Estimated time of restoration prediction that tells customers how sure it is, and updates the moment the crew learns more
Restoration times customers can plan around, updated the moment the crew learns more.
See one case, screen by screen ↓When a recloser locks out, the first thing every customer asks is “when is it back?”. Most outage management systems answer with a default: a fixed number of hours from the cause code and the device — often two hours for an unknown overhead fault — and the dispatcher overrides it by hand when the crew radios in. At the 14:52 lockout in Kestrel Bay that rule would have told 2,184 customers their power was back at 16:52, 1 h 03 earlier than the time now expected.
The facts that would fix the time arrive in pieces: the crew's GPS, its arrival on site, the lineworker's damage report. They reach the outage system late or not at all, so the time moves only when someone remembers. Updates are all or nothing — every small change texted, or nothing at all. Life-support customers are called by hand from a spreadsheet, often after the bulk text. And few utilities measure how many restorations landed inside the window customers were told, so the defaults are never corrected.
Estimated split for one storm outage of about 2,000 customers, in minutes from each fact to customers being told, by hand and with the solution.
Six specialist agents read the outage, track the crew, read the damage and re-predict the time; the outage manager approves any big change.
Tuesday 13 October 2026. A thunderstorm cell crosses Kestrel Bay on Lumina Utilities' east network at 14:40. At 14:52 recloser R-MR12-3 on feeder MR-12, fed from Marlow Road substation, locks out. Here is what happens next, screen by screen, in the working solution.
Rhea owns the times customers are told on the east network. Eight outages are live, 4,816 customers off, each a chip with its own small fan so she sees which times are tight and which are still guesses. Kestrel Bay, OUT-26-04187, is selected: five rows, one for each fact that changed the time, and a band that has narrowed from ±3 h 00 to ±40 min. The expected time is 17:55, in a window of 17:15–18:35.
Default rule would have said 16:52 — “1 h 03 too early”
The incident reader reads the lockout with 2,184 smart-meter last-gasp messages and 37 calls. The cause is not known yet, so the predictor starts from six years of storm lockouts on overhead feeders — 1,946 restorations — adds 25 minutes because storm jobs run 22 % longer and 35 minutes because the nearest free crew is 21 km away at Halden depot. The first time is 18:20, between 15:20 and 21:20.
14:58 text: “We expect it back between 3:20 pm and 9:20 pm. A crew is being assigned and we will text you when we know more.”
Tomasz assigns Crew 14, Dario Mensah and Ines Kowal, with crossarm and conductor kit on board; the band narrows by 55 minutes. The crew tracker follows the truck every 30 seconds: on Quarry Road, 9 km from site, storm debris adds 6 minutes of travel. The time moves to 18:10 ±1 h 50. Each change is under the 30-minute push rule, so customers are not texted.
“Change under the 30-minute push rule: customers not texted”
At 15:46 Crew 14 enters the geofence at pole KB12-0447 on Cove Street, and the patrol from the recloser starts. Storm patrols on MR-12 take 10 to 25 minutes. The window has halved since the last text, from 6 hours to 3, so customers are told the crew is on site: 18:05, between 16:35 and 19:35.
15:47 text: “Our crew is on site in Kestrel Bay and checking the line. We now expect power back between 4:35 pm and 7:35 pm.”
Dario files damage report DR-14-0613 with three photos and a 48-second voice note: crossarm broken at the centre bolt, phase B down but not burnt, an oak limb on the span already cleared, transformer TX-KB-118 not damaged. Make safe was done at 16:02, earthed both ends. The damage reader reads the form, the photos and the voice note and names what is broken.
Voice note: “Crossarm's snapped at the bolt on 0447, B phase on the ground, limb's off now. Transformer looks fine. We've got a crossarm on the truck. Call it an hour and a half, give or take.”
The damage reader maps the report to four repair tasks with durations from past jobs: make safe and clear the limb 20 minutes, replace the crossarm 65, re-sag phase B 15, test and close the recloser 10 — 110 minutes from 16:05. The predictor matches 212 similar jobs (crossarm replaced, one crew, daylight): P10 78, median 104, P90 151 minutes. Daylight lasts until 18:41, so no lighting set-up is needed. New time 17:55, band 17:15–18:35.
The predictor never shows a time without its band, and never narrows the band past what the evidence supports.
The window has halved again, so the customer notifier pushes — and queues the 11 life-support customers in Kestrel Bay for the care team before the bulk text. Noor's team calls them between 16:05 and 16:08: oxygen concentrators, home dialysis, ventilators, a feeding pump, a suction pump. Nine are reached; Mrs J. Whitlow and Mrs H. Kaur go to voicemail, with a retry at 16:30. Then 2,184 customers get the new window: 1,602 by SMS, 388 by email, 171 in the app and 23 by voice.
16:06 text: “Lumina Utilities: the crew found a damaged pole fitting on Cove Street and is repairing it. We expect power back between 5:15 pm and 6:35 pm.”
In Millbrook a vehicle snapped a pole, and the police now keep Millbrook Road closed until 18:30, so the pole truck cannot reach the site. The new window, 8:30 pm to 11:50 pm, is 2 h 10 later than the 864 customers were told at 15:13. The notifier drafts the text and holds it. Rhea sees both messages side by side and chooses Approve and send — the 4 life-support customers are called first — or Keep the 20:00 time, and the predictor re-checks at 16:30.
“Later by more than 2 h — waits for the outage manager”
Ferngate is at risk: Crew 9 has located an underground cable fault 140 m from the breaker but has no jointer. Crew 11 — Iris Delacroix and Pavel Novak, cable jointers — finish a joint at Northgate at 16:20, 14 minutes away by GPS. From 96 similar faults, a jointer on site within an hour of locating cuts the median by 1 h 50. With Crew 11, Ferngate moves from 21:40 ±3 h 20 to 19:50 ±1 h 10.
“Dispatcher decides · Tomasz Wren” — the crew tracker never assigns or moves a crew.
The accuracy reviewer compares every window customers were told with the smart-meter restore events. Over 30 days, 81 % of 1,284 outages were restored inside the window; the first time given had a median error of 38 minutes, against 2 h 17 under the default rules. But underground cable and vehicle strike bands are too narrow — only 63 % and 58 % land inside an 80 % band — so it proposes widening underground cable bands by 35 minutes until jointer arrival is known. Rhea decides.
“Approve recalibration · outage manager decides” — the reviewer never changes the prediction on its own.
The same storm afternoon, seen by the five people who own the time — what it looked like with default times, and what it looks like now.
Six specialist agents read the outage, track the crew, read the damage and re-predict the time; the outage manager approves any big change.
Reads each new outage with the protective device that operated, the smart-meter last-gasp messages and calls behind it, and the weather; counts the customers, the life-support customers and the critical sites.
Predicts the expected time and the 80 % band from 412,860 past restorations by cause, device, area, time of day, weather and crew, and re-predicts when the crew is assigned, moves, arrives or reports damage — listing what moved the time and citing each fact.
Reads the position of each crew on an outage every 30 seconds, computes travel time with live traffic, and flags arrival inside the 150 m geofence and delays over 5 minutes.
Reads the lineworker's damage report, photos and voice note, names what is broken and maps it to repair tasks, taking each task's duration from past jobs.
When the window moves more than 30 minutes or halves, drafts the update in the approved wording, queues life-support calls for the care team first, then sends by SMS, email, app and voice.
Each morning compares every window customers were told with the time power came back, reports the share inside the window by stage, cause and area, and proposes a recalibration where a cause lands under 70 % inside an 80 % band.
Approves overrides and later times. The agents propose; a named person decides.
Ask about any outage, crew or restoration time — or say what to do. Every answer points to the facts behind it.
Why did Kestrel Bay's time move to 17:55?
Because the crew reported the damage. At 16:04 Dario Mensah reported a broken crossarm on pole KB12-0447, one phase down and the transformer undamaged. That is four tasks worth 110 minutes from 16:05, and 212 similar jobs put 8 in 10 between 17:15 and 18:35. The band was ±1 h 30 when the crew arrived at 15:46; now it is ±40 min, so customers were texted.
Which restoration times could miss tonight?
Two are at risk. Ferngate, 1,036 customers: the cable fault needs a joint and Crew 9 has no jointer; underground cable bands land only 63 % of the time. Millbrook, 864 customers: the police closed the road until 18:30, and the new time of 22:10 is held for your approval.
Send Crew 11 to Ferngate
Crew moves are Tomasz Wren's decision, so I have put it in front of him with the effect: Crew 11 are cable jointers finishing at Northgate at 16:20, 14 minutes away by GPS. With them on site, Ferngate moves from 21:40 ±3 h 20 to 19:50 ±1 h 10.
Call Millbrook's life-support customers before the update
Done. The care team has called the 4 life-support customers in Millbrook and told them the road is closed and the repair runs late. The text to all 864 customers still waits for your approval on Customer updates.
12 screens from the working solution, on its sample data. Pick one to see it large.
How restoration times are predicted, shown and pushed are settings, chosen by the utility. Changes are recorded in the audit trail.
| Setting | Default | Choose from |
|---|---|---|
| Confidence of the band shown | 80 % | 70 % to 90 % — 80 % means 8 in 10 restorations should land inside it |
| Minimum similar jobs | 25 | 10 to 100 — below this, fall back to the cause-code default, labelled as such |
| Re-predict on crew GPS | On | On · Off — every 30 s while a crew is travelling |
| Text when the expected time moves | 30 min | 10 to 60 minutes, or when the window halves |
| Widest window shown | 6 h | 3 to 8 hours — wider says “we are assessing” with the next update time |
| Quiet hours for texts | 22:00 – 07:00 | Restored notices and life-support always go |
| Call life-support customers first | On | On · Off — the care team calls before the bulk text |
| Later than the last text by more than | 2 h | 1 to 4 hours — waits for the outage manager, Rhea Castellanos |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
“demo” = seen in the working solution, on its sample network data · “target” = the design goal, measured in the live solution · “estimated” = our estimate: $1.1M a year = about 1.7 million customer-interruptions a year at the demo utility × 71 fewer repeat calls per 1,000 customers off (212 → 141) × $9.54 a call; the minutes split above is also our estimate · “typical” = published figures: J.D. Power Electric Utility Business Customer Satisfaction Study 2011 (729 when restored by the time given, 576 when later); EIA Today in Energy, 2024 (US customers averaged nearly 9 hours without power from major events); arXiv 2505.00225, a 2025 preprint (update-sequence ETR models across six operating companies, median 36.9 % less customer-weighted error than the utilities' own ETRs); E Source ($9.54 a call). Lumina Utilities and the people, places, feeders, crews and customers named on this page are fictional — characters and sample data in the working solution.
An estimated time of restoration is the time a utility tells customers their power should be back. ETR prediction replaces the outage system's default — a fixed number of hours by cause code — with a time and an 80 % window learned from past restorations, updated as the facts arrive. Estimated Time of Restoration Prediction does this with six agents, and the outage manager approves any big change.
The restoration predictor matches each outage to past restorations by cause, device, area, time of day, weather and crew, and gives the expected time with its band. It re-predicts when the crew is assigned, while it travels, when it arrives inside the geofence and when the lineworker reports the damage. Each revision lists what moved the time and cites the fact.
In the working solution's last 30 days, 81 % of 1,284 outages were restored inside the window customers were told, against 46 % for the same outages under last year's default rules. The design goal is 8 in 10. The accuracy reviewer checks whether the 80 % band is honest by cause each morning and proposes recalibration where it is not.
No. It texts when the expected time moves more than 30 minutes or the window halves, keeps quiet hours from 22:00 to 07:00 except for restored notices and life-support, and says “we are assessing” when the window is wider than 6 hours. A time more than 2 hours later waits for the outage manager. All of these are settings.
The incident reader counts the life-support customers on every outage, and the customer notifier queues them for the care team before every bulk text. In the sample case the care team called all 11 Kestrel Bay life-support customers between 16:05 and 16:08, with each call recorded as reached or voicemail and a retry time.
It reads the incidents your outage system opens, with the device, cause and default time, together with smart-meter last-gasp and restore events, crew vehicle tracking, field damage reports and your restoration history. It never changes an outage record or its cause code.
Yes. The outage manager approves overrides, any time more than 2 hours later than the last text and every recalibration. The dispatcher accepts or dismisses every crew recommendation. The agents never assign a crew and never decide it is safe to re-energise.
The Agentic Solution Engine builds and deploys it from your requirements and documents — your outage procedures, message wording, push rules and a history of past restorations — and it goes live once every quality gate has passed. We will walk you through it on your own past outages first.
We'll run Estimated Time of Restoration Prediction on a sample of your own past outages and the times you told customers.