基于生成式多对抗强化学习的高比例新能源电网日内优化调度

With the continuous increasing proportion of new energy sources，the strong randomness on both the supply and demand sides has heightened the risk of power grid safe operation. The learning ability of reinforcement learning scheduling algorithms in dealing with the uncertainty of system state transition is still limited，and their anticipatory decision-making ability needs further enhancement. To address these challenges，the intraday optimal scheduling for power system with high renewable energy based on generative multi-adversarial reinforcement learning is proposed. A generative adversarial network as the target network for reinforcement learning is constructed to learn the reward feedback distribution of the power grid’s future operational status，so as to predict the operational trend within the scheduling period，ensuring the optimality of scheduling decision. During training，a hybrid experience cross-driving mechanism is employed，where experiences are evaluated based on scheduling performance and extracted in proportion，thereby reducing the training duration. The proposed method is tested on the SG-126 node power grid dispatching simulation platform，and the computational results validate the effectiveness and stability of the method.