Privacy and preparation
Energy operational dataset seasonal gaps: an extreme weather audit
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
Assess coverage by treating extreme weather as a stress test. Date every cold snap, heat wave, ice storm and high-wind event in your territory from the public storm record, then check whether the archive holds the exceptions those days produce: estimated and corrected reads, rebills, emergency work orders, outage tickets and approvals. A dataset whose storm days are thin, estimated or missing has a seasonal gap, whatever its date range says.
Checklist
Energy operational dataset
Key takeaways
- A complete date range can hide a seasonal gap: storm-day reads may be mostly estimates and exception records may never have been exported.
- Date each extreme condition from a public event record or your own operating triggers, and compare it with a mild control period in the same season.
- Check reading quality codes, correction trails, exception records, major event day flags and the gap between event time and posting time.
- Keep major event days as flagged rows; reliability views that exclude them remove the very days a storm-ready model needs.
- Document gaps you cannot fill, including deliberate safety and confidentiality exclusions, instead of patching them with synthetic or normalized values.
How do you assess dataset coverage using extreme weather cases?
Start from the weather, not from the data. List the severe conditions your service territory actually experienced over the period the dataset claims to cover, give each one a date range and a geography, and then ask a narrow question of every condition: on those days, did the operational records that normally follow such an event get created, kept and exported at the same fidelity as on an ordinary day?
That question exposes seasonal gaps that a records-per-month chart hides. A billing archive can show a steady row count through February while most of the readings in a cold week are estimates, the outage tickets sit in a different system that was never exported, and the corrections arrived weeks later in a batch that the extract cut off. The date range is complete; the coverage of the hardest days is not.
The output of the audit is a short coverage statement for each condition: present and complete, present but degraded, or absent, with the evidence behind each verdict. That statement is what a model team, a reviewer or a licensing counterparty needs before they rely on the archive for anything that must work in bad weather.
Why do energy archives thin out on the worst days?
Severe weather changes how records are made at the exact moment the records matter most. Meters cannot be read or fail to report, so billing systems substitute estimates. Crews work from radio, phone and paper during the event and post work orders afterward. Call volume spikes, and contact notes get shorter. Some of those effects are visible in the data if you know where to look; others leave no trace at all.
Meter data is the clearest case. In the Energy Services Provider Interface schema published in the Green Button Alliance's open-source implementation, each interval reading can carry quality codes that separate validated data from values replaced or approved by a human, values estimated from historical data using a reference day, values computed by linear interpolation, data that failed one or more checks, and values normalized for weather. A storm week heavy with reference-day estimates looks complete by row count and tells a model almost nothing about real consumption under stress.
Reliability reporting adds a second, deliberate thinning. Under the IEEE 1366 method that EIA describes for Form EIA-861, a major event day is a day whose distribution SAIDI exceeds a threshold computed from five years of daily values, and utilities report reliability both including and excluding those days. EIA's Electric Power Annual publishes the series with and without major event days side by side. Any dataset or metric built on the excluding-major-events view has removed the extreme days by design, which is correct for benchmarking and wrong for a model that has to handle storms.
Which weather conditions belong on the audit list?
Use a public, dated record of events for your territory rather than memory. NOAA's Storm Events Database, maintained by NCEI, is one such record: its bulk files give each event a type, begin and end date-times usually in local standard time, a time zone field, and a code saying whether the event was recorded for a county, a National Weather Service forecast zone or a marine area. NCEI notes that new events take about three months to appear, so the most recent quarter of any archive cannot yet be audited this way.
Add the operating thresholds your own company already uses, such as the temperature at which cold-weather procedures start or the heat index at which load forecasts are overridden, because those are the days your records behave differently even when no storm was declared. The table below maps condition types to the operational exceptions they usually produce and the records that should exist afterward. The pairings are a starting point for a domain reviewer, not a rule.
| Condition | How to date it | Exceptions it usually produces | Records that should exist afterward |
|---|---|---|---|
| Extreme cold or ice storm | Storm event type and dates for your counties or zones, plus your own cold-weather trigger | Missed and estimated reads, frozen-equipment work orders, demand spikes, curtailment or load relief requests | Reading revisions with quality codes, emergency work orders, curtailment notices, rebill adjustments |
| Extended heat | Storm event type and dates, plus your heat or peak-demand trigger | Peak-demand events, equipment overload orders, demand response dispatches | Interval data on peak days, dispatch logs, transformer replacement orders |
| High wind or thunderstorm | County-coded event dates and times | Distribution outages, restoration crews, damage claims | Outage tickets with start and restore times, crew assignments, claim records |
| Flooding | Zone or county event dates | Site access limits, inundated equipment, service disconnections | Access exceptions, equipment condition records, reconnection orders |
| Wildfire smoke or fire weather | Event dates plus any public safety shutoff periods your company declared | Planned de-energization, customer notifications, delayed field work | Shutoff and restoration timestamps, notification logs, deferred work orders |
| Mild control period | Days with no event in the same months | Ordinary operations | The baseline every other row is compared against |
What should the condition and exception coverage audit include?
The audit is a fixed set of checks run once per condition and once for a matched mild control period in the same season, so that a gap is measured against the archive's own ordinary days rather than against an ideal. Version it (for example v1.0, dated 2026-10-18) and keep the results with the dataset so later exports can be compared with the same checks.
At a glance
What should the condition and exception coverage audit include
Each line names what to measure, what evidence passes and what signals a gap. A domain reviewer should adjust thresholds for the systems in scope; the point is that every verdict is written down with its evidence.
| Audit line | What to measure | Evidence that passes | Gap signal |
|---|---|---|---|
| Event calendar | Conditions listed for the territory and period, with dates and geography | Every condition traced to a public event record or an internal trigger log | Conditions chosen from memory or only the famous storms |
| Record volume on event days | Records per day for each system against the mild control period | Volume moves the way operations did, for example more work orders | Flat or falling volume during a known event |
| Reading quality mix | Share of reads carrying estimated, interpolated, failed or weather-normalized quality codes | Estimates present but each later superseded by an actual or validated value | Estimates never corrected, or quality codes dropped in the export |
| Correction trail | Revisions and rebills dated after the event that refer to event-day reads | Full revision history with timestamps, linked to the original bill | Only final values exported, or corrections outside the extract window |
| Exception records | Emergency work orders, outage tickets, curtailment notices, claims | Each type present for each condition, with open and close times | A whole exception type missing because it lived in another system |
| Major event day flags | Whether event days are marked and whether any view excludes them | Flags kept as a field, raw rows retained | Event days silently removed by a reporting filter |
| Timestamps and time zones | Local versus UTC, daylight-saving handling, posting date versus event date | Event time and posting time kept as separate fields | Backlogged entries stamped with the day they were typed in |
| Linked context | Tariff, contract term and asset reference in force on the event date | Dated reference fields join without guesswork | Current tariff or asset record applied to historic days |
| Excluded content | What was removed for safety or infrastructure confidentiality | A list of removed classes with the reviewer who approved each | Removals undocumented, so absence looks like a gap |
Which dates and boundaries decide what an event-day record may be used for?
On event days, the correct answer to an administrative task usually depends on a date: which reading is final, which tariff or contract term was in force, when a work order was opened against when it was posted. Keep those dates, or a model trained on storm records learns to settle bills against facts that did not apply yet.
Extreme days also sit outside the assumptions equipment and processes were designed for. The joint FERC, NERC and Regional Entity review of the February 2021 cold weather event found that 81 percent of freeze-related generating unit outages occurred at temperatures above the units' stated ambient design temperature, and the event included 23,418 megawatts of manual firm load shed. Records from days like that contain exceptions an ordinary winter never produces, which is why their absence from a dataset matters.
For generator owners, NERC's current cold weather standard, EOP-012-3, creates its own dated records: a generator owner either implements freeze protection measures for its generator cold weather critical components or develops a corrective action plan, and keeps dated evidence of which it did. Such records can be useful context for an administrative coverage audit, while the operating detail inside them belongs to the safety boundary.
That boundary is firm. Switching orders, load shed sequences, isolation steps and anything that instructs a crew stay out of the training set, or are reduced to a masked reference that a procedure applied. Diagrams, settings and asset vulnerabilities are a separate infrastructure confidentiality question for the company's compliance and security staff. Record each removal in the audit's excluded-content line so that a deliberate exclusion is never mistaken for missing coverage.
Worked example (illustrative): one ice storm and one heat wave in a billing and work order archive
This example is fictional; the company, dates and figures are illustrative. A regional distribution company offers five years of billing, interval meter and work order records and describes the archive as covering all seasons.
The event calendar finds two conditions worth auditing in the counties served: a three-day ice storm in the second winter and a nine-day heat wave in the fourth summer. Each is compared with a mild control period of the same length in the same month of an adjacent year.
During the ice storm, interval records keep arriving at the normal daily volume, but about 60 percent of event-day reads carry a reference-day estimate code, against about 4 percent in the control period. Only about half of those estimates have a later validated value in the export, because the extract was cut at the end of the quarter and the corrections were posted after it. Emergency work orders appear, but most were posted two to five days late and carry the posting date in the only date field.
The heat wave passes most checks. Interval data is mostly validated, peak-day volume rises as expected and demand response dispatch logs exist. One gap remains: transformer replacement orders were tracked in a separate asset system that is not in the export.
The coverage statement therefore reads: extreme heat present and largely complete, with one missing record type; ice storm present but degraded, with corrections and true event times missing. The remedy is concrete: extend the extract window by a quarter, add the work order creation timestamp, and either add the asset system or narrow the claim to exclude transformer work.
When thin storm coverage is the honest answer
Sometimes the audit finds a gap that should be documented, not filled. Three cases come up often, and each calls for a written explanation rather than more data.
First, the territory never experienced the condition during the period. A five-year archive from a mild coastal area may hold no ice storm at all; the honest statement is that the condition is out of scope. Second, the records were removed on purpose, for safety or infrastructure confidentiality, and the excluded-content line shows who decided. Third, the records were lawfully destroyed under a retention schedule before anyone thought to keep them.
In none of these cases should the gap be patched with synthetic storm records presented as real, or with weather-normalized values offered as observed consumption. The Datasheets for Datasets framework asks dataset creators whether a dataset is a sample of a larger set, whether it is representative and how that was checked, and whether information is missing from individual instances and why. A coverage statement that names a missing condition answers those questions; a filled-in gap hides the answer.
Narrowing the coverage claim before anyone relies on it
Some findings mean the dataset description should get smaller, or a specialist should look before it is used.
- Narrow the claim to the conditions that passed when one or more conditions are degraded and cannot be repaired from source systems.
- Drop the phrase all seasons or all weather from any description when a single named condition is absent or degraded.
- Bring in an energy operations reviewer when event-day records mix administrative facts with operating instructions that need screening.
- Ask the compliance or security owner before including any outage, asset or location detail tied to a severe event.
- Ask counsel when event-day records include customer communications, regulator filings or third-party contractor data.
- Pause and re-run the audit after any change to the extract window, the systems included or the exclusion rules.
Describing weather coverage when licensing energy records
An audited coverage statement is useful in any licensing conversation, because it says plainly which hard conditions the archive represents and which it does not. Rights are a separate question: access to the systems, customer privacy permissions, ownership of the records and contract terms with vendors are four distinct checks, and removing customer identities does not by itself create a right to license.
Energy companies can license selected operational records directly to SourceX, which handles rights review, contracting, preparation and delivery. As a general guide, buyers look for several years of consistent history in one system, which is exactly where a weather audit shows whether those years hold up on the worst days. An energy company keeps control of that step: it approves exactly which records and terms are in scope, and signs, before delivery begins.
Frequently asked questions
What is a seasonal gap in an energy operational dataset?
It is a stretch of the year, or a type of weather condition, where the dataset's records are missing, thinner or lower in quality than on ordinary days. The date range can look complete while storm-day reads are mostly estimates or exception records were never exported. A condition-by-condition audit is the way to find it.
Should major event days be removed before training a model?
Not by default. Reliability reporting separates major event days so utilities can be benchmarked, but a model that must handle storms needs those days. Keep the rows and carry a major event day flag as a field, so each use can include or exclude them on purpose.
Can estimated or weather-normalized meter values be used as labels?
Use them as labeled estimates, not as observed consumption. Keep the quality code with each value and the later validated reading when one exists. Tasks about billing corrections need both; tasks about real demand should use validated values only.
How many years of history does an extreme weather coverage audit need?
Enough for each condition you claim to appear at least once, ideally several times, with a mild control period to compare against. A short archive may simply not contain a rare condition, and the right response is to state that rather than stretch the claim.
Does matching public storm records to company data change who controls the data?
No. The public event record is used only to date and locate conditions. Rights in the company's own records still depend on ownership, customer permissions and contract terms, which are reviewed separately.
Sources
- NOAA NCEI's Storm Events bulk file format gives each event a type, begin and end date-times usually in local standard time, a time zone field and a county, zone or marine area code. Source
- NCEI's Storm Events Database notes that newly reported events take about three months to appear in the database. Source
- EIA's Form EIA-861 guidance describes the IEEE major event day threshold, computed from five years of daily SAIDI values, for reporting reliability with and without major events. Source
- EIA's Electric Power Annual publishes SAIDI and SAIFI both with and without major event days. Source
- FERC's release on the final February 2021 cold weather report says 81 percent of freeze-related generating unit outages occurred above the units' stated ambient design temperature. Source
- The FERC, NERC and Regional Entity staff report on the February 2021 event records 23,418 MW of manual firm load shed, the largest controlled firm load shed event in U.S. history. Source
- NERC's EOP-012-3 cold weather standard requires a generator owner to implement freeze protection measures for generator cold weather critical components or develop a corrective action plan, with dated evidence of either. Source
- The ESPI schema lets interval readings carry quality codes for validated, human-replaced, reference-day estimated, interpolated, failed and weather-normalized values. Source
- Datasheets for Datasets asks whether a dataset is a sample and representative of a larger set, how that was checked, and whether information is missing from instances and why. Source
See if your company qualifies
A short company assessment. No data uploads are needed.