Coupling Hydrological Model With Interpretable Machine Learning for Reliable Streamflow Modeling: Daily Dynamics and Extreme Events

Abstract

Reliable long-term daily and extreme streamflow simulation, essential for watershed sustainable development, remains challenge in changing environments due to the complementary limitations inherent in conventional physical-driven and data-driven models. This study proposed a physics-guided machine learning (ML) approach that coupled SWAT with interpretable ML to enhance streamflow simulation accuracy for both daily and extreme streamflow whilst maintaining physical interpretability. This study systematically compared SWAT and three SWAT-ML models (SWAT-DT, SWAT-LSBoost, and SWAT-RF) to modify systematic model residuals, incorporating Shapley additive explanations (SHAP) to quantify feature contributions to streamflow simulations, and apply it to the Taoer River Basin (TRB), China. Results demonstrated that coupled models achieved daily streamflow simulation with $\mathrm{KGE}$ values consistently above 0.94 and $\text{PBIAS}$ values for extreme streamflow within 17%. In comparison with the standalone SWAT, the coupled framework further cut runtime from nearly 200 h to a few minutes. Additionally, multi-model comparisons revealed the superior performance of SWAT-LSBoost in streamflow simulations, with SHAP further highlighting the predominant role of watershed hydrological process in governing coupled model. Thus, this approach enhanced modeling precision while strengthening the reliability and transparency of outputs, offering a scientifically robust foundation for decision-making in long-term water resources planning and flood-drought disaster mitigation strategies.