Mingyue: A Multi-Agent System for Human-AI Collaborative Decision-Making Under Safety Constraints in Complex Power Grids

To address the conflict between the high complexity and uncertainty of modern power systems and the extreme safety requirements for dispatch decision-making, this paper introduces ‘Mingyue,’ a multi-agent system for human-AI collaborative decision-making. Traditional dispatch methods rely on manual experience, which is inefficient and prone to errors. Conversely, applying general artificial intelligence, such as large language models (LLMs), directly to this safety-critical domain poses fundamental risks of factual hallucination and unreliable actions. The Mingyue system overcomes these challenges through three key innovations: 1) A unified knowledge representation framework for power dispatch that transforms vast, heterogeneous domain knowledge (e.g., regulations, case studies, and operational procedures) into a machine-readable and traceable knowledge base, providing a reliable foundation for all decisions. 2) A decoupled ‘Intent-Plan-Tool’ architecture that leverages LLMs for task planning while invoking a deterministic professional toolset for execution, fundamentally isolating the non-deterministic nature of AI from the precise control required in the physical world. 3) A hierarchical multi-agent collaborative mechanism for complex dispatch workflows, which automates end-to-end tasks through the coordination of specialized agents (e.g., safety supervision, decision-making, and operation agents). The system has been deployed in the Guangdong power grid, the world's largest provincial grid. Experimental results show that compared to traditional manual modes and standalone AI tools, Mingyue improves efficiency by over 85 % in key tasks such as risk analysis, emergency response, and operational ticket generation, significantly enhancing the safety and intelligence of power grid operations.