Collaborative Stability and Benefit Optimization of V2G Based on Multi-agent Graph Reinforcement Learning with Privacy Protection

The adoption of renewable energy sources (RESs) has reshaped power systems, causing voltage violations and power imbalances. Electric vehicles (EVs) add to this burden but also offer potential active and reactive power support. Distribution system operators (DSOs) need effective distributed control to manage uncertainties and address privacy concerns. Existing research has mainly relied on model-based methods to solve active and reactive power control problems. However, these studies often neglect the privacy issues of EV owners and lack effective incentives to encourage users to participate in power regulation. This paper proposes a supply-side and demand-side cooperative control framework based on Multi-Agent Twin Delayed Deep Deterministic Policy Gradient (MATD3). This framework uses Graph Convolutional Networks (GCN) as the state encoder and combines a regional-global joint supervision strategy to construct the safety control policy, while the Federated-EVs control scheme is employed to protect user privacy and enhance user satisfaction. Simulation results show that the proposed algorithm demonstrates superior performance and advantages in peak voltage control, reducing peak power loss, meeting user demands, and improving user benefits.