Predicting Extreme Weather-Induced Power Outages with Tabular Foundation Mode
Lead PI: Dr. Dongjin Song, UConn
Industrial Relevance:
The enhanced OPM system has the potential to significantly improve how utility companies anticipate, manage, and respond to outage risks. By enabling robust few-shot generalization and strong adaptation, this study addresses the long-tail nature of rare extreme events and the spatial heterogeneity of power systems. Specifically, this study will: 1) strengthen outage prediction across diverse geographic regions and extreme weather types, providing guidance for the next generation of OPM system development; 2) support proactive outage management by providing utilities with actionable insights for resource allocation, rapid response, and contingency planning, ultimately reducing downtime and improving service reliability for customers; 3) promote industry adoption of foundation models for operational decision-making by demonstrating the practical value of tabular foundation models in addressing rare-event prediction, distribution shifts, and complex multi-factor interactions in real-world power grids.
Objectives:
We will develop a unified framework that enables in-context learning for extreme outage prediction using tabular foundation models (e.g., TabPFN), thereby improving robustness to long-tail and rare extreme weather events through few-shot inference.
We will enhance model adaptability and task-specific performance via retrieval-based support set construction and fine-tuning technique (e.g., encoder-based fine-tuning), enabling few-shot generalization from limited data without full model retraining.
We will incorporate spatial relationships into the feature representation to effectively capture spatial dependencies.
We will conduct comprehensive evaluations across multiple extreme weather event types and diverse regions, comparing the proposed framework with existing OPM methods (e.g., Random Forest, XGBoost, and GNN-based approaches), with the goal of improving predictive performance, robustness and generalizability.