Large Language Model for Secure Operation of Power Systems

The integration of large-scale renewable energy, multi-criteria operational constraints, and complex grid topologies has intensified the challenges faced by the security monitoring process within power system dispatch. Dispatch guidelines, typically expressed in natural language, are difficult for conventional algorithms to interpret and apply in real time, while general-purpose Large Language Models (LLMs) lack domain-specific knowledge, risking inaccurate or unsafe recommendations. This study proposes an LLM-based monitoring framework that integrates domain-specific prompt engineering with fuzzy evaluation to address these limitations. The framework interprets dispatch guidelines, analyzes real-time power flow data, and converts semantic assessments into quantitative safety scores, enabling closed-loop decision-making. Validation on the IEEE 14-bus system demonstrates that the optimized LLM outperforms a general LLM in accuracy, logical consistency, and stability under complex multi-standard scenarios, while reducing reliance on manual intervention. The results highlight the framework’s potential to enhance monitoring efficiency and ensure intelligent, secure power system operation.