Abstract: To address the "curse of dimensionality" and computational efficiency bottlenecks associated with applying traditional Discrete Differential Dynamic Programming (DDDP) to reservoir group operation, this paper proposes an improved method (IDDDP) that integrates DDDP with machine learning. This method employs a Bidirectional Long Short-Term Memory network with an attention mechanism to establish a direct mapping relationship between "inflow-initial/final water level" and reservoir power output, replacing the traditional recursive calculations and thereby significantly reducing the computational burden. Taking the cascade reservoir system in the Wujiang River basin as a case study, comparative experiments were conducted for two typical scenarios—power generation scheduling and grid peak-shaving—under varying discretization precision and system scales. The results indicate that the scheduling schemes obtained by IDDDP are highly consistent with DDDP results in key performance indicators such as power generation and load rate, with relative deviations kept within an allowable engineering range. The computational time is reduced by one to two orders of magnitude, and the method maintains stable performance under both wet and dry typical hydrological year types. The proposed method significantly enhances computational efficiency while ensuring scheduling accuracy, providing a reliable and efficient new approach for the optimal operation of large-scale reservoir groups.