A Physics–Data Hybrid Xin’anjiang Flood Forecasting Model Based on LSTM Residual Correction and SHAP Interpretability

Flood forecasting in small mountainous catchments is challenging due to strong nonlinearity in runoff generation and short hydrological response times, which are often inadequately represented by conceptual models. The Xin’anjiang (XAJ) model, although widely applied, relies on simplified process representations that limit its ability to capture complex flood dynamics. In contrast, data-driven approaches such as Long Short-Term Memory (LSTM) networks offer high predictive flexibility but suffer from limited physical interpretability. To bridge this gap, we propose an interpretable physics–data hybrid framework (XAJ–LSTM), in which an LSTM network dynamically corrects residuals from the XAJ model while explicitly incorporating physically meaningful state variables. Model interpretability is enhanced using SHapley Additive exPlanations (SHAP), which quantify the contribution of different inputs to flood predictions. The framework is evaluated using 15 flood events observed between 2015 and 2018 in the Qiaodong Village catchment, a representative small mountainous basin in China. The results indicate that the XAJ–LSTM hybrid model significantly outperforms the standalone physical model, improving the Nash–Sutcliffe Efficiency (NSE) from 0.55 to 0.77 and effectively correcting peak flow errors. Moreover, the integration of physical state variables, particularly soil moisture, is crucial for improving predictive accuracy, whereas adding redundant runoff components introduces noise and degrades model performance. SHAP analysis further confirms that antecedent observed discharge and XAJ-simulated discharge are the dominant drivers of the LSTM-based correction. Overall, this hybrid framework improves flood forecasting accuracy while enhancing interpretability, offering a promising approach for physically informed modeling in nonlinear, data-limited catchments.