In order to solve the problems of complex prediction model, large prediction error and low generalization ability in the existing short-term power generation prediction methods, a short-term power generation prediction method based on deep belief network is proposed. Firstly, the environmental factors affecting PV output, the operation parameters of PV panels and the historical power generation data of PV power stations are comprehensively considered to train and learn the deep belief network. Then, the hidden layers of deep belief network are determined by reconstruction error. Finally, a case study of a photovoltaic power station's short-term power generation is carried out to verify that the prediction model can actively select the abstract characteristics of samples and automatically determine the hidden layers, and has a high prediction accuracy for short-term power generation. Comparing the Back Propagation (BP) neural network prediction model and the Long/Short Term Memory (LSTM) prediction model, the results show that the proposed method has low computational cost and high prediction accuracy, and that increasing the depth of the neural network is more effective than improving the neural network neurons for the prediction effect. This work is supported by National Natural Science Foundation of China (No. 51977133).