To address the issue of simulation bias in hydrological models caused by insufficient modeling or unmodeled processes under complex environmental conditions, this study proposes a residual correction and fusion method based on machine learning and Bayesian Model Averaging (BMA). Using the Hydromad model to generate initial inflow simulations and residual sequences, three machine learning models—Random Forest (RF), XGBoost, and Support Vector Regression (SVR)—are employed to capture the nonlinear characteristics of the residuals. The residual predictions from these models are then fused via BMA to achieve post-processing error correction and uncertainty assessment. Case study results demonstrate that the proposed method significantly improves model performance in peak flow response as well as rising and recession stages. After correction, the Nash–Sutcliffe Efficiency (NSE) increases to 0.983, the Kling–Gupta Efficiency (KGE) reaches 0.987, and the Percent Bias (PBIAS) is controlled within ±0.03%. This approach provides a more accurate and uncertainty-quantifiable technical support for hydropower station inflow forecasting and operational decision-making.