A GRU-based Xinanjiang model for watershed runoff simulation

To address the dilemma of limited accuracy in traditional physics-based hydrological models and poor generalization in pure data-driven models, this study proposes a novel hybrid hydrological model (XAJ-BiGRU) based on the concept of "modular deep replacement." This model uses the Xinanjiang model as its physical framework, employing a Bidirectional Gated Recurrent Unit (BiGRU) network to precisely replace the original model's error-prone evapotranspiration and runoff generation modules. Meanwhile, it retains physically meaningful modules, such as routing, as constraints, thereby constructing a "grey-box" model that balances high accuracy with physical interpretability. In a case study validation in the Niulanjiang River Basin, the proposed XAJ-BiGRU model demonstrated significant advantages in both runoff simulation accuracy and stability compared to the traditional Xinanjiang model and a pure data-driven model, effectively overcoming the structural deficiencies of the former and the overfitting issue of the latter. The research indicates that this deep integration strategy successfully combines the reliability of physical mechanisms with the flexibility of data-driven approaches, offering an effective pathway for developing next-generation intelligent hydrological forecasting models.