When applied to precipitation forecasting, the mean generating function - optimal subset regression (MGF-OSR) model is limited by its low accuracy and high error, while the back propagation (BP) neural network model has difficulty in learning for matrix selection. This paper proposes a new MGF-OSR-BP model, which uses a MGF to extend original data, an OSR to select the best series as the BP neural network input node and learning matrix, and the resultant data for training. The training procedure determines the number of hidden layers and uses an optimal number of hidden layers for model training. This paper uses the MGF-OSR-BP model to analyze precipitation data from Hangzhou, China, for 53 years, from 1956 to 2008. The 1956-2006 precipitation data are used as the training sample, and the 2007-2008 data are used as the test set data to verify the practicality of the forecast system. A fitting verification is performed using the forecasted data against field measurement data, and the results show that the forecast accuracy is better than that of the MGF-OSR model or the MGF stepwise multiple regression model.