CASE STUDY - FORTUNE 500 UTILITY
Knowing where the grid will actually break.
A Fortune 500 utility's forecast of storm outages, up to six days ahead. NEXUS built a solution in one week that beat the incumbent vendor forecast on accuracy at every horizon - so crews and equipment are staged where the grid will actually break.
The Task
Predict storm outage counts
Decision Enabled
Crew & equipment staging
Accuracy vs. Incumbent
0.86 vs. 0.82, built in a week
Crews staged before the storm, not after
A Fortune 500 utility has to decide where to stage repair crews and materials before a storm arrives. It already relies on an outage forecast from an established vendor, delivered four times a day, up to six days ahead, company-wide and for each of its 13 service centers. The question: could NEXUS match that full scope and improve on its accuracy, quickly?
One week, the full forecasting scope
NEXUS's pre-trained foundation model identified outage signals in the utility's data, then was tested head to head against the forecast the utility uses today.
Same scope
Outages and customers out, company-wide and for each of 13 service centers, four times a day, same day to six days ahead.
Same rules
Only data available before each forecast was issued, so the backtest matches live use.
Same scoring
Scored on the same hold-out period, at the same issue times and horizons as the incumbent vendor forecast.
More accurate at every horizon, built in a week
Measured Outcome
0.86 vs. 0.82
company-wide outage forecast accuracy, issued at 18Z for the next day, with a solution built in one week
Measured Outcomes
0.79 vs. 0.74
accuracy six days out, where NEXUS's lead is widest (averaged across issue times). NEXUS holds close to its same-day accuracy of 0.85 across the whole week, giving planners a more dependable forecast further ahead
One week from raw data to a forecast ahead of the incumbent, on the same days, issue times and metric. NEXUS followed the actual outage curve more closely, month after month from February to July.
Accuracy is 1 minus the average percentage error; higher is better.
One week to value, with five more to follow.
The pipeline runs end to end - weather download, point-in-time checks, NEXUS, scoring - and can deliver the same daily forecast file at the four issue times, alongside the current one. Each new use case swaps the target and keeps the rest: restoration time per outage, which circuits fail next, substation peak load, daily gas demand, equipment failure risk.
