Skip to content

Systems and records

Utility outage communication dataset: keeping changing estimates consistent

By SourceX Editorial · Reviewed by Noah Loul ·

8 sections

Short answer

A revised restoration estimate should rest on a dated operational change the message can name: a crew assessment, new damage, a switching step, a change in the affected area, or a safety hold. A useful utility outage communication dataset links every customer update to that event record and to the message it replaces, so an evaluator can check that each update was current, consistent across channels and explained.

System export

Utility outage communication

This page covers utility outage communication dataset: keeping changing estimates consistent.Illustrative

Key takeaways

  • Change a customer's restoration estimate only when a dated event changes the facts; otherwise reconfirm before the old estimate expires.
  • Link every message to an event version and to the message it supersedes, as alerting standards already do.
  • Check chain integrity, explained changes, channel agreement, scope, lapses, restoration truth and safety wording separately.
  • Grade drafted updates against the event snapshot at send time, never against later facts such as the final restore time.
  • Remove grid engineering detail and customer contact data before any outage history is reused.

What evidence should support changing a customer's restoration estimate?

A changed estimate should rest on a dated operational fact the message can name: a crew's damage assessment, a newly found failure, a completed switching step, a change in the affected area, or a weather or safety hold. If nothing in the event record changed, the estimate should not change either; the customer gets a reconfirmation before the old estimate expires, not a new guess.

This matters because customers act on these numbers. In its interim investigation of Tropical Storm Isaias, New York's Department of Public Service staff wrote that customers depend on estimated times of restoration to make health and safety decisions, including whether they need somewhere else to stay. Staff faulted a utility for letting estimates expire without updating them and said a status such as More Work Required is not an estimate at all.

An outage communication dataset is therefore two linked histories: the operational events from the outage management system and crew logs, and every customer-facing message across text, phone, web map and call center scripts. The value is in the link. With it, an evaluator can check whether each message was newer, consistent with the event state at the time and explained; without it, the messages are just text.

Which outage events justify a new estimate, and which only justify a reminder?

Treat estimate changes as event-driven. The table shows common outage events, whether each should move the estimate, the record that proves it and what the next message should say. Wording and thresholds will differ by utility; the structure is what matters.

Two habits cause most unexplained changes. One is recalculating an estimate on a fixed timer, so the number drifts while nothing on the ground has changed, which customers read as confusion. The other is pushing one system-wide estimate to every customer when only part of the area has been assessed. Recording the basis and the scope of every estimate makes both habits visible and gives a reviewer one question per change: which event moved it?

Which outage events justify a new estimate, and which only justify a reminder?
EventMoves the estimate?Evidence recordWhat the next update should say
First crew assessmentUsually, from a system-level default to a local estimateCrew arrival and assessment entry with timeCrew has assessed the cause; new estimate and its basis
Additional damage foundYes, laterDamage report naming the equipment type, not its engineering detailRepairs are larger than first assessed; revised estimate
Partial restoration by switchingYes for restored customers; maybe not for othersSwitching step completion and customer count by deviceRestored for some addresses; remaining area keeps its estimate
Nested outage discoveredYes, for a smaller groupChild outage linked to the parent eventYour address has a separate problem; separate estimate
Weather or safety holdPauses itHold entry with reason and review timeWork is paused for safety; next update time
Estimate about to expire with no new factsNoEvent state unchanged since last messageReconfirmation with the same estimate and next update time
Customer reports still out after restoreReopens itCallback or new call linked to the restored eventWe are checking your address; new estimate if needed

How should each customer update point back to the one it replaces?

Give every message an identifier, the event version it was based on and the identifier of the message it supersedes. Emergency alerting already works this way. The OASIS Common Alerting Protocol defines an Update message as one that updates and supersedes earlier messages named in its references field, a Cancel as one that cancels them, and specifies each reference as the sender, identifier and sent time of the earlier message.

Customer outage notifications rarely carry that structure in their exports, so the dataset has to reconstruct it. The field set below is enough to rebuild the chain for each customer or service area.

How should each customer update point back to the one it replaces?
FieldMeaningIllustrative valueWhy it is needed
message_idUnique identifier of the customer messagemsg-000412Lets later messages reference this one
supersedesMessage identifiers this one replacesmsg-000377Rebuilds the update chain
event_id and event_versionOutage event and the version of its record usedevt-88 v3Ties the message to the facts at that moment
scopeService area, device or customer group addressedgroup-B on tap 2Checks the right customers got the message
etr_value and etr_basisEstimate and the event that justified it+5 h; additional_damageMakes every change explainable
channel and sent_timeText, phone, web map or script, with timestampsms, 14:05Checks channels agree at the same moment
expires_atWhen the estimate stops being valid18:00Detects estimates left to lapse

Which checks belong in the event and message consistency brief?

The brief is a set of checks run over the linked event and message histories, either to audit past communications or to grade a model that drafts them. Each check has an input, a pass condition and a typical failure. Teams should agree the thresholds per check before running it.

Run the checks first on past storms whose outcome is known, to see which checks fail often and whether those failures are real. A check that fails on nearly every message usually points to a missing field, such as an absent supersedes link in the export, rather than to poor communication, and the fix belongs in data preparation.

Which checks belong in the event and message consistency brief?
CheckQuestionInputsPass conditionTypical failure
C1 Chain integrityDoes each update supersede the right earlier message?message_id, supersedes, sent_timeNo orphan or circular referencesTwo live estimates for the same customer
C2 Explained changeIs every changed estimate backed by an event?etr_basis, event versionsA new event version precedes each changeEstimate moves with no new facts
C3 Channel agreementDo text, phone, map and script agree?All messages in a time windowSame estimate and status across channelsMap says restored, text says pending
C4 Scope matchDid the message reach the affected customers only?scope, outage customer listRecipients match the event's customer setRestored customers told work continues
C5 No lapseWas any estimate left to expire?expires_at, next message timeA new or reconfirmed message before expiryEstimate passes with no update
C6 Restoration truthDoes a restored message match a restore event?Restore events, callbacksRestore event exists and no reopen followsRestored message, then customer still out
C7 Approved safety wordingIs safety content taken from the approved script?Message text, script libraryExact or approved variantModel paraphrases downed-wire guidance

How should a model that drafts outage updates be scored?

Freeze the event record at the moment each message was due, give the model only that snapshot, and grade the draft against the same snapshot. Using later facts, such as the final restore time, to judge an earlier message rewards hindsight, not consistency.

Research on restoration time prediction already treats an outage as a sequence of updates. One study models each outage as a series of record updates and proposes customer-informed regression metrics, and an earlier paper revises its duration estimate as each repair log entry arrives. The brief here is narrower and complementary: it does not score whether an estimate was accurate, only whether the messages were faithful to the information available when they were sent.

Report each check separately, by storm and by channel, and hold out whole storms rather than random messages, since messages within one event share wording and facts. A single blended score hides the failure that matters most to customers, usually a lapse or a contradiction between channels.

Sampling needs the same care. Large storms produce most of the messages but only a minority of events, so a sample drawn message by message will be dominated by a few storms. Stratify by event size and by channel so that small, routine outages, where automated messages often go out without anyone reading them, are represented too.

  • Unit of analysis: one message, grouped by event and storm.
  • Comparison: the utility's sent messages versus model drafts on the same snapshots.
  • Metrics: failure rate per check with intervals clustered by event.
  • Limits: the brief cannot judge whether an estimate was operationally reasonable; that needs an operations reviewer.

Illustrative worked example: three estimates for one feeder outage

This example is invented to show the method; it describes no real utility or storm. A feeder trips during a wind event at 10:00 and about 1,200 customers lose power. The first automated text at 10:20 gives a system default estimate of four hours, based only on the event's existence.

At 12:30 a crew assesses the cause: a broken crossarm and a downed conductor on one tap. The estimate moves to 18:00, and the message basis reads crew_assessment. At 14:00 switching restores roughly two thirds of the customers from an adjacent feeder. A correct update tells those customers they are restored and tells the remaining group, on the damaged tap, that their estimate stays at 18:00.

The exception: at 14:05 the web map still shows the whole area out with the old four-hour estimate, which expired at 14:00. Checks C3 and C5 both fail. A second subtlety is that the downed conductor triggers the approved safety message for the tap's customers; a drafted message that rewrites that guidance in friendlier words fails C7 even though nothing in it is false.

Every value here, from customer counts to times, is illustrative. In real data the same review would read the event versions and message logs rather than a narrative.

What must stay out of an outage communication dataset?

Keep the customer-facing story and drop the grid engineering. FERC's guidance describes critical energy and electric infrastructure information as specific engineering, vulnerability or detailed design information about critical infrastructure that goes beyond simply giving its general location. Feeder diagrams, switching orders, protection settings and equipment identifiers that reveal network layout should be excluded or generalized before any outage history leaves the utility.

Customer records need their own handling. The FCC has ruled that utilities may make automated calls and texts to customers about matters closely related to their service, such as outages, because customers provided their numbers; that covers sending the messages, not reusing contact details. Phone numbers, names, addresses, account numbers and medical-priority flags should be removed or replaced with stable pseudonyms before analysis.

Stop and route to specialists when:

  • Records include critical-facility or life-support customer designations.
  • A message history is part of a regulatory investigation or claim.
  • Safety scripts or crew safety holds would need rewording to fit a model task.
  • Event detail cannot be generalized without breaking the link to the message.

Can a utility license its outage communication history?

Utilities, co-ops and outage communication vendors hold exactly these linked histories, and they are uncommon. SourceX can take on that licensing for a utility or vendor, qualifying the company first and then handling rights review, contracting, preparation and delivery. Storm-season archives too large to move easily can remain in the company's own storage or travel on encrypted drives, since SourceX does not host multi-terabyte datasets.

Nothing is delivered until the company has approved the scope and terms and signed, and whether it qualifies is first checked through a brief questionnaire that asks for no outage records.

Frequently asked questions

What is an estimated time of restoration?

It is the utility's current estimate of when power will return for a given area or customer group. It is a forecast that should be revised as crews learn more, and each revision should be tied to the operational event that caused it.

Is a message like 'crews are assessing damage' an estimate?

No. It is a status. New York staff, reviewing Tropical Storm Isaias, said a status of More Work Required gives no meaningful restoration information. A dataset should record such messages as status-only so they are not scored as estimates.

Why do outage text messages, web maps and call center answers disagree?

They are often fed by different systems refreshed on different schedules. Linking every channel's message to the same event version, with timestamps, is what lets a team find and fix the lag.

Can past outage notifications be used to train a model?

Possibly, once contact details, account data and grid engineering detail are removed and the utility has confirmed it has the rights to reuse them. Counsel and grid security staff should review the specific records first.

Should the dataset include the final restoration time?

Yes, as a separate outcome field. It is needed to study estimate accuracy, but it must be hidden when grading whether an earlier message was consistent with what was known at the time.

Sources

  • New York DPS staff found customers depend on restoration estimates for health and safety decisions, faulted estimates left to expire, and said More Work Required is not an estimate. Source
  • The Common Alerting Protocol defines Update and Cancel messages by the earlier messages they reference, with each reference giving sender, identifier and sent time. Source
  • A 2025 study treats each outage as a sequence of updates for restoration time prediction and proposes customer-informed regression metrics. Source
  • An earlier study makes an initial outage duration estimate and updates it as repair log entries arrive. Source
  • FERC describes critical energy and electric infrastructure information as specific engineering, vulnerability or detailed design information that goes beyond general location. Source
  • The FCC ruled that utilities may make automated calls and texts to customers about matters closely related to service, such as outages, because customers provided their numbers. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies