Coupling SWAT and LSTM for Improving Daily Streamflow Simulation in a Humid and Semi-humid River Basin

Simulation of watershed streamflow is essential for the prevention and control of flood and drought disasters. To improve streamflow simulation, a coupled SWAT-LSTM model was constructed by combining a conceptual process-based hydrological model—Soil and Water Assessment Tool (SWAT)—with a machine learning model—Long Short-Term Memory (LSTM). The coupled model was applied to simulate the daily streamflow of the upper Huaihe River above the Xixian station from 1962 to 2010, which identified the optimal explanatory variables of the model and reduced streamflow simulation errors. Furthermore, four machine learning models, back propagation (BP) neural network, gated recurrent unit (GRU), support vector regression (SVR) and extreme gradient boosting (XGBoost), were chosen to assess the effectiveness of coupling SWAT with LSTM in streamflow simulation. Results showed that the coupled SWAT-LSTM model performed satisfactorily in streamflow simulation in the study area, with NSE reaching 0.90 and 0.85 in calibration and validation periods, respectively. The coupled model showed a significant improvement in simulating flood peak and average streamflow in each period, with mean NSE increasing by 0.24 compared to the standalone SWAT model. In comparison to other coupled models (i.e., SWAT-BP, SWAT-GRU, SWAT-SVR, and SWAT-XGB), the mean NSE of SWAT-LSTM exhibited an improvement of 0.02–0.16 during validation period. Furthermore, the coupled model effectively avoided the overfitting problem and had better generalization performance. The findings of this study offer new ideas for streamflow simulation of watersheds and provide references for water resources management and planning.