Research on short-term optimal scheduling of hydro-wind-solar multi-energy power system based on deep reinforcement learning

Recently, grid-integrated wind power and photovoltaics have experienced rapid growth. However, their uncertainty increases the difficulty of grid scheduling and operation. Short-term power generation decisions made by conventional scheduling methods, which are based on the output forecast information of wind and solar power often impact the power benefit owing to forecast errors during the actual operation. This study explores the application of deep reinforcement learning methods for short-term optimal scheduling of hydro-wind-solar multi-energy power system, considering forecast uncertainty. First, with the objective of maximizing power generation benefit from the multi-energy complementary system, the Deep Q Network (DQN) method in deep reinforcement learning is employed to construct the model framework of the short-term optimal scheduling of hydro-wind-solar multi-energy power system (collectively referred to as DQN model later). Thereafter, the impact of each learning parameter in the DQN model on performances is analyzed, and the model parameters are reasonably configured. Furthermore, the DQN model is driven by wind and solar power output forecast data to develop short-term power generation decisions. Finally, the actual wind and solar power output is used as input to implement short-term decisions, whereas the dynamic programming (DP) method is used to build the model as a comparison to evaluate the power benefit and computational efficiency of the DQN model. The multi-energy complementary system of hydro, wind, and solar power of the Jinping-1 Hydropower Station in the Yalong river basin is used as an example for the study. In the flood season, compared with the DP model and the historical actual operation process, the total power generation benefits obtained from the DQN model's implementation of short-term decision-making increases by 6.18% and 1.38%, respectively. In the dry season, the total power generation benefits obtained from the DQN model's implementation of short-term decision-making increases by 2.26% and 19.00% respectively compared with the DP model and the historical actual operation process. Additionally, the DQN model can significantly improve the decision-making efficiency. The DP model requires minutes or hours of computation time; the DQN model consumes less than 1s to compute, which is conducive to efficient decision making of the complementary system.