Accurate streamflow forecasting is crucial for water resources management and flood mitigation, yet it remains challenging due to the complex dynamics of hydrological systems. Conventional data-driven approaches often struggle to effectively capture spatio-temporal evolution characteristics, particularly the dynamic interdependencies among streamflow gauges. This study proposes a novel deep learning architecture, termed DynaSTG-Former. It employs a multi-channel dynamic graph constructor to adaptively integrate three spatial dependency patterns: physical topology, statistical correlation, and trend similarity. A dual-stream temporal predictor is designed to collaboratively model long-range dependencies and local transient features. In an empirical study within the Delaware River Basin, the model demonstrated exceptional performance in multi-step-ahead forecasting (12-, 36-, and 72 h). It achieved basin-scale Kling–Gupta Efficiency (KGE) values of 0.961, 0.956, and 0.855, significantly outperforming baseline models such as LSTM, GRU, and Transformer. Ablation studies confirmed the core contribution of the dynamic graph module, with the Pearson correlation graph playing a dominant role in error reduction. The results indicate that DynaSTG-Former effectively enhances the accuracy and stability of streamflow forecasts and demonstrates its strong robustness at the basin scale. It thus provides a reliable tool for precision water management.
Li et al. (Mon,) studied this question.