Analytica Data Science SolutionsContact
ESC

↑↓ to move↵ to open

Guide

Storm outage prediction: from a weather forecast to staged crews

How a storm outage prediction model turns a weather forecast into district damage estimates and a crew staging plan, up to five days before the storm.

By 21 November 2026

A utility usually knows a storm is coming days before it arrives. The weather forecast says where it’s tracking and how strong it should be. What the forecast doesn’t say is which districts will lose power, how much equipment will be damaged, and where crews, poles, transformers and wire need to be before the first pole comes down.

That gap is what a storm outage prediction model is for. This guide covers what the model has to answer, how we built one for Southern California Edison, the data it ran on, and how its output becomes a staging plan.

The operational problem: knowing where to stage crews

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

A weather forecast alone doesn’t support that decision. It describes the storm, not the grid’s exposure to it. The utility we worked with had weather forecasts and it had asset records, but nothing 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.”

Across the United States, most of the difference from one year to the next in how long customers go without power comes from major events such as hurricanes and other storms, according to the U.S. Energy Information Administration.

Hours without power per U.S. customer, 2015–2025

Average annual hours without power per U.S. electricity customer, with and without major event days, 2015 to 2025Source: U.S. Energy Information Administration, Electric Power Annual Table 11.1, IEEE method, converted from minutes to hours. 2015: 3.5 hours with major event days, 2.0 hours without; 2016: 4.5 hours with major event days, 2.0 hours without; 2017: 8.4 hours with major event days, 2.0 hours without; 2018: 5.8 hours with major event days, 2.0 hours without; 2019: 4.9 hours with major event days, 2.0 hours without; 2020: 7.6 hours with major event days, 1.9 hours without; 2021: 7.9 hours with major event days, 2.1 hours without; 2022: 5.6 hours with major event days, 2.2 hours without; 2023: 6.1 hours with major event days, 2.1 hours without; 2024: 11.0 hours with major event days, 2.2 hours without; 2025: 4.5 hours with major event days, 2.1 hours without.02468101220153.520164.520178.420185.820194.920207.620217.920225.620236.1202411.020254.5Hours per customer per year
The average time a U.S. electricity customer spent without power each year, for utilities reporting under the IEEE standard. Both bars count sustained interruptions only, not momentary ones. The longer bar includes major event days; the shorter bar leaves them out. A major event day is a day whose outage time exceeds a threshold set from the utility's own previous five years. Without those days the average stayed between 1.9 and 2.2 hours a year; with them it ranged from 3.5 hours in 2015 to 11.0 hours in 2024. EIA attributes most of the difference between years to major events such as hurricanes and other storms.Source: U.S. Energy Information Administration, Electric Power Annual, Table 11.1 (Form EIA-861), SAIDI converted from minutes to hours; and Today in Energy, 1 December 2025.

What an outage prediction model has to answer

For the people staging crews, a useful model answers four questions, in this order:

  1. Where will damage occur? Which districts in the service territory are likely to take damage from this storm.
  2. How badly? Within those districts, which grid assets are likely to be damaged.
  3. What will it take to restore? The crews, materials and hours the predicted damage will need.
  4. How far ahead? Early enough to move crews and materials before the storm, not after the outages start.

The first two are forecasts of damage. The third turns damage into resources, and it’s the one the operations team acts on. The fourth decides whether any of it is useful: a damage forecast that arrives after crews can no longer be moved doesn’t change the response.

Each answer is an estimate. A planner deciding how many crews to send to a district needs to know how far the number could reasonably be off, which is why a damage forecast should carry its uncertainty rather than a single figure.

Why the model is built in two stages

The model we built for Southern California Edison splits the damage forecast into two stages, so each stage answers one question.

  • 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 the districts stage one flags.

A third component, the resource model, converts the predicted damage into the crews, materials and hours required.

The split matches the questions above. “Is this district at risk?” is asked of the whole service territory. “Which poles and transformers in it?” only needs asking where the answer to the first question is yes, so stage two runs inside the flagged districts rather than across every asset the utility owns.

The two-stage storm damage model

Two-stage storm damage model over a service territoryA forecast cone crosses the utility's service territory, widening with lead time. Stage one scores every district for the likelihood of damage; the districts crossing the threshold are outlined. Stage two runs only inside those districts, scoring individual poles and transformers along the distribution feeders — and the predicted failures cluster under tree canopy, because in most storms something falls on the grid rather than the grid failing on its own. A resource model turns that into crews, materials and hours, up to five days before the storm arrives.SERVICE TERRITORYONE FLAGGED DISTRICTFORECAST TRACK · CONE WIDENS WITH LEAD TIMEEvery district scored. Most will be fine.FEEDER · POLES · CANOPYTrees and debris striking lines cause most storm damage.STAGE 01 · WHICH DISTRICTSSTAGE 02 · WHICH ASSETSCREWS + MATERIALS STAGEDFORECAST ISSUEDSTORMUp to five days of lead time to stage crews and materials.
Stage one of the storm damage model over a service territoryA forecast cone crosses the utility's service territory, widening with lead time. Stage one scores every district for the likelihood of damage; the districts crossing the threshold are outlined, and the one opened up below is outlined in red.SERVICE TERRITORYFORECAST TRACKCONE WIDENS WITH LEAD TIMEEvery district scored. Most will be fine.Stage two inside one flagged district, and the lead timeStage two runs only inside the flagged districts, scoring individual poles and transformers along the distribution feeders, and the predicted failures cluster under tree canopy, because in most storms something falls on the grid rather than the grid failing on its own. A resource model turns that into crews, materials and hours, up to five days before the storm arrives.ONE FLAGGED DISTRICTFEEDER · POLES · CANOPYTrees and debris striking lines causemost storm damage.FORECAST ISSUEDSTAGE 01 · WHICH DISTRICTSSTAGE 02 · WHICH ASSETSCREWS + MATERIALS STAGEDSTORMUp to five days of lead timeto stage crews and materials.
likelihood of damageflagged by stage 01predicted asset failureStage one scores every district in the service territory for the likelihood of damage. Stage two predicts damage to specific grid assets only inside the districts stage one flags, and the resource model turns that damage into the crews, materials and hours required, up to five days before the storm arrives.

Flagging a district means its score has crossed a threshold, and where that line sits is an operational decision as much as a modeling one. Set it low and more districts get flagged, so crews are staged for damage that doesn’t come. Set it high and fewer false alarms reach the planners, at the cost of districts that take damage without crews nearby. That’s the trade-off precision and recall describe, and the right balance depends on what a missed district costs the utility compared with an idle crew.

The data an outage prediction model uses

The Southern California Edison model drew on three kinds of data.

  • Weather. Feeds from multiple open sources, ingested and prepared automatically.
  • Grid assets. Static data on the number, age and location of poles, transformers and other equipment exposed to a given storm track.
  • Tree cover and soil moisture. The geospatial layer that turned out to matter most, because trees and debris striking lines cause most storm damage.

Turning a damage forecast into crew staging

A district-level damage map is not yet a staging plan. The resource model is the step that makes it one: it takes the predicted damage and converts it into the crews, materials and hours needed to restore each district.

The output then has to reach the people making the call. For Southern California Edison that took two forms:

  • 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 aren’t sitting in front of a dashboard at 4am.

The forecast horizon was up to five days. Five days is roughly how long it takes to move crews, which is why the horizon matters as much as the accuracy of the damage model.

From weather forecast to staged crews

Storm outage prediction pipeline, from data layers to the crew staging decisionThree data layers feed the two-stage damage model: weather from multiple open sources, grid asset records with the age and location of equipment, and tree cover and soil moisture. Stage one scores every district for the likelihood of damage. Stage two predicts damage to poles, transformers and other equipment, only inside the districts stage one flags. The resource model converts that damage into crews, materials and hours to restore. A dashboard with maps by district and automated severe-weather email alerts carry the result to the operations team, who decide where to stage crews and materials before the storm arrives. The forecast is issued up to five days ahead, roughly how long it takes to move crews.DATA LAYERSWeathermultiple open sourcesGrid assetsage and locationTree cover and soil moisturegeospatial layerTWO-STAGE DAMAGE MODELSTAGE 01Which districts take damageevery district scored forlikelihood of damageSTAGE 02How much, asset by assetpoles and transformersin flagged districts onlyRESOURCE MODELCrews, materials, hours to restorewhat the predicted damage will needDELIVERYDashboard and alertsmaps by districtsevere-weather email alertsTHE DECISIONStage crews and materialsin place before the storm arrivesflagged districts onlyFORECAST HORIZONUp to five daysFORECAST ISSUEDSTORM ARRIVESroughly how long it takes to move crews
The Southern California Edison platform, end to end. Weather, asset and tree-cover data feed the two-stage damage model: stage one flags the districts likely to take damage, stage two predicts asset-level damage inside them, the resource model turns that into crews, materials and hours, and the dashboard and email alerts put the result in front of the people who decide where crews and materials go.

Building the dashboards, feeds and alerts that put model output in front of an operations team is our decision-support platforms work. The models behind them are our AI and data science delivery work, and when a model needs an operator interface, the two are scoped together.

What we delivered for Southern California Edison

Over two years, on a subcontract, we built Southern California Edison a storm damage platform: the two-stage damage model, the resource model, the automated weather feeds, the dashboard and the email alerts described above.

forecast horizon
5 daysforecast horizon
model stages
2model stages
engagement
2 yearsengagement
reduction in outage times, client-reported
15%reduction in outage times, client-reported

The client reported outage times falling by 15%, with damage predicted up to five days ahead, enough lead time to stage crews and materials before the storm arrives. The full case study is Predicting where a storm will break the grid.

The same kind of event looks different from the federal side. For the Cybersecurity and Infrastructure Security Agency, as a subcontractor, we forecast how a hurricane moves through dependent critical infrastructure, from a port to a refinery to the region it supplies. That work is described in Predicting how a hurricane cascades through US critical infrastructure.

What a utility needs in place to start

The inputs to the Southern California Edison model point to what a utility needs to have before a project like this starts.

  • Asset records with locations. The number, age and location of poles, transformers and other exposed equipment. This is the part only the utility holds.
  • Tree cover and soil moisture data for the service territory.
  • Weather forecasts. The Southern California Edison model took its weather from multiple open sources.
  • A defined staging decision. Who decides where crews and materials go, how far ahead, and what they need to see to decide. The resource model is built around that decision, and the dashboard and alerts are built around the people who make it.

Whether the records a utility holds are complete enough to support a model is the first thing to check, before anything is built.

Questions about storm outage prediction

How far ahead can storm damage be predicted?

The model we built for Southern California Edison predicted damage up to five days ahead. Five days is roughly how long it takes to move crews, so that horizon leaves time to stage them before the storm arrives.

Does the model score every asset the utility owns?

No. Stage one scores every district for the likelihood of damage, and stage two predicts asset-level damage only inside the districts stage one flags.

Why does tree cover matter so much?

Because trees and debris striking lines cause most storm damage. That’s why tree cover and soil moisture sit alongside weather and asset data as inputs.

How does a damage forecast become a staging plan?

Through the resource model, which converts predicted damage into the crews, materials and hours each district will need. The results reach the operations team through a dashboard and email alerts.

Start with the decision

Tell us what you're trying to predict — storm damage, equipment failure, which employees are likely to leave. We'll look at the data you already have and tell you whether it can support the decision, before you've committed to anything.

Start a conversation