> ## Content Index
> Fetch the complete content index at: https://www.agilebrainsconsulting.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# How a utility turned outage restoration times into prediction windows customers can trust
- URL: https://www.agilebrainsconsulting.com/case-studies/utility-etr-outage-restoration/
- Published: 2026-10-23T14:49:29.000Z
- Updated: 2026-10-23T15:13:30.000Z
- Description: A single-point restoration estimate was set up to fail. In a proof of concept, Agile Brains replaced it with probabilistic prediction windows calibrated to how restorations really behave, and a policy that matches the window to the outage.
- Author: Lara Hill
- Tags: #case-study, Utilities

Every outage raises the same question: when will my power be back? We helped a large East Coast electric and gas utility answer it honestly, replacing a single-point estimated time of restoration (ETR) that was set up to fail with a probabilistic window calibrated to how restorations really behave.

## The challenge

Outage restoration times are heavily skewed. Most outages resolve quickly, but a long tail of complex events runs far longer. A single-point ETR, such as "power back by 4:00 PM," collapses that reality into one deterministic guess. It over-promises on routine outages and badly under-warns on complex ones.

Every miss carries the same costs: repeat "when will it be back?" calls flooding the contact center, eroded customer trust, and harder questions from regulators about storm performance. The utility didn't need a slightly better point estimate. It needed to stop treating restoration as a single moment.

![Diagram: a single-point estimate misses the actual restoration time, while a prediction window contains it.](https://storage.ghost.io/c/e4/2d/e42df399-b545-4436-aee1-6737869588bf/content/images/2026/10/case-utility-etr-window-web.jpg)

## How we helped

We reframed ETR as a probabilistic forecast. Instead of one timestamp, the model predicts a set of percentile bounds and expresses the answer as a window, such as "expected between early and late afternoon," using quantile regression with LightGBM rather than a conventional point model.

Getting there took disciplined data work. We tuned the training population to exclude isolated single-meter events, whose extreme variance would have inflated every window, and kept the signal that generalizes to real, multi-customer outages. Each prediction is anchored on the historical restoration profile of the specific cause of interruption, the single strongest driver, then adjusted for crew response time (the lever operations can control), time of day and crew-shift factors, and the number of customers affected.

![Diagram: the prediction is anchored on the cause-based historical baseline, then adjusted for crew response time, time of day and crew shift, and customers affected, producing a prediction window.](https://storage.ghost.io/c/e4/2d/e42df399-b545-4436-aee1-6737869588bf/content/images/2026/10/case-utility-etr-drivers-web.jpg)

**A policy, not just a model.** The model alone isn't the product; the operating policy is. We designed a segmented approach that matches the window to the event. Routine outages, the large majority, get a tight precision window built for confident customer communication. Complex outages get a wider safety window that puts reliability ahead of precision, so the utility stops missing its own ETRs on exactly the events where a miss hurts most.

And because no model should quietly over-promise, any outage whose predicted window grows too wide is automatically flagged for dispatcher review instead of going to customers. Human judgment stays in the loop where it belongs.

![Diagram: routine outages get a tight precision window, complex outages get a wider safety window, and windows that grow too wide are flagged for dispatcher review.](https://storage.ghost.io/c/e4/2d/e42df399-b545-4436-aee1-6737869588bf/content/images/2026/10/case-utility-etr-policy-web.jpg)

The work was built and deployed as a proof of concept on Vertex AI.

## What we're targeting

This proof of concept is moving through validation, so we're reporting the targets it was built toward rather than claiming results not yet confirmed in production:

- **Well-calibrated windows**, whose stated confidence matches real-world coverage
- **Tighter routine windows**, precise enough to share with customers confidently
- **Fewer missed ETRs on complex events**, with the safety window absorbing long-tail uncertainty
- **Lower repeat-call volume** into the contact center
- **Stronger storm-time customer trust**, measured through restoration-related satisfaction
- **A stronger regulatory storm-reporting position**, built on defensible, data-grounded ETR practice