A Spatiotemporal Pathformer‐Based Deep Learning Framework for Watershed Flood Forecasting

Abstract Effective flood forecasting is essential for implementing proactive flood management and risk reduction strategies. However, conventional artificial neural networks often fail to capture the complex spatiotemporal dependencies among hydrometeorological variables, resulting in system biases and time‐lag errors, especially during extreme flood events. This study introduces a spatiotemporal Pathformer‐based deep learning framework for multi‐step‐ahead flood forecasting that dynamically adapts to flood magnitude and duration. The model integrates a dual self‐attention mechanism and adaptive path selection, enhancing its ability to model nonlinear rainfall‐runoff relationships and long horizon dependencies. Using a case study in the Jianxi basin with 25,341 hydrometeorological records at a 3‐hr resolution, the spatiotemporal Pathformer's performance was evaluated across 1‐ to 7‐step forecast horizons. A comparative analysis with Long Short‐Term Memory (LSTM) and Transformer models demonstrates the spatiotemporal Pathformer's superior predictive accuracy and stability. It improves nash‐sutcliffe efficiency by 3.0% and 7.4%, increases Volume Efficiency by 3.4% and 9.6%, reduces root mean square error by 18.4% and 34.9%, and lowers mean absolute error by 17.5% and 36.1% compared to LSTM and Transformer, respectively. By effectively mitigating time‐lag errors and prediction bottlenecks, the spatiotemporal Pathformer ensures robust and reliable forecasting, even during extreme flood events. The application of SHapley Additive exPlanations analysis increases the model's interpretability, transparency, and trustworthiness by revealing the key hydrometeorological drivers behind its predictions. These results establish the spatiotemporal Pathformer as an advanced solution for next‐generation flood forecasting, with strong potential to improve real‐world applications in flood prevention and water resource management.