PatchX: A dual-branch embedding framework with cross-attention for streamflow forecasting

Accurate streamflow forecasting is critical for flood management, yet hydrological models based on deep learning face an architectural dilemma: channel-mixing models may obscure heterogeneous hydrological signals through shared representations, while channel-independent architectures fail to explicitly capture causal rainfall-runoff mechanisms. We propose PatchX, a hydrologically structured dual-branch framework that separately encodes antecedent streamflow and meteorological information. This separation preserves runoff continuity while reducing interference from heterogeneous inputs. Multi-scale meteorological embeddings represent different response lags, and asymmetric cross-attention uses the runoff-state representation to select and incorporate relevant meteorological variables and temporal scales. Evaluated across six diverse basins (1- to 7-day lead times), PatchX consistently outperformed baselines, surpassing iTransformer and PatchTST by up to 7.82% and 3.78% in KGE. Furthermore, SHAP-based interpretability and ablation studies revealed that PatchX mitigates "attention collapse" and constructs internal mathematical representations that more closely align with physical and hydrological principles, providing a robust paradigm for streamflow forecasting.