In the current PV output short-term forecast, BP or optimization BP neural network algorithm is used commonly, which has problems of single optimization algorithm, the lack of a variety of optimization algorithms for comparison and selection, and big forecast error. Therefore, based on local 5 kW small-scale distributed PV power station, considering the related factors that influence PV output such as solar radiation intensity, environmental temperature, wind speed and historical generation data of photovoltaic power station, this paper uses BP, GA-BP and POS-BP neural network algorithm respectively to construct short-term prediction model of PV output in sunny, cloudy and rainy weather conditions. Test results show that three kinds of neural network prediction models all reach certain prediction accuracy under three different weather conditions, among which GA-BP and POS-BP prediction models reduce the prediction errors compared to the traditional BP model, and POS algorithm has a better optimization effect on BP neural network prediction model and a stronger applicability compared to GA algorithm, and further reduces the prediction errors.