Application Of Neural Network On Estimating Daily Irradiation Exposure Of Global Radiation

An artificial neural network is a kind of nonlinear dynamic system,combining with the artificial intelligence theory such as mathematical statistics,neural calculation,symbolic logic and so on.The artificial neural network model can map non-linear relation without establishing complex mathematical models.Solar radiation,a green energy,is important for the development of economic and society.The solar radiation availability is affected by many environmental factors.for instance,there is a certain degree of latitude,longitude and altitude of regions,while the sunshine hours and atmospheric water content have a certain randomness.Thus,the total radiation sequence is of both certainty and randomness.Therefore,since the 1990 s,artificial neural network models have become an important method to estimate and predict solar radiation resources.However,different neural network model or an algorithm used to process the input and output data,is directly related to the accuracy of the estimated accuracy.In this paper,Generalized Regression neural network model and Elman neural network model are respectively used to estimate the daily irradiation exposure of global radiation of Yushan in Yantai,Shandong Province.At the same time,the influence of wavelet analysis on the Generalized Regression neural network model is discussed.The results are as follows:(1)The Generalized Regression neural network model is used to estimate the daily irradiation exposure of global radiation from 2000 to 2003 at Fushan Meteorological Station,Yantai city Shandong province.The cross validation method is adopted to determine the key parameter of Generalized regression neural network model(smooth factor).The input parameters of Generalized regression neural network model included sunshine duration,average pressure,average air temperature,daily maximum air temperature,relative humidity,and aerosol optical thickness.Results are promising with the average percentage error 15.8%,the root mean square error 2.18 MJ·m-2,and the goodness of fit 0.906.The optimized Generalized regression neural network presents the estimate better than LM-BP network.Aerosol optical thickness in the model has almost no influence on average percentage error,root mean square error and the goodness of fit.Therefore,using the Generalized regression neural network model with the meteorological observation data is a very effective method to estimate the daily irradiation exposure of global radiation of some region which has no radiation observation site.(2)The Generalized regression neural network model combined with wavelet analysis was used to study daily irradiation exposure of global radiation from 2000 to 2003 at Fushan Meteorological Station,Yantai city Shandong province.The input parameters included sunshine duration,average pressure,average air temperature,daily maximum air temperature,relative humidity,and aerosol optical thickness.The output value was daily irradiation exposure of global radiation.The results show that the average percentage error is 1.72% and the mean square error is 0.38 MJ·m-2.The goodness of fit is 0.998 when the input and output components were low frequency ones based on the 3-layers wavelet decomposition.The Generalized regression neural network based on wavelet decomposition fully improves the prediction accuracy and goodness of the generalized regression neural network without wavelet decomposition.(3)The Elman neural network model was used to study daily irradiation exposure of global radiation from 2000 to 2003 at Fushan Meteorological Station,Yantai city Shandong province.Similarly,the paper takes the variables including sunshine duration,average pressure,average air temperature,daily maximum air temperature,relative humidity,and aerosol optical thickness,as the input values,and daily irradiation exposure of global radiation as the output value.The results show that the average percentage error is 22.3%,the mean square error is 1.82 MJ·m-2,and goodness of fit is 0.876.Compared with the Generalized Regression neural network,the Elman network model decreases the difference between the estimate value observations and the root mean square error is reduced.The average percentage error is higher than the generalized regression neural network,which may be caused by the local recursive structure of the Elman network.