A physically guided and interpretable SWAT-BiLSTM framework with Bayesian optimization for bias correction in daily...

Accurate extreme streamflow simulation is essential for flood forecasting, water resource management, and water quality protection. However, process-based and data-driven models often suffer from limitations such as systematic bias, limited robustness, and poor interpretability, especially under extreme hydrological conditions. To address these challenges, this study proposed a novel hybrid modeling framework that integrated a process-based model, an optimized deep learning approach, and an interpretable analysis tool to improve streamflow prediction. Outputs from a calibrated Soil and Water Assessment Tool (SWAT) model were combined with meteorological data and fed into a Bidirectional Long Short-Term Memory (BiLSTM) network optimized by Bayesian optimization (BO). Feature factors selection was performed using a combination of random forest (RF) algorithm and correlation analysis, yielding two coupling schemes. SHapley Additive Explanations (SHAP) was further employed to provide interpretable insights into the behavior of the coupled model. Results indicated that the coupled models outperformed the standalone models in both accuracy and stability, with validation period R² and NSE improving by 14.7%-27.1% and 10.0%-35.0%, respectively. Among them, SWAT-S-BiLSTM achieved the best overall performance, with R² and NSE reaching 0.89 and 0.81, respectively. The improvement was particularly evident for the top 0.5% of flows, where the relative error reduced from -11.72% (SWAT) to -0.27% (SWAT-S-BiLSTM). SHAP analysis revealed pronounced hydrological lag effects, threshold behavior, and nonlinear responses of the input variables, and also provided insights into how the coupled model corrected the deficiencies of the physical model. Overall, the proposed framework enhanced predictive accuracy, interpretability, and scalability, offering a promising solution for streamflow forecasting under extreme hydrological events.