Large Language Model-Based Decision Support Technology for Operation and Dispatch of New Power Systems

Abstract: The significantly increased source-load uncertainty in new power systems poses fundamental challenges to conventional dispatch decision-making. Large language models (LLMs), with their semantic understanding and logical reasoning capabilities, offer a new pathway to address these decision-making difficulties. This paper analyzes the core challenges faced by dispatch decision-making and the limitations of traditional methods. Starting from LLM domain adaptation mechanisms, we propose a theoretical decision-support framework that integrates knowledge injection and collaborative reasoning. Three key technologies are elaborated: knowledge embedding and representation methods tailored for power grid dispatch, a collaboration mechanism between LLMs and conventional numerical computation engines, and a reasoning and decision-making paradigm based on Chain-of-Thought (CoT) and Retrieval-Augmented Generation (RAG). Finally, challenges such as reliability, interpretability, and real-time performance are discussed, along with future research directions.