Abstract: To improve the forecasting accuracy of flood process, we coupled the forecasted inflow by the GR4J model (modèle du Génie Rural à 4 paramètres Journalier) into the Long Short-Term Memory Neural Network (LSTM), to construct a GR4J-LSTM hybrid model, and compared it with GR4J and LSTM models. Based on the data set related to flood events in the flood season of Lushui Reservoir from 2012 to 2019, combined with the 3h precipitation forecast products from the European Centre for Medium-Range Weather Forecasts (ECMWF), the GR4J-LSTM hybrid model was driven to forecast the inflow of Lushui Reservoir during the forecast period of 3 to 12 hours, and finally the relative importance of the input variables was evaluated by the Mean Impact Value (MIV) algorithm. The results show that the GR4J, LSTM and GR4J-LSTM models can simulate and forecast inflow well and the GR4J-LSTM hybrid model performs best, not only can learn the runoff process of the GR4J model, but also improve the flood forecasting accuracy. The research results can provide a reference for the formulation of flood forecasting schemes.