Large Language Model-Assisted Zoning for Heterogeneous Automatic Voltage Control in Renewable-Rich Regional Power Grids

Abstract High penetration of wind and photovoltaic generation complicates reactive power and voltage regulation in regional power grids. A uniform automatic voltage control (AVC) mode may not meet differing regional requirements, while independent control by electrically coupled renewable energy stations can produce poorly coordinated reactive power responses. This paper proposes a heterogeneous AVC framework assisted by a large language model (LLM). First, the LLM uses structured operating features and AVC constraints to group renewable energy stations into control regions, assign control modes, and recommend initial parameters. Then, constant reactive power (CQP) and conventional constant-voltage (CV) strategies are applied in their assigned regions. In strongly coupled regions, an external CV loop determines station-group reactive power requirements, which a multi-agent deep deterministic policy gradient (MADDPG) policy allocates among stations. Finally, power-flow calculations validate the control results, and the validation feedback is used to refine the zoning, mode assignments, and parameters. Case studies on a modified IEEE 39-bus system show improved voltage regulation, more balanced reactive power utilization among stations, and fewer voltage violations than uniform CV control.