OutageGPT: Multi-Agent Retrieval-Augmented Generation Framework for Power Outage Analysis and Prediction

Abstract

Power system outage prediction models remain limited by data fragmentation, interpretability challenges, and operational deployment difficulties. With the rise of large language models (LLMs), we explore their potential to assist utilities in outage analysis and prediction—specifically through retrieval-augmented generation (RAG), which augments the model’s context with retrieved historical records. ``OutageGPT a multi-agent RAG framework, is introduced. It integrates a mixture-of-experts architecture and advanced prompting strategies to address diverse outage-related queries. In retrospective 2021 severe-weather test cases, OutageGPT outperformed a state-of-the-art open-source LLM queried directly without retrieval, with ground-truth values more frequently within its predicted ranges due to contextual grounding from historical data. While it posses some limitations, like underestimating extreme events and producing broad prediction intervals, it demonstrates promise while highlighting future needs, including multimodal integration and domain-specific foundation models for energy systems.