In order to improve the accuracy of short-term photovoltaic power generation forecasting, the deep belief network (DBN) is used to establish the forecasting models of each model function. By analyzing the characteristics of each model function, the power prediction of photovoltaic power generation model is established. The traditional power prediction of photovoltaic model based on neural network is difficult to train multi-layer network, thus affecting its prediction accuracy. DBN uses unsupervised greedy layer-by-layer training algorithm to construct a multi-hidden layer network structure with excellent performance in regression prediction analysis, which has become a research hotspot in the field of deep learning. The weight of DBN connection is optimized by particle swarm optimization combined with chaotic crossover (CC-PSO), which avoids the phenomenon of local optimal solution caused by random initialization and improves the prediction performance of DBN network. Finally, case tests show the effectiveness of the proposed model.