Abstract: Validation of global radiation models with measured daily data based on the meteorological data (including the extraterrestrial radiation) from 15 radiation stations in South China during 1981-2009.The daily global solar radiation was estimated using five empirical models (i.e., Angstrom-Presscott model, Bahel model, Bristow-Campbell model, Chen model and Hargreaves model) and 12 generalized regression neural network models (GRNN) with different input variable combinations. The performance of these global radiation models were evaluated using observed daily global solar radiation data during 2010-2014 at the 15 radiation stations. Finally, the global radiation models were developed by meteorological data from the adjacent stations, and then evaluated the applicability of these models using the observed data from the studied station. The results showed that the Chen model had the highest accuracy among the empirical models, followed by the Bahel model. The Bristow-Campbell model performed similarly to the Hargreaves model for most radiation stations. The Bahel model had the highest accuracy at 9 of 15 radiation stations when developing models by using the meteorological data from the adjacent stations, while the Chen model was most suitable for 7 stations. The Bristow-Campbell model and Hargreaves model had higher accuracy at 13 of 15 stations. But the Hargreaves model performed better at the Wuhan station and Ganzhou station, where the RMSE was decreased by about 14%. The GRNN model had the lowest average RMSE with input variable of sunshine percentage, which was better than the Bahel model and Chen model, but the difference in average RMSE at each station was less than 2%. When using the local meteorological data, the RMSE of the GRNN model was decreased by about 14% compared to that of the Bristow-Campbell model and the Hargreaves model. When using the data from the adjacent stations, the GRNN model performed similarly to the Bristow-Campbell model and the Hargreaves model due to the large differences in the smoothing factor at each station. Therefore, the Bahel model and the Chen model were considered to be more suitable for the estimation of global solar radiation in these areas considering the complexity of GRNN model development.

Key words: Global solar radiation, Generalized regression neural network (GRNN), Bahel model, Chen model, Solar duration hours