You are on page 1of 4

Prediction of Electricity Consumption Based on Genetic Algorithm - RBF Neural

Network

Wu Jian
Zeng Qing-wei, Xu Zhi-Hai
Dispatching and Communication Center
Network Center Jiangxi Electric Power Company
Nanchang University Nanchang , China
Nanchang , China e-mail: w_wujian@163.com
e-mail: zqw@ncu.edu.cn, zhxu@ncu.edu.cn

Abstract—In order to avoid the economic loss due to too much GA-RBFNN is very suitable for electricity consumption
or too little of electricity consumption, electricity consumption prediction by relevant features data.
needs to be predicted. In order to solve the drawbacks of BP
neural network, genetic algorithm and RBF neural network II. ELECTRICITY CONSUMPTION PREDICTION
(GA-RBFNN) is presented to forecast electricity consumption METHOD BASED ON GENETIC ALGORITHM-RBF
in the study, and genetic algorithm is introduced and tried in NEURAL NETWORK
optimizing the parameters of RBF neural network. The
electricity consumption data and relevant features data of a A. RBF Neural Network
certain province from September to December in 2007 are
The RBF neural network is a type of feed-forward neural
used as the experimental data. The experiment results indicate
that GA-RBFNN is very suitable for electricity consumption network, and radial basis function is used in the neural
prediction by relevant features data. network. RBF neural network is composed of an input layer,
a hidden layer and an output layer. In the study, the RBF
Keywords-electricity consumption; RBF neural network; neural network has 5 input nodes, m hidden nodes and 1
genetic algorithm; prediction output nodes, in which the hidden nodes are radial basis
function, and it is described as:
I.INTRODUCTION
In order to avoid the economic loss due to too much or
 x − ci 
too little of electricity consumption, electricity consumption hi = exp − 2 
i = 1,2,L, m (1)
needs to be predicted, then science planning will be realized  2d i 
for electricity consumption[1]. As electricity consumption is
affected by the factors including temperature, wind-force,
dampness, weather condition, etc, its forecasting model is where ci is the center of the ith RBF hidden node, di is
difficult to be expressed by mathematical model. Recently, the width of the ith RBF hidden node, m is the number of
BP neural network is commonly adopted to forecast RBF hidden node.
electricity consumption due to good nonlinear mapping B. Constructing the Training Data
capabilities and fault tolerant ability[2,3]. However, BP
neural network has some inherent drawbacks, such as local The influence factors of electricity consumption mainly
optimization solution. In order to solve the drawbacks of BP have maximum temperature, minimum temperature,
neural network, RBF neural network(RBFNN) is applied to wind-force, dampness and weather condition. In the training
forecast electricity consumption, the kind of neural network data, maximum temperature, minimum temperature,
is a feed-forward neural network, which has the global wind-force, dampness and weather condition are used as
optimal ability[4,5]. And genetic algorithm is introduced input feature data, electricity consumption data are used as
and tried in optimizing the parameters of RBF neural output data. SVM forecasting model trained by the training
network, the method is well suited for searching global sample is described as:
optimal values[6]. Therefore, genetic algorithm and RBF
neural network (GA-RBFNN) is presented to forecast yi = f (thi , tli , f i , si , qi ) (2)
electricity consumption in the study. The electricity
consumption data and relevant features data of a certain
where th i is maximum temperature; tli is minimum
province from September to December in 2007 are used as
the experimental data. The experiment results show that the temperature; f i is wind-force; si is dampness; qi is
proposed GA-RBFNN model has greater forecasting
accuracy than BP neural network. It is indicated that weather condition; yi is electricity consumption.

C. Parameters Optimization of RBF Neural Network with


Genetic Algorithm
978-1-4244-5848-6/10/$26.00 ©2010 IEEE
The forecasting result of RBFNN is affected by the

339
center of RBF hidden node c ,the width of RBF hidden node until stopping condition is satisfied.
d ,and the weight of RBF hidden node w ,so the III.CASE ANALYSIS
appropriate parameters c , d and w should be selected to
As shown in Tab.1, the electricity consumption data and
improve the forecasting result of RBFNN.
relevant features data of a certain province from September
GA employed in this study is used to implement efficient
to December in 2007 are used as the experimental data. The
search strategies for optimal parameters c , d and w .The electricity consumption data from 2007-9-1 to 2007-12-14
process of selecting the parameters of RBFNN with GA is are applied to training data, the electricity consumption data
summarized as followings: from 2007-12-15 to 2007-12-31 are applied to testing data.
Step 1 Code the parameters of RBF neural network Maximum temperature, minimum temperature, wind-force,
The chromosome is composed of the parameters dampness and weather condition are used as input feature
c , d and w . Genetic algorithm is started with 20 randomly data, electricity consumption data are used as output data.
generated chromosomes. In order to eliminate the influence of dimension, the
Step 2 Evaluate the fitness electricity consumption data and relevant features data are
The fitness of each individual chromosome is the normalized. Then, the data are used to train the
negative of the MAPE. GA-RBFNN forecasting model for electricity consumption
prediction. BPNN is adopted to compare with GA-RBFNN.
Step 3 GA operator
Fig.1 shows the electricity consumption forecasting data of
Chromosome is selected by evaluating the fitness value
GA-RBFNN and BPNN, and Fig.2 shows forecasting error
of each chromosome. Crossover could be implemented
of GA-RBFNN and BPNN, which indicates that
either across genes [7].Mutation is implemented by flipping
GA-RBFNN has greater forecasting accuracy than BPNN.
a bit at a binary locus. The crossover rate is set to 0.5. while
As shown in Tab.2, the experimental results show that
the mutation rate is set to 0.05. The new population is
GA-RBFNN is very suitable for electricity consumption
produced by using the selection, crossover and mutation
prediction by relevant features data including maximum
operation.
temperature, minimum temperature, wind-force, dampness
Step 4 Judge stopping condition
and weather condition.
The same procedures from Step 2 to Step 3 are repeated
TABLE I. THE ELECTRICITY CONSUMPTION DATA AND RELEVANT FEATURES DATA OF A CERTAIN PROVINCE FROM SEPTEMBER TO DECEMBER IN 2007

Electricity
Maximum Minimum Wind-force / Dampness / Weather
Date consumption
temperature /oC temperature /oC level g/m3 condition
/104kw.h
2007-9-1 29 23 2 80 Overcast 473.29
2007-9-2 27 21 4 65 Overcast 434.06
overcast and
2007-9-3 26 22 3 70 411.10
rainy
2007-9-4 27 20 4 40 Overcast 394.69
2007-9-5 28 21 2 20 Cloudy 410.31
2007-9-6 30 21 2 20 Fine 431.78

2007-9-7 30 22 2 10 Fine 456.61


2007-9-8 28 22 2 60 Overcast 467.74
……
2007-12-24 9 6 2 70 Sprinkle 545.07

2007-12-25 11 6 2 20 Snow 556.06


2007-12-26 11 7 2 60 Overcast 565.81
2007-12-27 11 8 2 60 Overcast 566.57
2007-12-28 9 6 2 70 Sprinkle 566.75

2007-12-29 10 5 2 20 Cloudy 561.85


2007-12-30 11 5 3 10 Fine 560.54
2007-12-31 9 1 3 10 cloudy 563.35

340
TABLE II. COMPARISON OF ELECTRICITY CONSUMPTION FORECASTING DATA BETWEEN GA-RBFNN AND BPNN

GA-RBFNN BPNN
Date Actual values Forecasting Forecasting
Error/% Error/%
values values
2007-12-15 575.87 562.28 2.31 527.85 8.18
2007-12-16 567.79 554.69 2.23 542.48 4.31
2007-12-17 587.22 577.00 1.74 558.44 4.90

2007-12-18 586.76 569.20 2.99 606.43 3.35


2007-12-19 576.57 581.05 0.76 602.04 4.34
2007-12-20 582.08 587.18 0.87 559.79 3.80
2007-12-21 569.30 583.28 2.38 595.52 4.47

2007-12-22 548.45 575.18 4.55 590.81 7.21


2007-12-23 545.60 574.37 4.90 562.22 2.83
2007-12-24 545.07 569.50 4.16 586.76 7.10
2007-12-25 556.06 563.37 1.24 582.38 4.48

2007-12-26 565.81 555.31 1.79 588.70 3.90


2007-12-27 566.57 548.08 3.15 546.97 3.34
2007-12-28 566.75 552.55 2.42 548.93 3.03
2007-12-29 561.85 572.39 1.79 587.70 4.40

2007-12-30 560.54 569.12 1.46 588.91 4.83


2007-12-31 563.35 574.76 1.94 552.31 1.88
MAPE/% 2.39 4.49

Electricity Consumption/
620
Actual
610 GA-RBFNN
BPNN
600

590

580

570

560

550

540

530

520
2007-12-14 2007-12-16 2007-12-18 2007-12-20 2007-12-22 2007-12-24 2007-12-26 2007-12-28 2007-12-30 Date

Figure 1. Electricity consumption forecasting data of GA-RBFNN and BPNN

ІV. CONCLUSION relevant features data of a certain province from September


to December in 2007 are used as the experimental data.
Genetic algorithm and RBF neural network is presented Maximum temperature, minimum temperature, wind-force,
to forecast electricity consumption in the study, and genetic dampness and weather condition are used as input feature
algorithm is introduced in optimizing the parameters of data, electricity consumption data are used as output data.
RBF neural network. The electricity consumption data and The experiment results show that GA-RBFNN has greater

341
Figure 2. Forecasting error of GA-RBFNN and BPNN

[4] A. Shahsavand, A. Ahmadpour,“Application of optimal RBF neural


forecasting accuracy than BP neural network. It is indicated networks for optimization and characterization of porous materials”,
that GA-RBFNN is very suitable for electricity Computers & Chemical Engineering, 2005, vol.29, no.10,pp.
consumption prediction by relevant features data. 2134-2143.
[5] Juan Ignacio Mulero-Martínez,“Analysis of the errors in the modelling
REFERENCES of manipulators with Gaussian RBF neural networks”,
Neurocomputing, 2009,vol.72,no.7-9, pp.1969-1978.
[1] CHEN Guochu,LIU Jun,“Mid and Long Term Electric Load Modeling [6] Byungwhan Kim, Sanghee Kwon, Dong Hwan Kim,“Optimization of
Based on Neural Networks”, Journal of Shanghai Dianji University, optical lens-controlled scanning electron microscopic resolution using
2009, vol ,12 ,no.1,pp.20-24. generalized regression neural network and genetic algorithm”, Expert
[2] TIAN You-wen, WANG Wen- tao, WANG Si- yu,“Forecasting the Systems with Applications, 2010, vol.37,no.1,pp. 182-186.
Electricity Requirements in Rural Area Based on BP Neural Network” , [7] Kyoung-jae Kim, Ingoo Han,“Genetic algorithms approach to feature
Journal of Shenyang Agricultural University,2005,vol.36,no.4, discretization in artificial neural networks for the prediction of stock
pp.506-507 price index”, Expert Systems with Applications, 2000, vol.19,no.2,
[3] Juan Ignacio Mulero-Martínez,“Boundedness of the nominal pp.125-132.
coefficients in Gaussian RBF neural networks”, Neurocomputing, 2007,
vol.71,no.1-3,pp.197-220.

342

You might also like