Major U.S. power outages · 2000–2016
What makes an outage last longer?
OutageLens explores 1,534 major power outages from Purdue University's LASCI dataset. We connect where and why outages happen with how long restoration takes, then test how useful those patterns are for early estimates.
Our objective: understand how outage cause affects duration—and determine whether information available near the start can predict restoration time.
This tour takes about one minute. Use the numbered menu or the arrow buttons to move through it.
Explore when and where
Map & timeline
⌖
Outages by State
Darker blue means more recorded major outages. Hover for a quick count; click a state for duration, its leading cause, and a full cause breakdown. This shows frequency—not grid reliability or outage risk per resident.
↗
Power Outages Over Time
The blue line is the total number of outages recorded each year. Hover to reveal colored cause-level lines and see what drove a rise or fall. The partial 2016 value only covers January through July.
◷
Longest Outages
A ranked list of extreme restoration times. These cases reveal the dataset's long tail and explain why a few multi-week outages can distort averages and raw-error metrics.
▦
Summary tiles
Headline counts and model results give orientation. Median duration describes a typical recorded outage more robustly than the mean; MAE is the model's average absolute miss.
Explore what drives outages
Causes & climate regions
ϟ
Cause ranking
Bars compare the number of outages assigned to each cause category. Severe weather is the largest category. Counts describe prevalence in this dataset; they do not prove that a cause makes an outage longer.
◎
Climate regions
Each tile shows a region's share of all recorded outages. These NOAA regions help compare broad climate patterns while avoiding noisy state-level categories in the prediction model.
≠
Hypothesis test
A permutation test compares severe-weather duration with other causes. The observed difference is unlikely under the no-effect hypothesis (p < 0.05), supporting a real association with longer duration.
!
Read association carefully
Cause, geography, season, and infrastructure can overlap. The visuals uncover patterns, while the test measures evidence—not a complete causal explanation.
Estimate restoration time
The prediction model
The model uses cause, climate and NERC region, season, start time, year, and population density—features knowable around outage onset. Gradient boosting predicts log-duration so rare, extremely long events do not dominate training.
Baseline → final bars
Shorter MAE bars are better. The final model averages a 38.6-hour absolute error, improving on linear regression's 43.4 hours. The gain is meaningful but modest.
Duration buckets
Precision and recall show performance for <1 day, 1–3 days, and 3+ days. This triage view asks a more actionable question than predicting an exact minute.
Log-scale R²
R² = 0.53 describes fit on the transformed target the model optimizes. Raw-duration R² is only 0.08 because a handful of extreme outages dominate that scale.
What it cannot know
The data lacks physical damage, crew availability, and dispatch information. Treat the model as a pattern-based estimate, not a real-time restoration promise.
Interpret responsibly
Fairness, uncertainty & next steps
Fairness comparison
The chart compares model RMSE for severe weather and equipment failure. Larger RMSE means less accurate predictions for that group.
Small-sample warning
Only 12 equipment-failure examples appear in the test set, versus 151 severe-weather cases. That group's error estimate is therefore highly uncertain.
Test conclusion
The fairness permutation test gives p = 0.0795, so we do not find statistically significant evidence of unequal error at α = 0.05. This is not proof that the model is fair.
How to explore
Start on the dashboard, hover over the trend, click states on the map, then compare the model and fairness tabs. The Data and Notebook sections expose the evidence behind every view.