NukeWorker Outage Schedule Scorecard

The Challenge

Nuclear utilities publish outage schedules months in advance, but the reality rarely matches the plan. Refueling outages routinely extend days or weeks beyond the published end date. Start dates shift. And when an unplanned shutdown occurs, there's no published estimate at all. Workers and stakeholders are left guessing.

We set out to answer three questions: When will the next outage start? How long will a scheduled outage actually take? And when a forced outage happens, how long will it last?

Two Products, Two Validation Methods

Our Multi-Stage Pipeline

Our predictions use a multi-stage pipeline that combines traditional cycle analysis with an ensemble of machine learning models. The first stage analyzes historical refueling patterns to generate a baseline forecast. The second stage runs that forecast through an ensemble of three independent machine learning models, each one trained to spot different patterns, and combines their answers using optimized weights. A final post-processing step snaps the prediction to each unit's preferred day of the week, since most plants consistently start outages on the same weekday.

What the Models Learn

Our ensemble of machine learning models is trained on 898 historical predictions spanning 27 years of operating data, learning from 40+ features across five categories:

The single most predictive feature is cycle length, the number of days since the last refueling outage. Owner fleet sequencing (how an operator coordinates outages across multiple units) is the second most important factor for start date prediction, followed by each unit's preferred day of the week for starting outages.

Rigorous Backtesting

Every model earns its place. We validate by hiding one year of data at a time, asking the models to predict that year using only past information, then rotating through every year from 2015 to 2026. This ensures the models are always tested on data they have never seen.

On recent outages (2023 onward), our ensemble predicts outage start dates within an average of 3.3 days of the actual date. Across the entire 27-year backtest the average is 4.7 days — the gap reflects how much the fleet’s scheduling discipline has matured over the past decade and how much the model has learned from it. Accuracy is most consistent at operators who manage complex multi-unit outage schedules.

Key Insights

The fleet is not homogeneous. Owner operating practices, reactor design, containment type, and geographic region all create measurable differences in outage timing. An Entergy unit behaves differently from a Constellation unit, and the model learns these patterns.

Fleet coordination matters. When an operator runs multiple units, they stagger outages in a sequence. The gap between outages within an owner's fleet is the second most important predictor of start date accuracy.

Day of the week is a hidden signal. Over 84% of fleet outages start on a Monday, Saturday, or Sunday. Most units are remarkably consistent. Constellation units almost always start on Monday, APS on Saturday, Dominion on Sunday. Our model learns each unit's preferred start day and uses it to fine-tune the final prediction.

Major maintenance leaves fingerprints. Steam generator replacements, vessel head inspections, and other extended campaigns happen on multi-year cycles. The models track time since the last major work to anticipate when the next extended outage is due.

Continuous Improvement

Our models aren't static. As new outage data comes in, the system recalculates predictions and accuracy scores automatically. The machine learning ensemble is periodically retrained on the latest data, and the leaderboards on this page reflect live, up-to-date accuracy measurements, not cherry-picked results.

Every unit's accuracy score is computed by comparing what we predicted against what actually happened, weighted toward recent performance. If a unit's behavior changes, our model adapts.