POWER GRID STABILITY THROUGH AN AI AGENT FRAMEWORK: UNCERTAINTY-AWARE FORECASTING DURING EXTREME EVENTS

Abstract

Extreme weather events are increasingly destabilizing power systems by simultaneously damaging infrastructure and driving volatile demand. Traditional load-forecasting pipelines—designed for stationary dynamics and abundant data—are brittle in high-impact, low-frequency (HILF) regimes characterized by scarcity and heavy-tailed, multi-modal uncertainty. We propose TL-MDN, a two-stage AI agent that pairs transfer learning with a Mixture Density Network (MDN) head to produce calibrated, multi-modal probabilistic forecasts under rare-event conditions. A deep sequential encoder (e.g., LSTM/Transformer) is pre-trained on long-horizon “normal” operations from large markets (e.g., ERCOT, ISO-NE) and then adapted with an MDN head using sparse data from specific events (e.g., polar vortex, hurricane). We outline an evaluation protocol emphasizing proper scoring rules—CRPS, prediction-interval coverage probability (PICP), and Winkler score—to assess both reliability and sharpness. We further position TL-MDN against emerging LLM-based forecasters, highlighting complementary strengths and hybrid opportunities. The proposed framework targets deployment-grade usability for system operators through calibrated uncertainty, interpretable scenario modes, and seamless integration into reserve scheduling, demand response, and storage dispatch.