The amount of solar radiation affected by the season, atmospheric conditions, cloud conditions, temperature, humidity and even dust and other weather factors with a strong time variability and randomness. For the prediction method for nonlinear radiation, at present, someone put forward many methods. However, there is still lack of selecting the unreasonable intelligent algorithm，network structure poor generalization ability and prediction accuracy is not ideal. Aiming at the deficiency of that a photovoltaic power station solar radiation intensity of the original hourly data is not obvious and the normal BP neural network can not be completely mapped its features, it put forward a forecasting model based on Wavelet Packet Neural Network . It used wavelet packet to transform the radiation intensity sequence multiscale decomposition and established several BP neural networks to forecast each frequency components, obtaining the complete prediction value with the wavelet packet reconstruction finally. Results show that the prediction accuracy was significantly improved to meet the expected results, demonstrating the effectiveness and practical value of the model.