)
500 Meters and 48 Hours: How Xcel Energy Rewired Wind Forecasting for Wildfire Decisions
When a utility de-energizes a circuit to prevent a wildfire, it is betting against the forecast. Get it wrong one way and you risk catastrophic ignition; get it wrong the other and thousands of customers lose power needlessly. For Xcel Energy, serving millions across Colorado's complex Front Range terrain, that bet was long constrained by regional weather models too coarse to tell a canyon wind event from conditions on the plains above it, with lead times too short for deliberate, defensible planning.
Working with NVIDIA and Accenture, Xcel Energy deployed NVIDIA's Earth2 downscaling stack, pairing a global ensemble model with CorrDiff, a generative AI model trained on high-resolution atmospheric data, and WindNinja, a sub-kilometer terrain-aware wind solver. The pipeline moves from today's ~27-kilometer global forecasts to 500-meter, hourly resolution. Validated against 246 weather stations and six EPA-defined terrain classes, it outperformed HRRR by roughly 25% on probabilistic accuracy at lead times two to three days longer, delivering 36 to 48 additional hours of warning.
That warning window is only useful if operators can act on it. WindVal, a visualization and post-event analysis platform built alongside the forecasting pipeline, allows meteorologists to interrogate model performance by station, terrain, and event, and gives operations a single view of forecast uncertainty during active and historical PSPS events.
Attendees will see how the Earth2 architecture was adapted for terrain-complex utility operations, how WindVal supports the decision process, and what it takes organizationally to act on a longer warning window.
