillon; Dhillon J.S. ; Kothari D.P (2007) : Multi-objective Short Term Hydrothermal Scheduling based on Heuristic Search
Technique , Asian Journal of Information Technology, 6(4), pp 447-454.
[7] Jaydeep Chakravorty; Sandeep Chakravorty; Samarjit Ghosh (2009): A Fuzzy Based Efficient Load Flow Analysis, International
Journal of Computer and Electrical Engineering , Vol. 1, No.4.
[8] Kothari D.P; Dhillon J.S (2004): Power System Optimization, Prentice Hall of India Pvt. Ltd, New Delhi.
[9] Marco Dorigo; Thomas Stutzle (2004): Ant Colony Optimization, MIT Press , London.
[10] Maurice Clerc (2007): Particle Swarm Optimization , ISTE Ltd., London.
[11] Po-Hung Chen (2008) : Pumped-Storage Scheduling Using Evolutionary Particle Swarm Optimization , IEEE Transactions on Energy
Conversion, vol. 23, Issue 1.
[12] Ruey-Hsun Liang ; Ming-Huei Ke ; Yie-Tone Chen (2009) : Co-evolutionary Algorithm Based on Lagrangian Method for
Hydrothermal Generation Scheduling , IEEE Transactions on Power Systems, Vol 24, Issue 2.
[13] Singeresu S. Rao (2003): Engineering Optimization, New Age International (P) Ltd., , New Delhi.
[14] Yee Ming Chen; Wen-Shiang Wang (2009) : Economic Dispatch with Environmental Considerations Using Marginal Rate of
substitution Decision Approach , Journal of Quality, Vol. 16, No.2.
ISSN: 0975-5462 7034