Output power of photovoltaic (PV) power generating system has the characteristics of time varying, intermittence and randomness due to the various meteorological factors such as season, solar radiation, temperature, humidity, etc. In this paper, a forecasting method is proposed based on the principle of similar days and BP neural network. By using historical weather information from the solar power station, meteorological feature vectors are established, and similar days are found based on Manhattan distance. According to the given different forecasting day, output power of three similar days would be chosen as inputs of the forecasting model, and then the output power of generating station can be predicted directly. A forecasting model is made based on a photovoltaic power station and the forecast error is calculated and analyzed. The results show the validity of the algorithm.