支持向量机方法在南京太阳总辐射推算中的应用

Application of support vector machine method in estimation of total solar radiation in Nanjing

利用1981、1996和2001年逐日南京站太阳总辐射和日照时数观测资料,建立了基于支持向量机(support vector machine,SVM)方法的太阳总辐射推算模型,预测了1982、1997和2002年的太阳总辐射,并把推算结果和采用线性的气候学方法所得到的推算结果分别与实测值进行对比。采用线性方法得到的1982、1997和2002年的太阳总辐射预测值与实测值间基于1:1线的决定系数(R^2)分别为0.800、0.859和0.838,均方根误差(RMSE)分别为3.250、2.649和2.925 MJ·m^(-2)·d^(-1)。采用SVM方法得到的1982、1997和2002年的R^2分别为0.894、0.938和0.936,RMSE分别为2.353、1.726和1.804 MJ·m^(-2)·d^(-1)。SVM方法得到的太阳总辐射预测值与实测值之间的误差较小,预测精度高于线性方法,更适用于实际太阳总辐射的计算。

A total solar radiation model based on the support vector machine(SVM)method is studied for calculating solar radiation in Nanjing.With a solid the oretical foundation,the SVMmethod is a kind of novel small sample learning method,which can be employed to deal with the problems of highly nonlinear classification and regression.Based on the daily total solar radiation and sunshine duration data in Nanjing in 1981,1996 and 2001,a model of total solar radiation estimation is established by using the SVMmethod,which is used to forecast total solar radiation in 1982,1997 and 2002.The prediction accuracy of total solar radiation using the SVMmethod is compared with that obtained by the linear climatological method.Furthermore,the above results are compared with the observed values.According to the simulated results of total solar radiation in Nanjing by using the linear and SVMmethods,it can be clearly seen that the errors of total solar radiation predicted by the linear method is relatively larger,and the simulation accuracy changes significantly with different seasons.Nevertheless,the observed values of total solar radiation and the simulated values by the SVMmethod in Nanjing are basically consistent,and the simulation accuracy of total solar radiation by the SVMmethod is higher all the year round.Scatter plots of observed and simulated total solar radiation based on the linear method in 1982,1997 and 2002 indicate that when total solar radiation is less than 10 MJ·m^(-2)·d^(-1),the scattered points of prediction values are dispersed.Especially,when the observed values of total solar radiation are less than 5 MJ·m^(-2)·d^(-1),the scattered points on both sides of the fitted line are uneven and the errors are larger.The determination coefficients(R^2)between simulated and observed values based on the 1:1 line for total solar radiation based on the linear method in 1982,1997 and 2002 are 0.800,0.859 and 0.838,respectively,and there is significant correlation.On the other hand,the scatter plots of observed and simulated total solar radiation in 1982,1997 and 2002 based on the SVMmethod showthat the scattered points of prediction values are relatively concentrated.The R^2between simulated and observed values based on the 1:1 line for total solar radiation based on the SVMmethod in 1982,1997 and 2002 are 0.894,0.938 and 0.936,respectively,and the correlation is more significant than that obtained by the linear method.The root mean squared error(RMSE)between the simulated and observed values for total solar radiation based on the linear method in 1982,1997 and 2002 are 3.250,2.649,2.925 MJ·m^(-2)·d^(-1),respectively.Based on the SVMmethod,RMSE of total solar radiation in 1982,1997 and 2002 are 2.353,1.726and 1.804 MJ·m^(-2)·d^(-1),respectively,which are less than those obtained by the linear method.Total solar radiation calculated by the SVMmethod can accurately reflect variation of solar radiation in Nanjing,so the method is more suitable for the calculation of actual solar radiation