Abstract: Accurate estimation of the water consumption rate is directly related to the precision and operability of hydropower and cascade power generation scheduling plans.For addressing the issue that the head-water consumption rate curve of hydropower stations often deviates significantly from actual results over long-term operation, making accurate estimation extremely challenging, this study proposes a data-driven interpretable machine learning-based method for estimating the water consumption rate.First, the Pearson correlation analysis method is employed to analyze long-term historical operational data of hydropower stations, extracting temporal features from the operational data and identifying influencing factors such as reservoir water level, head, and discharge flow.Then, a water consumption rate prediction model is constructed using Support Vector Regression (SVR), which excels in handling nonlinear relationships, with the identified factors as inputs.Finally, the SHAP interpretability algorithm is introduced to calculate the contribution values of each input factor, and the prediction model is iteratively calibrated based on these contribution values.A simulation was conducted using a hydropower station in the Changjiang River Basin as a case study.The results show that the proposed method achieves RMSE and MAE values of 0.061 8 m³/kWh and 0.041 2 m³/kWh, respectively, representing reductions of approximately 70% and 77% compared to the 0.21 m³/kWh and 0.18 m³/kWh obtained by the traditional interpolation method.The Nash-Sutcliffe efficiency (NSE) reached 0.972 5, an improvement of about 30% over the 0.75 of the traditional interpolation method.This approach effectively overcomes the deviation in water consumption rate caused by changes in operational conditions while enhancing the interpretability of both the model structure and prediction outcomes.The research findings provide a valuable reference for the analysis and prediction of water consumption rate characteristics in hydropower stations.