Physics-guided Kolmogorov–Arnold Network for extreme flood and drought flow prediction: A case study of the Upper Hanjiang River Basin

Study Region: Upper Hanjiang River Basin (UHRB), headwater catchment of the Middle Route of the South-to-North Water Diversion Project, China. Study Focus: Daily flow prediction underpins water resources management, yet existing deep learning methods struggle to reproduce hydrological extremes. This study proposes Deep Process Learning-Long Short-term Kolmogorov–Arnold Network (DPL-LSTKAN). Specifically, DPL-LSTKAN captures the temporal evolution of meteorological forcings and employs Kolmogorov–Arnold Network (KAN) to represent the intrinsic dependencies between meteorological forcings and runoff. Decoupled learning scheme is adopted to separately characterize extreme and normal flow regimes. In addition, Physics-guided gating mechanism dynamically assigns predictor weights according to the dominant hydrological processes. DPL-LSTKAN is applied to flow prediction in the UHRB, with particular focus on its accuracy and physical consistency under extreme floods and droughts. New Hydrological Insights for the Region: In the UHRB, hydrological extremes are controlled by antecedent catchment wetness rather than instantaneous rainfall, most strongly by soil water storage ( 𝑟 = 0 . 5 9 ). Flood and drought follow distinct mechanisms: fastflow responds near-monotonically to precipitation via saturation-excess pathways, whereas baseflow scales with subsurface storage and declines toward zero during dry spells. Because fastflow is released only as storage approaches capacity, a saturation onset precedes flood peaks, reproducing peak timing within one day ( P T E = 0 . 8 3 d a y ) while suppressing implausible low-flow errors. These controls support reservoir scheduling, flood warning, and ecological-flow management.