Meteorology-Driven Self-Correcting LLM Priors for Smart Grid Outage Prediction and Bayesian Fault-Chain Risk Analysis

The increasing frequency and intensity of extreme weather events, coupled with uneven spatial demand, are shaping complex spatio-temporal outage dynamics. This makes accurate outage prediction operationally meaningful. In this paper, a large language model (LLM)-based prediction framework is proposed, which combines rule guidance with spatio-temporal prompting and a multi-agent closed loop. First, a severity-classified historical set and an expert if-then rule are refined by an LLM. Second, a spatiotemporal prompt paradigm encodes sliding-window multivariate series into natural language, revealing causal and sequential structures. Third, a multi-agent LLM with chain-of-thought and memory performs a closed-loop of prediction, evaluation, and optimization, adaptively updating rules. Finally, LLM priors are injected into a Bayesian network–based fault-chain risk module to compute chain and system probabilities, and produce demand response dispatch alarms. The proposed method is verified based on the dataset of the east coast of Australia. Across multiple sites, this approach improves discrimination and calibration over strong baselines, and system-level risk aligns well with observed events, capturing both single-node extremes and multi-chain concurrency. The result is a robust, self-correcting workflow that links forecasting to actionable decision-making for energy management through risk-aware DR recommendations. © 2026 IEEE.