Aiming at the double characteristics of regularity and randomness of wind speed series, wavelet decomposition combined with radial basis function (RBF) neural network are used for short term prediction of wind speed. Aiming at the trend term extraction of different components with different frequencies in wavelet decomposition of wind speed signal, RBF network prediction for different components decomposed with wavelet and the corresponding synthesization method are studied, which includes three kinds of decomposition-combination prediction methods, i.e. prediction using all-high-frequency and low-frequency components, prediction using part-high-frequency and low-frequency components, and prediction using low-frequency component. The prediction performances and characteristics are analyzed. Prediction results, which are based on the data sampled from different dates and different sites, are analyzed in the short-term wind speed prediction by using different methods, and the conclusion is that the optimal prediction results can be obtained only when appropriate decomposition layers, appropriate combination of high-frequency and low-frequency components are used. The conclusions have profound guiding significance for wavelet decomposition-based short term prediction of wind speed.