The U.S. outage records for 2025 add up to 52.8% fewer reported customer-hours than the records for 2024. How much of that gap comes from a handful of severe days? Drop each year’s ten highest-burden days and the difference is still 41.3%.
That check took a few lines of Python. Deciding what to sum took more care. The data behind this comparison cover 11 calendar years, 2015–2025, and about 247.6 million source rows. Each row is a county-level observation, with a timestamp and a count of customers reported out at that moment.
What a snapshot can measure
Suppose a county reports 400 customers out at 08:00 UTC and 600 at 08:15. At the source’s nominal 15-minute cadence, those samples contribute (400 + 600) × 0.25 = 250 customer-hours. The second snapshot can represent a different set of customers. The data lack the customer IDs needed to reconstruct individual outage durations.
An earlier version of the analysis stitched county snapshots together when gaps were short. That could turn a moving set of small outages into one apparently long event. It could also put an entire event into the year it started. The revised pipeline assigns every valid snapshot to its own UTC day and year. It sums positive integer counts, uses 0.25 hours as a nominal weight, and leaves collection gaps unfilled. The manifest tracks off-grid timestamps and input hashes. A missing row leaves coverage unknown.
The build reads each CSV in 500,000-row chunks and writes daily and county-year summaries. When validating a published number, the small annual CSV is enough. Use the source archive and manifest if you need to audit the raw inputs.
How much does the 2024–2025 comparison move?
| Calculation | 2024 customer-hours | 2025 customer-hours | Change |
|---|---|---|---|
| All valid snapshots | 1,891.7 million | 892.7 million | −52.8% |
| Exclude snapshots above the modeled county customer count where a match exists | 1,664.9 million | 859.6 million | −48.4% |
| Counties with positive records in all 11 years | 1,606.7 million | 687.3 million | −57.2% |
| Exclude each year’s 10 highest-burden days | 1,311.6 million | 770.4 million | −41.3% |
The top ten days account for 30.7% of the recorded total in 2024 and 13.7% in 2025. Removing them narrows the gap. The county screen is another sensitivity check. Its denominator is a modeled 2022 customer count, so a flagged row merits inspection. The 2,311-county subset checks whether the result holds within the same group of counties across years. A positive record in every year still leaves coverage within each year uncertain.
Reproduce the table with Python
Download the annual CSV beside your script and run it with Python 3. The numbers below come from the published aggregates.
import csv from decimal import Decimal with open("annual.csv", encoding="utf-8-sig", newline="") as source: rows = {int(row["year"]): row for row in csv.DictReader(source)} def amount(year, column): return Decimal(rows[year][column]) def compare(label, before, after): change = 100 * (after / before - 1) print(f"{label}: {before:,.0f} -> {after:,.0f} ({change:+.1f}%)") for column, label in ( ("ch", "All valid snapshots"), ("screen1", "County customer screen"), ("common_ch", "Counties with records every year"), ): compare(label, amount(2024, column), amount(2025, column)) def without_top_days(year): return amount(year, "ch") - amount(year, "top10_ch") compare("Without the top ten days", without_top_days(2024), without_top_days(2025)) All four comparisons point in the same direction. The size of the decline depends on how you handle coverage, high observations and concentrated days. A reliability claim would need consistent reporting coverage or an independent benchmark. The most useful thing to publish alongside the headline is the rule for turning a single row into a number a reader can reproduce.
Data and methods. I publish the WattSizer outage methods and source manifest. The EAGLE-I data descriptor covers 2014–2022. The later files are listed separately in the manifest. AI was used to draft this write-up from the published analysis and check the reproducible example.
