A mutual information-based dual-layer attention framework for modelling human activity-driven runoff variation in arid basins

In arid regions impacted by human activities and lacking engineering data, accurately characterizing runoff variability remains challenging. This study proposes a runoff-variation modelling framework, MIDLA-ML, which integrates a mutual-information-based dual-layer attention mechanism with a physical baseline. It reconstructs a naturalized runoff baseline and extracts anthropogenic regulation signals from runoff series. Using the Daheihe River Basin as a case study, the results indicate that traditional hydrological models exhibit significant biases in the downstream area influenced by water diversion (with a negative NSE). Compared to the baseline model, the MIDLA framework significantly improves runoff reconstruction accuracy at the mid-downstream stations, with NSE increasing by 0.37–1.24. Mutual-information analysis indicates that naturalized runoff and short-term lag features contribute most to model performance, highlighting the importance of hydrological memory and human regulation signals. This framework provides an effective modelling approach for basins lacking engineering operation data.