LLMS FOR POWER SYSTEM OPERATIONS

Abstract

The rapid integration of renewable energy sources, proliferation of distributed energy resources (DER), and the growing complexity of interconnected smart grids have placed unprecedented informational and cognitive demands on power system operators. Traditional automation tools — energy management systems (EMS), supervisory control and data acquisition (SCADA), and rule-based expert systems — while reliable, lack the adaptive reasoning, natural language interface, and contextual knowledge synthesis required for emerging operational scenarios. This paper proposes a comprehensive framework for deploying Large Language Models (LLMs) across core power system operation tasks, including real-time fault diagnosis, short-term load forecasting, contingency analysis, anomaly detection, natural language-to-control-action translation, and automated incident report generation. A domain-specific fine-tuned model, PowerBERT, is developed by continual pre-training and supervised fine-tuning of a 340M-parameter transformer on curated corpora comprising IEEE technical standards, NERC reliability guidelines, SCADA operational logs, and historical fault records. Augmented with Retrieval-Augmented Generation (RAG) using a power system knowledge base, the proposed LLM + RAG system achieves 97.8% classification accuracy, a load forecasting MAPE of 2.1%, and 89 ms average response latency across six operational task categories. Comprehensive evaluations on IEEE 39-bus, 118-bus, and real-world Indian and Saudi grid datasets confirm that the proposed framework significantly outperforms support vector machines, random forests, deep learning baselines, and general-purpose LLMs. Expert evaluation of natural language outputs further demonstrates near-human fluency, accuracy, and actionability, supporting the vision of LLMs as the next-generation cognitive layer in grid control rooms.