OutageDiT: A Generative Foundation Model for Power Outage Forecasting and Scenario Simulation

OutageDiT is designed to address the challenge of predicting power outages, which are often rare and heterogeneous across different regions. Traditional models trained on single-region data often lack sufficient examples of extreme events to provide reliable forecasts. To solve this, the authors developed a foundation model trained on nationwide outage and weather records. The model uses a condition encoder to process 14 days of historical outage data, weather, and calendar covariates, mapping them into horizon-aligned states. A shallow flow decoder then uses these states to generate complete seven-day outage trajectories at a quarter-hour resolution, allowing for both point forecasting and probabilistic scenario simulation.

OutageDiT consistently outperformed strong baselines, including supervised probabilistic forecasters and fine-tuned time-series foundation models, across national benchmarks. The model demonstrated superior accuracy in both normal and event-driven scenarios. A key strength of the architecture is its ability to maintain temporal dependence and forecast uncertainty, which are critical for operational planning. Furthermore, the model showed strong zero-shot transfer capabilities, outperforming competition-winning models on held-out data from Michigan without any regional fine-tuning.

Effective power-outage planning requires more than just a single point estimate; it requires a range of plausible scenarios that account for uncertainty in timing, magnitude, and duration. By providing a generative framework that can simulate these trajectories, OutageDiT bridges the gap between traditional forecasting and prescriptive operational planning. Its ability to perform well even in regions with little to no local training data makes it a highly scalable tool for grid operators preparing for extreme weather events.