Predicting streamflow in heavily regulated plain river networks using physics-guided spatiotemporal deep learning models: A case study of the Four Lakes Basin

Study region The Four Lakes Basin is a typical agricultural plain river network subject to intensive regulation by artificial sluices and pumping stations. Study focus This study proposes a hybrid streamflow prediction model, SWAT-GNN-LSTM, that integrates physical processes with data-driven methods. Using SWAT-simulated hydrological states as physically informed boundary conditions, the model jointly extracts spatial topology and time-lag effects to infer and reconstruct regulated streamflow indirectly. New hydrological insights for the region (1) The model effectively mitigates peak-flow misalignment and extreme errors caused by intensive cross-boundary pumping and internal regulation, with average NSE values of 0.77 in validation and 0.66 in independent testing across six regulated nodes. (2) Ablation experiments confirm that the spatiotemporally coupled architecture successfully reconstructs the spatial transmission and routing delays of operation-related flow responses. (3) Feature sensitivity analysis shows that the contributions of wind speed, evapotranspiration, and temperature unexpectedly exceed that of precipitation, indicating a substantial shift in the dominant driving factors of the unnatural hydrological cycle under intensive agricultural water diversion. (4) Confidence intervals generated using MC Dropout quantitatively capture plausible streamflow fluctuations under typical flood-control operation periods, thereby objectively characterizing operational margins for decision-making. This study provides a reliable approach for streamflow simulation in complex regulated basins that lack fine-resolution operational data.