Abstract: Hourly total solar radiation data has become the basic requirements of the meteorological service for multi-industry.However, radiation observation sites in China are sparse.Using hourly total solar radiation data in different seasons and corresponding meteorological data from 2003 to 2012 in Haikou weather station, an hourly total solar radiation prediction model based on a wavelet back propagation (BP) neural network method was developed.The model was tested using observed data in 2013 and compared with that based on a stepwise statistical regression model developed by the same data.The results indicate that the model based on the wavelet BP neural network method has higher accuracy, but the level of accuracy is different in different seasons.The accuracy is the highest in winter and lowest in summer.The index of weather type is favorable to enhance the forecast accuracy in different seasons.The root means squared error (RMSE) between predicted and observed total solar radiation for the wavelet BP neural network model with index of weather type are 0.32 MJ·m^-2 in spring, 0.47 MJ·m^-2 in summer, 0.35 MJ·m^-2 in autumn and MJ·m^-2 in winter, and its forecast accuracy increases by 28.8%, 16.3%, 17.9% and 20.4% respectively compared with that for the stepwise regression model.It suggests that the model based the wavelet BP neural network method is suitable in Hainan.

Key words: Wavelet neural network, Stepwise regression, Hourly total solar radiation, Prediction