You are on page 1of 6

Genetic Algorithm Applications to Stochastic Thermal Power Dispatch

K Selvi, Member
Dr N Ramaraj, Member
S P Umayal, Non-member

This paper presents a genetic algorithm (GA) based effective method for the optimal scheduling of thermal generation
incorporating the uncertainties in the system production cost data. A stochastic model of system production cost
equation is formulated, with production cost coefficients and generator outputs as random variables. Minimization of
total operating cost for thermal units in the system subjected to recognized constrains is solved using simple genetic
algorithm (SGA). The advantages of this method lies in its faster convergence towards the global solution. The
algorithm gives fairly accurate results. The effectiveness of the method has been demonstrated by analyzing sample
systems and the results are presented.

Keywords: GA; Stochastic; Variance

NOTATIONS But in practice, a basic difficulty is that, the information


a i, b i, c i : cost coefficients random variables, of ith that is supposed to be known is not deterministic. Several
generator inaccuracies2,3,4 and uncertainties lead to deviation from
optimal operation. The cost functions are important input
ai , bi , c i , Pi data for economic dispatch procedure but are inaccurate in
most cases. These inaccuracies are deviations from input
PD , PL , F : expected values
information given on the data sheets, from their values during
Bij : loss formula coefficient actual operation. They are caused through two reasons,
namely, (i) Inaccuracies in the process of measuring and
c v Pi : coefficient of variation of random variable Pi forecasting of input data; and (ii) change of input performance
E (.) : expectation during the period between measuring and operation.
F : system production cost, a random variable As a consequence, instead of having a system production cost
curve representing the bogey performance of the unit, what
N : number of generators
really exist is narrow band, such as, in Figure 1 which the real
PD : received load, MW curve cal lie. Cost functions may therefore only be known is a
Pi : active power generation, a random variable statistical distribution. From such a data, it is not possible to
get a deterministic minimum. Thus, a limitation of the
Pim , Pi M : lower and upper limit of Pi , MW conventional economic dispatch method is that, the actual
schedules determined with deterministic cost functions may
PL : transmission power loss, a random variable
vary. With rising fuel costs, there is growing interest to
Var Pi : variance of Pi account for these deviations, since the effect of inaccuracies
: Lagrange multiplier result in an increase in the overall cost.
λ
INTRODUCTION
Despite extensive research focusing on thermal power
dispatch problem, much of the efforts today has involved the
Cost, $/h

deterministic models applicable to study state conditions.


That is, conventional economic dispacth1 takes into account,
incremental production cost equations as deterministic. It
means that all input information is known with full certainty
and that the optimal plans of dispatch are always realized
exactly.
K Selvi, Dr N Ramaraj and S P Umayal are with Thigarajar College of
Engineering, Madurai 625 015.
This paper (remodified) was received on October 22, 2003. Written discussion Power output, MW
on this paper will be received until August 31, 2004. Figure 1 Production cost as a function of power generated

Vol 85, June 2004 43


Sauer and Heydt2 discuss the theoretical background in statis- A specific way of reducing a stochastic model to a
tical areas including transformation to normality. Viviani and deterministic one is to take its expected value11 and therefore,
Heydt5 incorporated the effect of uncertain parameters into the expected value of operating cost becomes
the optimal power flow problem. Parti, et al 6 presented an
economic dispatch of thermal generation with explicit F =E ∑
LM ecN
+ bi Pi + ai Pi2 jOPQ
recognition of uncertainty in the system production cost data.
Dhillon, et al 7 formulated the problem as multiobjective one.
N i =1
i

Schultz8 formulated a two-stage program with mixed integer


recourse including a risk term, reflecting the probability that
= ∑
N

i =1
bg b g b g b g e j
E c i + E bi E Pi + E ai E Pi2
the preselected cost threshold is exceeded.
e j
N
For a long time, the Lagrange multiplier method6,9 is the most = ∑ c i + bi Pi + ai var Pi + Pi 2
i =1
promising technique used for economic power allocation, its
main drawback is that, the duality gap which results from the
dual problem formulation and also its inability to obtain = ∑
N LMc + bi Pi + ai c Pi 2 + ai Pi 2
OP
feasible solution for some problem. Recently, a wide variety of i =1 N i vPi2
Q
direct and indirect methods are used for dispatch problems. In
power systems, GA’s have been tried in areas like load flow, N LMc FG IJ OP
economic dispatch10 etc. Unlike the Lagrange method, which
F = ∑
i =1 N i + bi Pi + ai Pi 2 1 + c
H vPi2 KQ (3)
requires the derivative of the cost characteristics, simple where c vPi is the coefficient of variation random variable Pi. It
genetic algorithm (SGA) facilities the use of objective function
is the ratio of standard deviation to the mean and is a measure
and constraints, both of which are non-linear and generates an
optimal or near optimal solution. This paper deals with of relative dispersion or uncertainty in the random variable. If
generalizing the existing load dispatch procedures considering c vPi = 0 , implies no randomness or in other words, complete
statistical uncertainty in the system production cost, and certainty about the value of random variable.
evaluates the potential economic significance using SGA. Representation of System Losses
STOCHASTIC MODEL The expected value of loss using B matrix is given by
A method of obtaining a stochastic model is to take a
PL = E PL = E ∑ ∑ Pi Bij P j
LM OP
deterministic model and transform it into a stochastic model
by (1) introducing random variables as inputs or as N i j Q
coefficients or as both; and (2) introducing equational errors as
= ∑ ∑ Bij Pi P j + ∑ Bii var Pi (4)
disturbances. Since this type of model is only an approximation, i j i
what is important in this approach is to make the randomness
reflect a real situation. The variance of transmission loss has been neglected for
convenience of analysis, since it is usually less than 2% of the
Mathematical Formulation total power transmitted.
The operating cost per unit time in a thermal power system is Equality and Inequality Constraints
generally approximated by a second order polynomial of the 1. The load demand constraint
generated active power1.
N
N
FH IK ∑ Pi = PD + PL (5)
F = ∑ ci + bi Pi + ai Pi2 i =1
i =1
(1)
2. Limit constraints
and the power balance equation is
Pi m ≤ Pi ≤ Pi M (6)
N
∑ Pi = PD + PL Expected Deviation
(2)
i =1 With generator outputs Pi treated as random variables, the
expected value associated with deficit or surplus of generation
A stochastic model of function F is formulated by taking the
can be treated as deviation proportional to the expectation of
deterministic production cost coefficients ai , bi , c i and the square of unsatisfied load demand.
generator active power Pi as random variables. Any possible
deviation of operating cost coefficient from their expected LMF I 2 OP
MNGH P JK
N
values is manipulated through the randomness of generator
power output7 Pi. The randomness of Pi implies that power
Expected deviation =E D + PL − ∑ Pi
i =1 PQ
balance equation (2) is not a rigid constraint to be satisfied.
Whatever the proposed solution will be, only deficit or N 2 N
= ∑ E Pi − Pi = ∑ var Pi (7)
surplus probabilities of generated power can be estimated. i =1 i =1

44 IE (I) Journal—EL
Deterministic Equivalent Table 1(a) Deterministic schedule for two plants at various loads
Lagrange multiplier solution to the combined system cost F PD, MW 100.0000 120.00 140.0000 160.0000
(comprising expected production cost and expected cost of
deviations with a penalty term p) leads to the set of co- P1, MW 80.0128 100.02 120.0288 140.0392
ordination equation as: P2, MW 20.0000 20.00 20.0000 20.0000

eb j FGH d2 B P i + 2 B c P IJK
FT , $/h 183.2200 220.84 259.2600 298.4800
i + 2 ai Pi + λ ∑ ij j ii
j vPi2 i

+ 2 b a + pg c
Table 1(b) Stochastic schedule for PD = 120 MW for two plants
i P =λ ; i = 1, . . . . . . . . , N (8)
vPi2 i
ε × 104 0.750000 0.8000 0.8500 0.9000 0.9500 1.0500
The first two terms of (8) are incremental production cost and
the incremental transmission cost similar to those in conven- P1, MW 70.025400 78.6568 84.4597 89.1495 93.1953 100.1012
tional co-ordination equations of deterministic dispatch. The
P2 , MW 50.955300 42.5805 36.9669 32.4399 28.5417 21.9028
third term of equation (8), results in increasing the incremen-
tal cost of received power λ . This is a risk premium ( ε ) since FT , $/h 231.151217 228.7889 227.4807 226.5871 225.9330 225.0672
it depends upon fluctuations of Pi.
Table 1(a) shows the deterministic schedule obtained for
IMPLEMENTATION OF GA various loads, using GA. Table 1(b) shows the expected
The steps involved in SGA are same as Siwhvin Bakirtzis, et al10. schedule (stochastic) for different values of ε for PD = 120
Optimal schedule of N generators are determined so as to MW using GA. It can be seen that, the expected schedules at
minimize the total operation cost (equation 3) subject to: equa- ε max are different from the deterministic schedules for the
N
tion (5), equation (6), and the binding constraint ∑ var Pi ≤ ε Plant 1
i =1 140
where ε (risk premium) is the maximum tolerable level 130
chosen for which the constraint is binding at the optimal
120
solution.
Output, MW

110
Implementation of the problem in GA1 starts from the
parameter encoding. Binary mapping is used for representing 100
the chromosome string. The random variables Pi are coded as 90
chromosome of the population. Decoding the encoded
80
chromosome string and computing the chromosome’s fitness
value using the decoded parameter accomplish evaluation of 70
chromosome. Applying GA operators, the cycle is repeated 60
till the specified generation count occurs. In that way, 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4
algorithm reports the best value. Risk E 
→ ×10 4
Plant 2
70
ILLUSTRATIONS AND DISCUSSION 65
The effectiveness of the proposed algorithm is tested on two 60
sample systems: 2 plants and 8 plants system. The software is 55
developed in MATLAB 5.1 and run on Pentium-III processor. 50
Output, MW

For deterministic cases, all c vpi are equal to zero. Coefficient 45


of variation of generator output for the stochastic cases can be 40
evaluated by Monte-Carlo method, which is not discussed in 35
this work. 30
25
Sample 1: Two Plants System 20
A two-plant system including transmission losses is selected. 0.5 1 1.5 2
The data is also follows: Risk E 
→ ×10 4
a1 = 0.001 ; b1 = 1.7 ; a 2 = 0.002 ; b 2 = 2.0 ; B11 = 0.0002 ; cvp1 = 01
. P D = 100 MW P D = 140 MW
B12 = 0 ; B21 = 0 ; B22 = 0 ; c vP1 = 0.1; c vP 2 = 0.1 ; cvp2 = 01. P D = 120 MW P D = 160 MW

20 MW ≤ Pi ≤ 200 MW Figure 2 Expected generation schedule with respect to risk for 2-plant
Both units are connected and operating. system

Vol 85, June 2004 45


Table 2 GA parameters for 8 plant system Sample 2: Eight Plants System

Parameter Value A large system is used to illustrate and to evaluate the possible
economic significance of the proposed method. All the plants
Population size 35.000
are participating in the dispatch. The data is same as reference 4.
Maximum generations 30.000 Table 2 shows the GA control parameters for PD = 1380 MW,
Cross over probability 0.750 at ε max = 1840 for stochastic case. CPU time needed for 30
generations for this case, in obtaining deterministic schedule is
Mutation Probability 0.003
3.0 minutes and for stochastic schedule is 3.2 minutes in
Penalty factors Pentium III Processor.
Binding constraint 22000 In Figure 3(a) and 3(b), generation schedules for various loads
Power balance constraint 17000 are plotted for all the plants. Plant 5 remains at its maximum
load of 310 MW throughout. There is shift in generation from
generators with large coefficient of variation to generators
specified c vPi . Figure 2 shows the expected generation with small coefficient of variation. Also, plant 2 remains
schedule for different values of ε and depicts that risk ε is throughout at its maximum loading of 210 MW in case if
proportional to power demand. deterministic schedules (Figure 3(a)) but it is not so for the

Plant 1 Plant 2
190 215

180
→

210
→

170

Output, MW

205
Output, MW

160

150 200

140
195
130

120 190
1200 1250 1300 1350 1400 1450 1200 1250 1300 1350 1400 1450
Load, MW 
→
Load, MW 
→
Plant 3 Plant 4
130 330
120
→

320

110
→

310
Output, MW

100

300
Output, MW

90

80 290
70
270
60
260
50
1200 1250 1300 1350 1400 1450 1200 1250 1300 1350 1400 1450
Load, MW 
→ Load, MW 
→
c vp2 = 0.17
Deterministic
c vp6 = 0.11
Stochastic
c vp8 = 0.13

Figure 3(a) Generation schedule for plants 1 to 4 for 8 plants system

46 IE (I) Journal—EL
Plant 5 Plant 6
311 180
310.8 175
→

→
310.6
170


310.4
165
Output, MW

Output, MW
310.2
160
310
155
309.8
309.6 150

309.4 145

309.2 140
309 135
1200 1250 1300 1350 1400 1450
1200 1250 1300 1350 1400 1450
Load, MW 
→
Load, MW 
→

Plant 7 Plant 8
90 125

85
→
120
→

80

115
Output, MW
Output, MW

75 110

70 105

65 100

60 95
1200 1250 1300 1350 1400 1450 1200 1250 1300 1350 1400 1450
Load, MW 
→
Load, MW 
→

Excess cost
4
3.8
→

Deterministic
3.6

3.4 Stochastic
Excess cost in $/h

3.2
c vp2 = 0.17
3
c vp6 = 0.11
2.8 c vp8 = 0.13
Ref 4
2.6
Prop.
2.4
2.2
1200 1250 1300 1350 1400 1450

Load, MW 
→

Figure 3(b) Generation schedule for plants 5 to 8 and excess cost curve for 8 plants system

stochastic case. This has a definite effect on the excess cost The excess cost involved in deviation from deterministic
incurred in the deviation from deterministic schedules. schedule is compared with reference 4 in Figure 3(b). It gives a

Vol 85, June 2004 47


saving of nearly 1 to 2 $/hr at each load due to the optimal REFERENCES
schedule thus obtained by GA. The tendency of this curve to 1. L K Kirchmayer. ‘Economic Operation of Power System.’ Wiley, Eastern
decrease and then increase is attributable to fuel cost of plant 2 Ltd, 1958.
contributes a constant value to the total fuel cost F of the
sample system for various loads in deterministic case, whereas 2. P Sauer and G Heydt. ‘A Convenient Multivariate Gram Charlier Type A
in the stochastic schedules, this cost is variable. Hence, the Series.’ IEEE Transaction on Communications, vol CM-27, no 1, January
consideration of stochastic cost curve Figures 3(a) and 3(b) in 1979, p 247.
the economic dispatch problem may significantly alter the
3. A F Glimn, L K Kirchmayer, G W Stagg and V R Paterson. ‘Accuracy
optimal generator outputs.
Consideration in Economic Dispatch of Power Systems.’ Transactions of
CONCLUSIONS American Institutes of Electrical Engineers on Power System, vol 75, part III, 1956,
p 1125.
Conventional economic thermal power dispatch method
allocates generation schedule to the individual thermal 4. Z Z Liang and J D Glover. ‘Improved Cost Functions for Economic
generating units based upon deterministic cost function Dispatch.’ IEEE Transactions, 1991, PWRS 6, p 21.
ignoring inaccuracies and uncertainties. Such generation
5. G L Viviani and G T Heydt. ‘Stochastic Optimal Energy Dispatches.’ IEEE
schedules will result in the lowest expected cost, which is also
Transactions (PAS), vol PAS-100, no 7, 1981, p 3221.
associated with a relatively large variance, which can be
interpreted as risk measure. Deterministic schedules may, 6. S C Parti, D P Kothari and P V Gupta. ‘Economic Thermal Power Dispatch.’
therefore, render the solution non-optimal to a power control Journal of The Institution of Engineers (India), pt El, vol 64, October 1983, p 126.
engineer who may like to avoid the risk element. The method
proposed in this paper, has the desirable feature of 7. J S Dhillion, S C Parti and D P Kothari. ‘Stochastic Optimal Thermal Power
incorporating inaccuracies and uncertainties in the economic Dispatch.’ IEEE Transaction, vol 5, no 6, 1987.
dispatch procedure. SGA solutions to the deterministic and 8. R Schultz and S Tideman. ‘Risk Aversion via Excess Probabilities in
stochastic schedules of two sample system are presented. The Stochastic Programs.’ Institute for Maths, Mercator Uty Duiburg 2001, from
results thus obtained, gives a saving in excess cost incurred in www.mathematic.hu-berlin.de.
$/hr at each load, as compared with the Parti, et al 6, ensuring
the applicability if SGA to the stochastic thermal dispatch 9. J F Bard. ‘Short Term Scheduling of Thermal Electric Generators using
problem. The method proposed gives power control centers, Lagrangian Relaxation Operations.’ Res, vol 36, no 5, 1988, p 756.
the potential of recognizing statistical uncertainty in the 10. A Bakirtzis, V Pertridis and Kazarlis. ‘Genetic Algorithm Solution to the
system production cost data and thus obtaining a more Economic Dispatch Problem.’ IEEE Proceedings on Generation, Tranmission
realistic economy in system operation. With recent advances and Distribution, vol 141, no 4, July 1994, p 377.
in parallel computing, the online solution of optimal despatch
with stochastic cost functions of generators may soon be 11. W D Papoulis. ‘Probality, Random Variables and Stochastic Processes.’
possible. Mcgraw Hill Book Co, New York, 1965.

48 IE (I) Journal—EL

You might also like