Vous êtes sur la page 1sur 4

Daily Solar Radiation Prediction Based on Genetic

Algorithm Optimization of Wavelet Neural Network


Jianping Wang,Yunlin Xie ,Chenghui Zhu,Xiaobing Xu
School of Electric Engineering and Automation, HeFei University of Technology
HeFei, China
xylhfut@sina.com

Abstract—Daily solar radiation prediction is a nonlinear and process of daily solar radiation was decomposed into quasi-
non-stationary process. It’s hard to model with a single method. stationary at different scales by multi-scale characteristics of
A Genetic Algorithm Optimization of Wavelet Neural Network wavelet transform. Input weights, output weights, scale factors
(GAO-WNN) model was set in this paper. The nonlinear process and translation factors of this model were optimized by GA.
of daily solar radiation was forecasted by neural network and the After that gradient descent method was used to make further
non-stationary process of daily solar radiation was decomposed training of the model. Simulation results indicate that the
into quasi-stationary at different frequency scales by multi-scale GAO-WNN model can solve the nonlinear and non-stationary
characteristics of wavelet transform. Input weights, output problems of daily solar radiation. It is a doable and effective
weights, scale factors and translation factors were optimized by
method to predict daily solar radiation.
genetic algorithm. Gradient descent method was used to make
further training of the model with temperature, clearness index,
and daily radiation data. Simulation results indicate that the II. CONSTRUCTION OF DAILY SOLAR RADIATION WNN
method is satisfactory to the prediction of daily solar radiation. PREDICTION MODEL

Keywords- genetic algorithm optimization; wavelet neural A. Neuron number of input and output layer
network; daily solar radiation prediction 
The neuron number of input layer is 3. It represents
temperature, clearness index, and daily solar radiation. The
I. INTRODUCTION neuron number of output layer is 1 which represents the output
Prediction of daily solar radiation has a particular of networks.
significance for photovoltaic power generation system. It is
currently researched by many scholars. Because of different B. Construction of hidden layer
temperature, weather conditions, and other factors, daily solar
radiation is a strong nonlinear and non-stationary process [1]. Construction of the hidden layer includes the activation
Common prediction method has ARMA method [2]. This function and the number of hidden layer neuron.
method was based on stability hypothesis of daily solar 1) Selecting of activation function of hidden layer neuron
radiation. It didn’t match with the actual situation. Neural There is no common method to choose the basic wavelet
network method has great advantage of solving nonlinear function. According to the characteristics of daily solar
problems. Main prediction methods include Adaptive Neural radiation series, the continuous and symmetry Molet wavelet
Fuzzy Inference Scheme (ANFIS) [3], Radial Basis Function function was chose in this paper. The form of the function is as
Networks (RBFN) [4], and Recurrent Back-Propagation follows:
Network (RBPN) [5]. But the non-orthogonal of the neuron
activation function make it time-consuming [6] and it’s hard to − x2
define the number of hidden layer neuron theoretically. The Ψ ( x) = e 2
cos(5 x)
(1)
WNN has characteristics of less time-consuming and better A wavelet sequence can be received by dilation and
learning abilities by changing the sigmoid function with translation:
wavelet function [7-8]. Compared with the
traditional BP neural network, WNN has a faster convergence 1 x − bj
rate, and to some extent, avoid local optimal defect. However, Ψ a,b ( x) = Ψ( ) (2)
it can not avoid all the shortcomings of BP algorithm and there | aj | aj
needed to be further study of the initialization of the network Where: a j —Scale factor; b j —translation factor.
parameters. GA is a global optimality method, and it can search
in nonlinear and non differentiable space to achieve global a j , bj ∈ R , a j ≠ 0.
optimality [9-10]. It is useful to train the model of daily solar
radiation. But, the main defect of GA is it’s limitation in detail 2) Number of hidden layer neuron
search capability. Daily solar radiation data from 2000 to 2004 can be seen in
Fig. 1.
The global optimality of GA and favorable local properties
of WNN were combined in this paper to construct the
predication model of daily solar radiation. The non-stationary

978-1-4244-8165-1/11/$26.00 ©2011 IEEE


602
2

1.5

0.5
0 500 1000 1500 2000
day Figure 4. Structure of daily solar radiation wavelet neural network prediction
model
Figure 1. Average daily solar radiation

Use one-dimensional wavelet to decompose daily solar According to the results of the wavelet decomposition, the
radiation series with different scales. The approach series at number of hidden layer neuron is 10.Daily solar radiation
different scales can be seen in Fig. 2, and detail signal at WNN prediction model is based on BP neural network
different levels can be seen in Fig. 3. topology. The structure of the model can be seen in Fig. 4.

III. PARAMETERS ADJUSTMENT OF SOLAR RADIATION GA-


2
WNN PREDICTION MODEL
a5

1
0 A. Parameters adjustment use GA optimization
0 500 1000 1500 2000
2 The steps to optimize the parameters of WNN are as
follows:
a4

1
0
0 500 1000 1500 2000 a. Put the input weights v ijh , output weights w oj1 , scale
2
factor a hj , and translation factor b hj together orderly as a
a3

1
0 solution of the problem. The character string is coded with real
2
0 500 1000 1500 2000 number, and the order of the string is as follows:

v11h ⋅⋅⋅v110
h
⋅⋅⋅ v31h ⋅⋅⋅ v310
h
w11o ⋅⋅⋅ w101
0 h
a1 ⋅⋅⋅ a10h b1h ⋅⋅⋅b10h
a2

1
0
0 500 1000 1500 2000
2 where: h — hidden layer; o—output layer;
b. Calculate the value of evaluation function, and order them.
a1

1
0 Use equation (3) to select the individual;
0 500 1000 1500 2000
N
pi = f i / ¦ fi (3)
Figure 2. One-dimensional wavelet decomposition of daily solar radiation i =1

where: f i —evaluation of individual i .


Adjusts f i by using equation (4) as follows:
0.1
0 ­°k f i , f i ≥ k f i
d5

-0.1 fi ' = ® (4)


0.2
0 500 1000 1500 2000
°̄ f i , fi < k f i
0 where: f i — average of evaluation; k ∈ [1,10] . The purpose
d4

-0.2
0 500 1000 1500 2000 of amendment is to reduce the impact of some individuals with
0.2 high adaptation value. It can slow the convergence speed and
0
d3

-0.2
enlarge the search space.
0 500 1000 1500 2000 The expectation replicate value ni can get by using equation
0.2
0 (5) as follows:
d2

-0.2
0 500 1000 1500 2000 ni = fi ' / f i (5)
0.2
0
c. Manipulate Gi and Gi +1 by using crossover probability PC.
d1

-0.2 The other individuals are replicated with expectation replicate


0 500 1000 1500 2000 value;
d. Using mutation probability PM manipulates individual G j
Figure 3. One-dimensional wavelet decomposition detail signal of daily to get a new individual G 'j ;
solar radiation

603
e. Insert the new individual in the genetic population and According to:
calculate the value of adaptive function; ∂E
f. If the optimum individual was found, the process is over.
= − ( d p − o p ) (14)
∂o p
Otherwise return c;
∂E
In order to get the input and output weights, Scale factors, = −(d p − op ) f ' (netk )wjk (15)
and translation factors, it’s needed to decode the optimum ∂zk
individual.
∂zk 1 1
=− [Ψab, (netj ) + Ψ'ab, (aj )](16)
B. Parameters adjustment using gradient descent method ∂aj | aj | | aj | | aj |
The output of daily solar radiation WNN is:
∂zk 1
=− Ψ'a,b (bj ) (17)
m n ∂bj | aj | | aj |
o =f [¦ w jk Ψ a ,b (¦ vij xi )] (6)
j =1 i =1 The result of adjustment is as follows:
Definition of wavelet neural network error function:
Δvij = η (d p − op ) f ' (netk )wjk Ψa,b '(net j ) xi (18)
100
1 1
(d − o)2 = ¦ (d p − o p ) 2
n
E= Δ w jk = η ( d p − o p ) f ' ( net k ) ¦ v ij x i (19)
2 2 p =1 (7)
i =1

1 '
Where: d p — the expectance output; o p —the actual output θ[Ψa,b (net j ) + Ψa,b (aj )](dp − op ) f ' (netk )wjk
| aj |
After the optimization of individual, use the gradient Δaj =
descent method to train the input and output weights, Scale | aj | | aj |
factors, and translation factors: (20)
∂E ∂E ∂net j 1
Δvij = −η = −η (8) Δb j = θ Ψ 'a ,b (b j )(d p − o p ) f ' (netk ) w jk
∂vij ∂net j ∂vij | aj | | aj |
(21)
∂E ∂E ∂netk
Δw jk = −η = −η (9)
∂w jk ∂netk ∂w jk IV. NUMERICAL EXAMPLE
Clearness index and average daily temperature data used
∂E ∂E ∂z k in this paper can be seen in Fig. 5(a) and Fig. 5(b).
Δ a j = −θ = −θ (10)
∂a j ∂z k ∂a j 0.8
∂E ∂E ∂zk
Δbj = −θ = −θ
Clearness Index

(11) 0.7
∂bj ∂zk ∂bj
0.6
n m
Where: net j = ¦v ij xi ; netk = ¦ w jk zk ; 0.5
i =1 j =1
0.4
zk = Ψ a ,b (net j ) ; η —Learning factor; 0 500 1000 1500 2000
(a)
θ —Momentum factor.
For the output layer defines: 30
Temperature(°C)

20
∂E ∂E ∂op ∂E '
δ =
o
k = = f (netk ) (12) 10
∂netk ∂op ∂netk ∂op
0
For the hidden layer defines:
0 500 1000 1500 2000
(b)
∂E ∂E ∂zk ∂E
δ jz = = = Ψa,b ' (net j ) (13)
∂net j ∂zk ∂net j ∂zk Figure 5. (a) Clearness index ;(b)Average daily temperature

604
number of optimum individual
10 As shown in Fig. 7(b), the relative error is less than 2%.

8 V. CONCLUSIONS
6 In this paper GA and WNN were combined to construct the
daily solar radiation prediction model. Simulation results
4 indicate that:

2
1 ˅ Using the combination of wavelet analysis and the
neural network to predict daily solar radiation can solve the
0 nonlinear and non-stationary problems. After trained with a
0 500 1000 1500 2000 certain number of data, the model can predict daily solar
Figure 6. Optimum individual obtained
radiation well.
2˅Using the combination of global optimality of GA and
The genetic manipulation population size S is 50, crossover favorable local properties of WNN to train the model, not only
probability PC is 0.3, mutation probability PM is 0.005, and provides an effective way for the initialization of parameters,
maximum genetic time GM is 2000. The result of genetic but also can find the global optimum quickly. It’s an accurate
algorithm optimization is shown in Fig. 6. and reliable method to predicate daily solar radiation.
After genetic algorithm optimization, decode the optimum
individual and get the initialization value of input and output ACKNOWLEDGMENT
weights, scale factors, and translation factors. Normalize the
input data and use the gradient descent method to train the The authors wish to thank the data providers that have led
model again. Let 100 average daily solar radiation data as the to the analysis being possible in the presentation of the paper.
expectation of the model. The predication result can be seen in
Fig.7 (a), and relative error in Fig. 7(b). REFERENCES

[1] Shuanghua Cao, Jiazong Cao. Study of Chaos Optimization Neural


Networks for the Forecast of Hourly Total Solar Irradiation[J]. ACTA
Prediction result ENERGIAE SOLARIS SINICA.2006:pp.164-169(in Chinese).
1.5 [2] A. Moreno-Muñoz, J. J. G. de la Rosa, R. Posadillo and V. Pallarés.
Short term forecasting of solar radiation [J]. IEEE, 2008: pp.1537-2541.
actual value [3] A.Mellit, A.Hadj Arab, N.Khorissi, H.Salhi.An ANFIS-based
prediction value Forecasting for Solar Radiation Data from Sunshine Duration and
Ambient Temperature [J].Power Engineering Society General Meeting,
1 IEEE, 2007: pp.1-6.
[4] A.Mellit, M.Benghanem , M.Bendekhis. Artificial Neural Network
Model for Prediction Solar Radiation Data: Application for Sizing
Stand-alone Photovoltaic Power System [J].Power Engineering Society
General Meeting, IEEE, 2005: pp. 40-44.
0.5 [5] Shuanghua Cao. Total Daily Solar Irradiance Prediction using Recurrent
0 10 20 30 40 50 60 70 80 90 100 Neural Networks with Determinants [J]. Academic discipline project of
(a) shanghai municipal education commission, China (Project No. J50502),
IEEE, 2010:pp.1-4(in Chinese).
relative error
2 [6] CHEN Guiyou, LUO Donghua, Han Zhaoyou.Study of the recognition
of tyre number based on wavelet neural network algorithm[J].Systems
Engineering and Electronics,Vol.30,No.4,Apr.2008:pp.752-754(in
1.5 chinese).
[7] Yu Jin, Guangju Chen, Hong Liu. Fault diagnosis of analog circuit based
on wavelet neural network [J].Chinese Journal of Scientific Instrument,
1 Vol. 28, No.9, Sep. 2007:pp.1600-1604(in chinese).
[8] Xinhui Hu, Li Chen. Research and application of wavelet neural
0.5 network based on genetic algorithm [J].Journal of Northwest University,
Apr., 2009, Vol. 39, No2: pp.203-207(in chinese).
0 [9] Yongsheng Wang, Jie Wang, Zehui Li, Hongda Fan. Forecasting chaotic
0 10 20 30 40 50 60 70 80 90 100 time series based on improved genetic WNN [J]. Computer
Applications, Vol.28, No.9, Sep. 2008: pp. 2363-2365(in chinese).
(b)
[10] DONG Fugui, ZHANG Shiying, TAN Zhongfu, ZHANGWenquan.
Application for Electricity Price Forecasting Using Improved Wavelet
Figure 7. (a).Predication value and actual value;(b) Relative error of Neural Networks Based on GA[J]. Computer Engineering, Vol.31, No.4,
predication value and actual value 2005:pp.32-34 (in chinese).

605

Vous aimerez peut-être aussi