International Journal of Applied Information Systems (IJAIS) – ISSN : 2249-0868

Foundation of Computer Science FCS, New York, USA

Volume 10 – No.4, January 2016 – www.ijais.org

Short Term Electric Load Forecasting using Neural

Network and Genetic Algorithm

Mahrufat D. Olagoke A.A. Ayeni, PhD Moshood A. Hambali

Computer Science Telecommunication Science Computer Science
Department, Department, Department,
University of Ilorin, University of Ilorin, Federal University Wukari,
Ilorin-Kwara State Ilorin-Kwara State Wukari-Taraba State

ABSTRACT (STLF): This is usually from one hour to one week. Medium
The major predicament with electricity as a means of Term Load Forecasts (MTLF): This is usually from one week
transporting energy is that it cannot be stored unlike gas, oil, to a month. Long Term Load Forecasts (LTLF): This is longer
coal or hydrogen. Due to this, the electric power company than a year.
faces economical and technical problems in planning, The forecasts for different time horizons are important for
operation and control of electric power system. For the different operations with an electric power company. The
purpose of optimal planning and operation of an electric natures of these forecasts are different as well. For instance,
power system, there is need for appropriate evaluation of the for a particular region, it is possible to predict the next day
present and future electric load. Electric load forecasting is load with an accuracy of approximately 1-3%. However, it is
used by electric power company to anticipate the amount of impossible to predict the next year peak load with similar
energy needed to meet up with the demand. Various statistical accuracy, since accurate long-term weather forecasts are not
and artificial intelligence techniques have been applied to available [3]. STLF can help electric load planners to estimate
short term electric load forecasting in the past but were load flows and make decisions that prevent overloading.
hampered with some drawbacks. This paper presents another Timely implementations of such decision lead to the
approach for short term load forecasting with lead time of a improvement of network reliability and reduce occurrences of
day ahead (1-24 hours) using artificial neural network (ANN). equipment failures and blackouts. MTLF is useful in unit
The hidden layer in ANN model was generated using genetic maintenance and determining the quantity of fuel to purchase
algorithm instead of the usual practice of trial and error; the in power plants. LTLF used to supply electric power company
ANN model was trained by Levenberg Marquardt. The data management with prediction of future needs for expansion,
(daily load data of 330/132/33KV substation, Ganmo, Kwara equipment purchases and inter-tie tariff setting [3].
State for the month of May, 2014) used in training and
validation of the neural network was obtained from the For STLF, several factors should be considered such as time
Transmission Company of Nigeria, National Control Centre, factor, weather data and possible customer’s classes. The time
Osogbo, Osun State, Nigeria. The model for short term load factors include the time of the year, the day of the week and
forecast was designed and implemented with MATLAB the hour of the day. There are significant differences in load
package. The result was evaluated by Mean Absolute between weekdays and weekends. The load on different
Percentage Error (MAPE) of 4.705 for the forecasted day. weekdays also can behave differently. For example, Mondays
and Fridays being adjacent to weekends may have structurally
General Terms different loads than Tuesday through Thursday. This is
Short Term Electric Load Forecasting, Generic Algorithm, particularly true during the summer time. Holidays are more
Artificial Neural Network difficult to forecast than non-holidays because of their relative
infrequent occurrence. Weather conditions influence the load.
Keywords In fact, forecasted weather parameters are the most important
Short Term Electric Load Forecasting, Artificial Neural factors in STLF. Various weather variables could be
Network, Genetic Algorithm and Levenberg Marquardt considered for load forecasting. Temperature and humidity are
the most commonly used load predictors. Most electric
1. INTRODUCTION utilities serve customers of different types such as residential,
The key function of an electric power company is to supply commercial and industrial. The electric usage pattern is
customers with high quality electric energy in secured and different for customers that belong to different classes but is
economical manner. In order to do so, an electric power somewhat alike for customers within each class. Therefore,
company faces economical and technical problems in most utilities distinguish load behaviours on a class-by-class
planning, control and operation of electric power system. For basis [3].
the purpose of optimal planning and operation of electric Many methods have been employed for load forecasting in the
power system, there is need for proper evaluation of present past; they are based on various statistical methods such as
day and future electric power load [1, 2]. time series, regression, Box-Jenkins model, exponential
smoothing and so on. However, the statistical models provide
An electric load forecasting is used by an electric power
physical transparency in interpretation of data and reasonable
company to anticipate the amount of electric energy needed to
accuracy in STLF but with the associated problems, such as
supply so as to meet up the demand. Electric load forecasts
limited modeling and heavy computational effort make them
can be divided into three categories based on the planning
less preferable over intelligent techniques [2].
perspective of the duration: Short Term Load Forecasts

22
 International Journal of Applied Information Systems (IJAIS) – ISSN : 2249-0868
Foundation of Computer Science FCS, New York, USA
Volume 10 – No.4, January 2016 – www.ijais.org

In recent years, research has converged towards methods that methods cannot properly represent the complex non-linear
use artificial intelligence (AI) such as fuzzy logic, artificial relationships that exist between the load and series of factors
neural networks and expert system. Among the AI methods, that influence it [1].
artificial neural networks (ANN) have received a great deal of
attention by researchers in this area of research due to its As from 1990s, researchers began to use different approaches
flexibility in data modeling [2]. for STLF other than classical approach. The emphasis shifted
to the implementation of various AI techniques for STLF. AI
ANN is a mathematical model or computational model based techniques such as neural network, fuzzy logic and expert
on biological neural system. It consists of an interconnected systems have been applied to deal with the non-linearity, large
group of artificial neurons and processes information using a data sets requirement in implementing the STLF system and
connectionist approach to computation. ANN has been other difficulties in modeling of classical methods used for the
particularly successful in predicting the behaviour of large application of STLF [1]. Among the AI techniques available,
complex and nonlinear systems with a large number of different models of neural network have received a great deal
variables [4]. An ANN model takes a set of inputs and of attention by researchers in area of STLF due to its
produces one or more output and does not require any pre- flexibility in data modeling. Park [8] were among the first
assumptions on logical or analytical form between the input group of researchers who used the ANN approach for STLF.
and output variables. Its mapping capability is obtained
through the architecture of the network and training of its Srinvasan, Liew and Chen, [9] used an ANN based on back
parameters with experimental data by examining the patterns propagation for forecasting, and showed its superiority to
between the inputs and their corresponding outputs. A neural traditional methods. Chen, Tzeng, and Hwang, [10] applied a
network gains the knowledge to predict a system’s output for supervisory functional ANN technique to forecast for three
any given inputs. These features make a neural network an substations in Taiwan.
ideal model for predicting the behaviour of a complex Christopher and Francis, [11] developed a supervised ANN
engineering system. based model for STLF and evaluated the performance of the
However, their application to some real world problems has model by applying the actual load data of the Kenya National
been hampered by their long training time; there is no clear grid power system.
guideline for determining the number of hidden nodes in Amera, Lamya and Sozan, [12] analyzed and discussed a
hidden layer(s) etc., it is the usual practice to use trial and comprehensive approach for STLF using ANN. Their
error to find suitable neural network architecture for a given proposed architectures were trained and tested using previous
problem. This method is not only time consuming but may not two years actual load data obtained from Duhuk ELC control
generate optimal neural network architecture. One of the in Iraq. In their study, four ANN models were implemented
methods to overcome these difficulties in designing ANN is and validated with reasonable accuracy on real electric load
by using genetic algorithm (GA) [5, 6]. GA is a global search generation output data.
algorithm which operates on a population of rules. Based on
the mechanics of selection and natural genetics, it promotes Samsher and Unde, [1] carried out STLF for P.D.V.V.P COE,
over time the rules that perform well in a given environment Ahmednager college campus using ANN technique and was
and introduces into the population new rules to be tried. Rules implemented on MATLAB-10.
are coded as binary strings of finite length. The measure of the Adepoju, Ogunjuyigbe and Alawode, [13] used a multi-layer
rules’ performance is defined by their fitness function. perceptron (MLP) networks with back propagation to adjust
The aim of this research work is to develop an ANN model the weights of the neural network. The neural network was
for short term electric load forecasting with lead time of one trained using data from the Nigeria power company.
day ahead using genetic algorithm to design hidden nodes of
ANN model. Mohammed and Sanusi, [2] developed a multilayer feed
forward ANN model for 132/33KV substation, Kano, Nigeria
2. STATEMENT OF THE PROBLEM using Levenberg- Marquardt optimization technique to train
An electric power company faces economical and technical the network.
challenges in planning, control and operation of an electric
power system, for the purpose of optimal planning and Slobodan et al, [14] presented a method for STLF, based on
operation of an electric power system, there is need for proper ANN targeted for used in large-scale system as distribution
management system (DSM). The system comprises of a
evaluation of present day and future electric power load [1,
preprocessing unit (PPU) and a feed forward ANN ordered in
2].Various methods have been proposed to solve the
demanding task of STLF, but are associated with some a sequence. Functionality of the proposed method was tested
drawbacks. This paper therefore, proposes a new method for on recorded data from Serbian electrical utility.
STLF using ANN and GA. ANNs have been integrated with several other techniques to
improve their accuracy. Nima and Farshid, [15] presented a
3. LITERATURE REVIEW hybridization of a neural network with a novel stochastic
For years the problems of improving the accuracy of load search technique for STLF.
forecasting has been an important topic of research. Research
in this area in the last years has resulted in the development of Zohreh, hadi, and Mahdi, [16] used adaptive neural-fuzzy
numerous forecasting methods [7]. These methods are mainly inference system (ANFIS) to study the design of STLF
classified into two categories: Classical approaches and systems for the east of Iran.
artificial intelligence (AI) based techniques. Classical Chaturvedi and Sinha, [17] Presented the design of STLF
approaches are based on various statistical modeling methods system for 33KV substation of Dayalbagh Educational
such as time- series, regression, exponential smoothing, Box- Institute, Agra using Neural-wavelet based hybrid model.
Jenkins model and Kalman filters. However, these classical

23
 International Journal of Applied Information Systems (IJAIS) – ISSN : 2249-0868
Foundation of Computer Science FCS, New York, USA
Volume 10 – No.4, January 2016 – www.ijais.org

This research work therefore, proposes a new method for of the training set.
STLF using ANN and genetic algorithm. The hidden layer of
neural network will be designed using GA and trained by 4.3 Network Structure Design
Levenberg Marquardt. The next step after splitting the training and validation data set
is to design the structure for neural network. This has to do
4. METHODOLOGY with selecting network topology and determination of the
4.1 Data Collection and Preparation input nodes, output nodes, number of hidden layers and
The historical data (load data and temperature) were collected number of hidden nodes. Network topology is usually
for this research. The historical load data was obtained from determined based on the type of task to be performed by the
the transmission company of Nigeria, National control centre, network proposed. Multilayer feed forward neural network
Osogbo, Osun State, Nigeria. And the temperature was has been successfully applied for prediction (regression and
obtained from the internet. The data collected is the actual classification) problems [4]. Multilayer feed forward will be
load data of the 330/132/33 KV substation, Ganmo, Kwara used in this research. The number of input nodes is usually set
State for the Month of May, 2014. The first twenty one days equal to the number of input variables.
load data and temperature will be used for training the The following are the input variables for this research
network and the load data and temperature for the remaining
days in the Month will be used for the network validation. (i) Hour of the day
Table 1 shows the sample of a day load dataset used for (ii) Temperature (average temperature)
training the network:
(iii) Day of the week
Table 1. Sample of a Day Load Dataset
The output of the neural network represents the forecasted
Date Hrs Load Temp load data for the forecasting day. Determining the number of
1/5/2014 1 25.000 28.50 hidden layer(s) and number of nodes in the hidden layer(s) is
a critical decision in the design of neural networks. Too many
1/5/2014 2 13.000 28.50 hidden neurons lead to many trainable weights, which can
1/5/2014 3 0.000 28.50 make a neural network to become erratic and unreliable. On
the other hand too few hidden neurons restrict the learning
1/5/2014 4 0.000 28.50 ability of a neural network and deteriorate its approximation
1/5/2014 5 0.000 28.50 performance [18]. However, there is no clear guideline for
determining the number of neurons in hidden layer(s). The
1/5/2014 6 0.000 28.50 usual practice is by using trial and error which may not yield
1/5/2014 7 0.000 28.50 an optimal network design and the process is time consuming.
One of the ways of overcoming these difficulties in designing
1/5/2014 8 0.000 28.50 ANN is to use GA [19, 6].
1/5/2014 9 17.400 28.50 Using GA to design ANN encompasses two major attracting
1/5/2014 10 56.000 28.50 features. First, it automates the design of the network which
would have done manually. Second, the design process is
1/5/2014 11 57.000 28.50 analogous to a biological process in which the neural network
1/5/2014 12 24.000 28.50 blueprints are encoded in chromosome. The following are the
GA steps for selecting the optimum topology for the neural
1/5/2014 13 31.000 28.50 network model.
1/5/2014 14 31.000 28.50 1. Set the initial population (neural network parameters
1/5/2014 15 30.000 28.50 are represented by the chromosome).
1/5/2014 16 40.000 28.50 2. Evaluate the fitness function for each individual in
45.000 the population.
1/5/2014 17 28.50
45.000 3. Select the first two individuals with highest fitness
1/5/2014 18 28.50
value in the current generation and apply cross over
1/5/2014 19 18.000 28.50 and mutation operators.
1/5/2014 20 0.000 28.50 4. Repeat from step 2 until all individuals in the
57.000 population met the convergence criteria.
1/5/2014 21 28.50
1/5/2014 22 53.000 28.50 Decode the converged individuals in the final generation and
obtain the optimized neural network structure with optimal
1/5/2014 23 62.000 28.50 number of input neurons and hidden layer neurons.
1/5/2014 24 59.000 28.50
4.4 Network Training
After the network has been designed, the next step is to train
4.2 Data Preprocessing the network. Training of an artificial neural is an iterative
Scaling of raw input data is usually necessary to reduce the process that has to do with adjusting of the connection weight.
bias caused by different measuring unit of original input Back propagation algorithm has been widely used in the past
variables. The approach used for scaling the network input as a basic learning algorithm for training feed-forward neural
and target was to normalize the mean and standard deviation networks, but takes a very long time in training due to the

24
 International Journal of Applied Information Systems (IJAIS) – ISSN : 2249-0868
Foundation of Computer Science FCS, New York, USA
Volume 10 – No.4, January 2016 – www.ijais.org

nature of gradient descent [20]. Several techniques have been discard the weights, increase λ using v and go to
used to improve the performance of back propagation, among step 4
them is by incorporating Levenberg Marquardt. Levenberg
Marquardt will be adopted for training neural network in this 8. Else decrease λ using v and stop.
research. Levenberg Marquardt is numerical optimization 4.6 Network Validation
based technique.
After the network has been properly trained, it needs to be
For Lenvenberg-Marquardt optimization the performance validated on its performance of generalization. New data that
index to be optimized is defined as: are not used during training process are feed in to the network
to see whether it can predict well at these “unseen” data
F(w) = ] ------(1)
T
5. RESULT AND DISCUSSION
Where w = [w1,w2, …. wn] consists of all weights of the The data set is preprocessed before training and validation
network, dkp is the desired value of the kth output and the pth with function mapstd. This function allows the inputs and
pattern, Okp is the actual value of the kth output and the pth targets to have a mean of zero and standard deviation of unity.
pattern, N is the number of the weights, P is the number of the
patterns and k is the number of outputs. Levenberg Marquardt 5.1 Training, Validation and Testing of the
combines the speed of Gauss-newton’s method and the
stability of error back propagation algorithm during training
Network
steps, Levenberg Marquardt learning rule utilizes a scalar Neural network parameters used in the experiment, is
summarized in table 2.
(combination coefficient)  to control the shift between
Gauss-netwon method and error back propagation algorithm Table 2: Summary of Network Parameters
during training steps (equation 2). When  is large, the
learning process approximate error back propagation Network topology: Multi-layer feed forward neural network
algorithm. When  is small, the learning process essentially
Input variables: day of the week, hour of the day and the
follows the Gauss–netwon rule. The value of  increases, if
average temperature
the error increases after an iteration step and decreases when a
step reduces the error note that J is the jacobian matrix
Number of hidden neuron: automated (using GA)
defined on error vector e.
W = (JTJ+I)-1JTe ______ (2) Number of epoch – 91

The Jacobian matrix is defined as Training algorithm – Levenberg –Marquardt (trainlm)

Transfer function used in the hidden layer – sigmoid function

Transfer function used in the output layer - sigmoid function

Sigmoid function is defines as:

YT = 1/ (1+ e-y)

Where YT is the transformed (i.e. normalized) value of Y

In neural network the jacobian matrix is a N- by- W matrix,

where N is the number of entries in our training set and W is 5.2 Results Analysis and Evaluation
the total number of parameters (weights + biases) of the Performance plot (fig. 1) indicates the behaviour of the
network. network during training, validation and testing. The
performance goal was met at 91 epochs (number of
4.5 General Levenberg-Marquardt iterations). This figure shows the mean absolute error (MAE)
Algorithm versus epochs; revealed that, mean absolute error was
Each learning iteration in Levenberg Marquardt contains the decreasing during the learning process.
following basic steps
1. Compute the jacobian matrix
2. Compute the error gradient g = JtE
3. Approximate the hessian using the cross product
jacobian H = JTJ
4. Solve (H + λI) = g to find 
5. Update the network weights w using 
6. Recalculate the sum of squared errors using updated
weights
7. If the sum of the squared error has not decreased,

25
 International Journal of Applied Information Systems (IJAIS) – ISSN : 2249-0868
Foundation of Computer Science FCS, New York, USA
Volume 10 – No.4, January 2016 – www.ijais.org

Fig. 3: Training State Plot

Fig. 1: Performance Plot
The figure 3 consists of three different graphs. The first plot is
the learning function versus number of epochs. This shows the
trend of the gradient values as the number of computational
iterations increases. It is necessary in monitoring the manner
in which the training progresses. The second plot is that of the
learning rate (mu) against increasing number of epochs. This
plot is essential in monitoring the rate at which the computed
network error reduces during the progress of the training. The
final plot here is that of the validation checks carried out
automatically any time a sudden change is observed in the
network gradient computation is carried out.
Figure 4 shows the comparison between actual load data
obtained from the power company which is very close to the
results obtained from the trained neural network model output
data.
Fig. 2: Regression Plot
5.3 Result Evaluation
This graph (fig. 2) shows the closeness between the outputs Mean absolute percentage error (MAPE) is used in the
and the target. A R value of one (R = 1) indicates a perfect evaluation of the results and is defines as:
relation between the target and the output. From the
regression plot R = 0.90215 for training, R= 0.8337 for
validation, R = 0.82219 for testing and R= 0.86856 for all.
MAPE = |xi – yi| x 100
This shows that the target and output values are very close. xi
This means that the network predicted the output Where xi represent the actual values and yi represent the
satisfactorily. predicted values.
The MAPE for the forecasting day is 4.705% and the MAPE
for the historical load is 4.88%

5.4 Output Data

The hourly forecast for the next day with average temperature
of 300 C.

26
 International Journal of Applied Information Systems (IJAIS) – ISSN : 2249-0868
Foundation of Computer Science FCS, New York, USA
Volume 10 – No.4, January 2016 – www.ijais.org

Fig. 4: Comparison of forecasted load and Actual load

Table 3: Forecasted Load for a Day analogous to a biological process in which the neural network
blueprints are encoded in chromosomes. The training,
HOURS PREDICTED HOURS PREDICTED validation and testing of the ANN was conducted offline
LOAD LOAD using MATLAB. The results obtained in this work confirm
1 49.4 MWatts 2 55.2 MWatts the efficiency of ANN in short term electric load forecasting.
The model was able to determine the non-linear relationship
3 60.3 MWatts 4 64.2 MWatts that exists between the historical load data and temperature.
With forecasted average temperature for the forecasting day,
5 65.9 MWatts 6 63.9 MWatts the model could make a prediction of the next day (hourly)
load.
7 56.8 MWatts 8 47.6 MWatts
9 41.8 MWatts 10 41.4 MWatts 7. FUTURE WORK
For future works, the error in the network can be further
11 47.3 MWatts 12 58.7 MWatts reduced if a larger set of data is used for the network training.
13 65.0 MWatts 14 64.4 MWatts Also, additional information such as customer class can be
included in the network so as to obtain a more representative
15 58.7 MWatts 16 50.0 MWatts forecast of future load.
17 39.7 MWatts 18 28.9 MWatts 8. REFERENCES
19 18.1 MWatts 20 7.7 MWatts [1] Samsher Kadir Shiekh and Unde M. G. 2012. Short Term
Load Forecasting Using ANN Technique. International
21 2.1 MWatts 22 11.3 MWatts Journal of Engineering Sciences and Emerging
Technologies Vol 1.Issue 2, Pp: 99-107.
23 19.7 MWatts 24 27.6 MWatts
[2] Mohammed Buhari and Sanusi Sani Adamu 2012. Short-
Term Load Forecasting Using ANN. Proceedings of the
5.5 Contribution to Knowledge International Multi-conference of Engineers and
STLF play a vital role in planning and operation of electric Computer Scientist IMECS 2012, Hong Kong Vol. 1 Pg.
power system, it can help an electric load planners to estimate 978-988.
load flows and to make decisions that can be prevent [3] Eugene A. Feinberg and Dora Genethliou 2006. Book:
overloading. Timely implementations of such decisions can Title / Chapter Applied Mathematics for Power Systems.
lead to the improvement of network reliability and reduces “Chapter 12 Load Forecasting” Weather (2006), Issue:
occurrence of equipment failures and blackout. August, Publisher: Springer, Pg 269 -285
6. CONCLUSION [4] Zhao H., Ren Z. and Huang W. 1997. Short Term
A new method for short term electric load forecasting was Forecasting Considering Weekly Period Based on
developed using ANN optimized with GA. The use of GA to Periodical Auto Regression. Proceedings of the Chiness
design ANN encompasses two major attracting features. Society of Electrical Engineers.
Firstly, it automates the design of the network which would
[5] Geoffrey F. Miller, Peter M. Todd and Shailesh U.
have been done manually. Secondly, the design process is

27
 International Journal of Applied Information Systems (IJAIS) – ISSN : 2249-0868
Foundation of Computer Science FCS, New York, USA
Volume 10 – No.4, January 2016 – www.ijais.org

Hegde 1989. Designing Neural Networks Using Genetic International Journal of Soft Computing and Engineering
Algorithms. Proceedings of the Third International (IJSCE), Vol. 3, Pg. 56-58.
Conference on Genetic Algorithms, Edited by J. David
Schaffer Pg. 379-384. [13] Adepoju, G. A, Ogunjuyigbe, S. O. A and Alawode, K.
O. 2007. Application of Neural network to Load
[6] Belal Al- Khateeb and Maha Mahmood 2013. A Forecasting in Nigerian Electrical Power System. The
Framework for an Automatic Generation of Neural Pacific Journal of Science and Technology. Vol. 8
Networks. International Journal of Computer Science Number 1. Pg 68-72.
Issues. Vol. 10 issue 3 Pg. 59-62.
[14] Slobodan Ilic, Aleksandar Selakov, Srdan Vukmirovic,
[7] Gwo-Chung Liao and Ta- Peng Tsao 2006. Application Aleksandar Erdeljan and Filip Kulic 2013. Short Term
of a Fuzzy Neural Network Combined With Chaos Load Forecasting in Large Scale Electrical Utility Using
Genetic Algorithm and Simulated Annealing to Short Artificial Neural Network, Journal of Scientific and
Term Load Forecasting. Evolutionary Computation IEEE Industrial Research Vol. 72, Pg 739-745.
Transaction Vol. 10(3) Pg 330-340.
[15] Nima Amjady and Farshid Keynia 2011. A New Neural
[8] Park, D. C., El- Sharkawi, M. A., Marks, R. A. II, Atlas, Network Approach to Short Term Load Forecasting of
L. E. and Danborg, M. J. 1991. Electric Load Electrical Power Systems. Energies ISSN 1996-1073.
Forecasting Using an Artificial Neural Network. IEEE
Transactions of Power Engineering Vol. 6 Pg. 442-449. [16] Zohreh Souzanchi k, hadi Fanace T, Mahdi Yagbobi
2010. A Multi Adaptive Neuro Fuzzy Inference System
[9] Srinvasan D., Liew A. C. and Chen S. P. 1991. A Novel For STLF by Using Previous Day Features. International
Approach to Electric Load Forecasting Based on Neural Conference on Electronics and Information Engineering.
Networks IEEE International Joint Conference of Neural
Networks, Singapore. [17] Chaturvedi D.K and Premdayal S. A. 2013. Neural-
Wavelet Based Model for Short Term Load Forecasting.
[10] Christopher A. Moturi and Francis K. Kioko 2013. Use Control Theory and Informatics. Vol. 3 No. 2 Pg. 42-45.
of Artificial Neural Networks for Short Term Electricity
Load Forecasting of Kenya National Grid Power System. [18] Vladimir, C and Mulier, F. 1998. Learning from Data:
International Journal of Computer Applications Volume Concepts, Theory and Methods, John Wiley & Sons. Inc,
63 Number 2 Pg 25-30. New York.

[11] Chen, C. S, Tzeng, T. M and Hwang, J. C. 1996. [19] Goldberg, D.E. 1989. Genetic Algorithms in Search,
Application of Artificial Neural Networks to Substation Optimization and Machine. Learning, Addison-Wesley,
Load Forecasting. Electric Power Systems Research, Reading, Massachusetts.
38,153-160. [20] Demith H. and beale M. (1998). Neural Network
[12] Amera Ismail Melhum, Lamya Abdullateef Omar and Toolbox, for Use with MATLAB, the Math Works
Sozan Addulla Mahmood 2013. Short Term Load www.mathworks.com
Forecasting Using Artificial Neural Network.

28