In order to schedule the large-scale grid-connected PV generation, the weather-related factors of PV power output are analyzed. PV short-term power output forecasting system is proposed in accordance with the four seasons based on LVQ-GA-BP neural network, whose input parameters are light intensity and temperature. The samples about PV output and weather-related factors are classified by learning vector quantization (LVQ) neural network. Then, the classified samples are trained to get the PV short-term output forecasting system based on GA-BP neural network in purpose of increasing forecasting accuracy. Secondly, we propose GA algorithm is an alternative to traditional learning algorithm to optimize BP neural network weights and thresholds, improving forecasting network training speed. At last, the LVQ-GA-BP forecasting system and the traditional forecasting system are compared and analyzed. The results show that the proposed forecasting system not only improve the PV output forecasting accuracy, but also raise the BP neural network training speed, which has potential value in engineering applications.