Find your setup

WattSizer research · Methods · 2015–2025 records

How we measured recorded outages—and where the evidence stops

Definitions, calendar attribution, coverage limits, sensitivity screens, input hashes and downloadable evidence for the WattSizer outage research series.

The unit and the calculation

Reported customer-hours = Σ(customers_out × 0.25 hours) over positive, finite integer snapshots. This is an approximation using the source’s nominal 15-minute cadence. Some timestamps are offset from quarter-hour boundaries; they are counted under the same approximation and exposed in the manifest. Irregular collection gaps are not interpolated. Missing, nonpositive and nonintegral values contribute nothing. We retain the distinction between no source record and a verified zero outage.

Calendar attribution and event definitions

Every snapshot contributes to its own UTC date, month and calendar year. Peaks retain their actual timestamp. The former six-hour county-event stitching rule is retired: persistent small outages in different neighborhoods can otherwise produce implausible month-long or year-long “events.” This series makes no event-count or longest-household-outage claims.

How this edition differs from the previous report

The earlier edition summed only snapshots inside its self-defined “significant events” and attributed cross-year events to their starting year. This edition includes all positive integer snapshots and assigns each to its actual date. Its totals therefore measure a different, explicitly stated quantity; they should not be compared with the withdrawn event-filtered series as evidence of a new trend. Legacy state and disaster URLs now use the corrected definitions or explain why the proposed calculation is unavailable.

Source windows and record counts

File yearSource rowsPositive integer rowsFirst timestamp UTCLast timestamp UTC
201512,939,15712,939,1572015-01-01 00:00:002015-12-31 00:00:00
201613,306,02413,306,0242016-01-01 00:00:002016-12-31 00:00:00
201715,078,36414,915,7162017-01-01 00:00:002017-12-31 00:00:00
201821,776,80619,748,9782018-01-01 00:00:002018-12-31 00:00:00
201924,074,12222,707,3512019-01-01 00:00:002019-12-31 00:00:00
202025,545,51723,443,9972020-01-01 00:00:002020-12-31 00:00:00
202124,826,10223,721,3642021-01-01 00:00:002021-12-31 00:00:00
202225,796,46524,515,9892022-01-01 00:00:002022-12-31 23:45:00
202326,101,05124,794,3812023-01-01 00:00:002023-12-31 23:45:00
202427,701,33426,616,1202024-01-01 00:00:002024-12-31 23:45:00
202530,482,14127,369,8162025-01-01 00:00:002025-12-31 23:45:00
A first and last timestamp do not prove continuous monitoring of every county or utility. Bounds are taken over positive records.

Coverage and valid comparisons

There are 2,311 FIPS codes with at least one positive record in every supplied annual file. The common-record sensitivity uses exactly that set. It does not establish fixed utility coverage or complete reporting. The raw count of counties with records is not a customer coverage percentage. County-equivalent boundaries and FIPS codes can change. Puerto Rico has no positive records before 2021 in these inputs; a Maria 2017 burden cannot be calculated from them.

Customer denominators and anomaly screens

MCC county identifiers are normalized to five characters before joining. MCC estimates modeled customer accounts as of 2022; they are not verified annual exposure counts. We calculate two sensitivity series: omit a snapshot when its value exceeds 1× or 2× the matched MCC estimate. Unmatched records remain in those series and are not treated as verified. Exceeding an estimate is a flag for review, not proof that an outage did not occur. For example, Cochise’s 2024 source peak is 718,642 customers at 2024-07-03 22:30:00 UTC, compared with 80,227 modeled MCC accounts. That disparity requires source investigation. We do not publish an unqualified national peak ranking.

The county-adjusted comparison

Counties with positive records in both 2024 and 2025, at least 10,000 modeled MCC customers, and no snapshot above that fixed modeled count in either year. The proxy is two-year customer-hours divided by modeled customer accounts and by two. A record without outages or a missing record does not confirm full monitoring. Passing the eligibility screen does not validate source geography or every remaining value. These selected rankings are not SAIDI, SAIFI, a household outage distribution or national reliability rankings.

Daily concentration and monthly patterns

Daily concentration ranks UTC-day customer-hour sums separately within each year. The ten highest-burden days need not be adjacent or independent. The screened concentration reranks dates after the screen. Monthly shares pool actual UTC months and are not corrected for month length or monitoring changes; they describe recorded burden, not a monthly hazard probability.

Reproduce and inspect

The build uses scripts/build_outage_analysis.py followed by scripts/build_outage_reports.py. The analysis records input SHA-256 hashes and source windows. Hashes identify the local inputs used; they do not certify that the files match an upstream release. Cache reuse requires matching input and MCC hashes. The manifest records the analysis version. Download the source manifest, annual totals, daily totals, county-year records, and the findings with denominators.

Sources and reproducibility

WattSizer calculations from the local EAGLE-I annual CSVs. Read the full methods and quality review. Original sources: Figshare historical data, Brelsford et al., Scientific Data (2024; 2014–2022 descriptor), and ORNL’s 2025 release. The descriptor explains the collection system; it does not describe observations from 2023–2025. The source manifest contains local input SHA-256 hashes, file sizes and timestamp bounds.