Reservoir operation rule function is a functional mapping method that reflects specific operational rules to guide reservoir operations. With the deepening of research, the scheduling function has changed from linear to nonlinear forms, from single-objective functions to integrated multi-objective functions, from ignoring uncertainty to incorporating uncertainty analysis, from using fixed parameters to parameters that dynamically change over time, and the system has been continuously refined. Different dispatching rule functions have different actual dispatching effects (flood control, power generation benefits, water resource utilization rate and reliability). Reservoir optimal dispatching aims to identify the best rule functions to achieve the specified objectives. Through advanced computing technologies, the complex relationships between dispatch-related factors and decision variables can be extracted from large amounts of historical reservoir operation data, thus facilitating the extraction of reservoir dispatch rules. With the construction of more and more reservoirs, the hydraulic connections between reservoir groups are complex, with mutual influences occurring across both spatial and temporal scales. These complex correlation factors need to be considered when extracting dispatch rule functions. Compared to conventional operation charts, the operation rule functions extracted through data mining methods offer greater convenience in reservoir group operations. This study adopts a machine learning model based on an autoencoder and self-attention mechanism, focusing on key reservoirs along the Yangtze River and its tributaries. A combined model of these techniques is proposed, and the natural gradient boosting scheduling comparison model is applied to optimize and compare the extracted scheduling rules. The applicability of the model is verified through analysis, and the results are further explained using SHAP (SHapley Additive exPlanations) theory, which provides interpretability for the model results.