Professional Documents
Culture Documents
Abstract--- Electrical load forecasting is one of the important filtering [6], [7], time-series method such as ARMA (Auto
parts for smart grid system. The reliable prediction of the load Regressive Moving Average) [8], ARMAX (Auto Regressive
demand contributes to the efficient and economical operations Moving Average with exogeneous inputs) [9], and Box-
and planning. The artificial neural network is used extensively in Jenkins [10]. The statistical models have disadvantages that
load demand forecasting. The nonlinear nature of the electrical
cannot adjust rapidly with the abrupt changes of loads.
load demand conforms to the ability of the artificial neural
network in calculating the nonlinear relationship of inputs and Another method used an ANN (Artificial Neural
outputs. Among many models of neural networks, radial basis Network). The ANN was first introduced in 1943 by
neural networks yield superior performance in small error and McCulloch and Pitts as a simple model of the human brain.
fast simulation time.
However, it is challenge to design the radial basis neural
The ANN architecture is comprised of brain cells, or so called
networks. The excessive numbers of hidden neurons lead to neurons, connected together forming a network as shown in
lacking of generalization or so called overfitting problems. This Fig. 1. The ANN computes the nonlinear relationship between
paper proposes an approach to design the radial basis neural inputs and outputs by adjusting the weights and biases of the
networks that use as least numbers of hidden neurons as connections. The ANN was developed more than 5 decades
possible. The error criterion is optimized based on modified ago, and has been used in many applications such as pattern
genetic algorithm as the numbers of hidden neurons are recognition [11], prediction [12], controls [13], signal
incrementally increased. Simulation results of short term load processing [14], etc.
forecasting are calculated in Matlab, and compared to the
orthogonal least square error method. The proposed approach The artificial neural network has also been used widely in
gives better results with the same numbers of hidden neurons. the electric power system since its ability to compute a non-
linear relationship between inputs and outputs conforms to the
Index Terms—Artificial neural network, hidden neuron, non-linear characteristics of the electric power system. It
genetic algorithm, radial basis function
helps strengthen the power system infrastructure by
I. INTRODUCTION monitoring and assessing the stability and security [15]-[21].
Smart grid is the upgraded electrical network that operates It is used to analyze and enhance the power flow control [22]-
in more efficient, reliable, and secure manner with the [24], and estimate system states [24]. It is used to screen,
contribution of communication, information, control, and identify, and analyze possible faults during abnormal
management technologies. The electrical load forecasting situations [26]-[29]. The short-term load forecast with ANN
provides the intelligence to the smart grid operations and mostly used feedforward models [30]-[32].
planning. The load prediction is advantageous to dispatchers MATLAB is a computing software tool that is used widely
in decision making for generating electric power, load in many applications. The user-friendly neural network
switching, and fuel allocation [1]-[3]. Load forecasting is toolbox in MATLAB provides functions, applications, and
basically categorized into 3 types based on the period of user guides to design and model the ANN. There are multiple
prediction. The LTLF (Long-term Load Forecast) involves the models of the ANN in MATLAB. Each model has different
forecasting period of one year to several years. The MTLF characteristics, and yields various results. The comparative
(Medium-term Load Forecast) relates to the forecasting period study is essential to select the model properly. Previous works
of a few weeks to several months. The STLF (Short-term focused on the comparison between two models. [33]
Load Forecast), which is of interest in this paper, refers the compares the feedforward model with the radial basis neural
forecasting period of hours to one week. network. [34] compares the feedforward model with the
Traditional approaches to load forecasting use statistical recurrent neural network. This is not sufficient to recognize
models, which involved mathematical equations to predict
future values of loads. The statistical models include
exponential smoothing [4], regression method [5], Kalman
Input layer simulate the networks with the test data without the desired
Hidden layer Hidden layer
Output layer
outputs.
The neural network architecture is designed such that the
forecasting method is based on the iterative forecasting that
….
….
….
predicts the load demand in a series manner [37]. This method
….
166
2014 IEEE Conference on Technologies for Sustainability (SusTech)
results for both error and time. TABLE 3 summarizes the adjusted when each input is presented. Based on the
performance ranking including the times used in the case performance ranking, the FFTD and recurrent networks are
study 1 and 2. The radial basis neural network and recurrent simulated in the on-line training mode with test data of case
radial basis neural network give better results compared to the study 2 for 200 iterations or passes as shown in Fig. 5. With
other networks. sufficient iterations, the networks perform excellently
compared to the off-line training. Although the error is very
small, the simulation time is large as the iteration numbers
increase. Fig. 6 depicts the MSE and the time used of the
recurrent network in the on-line training mode. The MSE
decreases as the numbers of passes increase. Nonetheless, the
simulation time also increases. The radial basis neural
network, which is a nonrepetitive training network, can be
trained on-line by repeatedly training the network with a
group of training example and updating the weights and
Fig. 3 simulation results of case study 1
biases until the error is below the criterion [39].
Case study 1
Models
MSE Time (s.)
Feedforward 0.0046 1.5812
Radial basis 7.4731e-5 0.078
Linear 0.0020 0.4128
Focused
1.0235e-4 0.5092
time-delay
Recurrent 9.559e-5 1.2544
Recurrent
2.5933e-4 0.094
radial basis
Fig.6 The MSE and simulation time in on-line training mode
TABLE II SIMULATION RESULTS of CASE STUDY 2
167
2014 IEEE Conference on Technologies for Sustainability (SusTech)
numbers. As a result, each gene carries 0 or 1, as shown in algorithm is created by randomization of the real
Fig. 7. numbers. In this case the variable of interest is the
gene input weights of radial basis neural networks, which
their input weights are the subset of the input data.
0 1 0 0 1 0 1 1
Starting with one hidden neuron, each member of the
searching population is a vector of elements of the
Chromosome
input data obtained from randomization. Each vector
Fig. 7 Chromosome in the searching population of the genetic algorithm has the length n for the input matrix n m .
c) Find the fitness value of each member. The fitness
The genetic algorithm procedure is described as follows [40]: value is the MSE of the outputs and the desired
Initialize the population randomly. The size of the outputs. The outputs of the radial basis neural
population is set arbitrarily. Large population network are found in (2). The outputs of the
guarantees the convergence of the optimization. recurrent radial basis neural network are found in (3).
Nonetheless, it comes at a price. The simulation P, Y, LW, IW, b1, and b2 are input, output, layer
times are prolonged substantially.
weight, input weight, biases of the hidden layer and
At each generation, find the fitness value of all
the output layer, respectively.
members.
The members that give the lower fitness values are
Y LW (exp( IW P b1 ) 2 ) b2 (2)
passed to the next generation.
The rest of the population is selected to form parents.
The children of the next generation are produced by Y (t ) LW (exp( IW P b1 LW 'Y (t 1) ) 2 ) b2 (3)
crossover and mutation of the parents. The crossover
relates to the process of combination of two or more d) The members with lower fitness values are passed to
parents to form the children. The mutation refers to the next generation.
the process of the change in an individual parent. e) Select some members to be parents based on their
Then, replace the current population with the fitness values. The selection method is the stochastic
children. One generation is complete. uniform method [41].
The process is iterated until it reaches the stopping f) The children are produced by the crossover and
criterion, which is the predetermined maximum generation. mutation of the parents. As shown in Fig. 9, one of
The iteration ends when the generations exceed the specified each pair of parents is flipped from gene p1 to gene
value.
n-p1. Gene 1 to gene p1 and gene n-p1 to gene n are
B. Proposed Algorithm switched position where p1 is a randomized number.
The genetic algorithm optimization that uses strings of The mutation process is the exchange of two
genes or chromosomes is analogous to the format of input randomized genes p1 and gene p2 as shown in Fig.
weights of the radial basis function neural networks. Instead 10.
of number 0 or 1 in the binary system, a chromosome contains g) Replace the current population with the children.
elements, which are the subset of the input data when there is h) Increase one generation
one hidden neuron. If more hidden neurons are necessary, the i) Check whether the generation exceeds the
size of input weights is increased to k chromosomes or vectors predetermined value. If no, go back to step c). If yes,
for k hidden neurons. Each vector has n genes according to proceed to step j).
the row number of the input matrix. The proposed approach j) Check whether the MSE is still larger than the
uses the incremental growing of the hidden neurons combined desired error. If no, the procedure ends. If yes,
with the modified genetic algorithm to design the radial basis proceed to step k).
neural networks. The numbers of hidden neurons are as least k) Increase one hidden neuron.
as possible that still acquires the desired error. The proposed l) Go back to step b). The population is initialized
algorithm is depicted in Fig. 8 and described as follows: again. Each member of the initial population
contains 2 rows of vectors corresponding to the
a) In the beginning the number of hidden neuron is set
numbers of hidden neurons.
to 1.
b) The next step is to create the initial population. With the proposed approach, the network is built such that
Typically, the initial population of the genetic the hidden neuron is incrementally increased until the error of
the outputs and the desired outputs is met.
168
2014 IEEE Conference on Technologies for Sustainability (SusTech)
j)
TABLE 5 summarizes the comparative results of the proposed
Yes
MSE > desired error
approach and the OLS method with the recurrent radial basis
No neural network. Fig. 12 illustrates the simulation results of the
End
test data. The MSE is 4.519e-5. The error of both RBN and
Fig.8 Proposed algorithm RRBN networks are decreased from the results shown in table
1. This is because the data points are increased from 16 to 40
1 p1 n-p1 n
points. The results ensure one of the properties of neural
0.9 0.97 1.0 0.95 0.96 0.85
0.88 0.88
networks. The more the data points are given, the more the
accuracy is obtained.
0.88 0.96 0.85 0.95 1.0 0.88 0.9 0.97 TABLE IV MEAN SQUARED ERRORS OF RBN NETWORKS
Numbers of Orthogonal
Fig.9 Crossover method The proposed approach
Hidden neurons Least Square
1 0.0012 0.0034
1 p1 p2 n
2 2.3447e-4 0.0029
0.88 0.9 0.97 1.0 0.95 0.88 0.96 0.85
3 3.2659e-4 0.0020
4 5.1468e-5 0.0012
5 3.2527e-6 6.8513e-4
0.88 0.9 0.88 1.0 0.95 0.97 0.96 0.85 6 9.2695e-7 6.2697e-4
7 1.0502e-8 2.1013e-4
Fig.10 Mutation method 8 5.2518e-12 1.1743e-4
9 0 4.9304e-32
to forecast the load demand in the case study 1, which is the Numbers of
The proposed approach
Orthogonal
average-workday load demand of medium-sized businesses in Hidden neurons Least Square
1 0.2266e-3 0.0037
winter. The networks are simulated by separating the input
2 0.0175e-3 0.0032
data to training and test sets. There are 24-hour data of the 3 0.0219e-3 0.0020
training and test sets. 4-hour data is required as historical data 4 0.0052e-3 0.0013
to predict the load demand in later hours. Therefore, there are 5 0.0003e-3 7.9012e-4
20-hour predictions for both the training and test data. The 6 2.0664e-8 7.8934e-4
7 6.5128e-9 6.3322e-4
input variable is the load demand, which is sufficient to verify
8 3.6095e-13 3.2801e-4
the proposed algorithm. 9 4.9304e-33 6.1630e-33
169
2014 IEEE Conference on Technologies for Sustainability (SusTech)
REFERENCES
170
2014 IEEE Conference on Technologies for Sustainability (SusTech)
171