Research on Reactive Voltage Optimization in Distribution Networks Using Graph Neural Network-Based Reinforcement Learning

With the increasing integration of distributed renewable energy, distribution networks are facing increasingly severe issues such as voltage violations and increased network losses. Although there has been considerable research on reactive power voltage control using reinforcement learning, few studies have successfully maintained good performance when distribution network topology changes or data is missing. In this paper, a reactive voltage control strategy for distribution networks is proposed, which integrates Graph Neural Networks (GNNs) and the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm. This method effectively addresses the shortcomings mentioned above. First, the reactive voltage control problem in the distribution network is modeled as a Markov Decision Process (MDP). The GNN is used to extract topological features of the distribution network, thereby enhancing the agent's perception of the system state and robustness, enabling the optimization of reactive power output for photovoltaic inverters and Static Var Compensators (SVCs). Finally, simulations are conducted on the improved IEEE 33-node system, and the results demonstrate that the proposed method performs excellently in reactive voltage control for distribution networks.