In view of the traditional BP neural network model used in Photo Voltaic (PV) power generation prediction, there are shortcomings of low prediction accuracy and slow convergence speed. In this study, an improved BP neural network model is proposed to improve the defects of traditional BP model by adding momentum term and adaptively selecting the best hidden layer. Firstly, the correlation of meteorological factors for PV power output is analyzed, and six meteorological factors that can affect the PV power are extracted as input of the network model. Then, an improved BP network model is established, combined with the historical data of PV power output, to directly predict the generation of data. Finally, according to the prediction results under different climate types, the feasibility and effectiveness of the model for power prediction are analyzed and verified.