Data-driven model-free graph multi-agent deep reinforcement learning for voltage-voltage-ampere reactive control in active distribution networks

The high penetration of distributed generation (DG) and renewable energy sources (RES) exacerbates voltage-voltage-ampere reactive (volt–VAR control, VVC) due to fast uncertainty-driven voltage fluctuations, topology-coupled interactions across feeder regions, and the limited real-time practicality of model-dependent optimization and iterative power-flow solvers. This paper formulates VVC as a Markov game and proposes an integrated, data-driven, topology-aware multi-agent learning architecture. From an artificial-intelligence perspective, we develop a centralized training with decentralized execution (CTDE) multi-agent graph soft actor–critic framework (MAGSAC), where graph convolution is embedded into Actor–Critic networks to enhance topology-aware coordination and stabilize multi-agent learning under partial local observations. To avoid repeatedly invoking iterative power-flow solvers during interaction, an offline-trained power flow neural network (PNN) is embedded into the environment to provide fast voltage estimation. It is trained using Latin hypercube sampling and a Levenberg–Marquardt optimizer, reducing reliance on explicit network parameters. A polygon voltage fortress reward is further introduced for fine-grained penalty shaping near operational limits. From an engineering-application perspective, the proposed architecture enables coordinated reactive-power regulation of inverter-interfaced DGs and compensators under high DG variability, improving voltage compliance and operational economy. Case studies on modified IEEE 33-bus and IEEE 69-bus systems show that MAGSAC achieves zero voltage violation on the 33-bus typical-day test and reduces network losses by 38.13 % versus no control. On the 69-bus system, it attains the lowest voltage-violation ratio among compared strategies and reduces daily loss by 27.18 %. Additional tests under topology changes and measurement-data loss support robustness and generalization under the studied settings.