OutageDiT: Generative AI Model Forecasts Power Outages and Scenarios

Power outage planning is a critical task that requires anticipating future scenarios, including the magnitude, timing, and duration of outages, while accounting for temporal dependencies and inherent uncertainties. A major challenge is the rarity of severe events, leading to limited data for training robust models in any single region. To address this, researchers introduce OutageDiT, a generative foundation model designed for forecasting power outages and simulating scenarios. Trained on extensive outage and weather records across the United States, OutageDiT generates detailed seven-day outage trajectories at a quarter-hour resolution. The model employs a condition encoder to process historical context and known future covariates, and a shallow flow decoder to generate complete trajectories. This architecture allows OutageDiT to support not only point forecasting but also uncertainty quantification and conditional event simulation within a single deep generative framework. Empirical results show that OutageDiT significantly improves forecast accuracy and scenario quality compared to existing baselines and demonstrates zero-shot transferability to previously unseen regions, positioning it as a crucial tool for operational planning under uncertainty.