Hybridization of Artificial Neural Network and Lagrange

Multiplier Method to Solve Economic Load Dispatch Problem

Mohammad Mohatram
Waljat College of Applied Sciences, Muscat, Oman
[email protected]

Abstract: Lagrange multiplier methodalso known as O- ANN is composed of elements that carry out actions in a way
iteration method has been in use to find the solution of the which is similar to the biological neuron [8]. ANN may be
economic load dispatch problem for many years. Though it defined as an interconnected group of artificial neurons that
produces results with great accuracy but it suffers from the uses a mathematical model for information processing based
disadvantage of convergence as it is highly sensitive to the on a connectionist approach to computation [19]. Normally, an
initial value of O and at times takes more time to produce ANN is characterized by its architecture, i.e., pattern of
results. To overcome this drawback, in this paper an connections between the neurons, the training or learning, i.e.,
innovative technique to solve the ELD problem, by the method of determining/updating the weights and its
hybridization of artificial neural network (ANN) and LMM activation function which control the output generated by
is proposed. To formulate the problem, generation cost is ANN within a specified limit [18].
expressed as a function of active power by exponential
functions. The equality constraint of power balance and A hybrid technique involves two or more methods to coalesce
inequality constraint that fix the maximum and minimum their strengths and overcome each other’s weakness. Usually,
active power are duly considered. ANN trained by one or more methods are employed to find a solution which is
backpropagation algorithm is used to find the optimal value closed to optimal solution, and then other method is used to
of λ which is then used to find solutions using LMM. get the final optimal solution.
Training data, i.e., input and target patterns for ANN are
obtained by LMM. Using proposed hybrid method, λ- Abdelaziz et al. [1] introduced a hybrid approach of Hopfield
iterations are eliminated and promising results are obtained neural network and quadratic programming to solve the ELD
in much less computation time. Hence, the proposed method problem with ramp rate and transmission losses. First, the
can be useful in real time solution for seeking online Hopfield neural network was used as a base search procedure
solution of ELD problems. to find a near optimal solution without ramp rate constraints,
and then quadratic programming was used to find the optimal
Keywords: Economic Load Dispatch, Artificial Neural solution with ramp rate constraints. Bhattacharya and
Network, Lagrange Multiplier Method, Hybrid Technique, Chattopadhyay proposed a hybrid technique using differential
Exponential Functions, Equality and Inequality Constrains. evolution with biogeography for solution of economic load
dispatch and tested successfully on systems with 3, 10, 38, 40
I. INTRODUCTION units. It was highlighted that the proposed method is best
suited for complex power systems [2]. Dubey et al. presented
In a power system comprising of large number of generating hybrid swarm intelligence combining particle swarm
units, economic load dispatch (ELD) is the procedure of optimization and gravitational search Algorithm and tested the
dispatching electric power at minimum cost to the consumers. technique on systems comprising of 6, 18, 20 and 54 units.
The problem is solved by considering the important The proposed technique is reported to be superior in solution
constraints of a typical power system [4, 21]. quality and computational efficiency [7]. Ashouri and
Hosseini also hybridization of particle swarm optimization and
Over the years, many methods have been proposed to solve
gravitational search Algorithm and compared the experimental
the problem of ELD. According to, Lagrange multiplier
results with traditional and intelligent methods. It is reported
method (LMM) one of the widely used methods, the cost of
that the approach yields accurate results quickly [11].
generating electricity is minimum when incremental cost of
each generating unit multiplied by its penalty factor is the Malik in his approach hybridized Genetic Algorithm and Ant
same for all units in the grid [5, 9]. Penalty factor depends on Colony Optimization. The method was tested on systems with
the distance of the generating unit from the load center. High 3, 5 and 6 units and it was reported that solution time is fast
penalty is imposed on the generating unit which is far away and cost of generation is less [13].Younes presented a novel
from the load center. hybrid approach for economic power dispatch and accentuated
that the approach is a powerful tool for optimizing feasible

978-1-5386-0514-1/17/$31.00 ©2017 IEEE

solutions for the non-convex ELD problems [14]. A Subject to the constraint that power generation should equal
hybridization of O-iteration and Hopfield neural network was total power demands, i.e.,
proposed efficiently by Liang [16] in which re-dispatch to
dynamic generation allocation was suggested. In his approach, + = + + ⋯+
first O-iteration method was used to obtain the ELD solution
as the base case. Then, the problem was linearized about this + =∑ (equality constraint) … (2)
base case and solved using the Hopfield neural network re-
Satisfying the inequality constraints, expressed as follows,
dispatch approach. Maa and Michael [3] proposed linear
programming and quadratic programming based neural
≤ ≤
network approach to solve ELD problem. The authors
compared their methods with classical method and as per their for = 1, 2, … , (inequality constraint) … (3)
claim the error was decreased to approximately 1%.
Comparison of different ANN models for ELD problem Where is the operating cost corresponding to active power
solution was also presented by Hotaet al. [15]. output for the ith generating unit. is the maximum
active power output and is the minimum active power
Pandian and Thanushkodi presented the solution of economic
load dispatch problem considering transmission losses by a output of ith generating unit limited by the flame instability of
Hybrid EP-EPSO algorithm for solving both smooth and non- the boiler respectively [4, 21].
smooth cost function [20]. A comprehensive study of
The power demand is the sum of all customers’ actual
economic load dispatch by various hybrid optimization
power consumption and is calculated as,
techniques is presented by Dipankar et al. The authors
highlighted the advantages and disadvantages of various
approaches used by several researchers in the past [17]. = . … (4)
It is evident from the past research that to seek the solution of
ELD problem many methods have been proposed, each Where is the power demand of ithcustomer. The
method has its own advantages and disadvantages. This paper transmission losses which usually lie in the range from 5% to
presents a hybrid technique in which ANN and LMM are 20% of the total power demand are calculated using the
integrated with the aim that the advantages offered by these famous George formula [5].
methods can be exploited to solve the ELD problem.

II. PROBLEM FORMULATION = . … (5)

The electric power demand depends on many factors such as
environmental, social, economic, etc., and therefore, it
Where B is one of the elements of B-matrix.
fluctuates over a wide range during a day in winter to a day of
summer. Generally, the cost functions for various power For ith generating unit, using exponential function the
generating units in a typical power system are not the same. variation of w.r.t. active power output is expressed as,
Thus, to have the minimum cost of generation for a particular
= exp( ). for = 1, 2, … , (6)
power demand, the power among the generating units is
distributed in such a way that overall generation cost is The coordination equation which is the outcome of LMM in
minimum [4]. ELD solution is given as [9],
Consider a power system with n generating units. Let + = . … (7)
, ,…, be the operating costs of individual plants for the
corresponding active power outputs
, ,…, respectively. Let be the total operating cost Where O represents the incremental cost involved in the
of the entire system, be the total power demand, and be dispatch of power.
the total transmission losses, then the problem of ELD
For exponential cost function, Eq. (7) yields,
considering transmission losses can be stated as follows [21],
( ) ∑
Minimize,
= … (8)
= + + ⋯+ = (objective function) … (1)
Where, αi and βi are the constants computed on the basis of
physical and environmental operating conditions of the
system.

515
 2017 International Conference on Infocom Technologies and Unmanned Systems (ICTUS'2017),
Dec. 18-20, 2017, ADET, Amity University Dubai, UAE

III. HYBRIDIZATION OF ANN AND LMM (optimal) value of O corresponding to a given power demand,
PD. Training and test patterns, i.e., input-output combinations
A new way for achieving the solution of ELD problem is
using a hybrid technique in which an ANN is integrated with of power demand and optimal value of O are generated by
conventional iterative method of ( -iteration) LMM. ANN is using the conventional LMM, by changing the power demand
trained according to backpropagation algorithm. As shown in in steps of 100 MW from minimum to maximum generation
capacity taking into consideration the generator power limits
Fig. 1 ANN is used to find optimum value of O.
and the transmission losses. During the training process, the
weights of the ANN are modified in such a way that the error
between the actual value of O and the target value (optimal
Power Demands

value) of O is reduced to minimum. Rate of convergence of

Optimal O
ANN during training is influenced by the learning rate
Neural Network parameter (η), and momentum factor (α). Their values lie in
between 0 and 1. The number of training patterns, hidden
layers, and neuron in each hidden layer has been selected
through experimentation for different combinations. The
parameters selected are: Number of training patterns = 24,
Fig. 1. ANN used to generate training and test pattern Number of hidden layer = 1, Number of neurons in hidden
layer = 4, η = 0.85, and α = 0.90 and 0.95.
The overall calculations are worked out in two stages as
shown in Fig. 2. In the first stage, ANN is used to find the best

ln ߣ − ln(ߙ௜ ߚ௜ ) − 2 ∑௡௝ ‫ܤ‬௜௝ ܲீ௝

௝ ௜ PG1
ܲீ௜ =
ߚ௜ + 2 ‫ܤ‬௜௜
PG2
Satisfying the inequality constraints,
PD TrainedA Optimal ௠௜௡ ௠௔௫
NN ܲீ௜ ≤ ܲீ௜ ≤ ܲீ௜ , PG3
O
for ݅ = 1,2, … , ݊. ⋮

PGn

Fig. 2. ANN and LMM hybrid approach to solve ELD problem

The optimal value of O found in the first stage is used in the iii) Use trained ANN to find out the optimal value of O for an
second stage to find the ELD solutions using the conventional unknown power demand.
LMM. As the value of O is optimal, convergence is extremely
fast and results are obtained within no time. The iv) Solve Eq. (8) by substituting the optimal value of O as
computational results are compared with those obtained from obtained in step (2) to get final values of generations.
the classical LMM, it is ascertained that the proposed method
produces the desired optimum ELD solutions in a very fast IV. TEST SYSTEM: THIRTEEN GENERATORS
and effective manner and that too with appreciable accuracy. Consider a system consisting of 13 generating units with the
data given in Table 1 [6]. The data for quadratic cost function
3.1 Algorithm in Table 1 is converted into exponential cost function and is
i) Generate the training and test patterns using Eq. (8) presented in Table 2 along with the active power limits of
considering the inequality constraints, i.e., power generators using MATLAB built-in function p = polyfit(x, y,
generation limits of the generators; this includes power n) which finds the coefficients of a polynomial p(x) of degree
demand, PD as the input and the corresponding optimal n that fits the data, p(x(i)) to y(i), in a least squares sense [10,
values of O as the output. 12].Table 3 presents the B-coefficients used to calculate the
transmission losses. Variation of cost as quadratic and
ii) Use the training patterns as obtained in step 1 to train the exponential function w.r.t. active power generation for
ANN. different generating units in the system is shown in Fig. 3.

516
 TABLE 1: DATA FOR QUADRATIC COST FUNCTIONS

Coefficients of Quadratic Cost

Unit Number Function
αi βi γi
1-6 0.00324 7.74 240.0
7-8 0.00284 8.60 126.0
(b)
9-10 0.00284 8.60 126.0
11 0.00028 8.10 550.0
12 0.00056 8.10 309.0
13 0.00056 8.10 307.0

TABLE 2: GENERATOR ACTIVE POWER LIMITS

Unit Coefficients of ECF

(c)
Number αi βi MW MW
1-6 9245 0.000837 60 180
7-8 13021 0.00066 40 120
9-10 13021 0.00066 55 120
11 117160 0.000069 0 680
12 58580 0.00014 0 360 (d)

13 58580 0.00014 0 360

V. RESULTS AND DISCUSSION

The code for the algorithm is developed in C-language. Input
to the ANN is the power demand whereas optimal value of λ
is the output. Training and test patterns are generated using
LMM. During training mode load demand is varied from 550
MW to 2850 MW with an interval of 100 MW.
(e)

(f)

Fig. 3. Graphical variation of cost w.r.t. power generation

(a)

TABLE 3: B-COEFFICIENTS (Bmnu 10-4)FOR 13-UNIT SYSTEM

Units 1 2 3 4 5 6 7 8 9 10 11 12 13
1 0.5 0.1 0.05 0.05 0.1 0.05 0.15 0.05 0.05 0.1 0.05 0.05 0.15
2 0.1 0.4 0.1 0.05 0.15 0.05 0.05 0 0.1 0.05 0.05 0 0.05
3 0.05 0.1 0.3 0.05 0.05 0 0.1 0 0.05 0 0.05 0 0.05

517
 2017 International Conference on Infocom Technologies and Unmanned Systems (ICTUS'2017),
Dec. 18-20, 2017, ADET, Amity University Dubai, UAE

Units 1 2 3 4 5 6 7 8 9 10 11 12 13
4 0.05 0.05 0.05 0.44 0.15 0.1 0 0 0.05 0 0.05 0 0
5 0.1 0.15 0.05 0.15 0.4 0.15 0.1 0.05 0 0 0.05 0 0
6 0.05 0.05 0 0.1 0.15 0.4 0.1 0.05 0 0.05 0 0 0
7 0.15 0.05 0.1 0 0.1 0.1 0.3 0.15 0.1 0 0.05 0 0
8 0.05 0 0 0 0.05 0.05 0.15 0.4 0.1 0.1 0 0.05 0
9 0.05 0.1 0.05 0.05 0 0 0.1 0.1 0.3 0.1 0.05 0.05 0
10 0.1 0.05 0 0 0 0.05 0.1 0.1 0.5 0.15 0.1 0.05 0
11 0.05 0.05 0.05 0.05 0.05 0 0.05 0 0.05 0.15 0.4 0.15 0.05
12 0.05 0 0 0 0 0 0 0.05 0.05 0.1 0.15 0.5 0.1
13 0.15 0.05 0.05 0 0 0 0 0 0 0.05 0.05 0.1 0.7

TABLE 4: PERFORMANCE OF NEURAL NETWORK DURING TRAINING MODE

Load Optimal Lambda ($/MWh) % Error Load Optimal Lambda ($/MWh) % Error
Demand LMM ANN Demand LMM ANN
(MW) (MW)
550 8.13294 8.16186 -0.00526 1750 8.80117 8.80259 -0.00008
650 8.24352 8.21861 0.00383 1850 8.85450 8.84909 0.00029
750 8.29364 8.27546 0.00242 1950 8.90831 8.89526 0.00067
850 8.34244 8.33217 0.00121 2050 8.96101 8.94196 0.00093
950 8.39155 8.38850 0.00032 2150 9.01216 8.99051 0.00101
1050 8.44129 8.44420 -0.00028 2250 9.06315 9.04277 0.00091
1150 8.49134 8.49905 -0.00067 2350 9.11257 9.10136 0.00048
1250 8.54186 8.55287 -0.00088 2450 9.16074 9.16966 -0.00036
1350 8.59284 8.60547 -0.00094 2550 9.22049 9.25151 -0.00122
1450 8.64414 8.65675 -0.00087 2650 9.30838 9.35024 -0.00158
1550 8.69607 8.70665 -0.00068 2750 9.44227 9.46650 -0.00088
1650 8.74846 8.75521 -0.00041 2850 9.65591 9.59513 0.00213
**************** Result for mean square error = 2.367924e-06 ***************

(a) (b)

Fig. 4. Convergence characteristics during training mode

518
 TABLE 6: Results of hybrid method during non-training mode

Load Total
Lambda Total Loss
Demand Generation
($/MWh) (MW)
(MW) (MW)
645.0 8.21373 648.770 3.857
804.0 8.31060 809.747 5.798
1299.0 8.55424 1315.523 16.511
1526.0 8.66335 1549.885 23.840
1761.0 8.79499 1794.037 33.068
1857.0 8.85132 1894.343 37.331
1947.0 8.89998 1988.588 41.586
2078.0 8.95423 2126.042 48.050
2215.0 8.98471 2270.152 55.166
Fig. 5. Neural network performance curve 2678.0 9.33840 2756.504 78.505
The network was trained till the value of mean square error 2845.0 9.49865 2930.316 85.316
reduced to 2.367924e-06. Convergence characteristics of the
ANN are shown in Fig. 4 and Fig. 5. Fig. 4 shows the VI. CONCLUSION
variation of squared error w.r.t. number of iterations whereas Over past few decades, a trend has been evolved in which
Fig. 5 presents mean square error w.r.t. power demand.Results hybrid approaches were employed by integrating one or more
obtained during training mode are provided in Table 4 which compatible approaches with the aim to minimize their
shows that the value of O as obtained by ANN is individual weaknesses and to combine their strengths to
approximately the same as that obtained by LMM. Table 5 produce accurate, fast and cost effective solutions. In this
presents the result of hybrid and LMM method. It is to be paper, a hybrid approach in which ANN is integrated with
noted from Table 5 that the values of λ and power generation LMM to find the solution of ELD problem has been presented
obtained by hybrid method and LMM are very close to each successfully. The practical constraints involved in ELD
other.In non-training mode, different values of load demand problem were duly considered.
are selected randomly as tabulated in Table 6. It is clear that
the proposed hybrid method works well for these inputs too. The proposed approach was tested on a system consisting of
The proposed method provides consistent results. 13 generating units. It generates the results quickly as the
value of λ which is required initially in LMM is set close to its
TABLE 5: COMPARISON OF RESULTS OF HYBRID optimal value using ANN. Further, the results showed that the
METHOD DURING NON-TRAINING MODE and LMM total generation cost for LMM and proposed hybrid method
METHOD were approximately the same. The results
Load Total Total For future works, it is suggested to extend the problem for real
Lambda
Demand Method Generation Loss power system comprising of large number of power
($/MWh)
(MW) (MW) (MW) generating units, e.g., 100 - 120 or even higher units. The time
LMM 755.080 5.058 8.29364 to train network using backpropagation algorithm is large but
750.0 once the network is trained, it gives results in much lesser
Hybrid 755.074 5.052 8.28693
time. Therefore, this work may be extended further such that
LMM 1265.197 15.133 8.54186
1250.0 the training time can be reduced.
Hybrid 1265.025 15.128 8.54186
LMM 1626.444 26.559 8.72211 REFERENCES
1600.0
Hybrid 1626.465 26.560 8.71807 [1] A.Y. Abdelaziz, et al., “A hybrid HNN-QP approach for
dynamic economic dispatch problem, ” Journal of Elect. Power
LMM 1887.052 37.012 8.85450 Systems Research, vol. 78, no. 10, pp. 1784-1788, 2008.
1850.0
Hybrid 1887.076 37.013 8.86571 [2] Bhattacharya and Chattopadhyay, “Hybrid differential
LMM 2149.283 49.164 8.98659 evolution with biogeography based optimization for solution of
2100.0 economic load dispatch”, IEEE Transaction on Power System,
Hybrid 2149.206 49.160 8.97045 vol. 25, no. 4, 2010.

519
 2017 International Conference on Infocom Technologies and Unmanned Systems (ICTUS'2017),
Dec. 18-20, 2017, ADET, Amity University Dubai, UAE

[3] C. Y. Maa and A. S. Michael, “Linear and quadratic Engineering, Oman, October 12-13, pp 87-93, 2008.
programming neural network analysis, ” IEEE Transaction on [13] M.T. Nadeem, et al., “A new hybrid approach for the solution
Neural Networks, vol. 3, no. 4, pp. 580-594, 1992. of nonconvex economic dispatch problem with valve-point
[4] D. P. Kothari & I. J. Nagrath, in Power System Engineering. 2nd effects”, Journal of Electrical Power System Research, vol. 80,
Ed., New Delhi, India, TMH Publication, 2008, pp. 1-375. no.9, pp.1128-1136, 2010.
[5] E. E. George, “Intrasystem transmission losses, ” AIEE [14] M. Younes, “A novel hybrid FFA-ACO algorithm for
Transactions, vol. 62, pp. 153-158, 1943. economic power dispatch” Journal of Control Engineering
[6] G. A. Bakare, et al., “Genetic algorithms based economic Applied Informatics, vol. 15, no. 2, pp. 67-77, 2013.
dispatch with applications to coordination of Nigerian thermal [15] P. K. Hota, et al., “Comparison of ANN models for economic
power plants, ” IEEE Power Engineering Society General load dispatch problem, ” Journal of the Institution of Engineers
Meeting, vol. 1, pp. 551-556, 2005. (India), vol. EL-81, no. 2, pp. 49-53, 2000.
[7] H.M. Dubey, et al., “Economic load dispatch by hybrid swarm [16] R. H. Liang, “A neural-based redispatch approach to dynamic
intelligence based gravitational search algorithm”, generation allocation, ” IEEE Transactionon Power Systems,
InternationalJournal of Intelligent System Applications, vol. 5, vol. 14, no. 4, pp. 1388- 1393, 1999.
no. 8, pp. 21–32, 2013. [17] S. Dipankar et al., “Study of Economic Load Dispatch by
[8] H. H. Mohamed, in Fundamentals of Artificial Neural Various Hybrid Optimization Techniques”, S. Bhattacharyya,
networks, New Delhi, India, PHI publication, 1998. et al. (eds.), Hybrid Soft Computing Approaches, Studies in
[9] L. K. Kirchmayer, in Economic Operation of Power System. Computational Intelligence, 611, Springer India, 2016.
New York, John Wiley and Sons, Inc., 1958, pp. 96-216. [18] S. Haykin, in Neural Networks: A Comprehensive Foundation,
[10] MATLAB, Optimization Toolbox for Use with MATLAB R 2nd Ed., Delhi, India, PHI Publication, 2005.
2010, Math Works, 2012. [19] S. Kumar, in Neural Networks: A Class Room Approach, New
[11] M. Ashouri and S. M. Hosseini, “Application of new hybrid Delhi, India, TMH Publication, 2008, pp. 41-186.
particle swarm optimization and gravitational search algorithm [20] S. M. V. Pandian, K. Thanushkodi, “Solving economic load
for non-convex economic load dispatch problem”, Journal of dispatch problem considering transmission losses by a Hybrid
Advance Computer Research, vol. 4, no. 2, pp.41–51, 2013. EP-EPSO algorithm for solving both smooth and non-smooth
[12] Mohammad Mohatram et al., “A New Approach to Economic cost function”, Int.Journal of Computer and Electrical
Load Dispatch using an Exponential Cost Function” Engineering, vol. 2, no. 3, pp. 1793-8163, 2010.
InternationalConference on Harnessing Technology: [21] X. F. Wang, et al., in Modern Power System Analysis, New
Challenges & Opportunities in GCC states, Energy, York, Springer Science Publication, 2008, pp. 71-89.
Environment & natural Resources, Caledonian College of

520