Abstract: The maintenance capacity of hydropower units is severely constrained by complex factors related to power generation, transmission, and consumption.Quantifying the influence of these factors is essential for improving the rationality of dispatch planning.This study proposed an analytical method based on explainable machine learning to assess the factors affecting hydropower maintenance.First, spatiotemporal variation characteristics were extracted from long-term time series of cascade hydropower maintenance capacity.A random forest algorithm was then used to establish a nonlinear relationship model between maintenance capacity and key influencing factors, such as water inflow, peak shaving, holidays, ecological dispatch, power exchange, line maintenance, and power supply security.Furthermore, a cooperative game theory-based method was developed to quantify the interaction effects among these factors and to identify the dominant drivers for each hydropower station.The results indicate that: ① The random forest model achieves the best performance in fitting the relationship between the multiple factors and maintenance capacity, with an R² of 0.94, MSLE of 44.72, and SMAPE of 0.02;② Water inflow, peak shaving, and holidays are the key factors, with influence weights of 52.89%, 20.95%, and 10.86%, respectively; ③ Influencing factors vary significantly among stations within a cascade, which depends on their specific operational roles in power grid.The proposed method improves the rationality and refinement of maintenance and generation scheduling for large-scale cascade hydropower systems.