Methodology
How every number is made
Every figure on this site is computed from public federal records. Nothing is estimated by hand, and no text is written about a place beyond sentences filled in from its own numbers.
The outage record
Outages come from EAGLE-I, the Environment for Analysis of Geo-Located Energy Information at Oak Ridge National Laboratory. Every 15 minutes since late 2014 it has read the public outage maps of US utilities and recorded how many customers were without power in each county. This site uses version 4 of the published record (doi:10.6084/m9.figshare.24237376.v4), 232,535,445 county readings from 2015 to 2025.
The record covers about 92% of US electricity customers by 2022, up from much less in 2015. Utilities without a public outage map, many small co-operatives and city utilities among them, are missing. Outages shorter than 15 minutes are mostly missed. A county with no reading for a run may have had no outage or may not have been collected; the record does not say which. All of this means the figures here are a floor, not a full count.
Cleaning
- Broken runs. A collection run with more than three times the day's usual number of rows is dropped whole (9 days have thin collection).
- One-run spikes. A reading of 500 or more customers that is more than four times both neighbouring readings is replaced by the larger neighbour. Real outages of that size last longer than 15 minutes.
- Short plateaus. A reading more than three times, and at least 20,000 above, the county's median over the surrounding twelve hours is replaced by that median. This removes collection glitches such as Los Angeles County reading 1.7 million customers out for two hours on 8 January 2025 between readings near 230,000.
- Ceiling. Customers out never exceed the county's customer count.
- Changed county codes (Alaska's Valdez–Cordova split, Kusilvak, Oglala Lakota, Bedford city) are mapped to the 2020 geography. Connecticut uses its eight legacy counties, as the record does; 2025 readings under the new planning regions are left out.
Customers
Each county's customer count comes from ORNL's modelled county customers (2022). Where that model is implausible (more than a factor of two away from a population-based estimate), the 2024 EAGLE-I customer count is used, and failing that the estimate itself: Census 2020 population times the national median of customers per person. Each county page says which source it uses.
Definitions
| Measure | How it is computed |
|---|---|
| Customer-hours | Customers out in each 15-minute reading × 0.25 hours, summed. |
| Hours without power | A county-year's customer-hours divided by its customers: the time the average customer spent without power. The headline figure averages well-recorded years from 2018, the first year with ORNL data-quality records. |
| Well-recorded year | A county-year with outage readings on at least 60 different days. |
| Outage event | Consecutive readings with at least 1% of the county's customers (and at least 25) out, joining gaps shorter than two hours. |
| Major outage | An event that peaks at 10% or more of the county's customers, or 50,000 customers. |
| Time to restore | Hours from an event's peak until customers out first fall to half (50%) and a tenth (90% restored) of the peak. |
| National outage | Hours with more than 600,000 customers out across the country, joining gaps shorter than twelve hours. The everyday background is 30,000–60,000. |
| State outage | Readings with at least 2% of a state's customers, or 25,000, out at once, joining gaps under six hours. |
Weather
Each outage is matched to NOAA Storm Events reports for the same county from twelve hours before the outage began to two hours after its peak. Reports filed by forecast zone are mapped to counties with the National Weather Service zone–county correlation file. When several kinds of weather match, tropical storms count most, then ice, wildfire, winter storms, tornadoes, wind and thunderstorms. Any National Hurricane Center best-track position within 350 km of the county's centre between a day before the outage and six hours after its peak names the cyclone. A match shows that the weather and the outage happened together, not that one caused the other.
Utilities
Utility reliability, prices and territories come from the Energy Information Administration's Form EIA-861, 2018–2025. SAIDI is the average minutes without power per customer a year; SAIFI the average number of interruptions. Both are reported by the utility for its whole territory in a state, with and without major event days, under IEEE Standard 1366 where the utility uses it. Price is residential revenue divided by residential sales for full-service customers.
Live readings
The live pages are built differently from the record. Every 15 minutes a collector reads ORNL's ODIN feed and the public files behind utilities' outage maps, adds customers out by county, and stores the reading. A utility read from its own map is not counted again from ODIN. County counts are capped at the county's customers, and a single reading more than four times both its neighbours (and at least 500) is replaced by the larger neighbour in replays. Large live outages use the record's definitions above. Coverage is narrower than EAGLE-I's, so live totals run lower than the record will for the same period; when EAGLE-I publishes the year, the record replaces the live readings. Sources and how they are read.
The light maps
Each light is a 2020 Census block-group centre of population; nearby centres are merged at the map's scale. During a replay, a county's share of customers out decides what share of its lights go dark. Outages are recorded by county, so which lights within a county go dark is illustrative.
Known limits
- Coverage grows over time, so early years undercount and trends before 2018 should not be read as change.
- Outage maps are sometimes down during the storms that matter most. Houston's CenterPoint map was offline as Hurricane Beryl came ashore in July 2024, and the record misses the first hours.
- Polygon-based utility maps are split across counties by area, not population.
- Nothing here is advice. The pages describe what was recorded.