As for the medium and long-term runoff forecasting factors selection, this paper introduced mutual information (MI) to select the subset of factors from numerous meteorological factors into back-propagation neural network (BPN) model. In the model, mean square error (MSE) and MI were presented as objective functions respectively to measure factors compound correlation for the purpose of selecting optimal forecasting factors. The study was applied to forecast flood season runoff of the Biliuhe reservoir. The results show that using MI to select the subset and combining MI with BPN model can identify the correlation between runoff and its affecting factors effectively. The methods of factors selection may provide a good reference for medium and long-term runoff forecasting.