Large Language Models in Power Systems: A Survey on Applications and Fine-tuning

Large language models (LLMs) are promising tools for knowledge representation, reasoning, and decision-making in industrial sectors. The employment of LLM in power systems is receiving increasing attention, but is still limited by the scarcity of domain-specific data, the absence of unified benchmarks, the complexity of integrating heterogeneous data, and strict physical time-varying constraints. This paper presents a focused survey on the applications and fine-tuning techniques of LLMs within power systems. The taxonomy of LLMs is first introduced, covering general foundation LLMs, industry foundation LLMs, domain-specific LLMs, and vertical scenario LLMs. The discussion then addresses the principles, categories of fine-tuning methods, evaluation metrics, and benchmark datasets that support model adaptation. Finally, the current progress, potential applications, and challenges of specialized LLMs in the power sector are examined, with emphasis on hybrid fine-tuning strategies, multimodal data integration, and deployment frameworks designed to ensure secure and reliable power system operation.