Machine learning model cuts power outage time predictions down by over a third

The problem with outage estimates#

When electrical grids fail during storms, utilities issue Estimated Times of Restoration (ETRs) to guide customer decisions on everything from food storage to medical device management. However, these predictions have historically been unreliable, treated as a single one-off forecast that ignores the wealth of information accumulated as crews assess damage, dispatch workers, and update estimates in real time.

A new approach to predicting restoration time#

Researchers reformulated the problem using Longitudinal Tabular Transformers (LTT), a neural network architecture designed to process sequential updates rather than static snapshots. The model consumes each revision to an outage estimate—from crew assignment through damage assessment and partial restoration—and refines its prediction at every step. This mirrors how outages actually evolve in the field, where fresh information flows constantly.

Real-world performance gains#

Testing on 242,928 outages across six utility companies, tracked through 10 million total revisions, LTT delivered substantial improvements:

• 36.9% median reduction in customer-weighted asymmetric error versus the utilities' published estimates during the same storms
• 11.3% improvement against the strongest conventional machine learning baseline at each company
• Monotonic error decline as revisions accumulate—the model's predictions improve with each update, but struggle most at the first revision when no history exists

Critically, LTT was the only method to improve both customer satisfaction metrics and reduce root mean squared error across all six operators, suggesting the gains are robust rather than tuned to a single measure.

Why it matters#

Accurate outage forecasts reduce customer panic, unnecessary trips to hospitals or shelters, and spoiled food. This work shows that applying modern deep learning to the temporal dimension of outage data—rather than discarding revision history—can unlock substantial real-world value in grid resilience and customer trust.