Modelling multi-component lagged hydrological responses in karst basins using a Bayesian-optimized dual-branch LSTM

Study region The Qingjiang River Basin (QJB), southwestern China. Study focus Accurate daily runoff prediction in karst basins remains challenging due to the distinctive dual-flow structure. To address this issue, we proposed a novel hybrid framework integrating baseflow separation method with Bayesian-optimized multi-lag rainfall feature selection. A dual-branch LSTM architecture (BOA-LSTM) was developed to separately simulate rapid and slow flow components based on their respective rainfall-response windows, while maintaining water balance through a combined streamflow output. The model was applied to the QJB using daily hydro‑meteorological data from 2006 to 2020. New hydrological insights for the region: Bayesian optimization automatically identified distinct rainfall-response windows for rapid flow (1–3 days) and slow flow (13–41 days). The BOA-LSTM outperformed other benchmark models (Transformer, LSTM, and VMD‑LSTM) across forecast horizons of 2–7 days, which mitigated the phase lag problem achieving an average RMSE reduction of 27.78%, with NSE values up to 0.94. Runoff decomposition revealed that slow flow contributed 22.4–100% of total flow, highlighting the dominant role of fissure-matrix storage in regulating hydrological release. Feature attribution analysis further shows that antecedent runoff and precipitation dominate daily predictions with positive contributions, while land surface temperature and canopy interception exert negative effects. This study bridges process understanding with data‑driven modeling, offering a framework that integrates hydrological mechanisms into deep learning‑based runoff prediction for karst basins.