Multi-Agent Deep Reinforcement Learning for Voltage Control in Regional Power Grid with Grid-Forming Energy Storage Stations

The increasing integration of stochastic renewable energy sources presents significant challenges to the voltage stability of regional power grids (RPGs). Traditional model-based control methods struggle with real-time requirements, while standard multi-agent deep reinforcement learning (MADRL) faces limitations in scalability and sample efficiency. To address these issues, this paper proposes a coordinated voltage control strategy integrating grid-forming energy storage stations (GFMSS) and an enhanced MADRL framework. First, a control-oriented model of GFMSS is established, explicitly embedding physical constraints into the learning environment. Second, to enhance scalability in mesh-structured grids, an attention mechanism is integrated into the critic network to dynamically filter redundant neighbor information. Third, a dual-track hybrid priority experience replay mechanism is designed, which employs a sliding window for concept drift adaptation and a permanent fault buffer for sparse sample retention, effectively mitigating catastrophic forgetting. Simulation results on modified IEEE 30-bus and 118-bus systems demonstrate that the proposed method reduces cumulative voltage deviation and achieves sub-second decision-making (0.14 s). Furthermore, the strategy exhibits superior transient resilience, enabling rapid voltage recovery within 1.2 s under short-circuit faults.