Machine Learning-Based Analysis of the Seasonal Effects of Three Gorges Dam Regulation on Discharge in the Middle Yangtze River

Quantifying the net hydrological impact of large dams amidst climatic and anthropogenic influences remains a major challenge. This study isolates the effect of Three Gorges Dam (TGD) regulation on discharge at Jiujiang Station in the middle Yangtze River (2009–2016) using a novel scenario-based framework. A Long Short-Term Memory (LSTM) network, optimized by the Sparrow Search Algorithm (SSA), simulated daily discharge with high accuracy (Nash–Sutcliffe Efficiency coefficient > 0.97). By comparing a “with-TGD” simulation against a “without-TGD” scenario—generated by replacing the dam’s regulated outflow with its reconstructed natural inflow—we quantified the net impact (ΔQ). Results show that ΔQ is substantially modulated by river–lake interactions. For example, in December, the backwater effect from Poyang Lake amplified the direct flow reduction by an additional −82.5 m3/s. The “peak-shaving” effect was context dependent: TGD regulation increased high flows (>30,870 m3/s) by an average of +372 m3/s while slightly decreasing low flows (<12,711 m3/s) by −31 m3/s. The impact exhibits strong seasonality alongside considerable intra-seasonal variability, reflecting multi-objective operations (flood control, power generation, water supply). This framework provides a transferable approach for attributing hydrological change in large regulated rivers and supports integrated water resources management.