From the Utility Management edition of October 29, 2026
Forecast Lead Time and the Timing of Crew Deployment for Weather-Driven Power Outages Using AI Weather Models
University of Illinois Urbana-Champaign (Golkar, Alves de Oliveira, Alipour) · Electric distribution utilities (storm restoration, mutual assistance) · Contiguous United States (county level)
- Publisher
- arXiv
- Original publication
- arXiv preprint v1, submitted 14 Oct 2026
- Source retrieved
- 2026-10-30
What happened
Using GraphCast and FourCastNet operational forecasts with ORNL EAGLE-I outage records (2022-2024, 2,491 counties), the authors find outage-occurrence skill (AUC-PR) falls from 0.1207 at 24h to 0.0646 at 120h, still above a 0.047 county climatology. Staging crews in the top 10% of counties captured 0.356 of high-burden outage at 24h vs 0.238 at 120h; the 72h-to-120h loss is significant. Committing at 72h with all crews beat 24h with 60% of crews; break-even availability 83.4%. A 17.5% skill gain from land-surface predictors produced no significant decision gain, joining antecedent predictors at valid time inflated skill 42-53% (data leakage), and 90% prediction intervals covered only 0.476 of tropical cyclone events.
Why it matters
Directly relevant to municipal utilities, public power agencies and co-ops that request mutual-aid crews before storms and are being sold AI outage-prediction tools; it gives a decision-value test rather than an accuracy claim.
Evidence and measured results
Retrospective modelling study: LightGBM hurdle model, temporal split (2022 train, 2023 calibration/selection, 2024 test), block bootstrap over 18 storm weeks, paired lead-time comparisons, SHAP attribution, conformalized quantile regression. Baselines are county climatology and a perfect-foresight oracle (0.893). Code and data on Zenodo. Not peer reviewed.
Limitations and uncertainty
Preprint; three years of data with a single training year; 2024 test year had an unusual hurricane season; western winter skill unexplained; county/6h resolution unsuited to feeder-level restoration; monetary values highly assumption-dependent (VoLL, restoration acceleration).
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-10-29; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Utilities must commit costly mutual-aid crews days before storms using uncertain forecasts.
- Stakeholders
- Storm restoration / emergency operations manager
- Public power or co-op general manager
- Joint action agency or mutual-aid coordinator
- Finance / rate staff
- Discovery questions
- How many days ahead do you currently request mutual-aid crews, and what drives that timing?
- What share of requested crews did you actually secure in recent events?
- Do you retain OMS outage records by event that could support a retrospective replay?
- Value hypothesis
- A forecast-informed staging rule may capture more outage burden than climatology at 1-3 day lead times; the study does not establish dollar savings for any specific utility.
- Potential engagement
- A retrospective storm-replay assessment of the utility's own events using archived forecasts, ending in a documented decision rule.
- Unsupported claims
- Guaranteed avoided outages (staging shortens restoration, it does not prevent outages)
- Specific dollar savings for a utility
- Feeder-level accuracy
- Reliable five-day crew commitments
Pre-sales engineering
Role takeaway
- Fit
- Strong fit for pre-storm planning analytics; weak fit for feeder-level restoration dispatch.
- Architecture and integration
- Combine AI weather forecast archives with OMS/EAGLE-I outage history and land-surface data in a batch scoring pipeline feeding emergency operations dashboards.
- Prerequisites
- Event-level outage history with consistent timestamps
- Access to archived AI weather forecasts
- Hazard labels for post-event review
- Constraints
- County and 6-hour resolution
- Western winter skill gap unexplained
- Gust field absent from the forecast archive
- Single-year calibration
- Security
- Outage and customer counts from OMS/AMI should be handled under existing data-classification rules; the model itself uses public data.
- Proof of value
- Replay at least two historical storms comparing 24h, 72h and 120h staging
- Run the valid-time vs issue-time correlation leakage diagnostic
- Report interval coverage by hazard class
Delivery
Role takeaway
- Work
- Retrospective replay, decision-rule drafting, integration of scores into storm briefings, and post-event review.
- Dependencies
- OMS data extracts
- Forecast archive access
- Emergency operations participation
- Ownership
- Storm restoration / emergency operations manager.
- Skills and adoption
- Analysts need forecast-verification and leakage-testing skills; operators need training on lead-time limits and uncertainty.
- Governance checkpoints
- Leakage check sign-off before any skill claim
- Incident-command approval of the decision rule
- Post-storm review of staging decisions
- Acceptance criteria
- Proposed: captured burden at 72h exceeds climatology on replayed storms
- Proposed: documented break-even crew-availability ratio for the utility
- Proposed: hazard-stratified interval coverage reported
- Risks
- Overconfidence in hurricanes
- Sparse or biased outage reporting
- Misreading skill gains as decision gains
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
Outage models built on AI weather forecasts need a gust field (the archive lacks one; adding gust raised wind-only skill about a third), issue-time feature joins, and hazard-aware calibration. Pipelines must lag reanalysis-derived features to forecast issue time.
Governance
Who approves, reviews and stays accountable for outcomes?
Procurement and validation should require decision-value metrics (captured burden, break-even crew availability) and leakage checks, not headline accuracy. Interval guarantees can fail under hazard-mix shift, so calibration evidence must be stratified.
Security and privacy
What data, permissions and controls need testing?
Limited direct security relevance; uses public county-level data. Operational deployments would ingest utility OMS/AMI outage data, which carries customer and grid-sensitivity controls.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Mutual-aid line crews are a constrained workforce; better timing affects crew fatigue and travel. Operators need training to interpret probabilistic outputs.
Procurement
What should contracts, pricing and exit terms secure?
RFPs for outage-prediction tools should require decision-value metrics, leakage diagnostics and hazard-stratified calibration evidence.
Operating model
Which teams own the service once it runs?
Supports a formal, documented mutual-aid commitment rule with post-event review rather than convention-based timing.
Publication history
- 2026-10-29Utility Management · Issue 0110 resources
Stable resource ID: uiuc-ai-weather-outage-crew-timing-2026