Abstract: The accurate estimation of water consumption rate is directly related to the accuracy and operability of hydropower and cascade power generation scheduling plans. In response to the serious deviation between the head water consumption rate curve of hydropower stations and the actual results due to long-term operation, which makes accurate estimation of water consumption rate extremely difficult, this paper proposes a water consumption rate estimation method based on an interpretable machine learning data-driven model. Firstly, the Pearson correlation analysis method is used to analyze the historical long-term operation data of hydropower plants, extracting time-series features from the operational data and determining influencing factors such as reservoir water level, head, and generation flow that affect the water consumption rate. Then, these extracted factors are used as inputs to construct a prediction model for the water consumption rate using Support Vector Regression (SVR), which is advantageous in handling nonlinear relationships. Finally, the SHAP interpretability algorithm is introduced to calculate the contribution of each input factor, and the prediction model is iteratively corrected based on the contribution values. The simulation was conducted on a hydropower station in the Yangtze River Basin. The results demonstrate that the proposed method achieves RMSE and MAE values of 0.0619 m³/kWh and 0.0412 m³/kWh, respectively, in predicting water consumption rate, representing reductions of approximately 40% and 60% compared to the traditional interpolation method (0.21 m³/kWh and 0.18 m³/kWh). Additionally, the Nash-Sutcliffe efficiency (NSE) coefficient reaches 0.9724, showing an improvement of about 20% over the traditional method (NSE = 0.75). The proposed approach effectively mitigates deviations in water consumption rate caused by variations in operational conditions while enhancing the interpretability of both the model structure and prediction results.