Abstract: With the continuous integration of large-scale distributed power sources into distribution networks, distribution networks face many challenges in terms of safety, stability and economy. And the existing deep reinforcement learning methods often exhibit limitations in generalization ability when training agents to cope with changing operating conditions due to insufficient generalization of collected data. Therefore a distribution network regional voltage optimization control strategy based on large language model-assisted deep reinforcement learning is proposed, combining large language model technology with deep reinforcement learning. Secondly, by guiding large language models to generate customized datasets for deep reinforcement learning agent training through prompt engineering, a multi-agent collaborative decision-making framework is constructed. Then, based on distributed partially observable Markov processes, dynamic control problems are modeled to reduce dependence on real-world data while improving agent generalization ability. Finally, the effectiveness of the proposed control strategy is verified on the improved IEEE 33-node system, with voltage deviation and network loss reduced by 60. 82% and 49.91%, respectively, exhibiting strong robustness under various operating conditions.

Key words: deep reinforcement learning, large language model, voltage control, multi-agent, dataaugmentation, data-driven