Call the repair crews early, or wait for a better forecast?
Before a big storm, power companies borrow repair crews from other utilities. Ask early and you get more crews but trust a fuzzier forecast; ask late and the forecast is sharper but the crews may be gone. This study puts numbers on that trade-off using AI weather forecasts for 2,491 US counties.
Why it matters
When a big storm is coming, power companies must decide days ahead whether to call in repair crews from other utilities, and calling early locks in more crews but relies on a shakier forecast. Using GraphCast and FourCastNet forecasts for 2,491 US counties, this study shows that committing five days out catches significantly less of the outage burden than committing three days out (0.238 vs 0.316 of it), and that committing three days out with every requested crew beats waiting until one day out unless a utility can still get more than about 83 % of them — one number it can check against its own records.
The paper
Forecast lead time and the timing of crew deployment for weather-driven power outages using AI weather models
Mohammad Kian Golkar, Luciano Alves de Oliveira & Mohamad Alipour (University of Illinois Urbana-Champaign). arXiv:2610.20918 [physics.ao-ph; cs.AI], submitted 14 Oct 2026. Read on arXiv · PDF
In plain language: the authors joined three years (2022–2024) of county-by-county power-outage records with archived forecasts from two AI weather models, GraphCast and FourCastNet, made 1, 2, 3 and 5 days ahead. A tree-based machine-learning model turns each forecast into a chance of a serious outage in every county. Then they asked a utility’s question: if I had sent crews to the 10 % of counties the model ranked riskiest, what share of the real outage burden would those counties have held? That share — the “captured burden” — falls as the forecast is made further ahead: 0.356 at 24 hours, 0.327 at 48, 0.316 at 72 and 0.238 at 120 hours (a perfect forecast would reach 0.893).
The big drop is between three and five days. Because waiting also means fewer crews are left, they work out the break-even point: committing at 72 hours with all crews matches committing at 24 hours with 83.4 % of them (95 % range 63.3–100 %). They also uncover a quiet data trap — soil-wetness-type predictors measured at the time the forecast is for, instead of the time it is made, secretly contain the storm’s own rain and inflate skill by 42–53 % — and show that the model’s 90 % uncertainty ranges covered only 47.6 % of hurricane events in 2024, a year with three major US landfalls after a calibration year with none.
Try it
Panels 1, 3 and 4 use numbers published in the paper (Tables 5–8, Section 3.6). Between the paper’s crew-availability points (40, 60, 80, 100 %), panel 1 joins values with straight lines — that interpolation is ours. Panel 2 is an illustrative toy storm on a made-up grid of counties that applies the paper’s staging rule (Eq. 8); its track errors and noise are invented for teaching, not taken from the paper.
1 · Commit early with every crew, or wait and get fewer?
2 · Where would the crews go? (illustrative toy storm)
3 · Is staging crews worth the money? (paper Eqs. 11–12, 72-hour commitment)
4 · The hidden-rain trap (temporal leakage)
| Occurrence skill (AUC-PR, 2024) | 24 h | 120 h |
|---|---|---|
| Land surface alone, joined at valid time (leaky) | 0.0775 | 0.0779 |
| Land surface alone, joined at issue time (correct) | 0.0452 | 0.0369 |
| Forecast meteorology alone | 0.1027 | 0.0625 |
| County climatology reference | 0.047 | 0.047 |
With the leaky join, land-surface predictors appear to beat the weather forecast at 120 h (0.0779 vs 0.0625); joined correctly they fall below the county climatology reference.
The maths in this page
Two-part (hurdle) outage model. Severity s = customers out ÷ housing units. Expected severity is split into the chance of an outage event (s ≥ 0.05 for at least 24 h) and its size if it happens, each fitted with LightGBM on 66 predictors:
Captured burden (Eq. 8). Rank counties by π, stage crews in the top K, and measure the share of observed severity they hold:
Fewer crews and break-even (Eq. 9). Getting only a fraction ρ of crews means staging the top ρK counties. The break-even ratio solves
Panel 1 solves the same equation by straight-line interpolation between the published table points, which gives about 83.3 % for 24 h vs 72 h.
Money (Eqs. 11–12). Avoided customer-interruption cost and total staging commitment:
with Hs = 178.8 million customer-hours in counties staged at 72 h, Σ Kt = 11,570 county stagings over 72 high-burden windows and d = 3 days, so D = 34,710 county-days (for m = 1). Break-even cost = AC ⁄ D. The 72 h vs 120 h difference is 72.5 million customer-hours. Central case in the paper (β = 1.2 kW, η = 0.25, VoLL = 1,000 USD/MWh): 53.7 million USD avoided, 21.7 million USD for the timing choice, 1,546 USD per county-day. The calculator recomputes these from the rounded published inputs, so it can differ in the last digit (e.g. $1,545).
Leakage check. A 72 h rain total ending at valid time t and one ending at issue time t − L share max(0, 72 − L) of their 72 hours; at L ≥ 72 h they share none, and the paper’s measured correlation drops from 0.782 (24 h) to 0.053 (120 h).
Toy map (illustrative). Panel 2 builds a fake storm, gives each county a true severity, then makes a “forecast” by shifting the track sideways by an amount that grows with lead time (about 0.9 squares per day, invented) and adding noise. It then applies Eq. 8 exactly. It only shows the idea; real numbers come from panels 1, 3 and 4.
Teaching demo built on published results — not the authors’ code (archived at Zenodo, doi:10.5281/zenodo.23326163) and not operational guidance.
How to use it
In panel 1, pick when you would commit early (48, 72 or 120 hours ahead) and when you would otherwise wait, then slide the share of crews you would still get by waiting. The dashed line is what committing early with every crew captures; the dot is what waiting captures. Where the waiting line crosses the dashed line is the break-even crew share — for 72 h vs 24 h it is about 83 %, so unless more than about 83 % of crews would still be free, early wins. The note under the stats says whether the paper found the gap between those two lead times statistically significant. In panel 2, drag the lead time from 24 to 120 hours and watch the green squares slide off the storm track and red “missed” squares appear; press New storm for another try. In panel 3, set your own value of lost load, how much faster pre-staged crews restore power, and crews per county to see the most a utility could pay per county-day and still come out ahead. In panel 4, step through lead times to see why a rain total that ends at the moment being forecast secretly includes the storm itself.