To improve the surface solar radiation forecasting accuracy in complicated weathers and to reduce the time-consuming cost, this paper proposes a method that using the China Meteorological Administration's weather classification criteria, to classify the historical meteorological data of Baiyun District, Guangzhou, into different weather types to forecast the surface solar radiation in sub models using support vector regression (SVR). Considering that the weather types are many, by analyzing the features of each sub model with extreme gradient boosting (XGBoost) algorithm and using Mann-Whitney test, series with similar feature importance are merged, thus achieving the secondary weather classification and reducing the complexity of forecasting model. Forecasting result by the model proposed shows that all the correlation coefficient, accuracy rate and qualification rate for the 12 successive months have exceeded the requirement of evaluation criteria, which guarantees a high accuracy. Besides, it totally consumes 10.633 hours to forecast, promising a faster forecasting method compared with the 13~34 hours of other models, which meets the demand for timely solar radiation forecasting in photovoltaic power plants.