Runoff forecasting is an important method to alleviate floods. Based on hydrological data from 1978 to 2010,combined with Long-Short Term Memory (LSTM), a runoff prediction model for the Bahe river basin was constructed.The results show that the combination of different neurons has an impact on the prediction effect of the LSTM model.Using the best combination of neurons can more effectively predict the change in runoff. The best model of Yuhe Dayu(the third) station is an LSTM model with 128 neurons in the first layer and 32 neurons in the second layer. The best model of the Bahe Luoli village (the fourth) station is an LSTM model with 128 neurons in the first layer and 8 neurons in the second layer. The best model of Bahe Maduwang station is an LSTM model with 8 neurons in the first layer and 2 neurons in the second layer. The best LSTM models at different sites can effectively predict the runoff changes of the three hydrological stations from 2006 to 2010. Among them, Dayuhe Dayu (the third) station has the best effect, while the other two stations have relatively poor results. There are differences in the forecasting effects of LSTM model for each quarter. The root mean square error of each site in the third quarter is relatively large, while the runoff forecasts for the first and fourth quarter are relatively accurate.