SafeVolt: Closed-Loop Large Language Model Framework for Safety-Aware Voltage Control in Active Distribution Networks

Voltage and reactive power control in active distribution networks is a safety-critical and highly dynamic problem, where traditional optimization methods often struggle to balance efficiency and robustness under complex operating conditions. Recently, large language models (LLMs) have shown promise in sequential decision-making tasks, but their direct application to power system control remains limited by the lack of physical grounding and safety guarantees. In this paper, we propose SafeVolt, a closed-loop LLM-based framework that integrates multi-candidate action generation, simulator-in-the-loop evaluation, and a fine-tuned expert judge for safety-aware decision making. In addition, a high-level rule distillation mechanism that converts successful control experiences into reusable operational axioms is introduced to enable iterative self-improvement. Experiments on a standard distribution network scenario demonstrate that the proposed method outperforms representative baselines, achieving substantial improvements in average reward, voltage violation rate, reactive power loss, and system stability. In particular, voltage violations and extreme events are substantially reduced, indicating enhanced operational safety. These results suggest that combining LLM reasoning with physical simulation and structured feedback provides a promising direction for reliable and adaptive power system control.