The Observatory
Compare

Are the windiest counties the ones losing power?

Wind gust across → · Outages per 1,000 residents up ↑

Gusts are the latest station reading and outages are now; a front that has already passed shows up as outages with calm wind. The storm response below follows this over time.

Outages per 1,000 residents →
0
0
0
0
0
0
0
0
0
Wind gust →
no datanot covered

Terciles of the counties that have both values; numbers are counties per cell.

Outages per 1,000 residents against Wind gust

One dot per county with both values. Dashed lines are the tercile breaks the map is coloured by.

No county has a value on both axes.

Rank correlation across counties

Too few counties with both values to say anything.

Over time, not across places

The map compares counties at one moment. The storm response follows each tracked county hour by hour and asks what happened to its outages in the day after a weather signal, which is the closer thing to cause and effect this data can show.

Storm response

All tracked counties, 90 days to 2026-11-18 · computed 7m ago · method v2 · region vintage 2026.5

chance outages rise within 24 h, by NWS alert

No active alert66% · 4,222 county-daysMinor80% · 48 county-daysModerate72% · 456 county-daysSevere80% · 185 county-daysExtremetoo few (1 county-days)

extra customers out per 10,000 residents at the 24 h peak, by observed gust

< 10 m/s (< 22 mph)17.1 · 4,112 county-days10–15 m/s (22–34 mph)26.9 · 928 county-days15–20 m/s (34–45 mph)15.6 · 290 county-days20–25 m/s (45–56 mph)59.1 · 17 county-days

chance outages rise within 24 h, by what was reported in the county that hour (NOAA storm reports)

No storm report62% · 102,249 county-daysHail only64% · 125 county-daysDamaging wind83% · 863 county-daysTornado90% · 30 county-days
How this is computed
alert class
Worst NWS alert severity active in the county at the hour. 0 = none, known only for hours the alerts feed ran or an archive backfill covered the county's state.
coverage gate
A window counts only if at least 12 of its 24 future hours have a power observation.
gust class
Observed wind gust at the hour; station measurement preferred over model, share measured reported per bin.
p increase
Share of windows in which customers out rose at all within 24 hours.
report class
What the NOAA Storm Prediction Center has on record in the county in that hour: none, hail only, damaging wind, or a tornado. A confirmed impact, not a forecast. Known only for hours the storm-reports feed covered.
response
Peak customers out over the next 24 hours minus customers out now, per 10,000 residents (raw count alongside).
uncertainty
95% intervals use distinct county-days as the sample size, not windows, because windows overlap.

By alert and gust: 59,672 county-hour windows over 4,912 county-days. By storm report: 2,060,316 windows over 103,267 county-days. Dropped: 323,455 for missing power data, 2,000,644 where the alerts feed did not run, 2,008,180 with no gust observation (those still count by alert). The dashed line is the no-alert or calmest-wind baseline.

Conditional averages, not a causal model. Counties without a tracked utility are absent, not zero, so the national pool over-represents the utilities this system scrapes. Storms hit neighbouring counties together, so even county-day intervals are narrower than the evidence supports. Severity is the NWS category, which mixes heat, flood and wind alerts. Storm reports exist only where someone was there to report: a storm over empty land is a 'no report' hour, which pulls that bin's response up.