This article proposes a real-time energy optimization model that integrates Conditional Generative Adversarial Networks (CGAN), Graph Neural Networks (GNN), and Deep Reinforcement Learning (DRL) to address the challenges of uncertainty, dynamism, and complexity in the power grid under large-scale integration of distributed energy. A double-layer framework of "off-line learning-on-line rolling optimization" is constructed. In the off-line stage, CGAN is used to learn the coupled distribution of source and load uncertainty to generate training scenarios, and GNN is used to extract the topological depth characteristics of power grid. The online phase employs a Deep Deterministic Policy Gradient (DDPG) agent to achieve millisecond-level adaptive decision-making. It transforms the power flow equations and safety constraints into soft penalty terms within the reward function, ensuring the feasibility of the solution. Experiments on the improved IEEE 33-node power distribution system show that compared with the stochastic model predictive control, the proposed model reduces the daily average operating cost by 13%, the frequency of voltage exceeding limit is significantly reduced, the new energy consumption rate is improved, and the average decision-making time is only 25 ms, which effectively takes into account the requirements of economy, safety and real-time, and provides a new paradigm for data-driven and physical information fusion for high-proportion new energy grid optimization.
Wu et al. (Sun,) studied this question.