3-minute read
A large U.S. electric utility set out to optimize storm outage response by aligning crew deployment with anticipated outage demand. Their existing model had significant prediction error, particularly during extreme storm events, limiting confidence in its outputs at times when teams were making their most critical staffing and resource decisions.
Logic20/20 evaluated the model, its weather inputs, and the surrounding workflow, then developed and operationalized an enhanced outage prediction model. The new model reduced prediction error by 24 percent across all storms and 38 percent for high-impact storms compared with the previous model. Applied retrospectively to 2025 storm events, the improved predictions showed the potential for approximately $20 million in annual storm-response savings.
We brought our expertise in
- Utility data & operations
- Storm outage prediction modeling
- Weather feature engineering & forecast validation
- Model performance & impact analysis
- Azure Machine Learning & MLOps
- Operational decision support & visualization
Powering some of the nation’s largest metropolitan areas
Nearly 11 million customers depend on this Fortune 200 electric utility for energy delivery across major metropolitan areas. Its six regulated transmission and distribution utilities and 20,000+ employees support communities across a broad and diverse service territory.
Outage predictions under pressure
When severe weather approaches, it is advantageous for utility teams to determine the number and location of crews required for outage response before the storm’s full impact is known. Crew allocation error carries consequences in either direction:
- Under-responding can force utilities to secure additional crews at higher, last-minute rates and can increase time-to-restoration (TTR).
- Over-responding can result in mobilizing and paying for more crews than outage response ultimately demands.
The utility had developed an outage prediction model to bring data-driven rigor to storm planning across its operating companies. Performance limitations had prevented the model from displacing the existing process, which remained heavily dependent on the judgment of individual subject-matter experts. Weather forecasts supplied core inputs to the outage prediction model, but their accuracy relative to observed conditions had not been fully assessed.
Take a unified approach to utility risk
From Siloed Programs to Unified Risk: A Guide for Utility Leaders provides a practical framework for connecting storm, wildfire, vegetation, and asset programs to support consistent, evidence-based risk decisions across the utility.
We will never sell your data. View our privacy policy here.
Reducing error across the outage prediction lifecycle
Logic20/20 began with a gap assessment spanning the utility’s data, outage targets, model design, weather inputs, and development workflow. The team identified the factors limiting prediction performance and systematically tested improvements against actual 2025 storm events.
Refining the model
Logic20/20 developed more than 30 new features representing outage drivers, including physics-informed weather interactions and historical outage rates by geographic area. The team also evaluated modeling approaches suited to the distribution of outage data, with particular attention to performance during high-impact storms.
Identifying weather forecast bias
Logic20/20 compared forecast conditions with observed weather and identified forecast bias, particularly the underprediction of severe wind conditions. Changes to the weather forecasting system fell outside the project scope, so the team documented recommendations for bias correction, monitoring, automated retraining and drift detection, and alternative forecast sources.
Integrating predictions into storm operations
Logic20/20 standardized model validation and experiment tracking and deployed the final model into the utility’s existing storm prediction workflow. The team also enhanced reporting with intervals showing the range of likely outage outcomes and impact analyses quantifying the cost implications of over- and under-response. These measures gave storm-response leaders a clearer basis for determining crew requirements.
Delivering measurable gains in prediction performance
The updated model reduced prediction error across the utility’s 2025 storm events, with the greatest reduction occurring during high-impact events:
- 38 percent reduction in prediction error for high-impact storms, compared with the existing model
- 24 percent reduction in prediction error across all storm events
- Approximately $20 million in projected annual storm-response savings, based on applying the updated model to actual 2025 storm events and estimating the cost implications of over- and under-response
Logic20/20 also established standardized validation, experiment tracking, and documentation so the utility could consistently evaluate future model enhancements against established performance measures.
Ready to optimize your storm response?
See how more precise outage predictions can inform crew planning, reduce the costs of over- and under-response, and support more confident storm-response decisions.
Elliot Dean is a Data Scientist in AI & Analytics at Logic20/20 with more than five years of experience applying data science in the energy sector. He specializes in machine learning, forecasting, and statistical modeling, with experience developing outage and demand forecasting models that inform utility planning and storm response. His work combines energy analytics and predictive modeling to support high-stakes operational and investment decisions.