Abstract: The physical mechanism and process of catchment generation and drainage in the river basin are extremely complicated,especially the variation of boundary conditions,which makes the construction of traditional hydrological models very difficult.And increasing the complexity of hydrological models does not necessarily bring about a significant improvement in forecast accuracy.The long short-term memory(LSTM)has a strong ability to fit non-linear relationships in complex multivariable time series and the ability to recognize historical data,which is suitable for simulating and forecasting complex time series processes such as runoff.In this study,the grey correlation analysis method was used to select appropriate forecasting factors,combined with LSTM,a G-LSTM forecasting model was established,and its application effect in short-term runoff forecasting is explored.The method was applied to the Cuntan section-Three Gorges Reservoir.By comparing the result of the this model with that of Xinanjiang model,BP neural network and LSTM model,the results showed that G-LSTM had an excellent ability to learn non-linear functions compared with the approximate mapping of traditional learning.Deterministic coefficients of the periodic and inspection periods were all above 0.9,which was significantly better than the simulation results of the other two models.G-LSTM can significantly improve short-term runoff forecast accuracy and is an effective method for runoff prediction.