Analyticadata science solutionsContact
← All work

Case study

Predicting where a storm will break the grid — five days out

The utility could see a storm coming. What it could not see was which districts would fail, how badly, or where to put its crews before the first pole came down.

Client
A major US investor-owned electric utility
Sector
Energy · Utilities
Role
Subcontracted delivery
Duration
2 years
reduction in outage times
15%reduction in outage times
forecast horizon
5 daysforecast horizon
model stages
2model stages

The problem

Storm response at a large utility is a staging problem. Crews, transformers, poles and wire have to be in roughly the right place before the weather arrives, because moving them afterwards is what turns a six-hour outage into a two-day one.

The utility had weather forecasts and it had asset records. What it did not have was anything connecting the two: no way to turn "a storm is coming" into "these districts will take damage, of roughly this magnitude, needing roughly these resources."

What we built

A two-stage predictive model, deliberately split so each stage answers one question and can be validated on its own.

Stage one — will it break?
Predicts the likelihood of damage occurring in a given district.
Stage two — how badly?
Predicts damage to specific grid assets within districts flagged by stage one.
Resource model
Converts predicted damage into the crews, materials and hours required — the output the operations team actually acts on.
Two-stage storm damage modelThree input streams — weather feeds, asset records and geospatial layers — feed stage one, which predicts whether damage will occur in a district. Districts it flags pass to stage two, which predicts damage to specific assets. A resource model converts that into crews, materials and hours, up to five days ahead.Weather feedsmultiple open sourcesAsset recordspoles, transformers, age, locationGeospatialtree cover, soil moistureSTAGE 01Will it break?Likelihood of damagein a given district→ flags districts at riskSTAGE 02How badly?Damage to specificgrid assets→ only where stage 01 firedOUTPUTCrewsMaterialsHoursFORECAST ISSUEDSTORMUp to 5 days of lead time — long enough to move crews, not just to watch
Splitting the model in two is the design decision worth noticing. Each stage answers one question and can be validated on its own — and stage two only ever runs where stage one fired, which is what made it cheap enough to run daily across every district.

The data

Weather feeds from multiple open sources, automatically ingested and prepared. Static asset data — the number, age and location of poles, transformers and other equipment exposed to a given storm track. And the geospatial layer that turns out to matter most: tree cover and soil moisture, because in most storms the grid does not fail on its own. Something falls on it.

How it was delivered

A cloud-hosted dashboard with interactive maps showing predicted damage by district, the resources required, and the hours to restore. Automated severe-weather alerts by email so the model reaches people who are not sitting in front of a dashboard at 4am.

Result

Outage times fell by 15%, with damage predicted up to five days ahead — enough lead time to stage resources rather than chase failures.

The number worth dwelling on is not the accuracy of the damage model. It is the five days. A model that is right on the morning of the storm is interesting; a model that is roughly right five days out is operationally useful, because that is how long it takes to move crews.

Tell us what decision you are trying to get right.

Not a discovery call about our capabilities. A conversation about the specific thing you need to predict, and whether the data you have can support it.

Start a conversation