International Journal of Soft Computing and Engineering (IJSCE)

ISSN: 2231-2307, Volume-1, Issue-6, January 2012

Application of Multi-Layered Perceptron Neural network

(MLPNN) to Combined Economic and Emission Dispatch
J.Hamidi
Electrical Engineering Department, Islamic Azad University, Sarakhs branch, Sarakhs, Iran.

Abstract— This paper presents a multi-layered perceptron result with other methods like the classical lambda, and
neural network (MLPNN) method to solve the combined genetic algorithms.
economic and emission dispatch (CEED) problem. The harmful
II. THE PROBLEM FORMULATION
ecological effects caused by the emission of particulate and
gaseous pollutants like sulfur dioxide (SO2) and oxides of The classic economic dispatch problem aims to supply the
nitrogen ( NOx ) can be reduced by adequate distribution of required quantity of power at the lowest possible cost [6]. The
load between the plants of a power system. However, this leads dispatch problem can be stated mathematically as follows:
to a noticeable increase in the operating cost of the plants. This
To minimize the total fuel cost at thermal plants:
paper presents the (MLPNN) method applied for the successful
operation of the power system subject to economical and n
environmental constraints. The proposed MLP NN method is F  MIN  ai  bi Pi  c i Pi 2  (1)
tested for a three plant thermal power system and the results are i 1
compared with the solutions obtained from the classical lambda
iterative technique and simple genetic algorithm (SGA) refined
Subject to the equality real power balance constraints:
genetic algorithm (RGA) method. n n

Index Terms— Economic dispatch, Emission dispatch, the

P  P
i 1
i
D
 P L  0 , P L   B i Pi 2
i 1
(2)

multi-layered perceptron neural network. The inequality constraint of limits on the generator outputs is:

I. INTRODUCTION Pmin,i  Pi  Pmax,i (3)

The combined economic and emission dispatch (CEED)

problem is to determine the optimal combination of power Where ai, bi and ci are the cost coefficients of the i-th
outputs for all generating units which minimize the total generator and n is the number of generators committed to the
fuel cost while satisfying load demand and operational operating system. Pi is the power output of the i-th generator,
constraints [1-2]. CEED problem is a multiobjective PD is the load demand and PL represents the transmission
mathematical programming problem which is concerned losses. However there is a large financial beneficial from the
with the attempt to improve each objective simultaneously classical dispatch strategy described above, it tends to
satisfying the standard load constraints and fuel constraints. produce high SO2 and NOx emissions. An alternative dispatch
The objective functions are made up of sum of linear or strategy to satisfy the environmental requirement is to
piecewise linear functions, each of which is a function of minimize operation cost under environmental constraints.
one or more variables from only one time step. Some Emission control can be included in conventional economic
constraints are made up of variables drawn from one time dispatch by adding the environmental cost to the normal
step whereas others span two or more time steps. dispatch. The emissions need to be converted as an
environmental cost and added to the generation cost. The
Currently neural networks have become the popular tool for objective function then becomes:
solving optimization problem involving complex objective
function and have been applied to many areas of power MinimizeC W 0 F W 1E S W 2 E N (4)
system engineering such as security assessment, load
forecasting, and economic dispatch. In [3-5] have presented Where ES is the SO2 emission function, EN is the NOx
method for economic load dispatch for units with a Hopfield emission function. W0, W1 and W2 are cost, SO2 emission and
Modeling Framework. In this paper, we have applied NOx emission weights respectively.
MLPNN approaches to CEED problem and compare our
In this paper, like fuel cost curves, the SO2 and the NOx
Manuscript received December 10, 2011. curves can be expressed as follows:
Javad Hamidi, Electrical Engineering Department, Islamic Azad
University, Sarakhs Branch, Sarakhs, Iran, +985116655937/
+9151193964., (e-mail: [email protected]).

242
Application of Multi-Layered Perceptron Neural network (MLPNN) to Combined Economic and Emission Dispatch

n However, the performance criteria of a MLP NN could be

 (d
2
Es  i  ei Pi  f i Pi ) (5) achieved by modifying the training data distribution. After a
i 1
successful learning phase, a MLP neural network will have
n
E n   ( g i  hi Pi  k i Pi ) an ability to generalize for the unseen data. During the
2
(6)
i 1
training phase the weights are optimized in order to
minimize a predefined error function. MLP neural networks
Where di , ei , fi , gi , hi and ki are parameters estimated on are trained using the back propagation (BP) algorithm which
the basis of unit emissions test results. is a gradient based supervised learning method. According
In this model, when emission weights are equal to zero, the to the algorithm, a mean squared error between the predicted
objective function becomes a classical economic dispatch and target values for the given input parameters is
problem. In this economic dispatch option, units are to propagated backward to adjust the interconnection between
minimize the total system production costs. When cost neurons in order to minimize the pre-defined error. In this
weight is set to zero, the problem becomes emission structure each neuron in a layer is mapping the sum-of-
minimization. In this case, units are to minimize the amount weighted input into an activation level that is determined by
of emissions. When weights are not zero in the objective an activation function. The most commonly used activation
function, the problem becomes minimizing the fuel cost plus functions are the sigmoid, the tangent hyperbolic, and the
emission at the same time. linear activation function. If a sigmoid function is used, the
III. OVERVIEW OF THE MULTILAYER NEURAL output of the kth neuron, Ok, in nth layer is determined by
NETWORK the following equation:

According to [7], the term multi-layered perceptron (MLP) 1

Ok  , net k   jW jk O j (7)
is used for the neural networks with a structure of input 1  exp(net k )
layer, one or more hidden layers and an output layer. Each
of the layers consists of inter-connected assembly of simple And Oj is the output of a neuron in the previous layer and
processing elements called neurons. These processing Wjk is weight between neurons (jth neuron in a layer and kth
elements are organized in a layered fashion. Each neuron in neuron in subsequent layer). In this structure the adjustable
a layer is connected to the neuron in the subsequent layer network parameters are optimized based on the BP
and so on. The interconnections between layers are called algorithm as follows:
weights. Despite of their simplified structure, neural
networks have ability to mimic human characteristics of
W jknew  W jk  W jk (8)
problem solving via learning and generalization. MLP can
be used to model non-linear systems due to their ability to
Where W jk is weight update for the connection Wjk.
learn the system behavior under inspection from samples. The weight update is calculated as follows:
For a successful application of MLP neural networks, one
E
should determine internal parameters, such as initial weights W jk   ( ) (9)
and network structure, to meet required performance criteria. W jk
In engineering neural network models, this is one of the
main problems, as an inadequate network would be unable Where η is learning rate which is chosen between 0 and 1
to learn. The problem of finding a suitable architecture and and E is the error (the cost function) defined as:
the corresponding weights of the network, however, is a very 1

2
complex task and main difficulties arise from the fact that E (t k  Ok ) (10)
the theory does not provide instruments to choice optimal
2 k
values for these parameters. This, in turn, results in the need Where tk is the target for the kth neuron in the output
to train an MLPNN multiple times, effectively using layer. Notice that Oi and tk pair constitutes the input and
different initializations and architectures. Consequently, the output parameters in the training data set.
practitioner is involved in a process that requires more
development time as well as experience and intuition. IV. IMPLEMENTATION OF MLP NN IN CEED
PROBLEM
Furthermore, this approach does not guarantee the
optimality of the obtained parameters for a given problem. To solve the constrained economic dispatch problem using
The common practice in the literature is to determine the the MLP Neural Network, a training set should be
number of neurons in the hidden layers by experience and constructed by combining the input and target pattern as
some rule of thumb. Therefore, there could be more than one
topology for an MLP to model a process successfully.

243
 International Journal of Soft Computing and Engineering (IJSCE)
ISSN: 2231-2307, Volume-1, Issue-6, January 2012

network (except the total load for input ) were normalized

pairs, i.e. (Oi and tk). Input patterns Oi‘s are consist of total with following equations :
power load sum (Pi) in (2) and the corresponding target
x x 1
patterns consist of the unit powers constraints Pi and the x , x norm  (11)
total cost in (1) . An additional set of the target parameters is Std x 1  exp(x )
applicable as target patterns in the case of environmental
economic dispatch problem, such as environmental Where x is the value of the component that must
constraints NOx, and SO2. The number of neuron in the normalize, x and Std x are , respectively , the mean and
input layer of MLPNN in 3 units ED problem is determined standard deviation of the same component. x norm is the
by the total load. The number of the neuron in the output
component value after normalization. Our input for network
layer of MLP is set by the power generated by 3 units, the
is total load and our output is P1, P2, P3, Ploss, NOx -emission and
total operation cost, power losses and pollutants of NO x
Tt-otal. That
Therefore, a MLP NN with a structure of 1–6–6–6 is chosen
Pi = power who generate in pant i.Where (i=1, 2, 3)
to solve the 3 units CEED problem. The MLPNN algorithm
Ploss = total losses for system.
is coded in MATLAB using Neural Network Toolbox with
NOx -emission = total weight of NOx for every hour.
an adaptive gradient descent BP algorithm. The learning rate
Tt-otal = total cost for system work.
is chosen as 0.01 and %5 increment is allowed. The
maximum allowed iteration is set to 1000. For 3 units CEED
We design a network who have a one input neuron (total
problem, training and test data pairs are produced by using
load ) and sex output neurons for (P 1 , P2 P3 , Ploss , NOx -
genetic algorithm (GA) and implemented in MATLAB [8-
emission and Tt-otal ). Hidden layers consist of two layers with
10]. As it is known that neural learning is sensitive to the
sex neuron in every layer. Our transfer function of hidden
initial weight settings, a fixed initial weight set should be
layers is ‗tansig‘, 'purelin' for output layer. Our back
used to fairly evaluate the performance of MLP neural
propagation error method is ‗train‘. Training parameters are:
network. If the initial weights are randomized at each
Epoch = 50, Learning rate = .05, Max number of iteration
training session, different final weights will be achieved and
=1000, Error tolerance=1e-10
accordingly the obtained solution will be different.
The system tested consists of three thermal units. The
V. SIMULATION RESULT parameters of this system and result of simulation are given
in Table (1,2).
For simulate our result, we consider a system with three
plants. Our simulation is performed in MATLAB neural
network toolbox. all components of the data presented to

Table 1: The parameters of tested system

Unit ai bi ci di ei fi Pimin Pimax

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

1 1243.5311 38.30553 0.03546 40.26690 0.00683 35 210

2 1658.56960 36.32782 0.02111 42.89553 0.00461 130 325

3 1356.65920 38.27041 0.01799 42.89553 0.00461 125 315

244
Application of Multi-Layered Perceptron Neural network (MLPNN) to Combined Economic and Emission Dispatch

Table 2: Comparison of test results obtained from conventional method, SGA and RGA, MLP NN

Power P1, P2, P3, PL, NOx Total

No Demand, Method MW MW MW MW Emission, Cost,

MW Kg/h Rs/h

Conventional 102.66 153.87 151.13 7.41 201.5 29922

SGA 99.47 147.26 161.08 7.69 201.35 29820

1 RGA 99.52 165.11 142.32 7.39 201.21 29812

400
MLP NN 99.03 158.44 135.12 7.45 201.19 29818

Conventional 120.00 192.81 190.08 11.88 312.00 39458

SGA 127.54 200.58 183.43 11.80 311.89 39441

2
RGA 127.52 193.50 190.62 11.70 311.33 39433
500
MLP NN 126.98 198.88 185.89 11.75 309.80 39400

Conventional 182.62 271.27 269.47 23.37 652.55 66690

SGA 190.11 274.71 258.21 23.29 652.04 66659

3 RGA 187.21 273.56 262.35 23.28 651.60 66631

700 MLP NN 180.28 290.48 252.17 22.93 649.28 66640

Figure 4: Comparison of NOx emission obtained from various methods for Figure 1: Comparison of total cost obtained from various methods for
power demands of 400MW power demands of 400MW

245
 International Journal of Soft Computing and Engineering (IJSCE)
ISSN: 2231-2307, Volume-1, Issue-6, January 2012

Figure 5: Comparison of NOx emission obtained from various methods for Figure 2: Comparison of total cost obtained from various methods for power
power demands of 500MW demands of 500MW

Figure 6: Comparison of NOx emission obtained from various methods for Figure 3: Comparison of total cost obtained from various methods for power
power demands of 700MW demands of 700MW

VI. CONCLUSION [4] T.Yalcinoz, B.J.Cory, M.J.Short, ― Hopfield Neural Network

Approaches To Economic Dispatch Problem, ‖ Electrical Power and
Energy Systems, Vol. 23,No. 6, August 2001 , pp. 435-442(8).
In this paper, solution to the economic dispatch problem as a [5] T.Yalcinoz, U.Hassan, ― Environmentally Constrained Economic
constrained optimization problem has been obtained using Dispatch via Neural Networks, ‖ Nigde University, and Nigde 51100,
Turkey.
MLPNN. The MLP method has advantage of being a quick [6] H. Altun, T. Yalcinoz, ― Implementing soft computing techniques to
method compared to the iterative optimization methods such solve economic dispatch problem in power systems, ‖ Expert Systems
as GA. once a MLPNN is trained, it provides solutions in a with Applications in Press, 2007.
[7] Y. S. Haruna, ― Comparison of Economic Load Dispatch Using
cycle. According to table response of net is acceptable and Genetic Algorithm and Other Optimization Methods, ‖ M. Eng.
in some case, better from other approaches such Degree Thesis, A.T. B. University, Bauchi, Nigeria, 2003.
[8] Yalcinoz T., Altun H. and Uzam M., ―Economic dispatch solution
conventional and SGA and RGA method. using a genetic algorithm based on arithmetic crossover, ‖ IEEE Porto
We propose, using of combined method of genetic and PowerTech2001, Porto, Portugal, 2001.
[9] S.M.R Slochanal, Dr M Sudhakaran, ― Application of Refined Genetic
neural network for better response for CEED. Three Algorithm to Combine Economic and Emission Dispatch, ‖, IE (I)
methods (conventional, SGA, RGA [10]) are came in Journal.ELvol 85, September 2004.
following table for comparison
. Javad Hamidi was born in Mashhad, Iran in 1983.
He received the B.Sc. degree in Electronic Engineering
REFERENCES
from university of Guilan, Rasht, Iran in 2006 and
M.Sc. degree in Power Electrical Engineering from
[1] Zwe-Lee Gaing, ― Particle swarm optimization to solving the University of Tabriz, Iran in 2009. His research
economic dispatch considering the generator constraints, ‖ IEEE interests include Power System Harmonics and Power
Transactions on Power Systems, vol.18, No.3, pp. 1187-1195, Aug. Electronic Converter and Application of modern
2003 control and Intelligent Algorithm in Power Systems.
[2] Ching-Tzong Su, Chien-Tung Lin, ― New Approach with a Hopfield
Modeling Framework to Economic Dispatch, ‖ IEEE Trans.On Power
Syst, vol. 15,No. 2, 2000.
[3] M. Basu, ― Fuel Constrained Economic Emission Load Dispatch Using
Hopfield Neural Networks, ‖ Electric Power Systems Research 63,
2002 , 51-57.

246