A BP neural network has problems associated with having many input variables and complex relationships, and poor stability. It is also easy for it to fall into local optimal solution. Thus an improved BP neural network short-term prediction model for photovoltaic power generation is established, one which combines Principal Factor Analysis (PFA) and Beetle Antennae Search Algorithm Majorization (MBAS). First, the model simplifies the dimension reduction analysis of photovoltaic historical power generation data and meteorological data, and uses the principal factor analysis method to analyze the correlation of the main factors affecting photovoltaic power generation. It chooses the principal factor as the input of the prediction model. Then, the optimal weight threshold of the BP neural network training is selected by using the spatial search of the MBAS algorithm. Finally, the simulation results show that the prediction accuracy of the optimized model can reach 92.5%, the image data has a high degree of fitting and it can be used to forecast photovoltaic power generation in various weather types. This work is supported by National Natural Science Foundation of China (No. 61563034).