One honest picture of every job site's hazards

A utility foreman deciding whether a crew should set up a bucket truck tomorrow would need to check a weather app, NWS alerts, a lightning map, FEMA flood maps, fire satellites and whatever the last crew said. Nobody has time for that at 6 AM, so the call is often a gut feeling.

If a source fails, the last good reading keeps being shown with its "as of" time, and the card is flagged when the reading gets old. A source with nothing yet shows "No data", never a made-up zero.

• Any flash within 10 miles of a site starts a stop-work alert.
• All clear = the latest strike + 30 minutes. Each new strike pushes it later.
• The countdown uses when the lightning struck, not when our server read the file, so late or out-of-order satellite data can never shorten it.
• The check runs every 20 seconds even with no lightning, because the all-clear is triggered by silence.

In a real replay of the August 4, 2025 storm, 21 of 55 sites got stop-work alerts. The busiest, Okeechobee Clean Energy Center, reset 177 times and cleared exactly 30 minutes after its final strike. Every reset is logged, so a safety manager can audit why a crew was stopped and when it was cleared.

A hybrid: the highest risk wins

1. The model learns storms and floods from the forecast, turning its estimated probability into a 0–10 weather score. An index of 7 does not mean a 70% chance.
2. Take the highest score from the model, OSHA heat-index bands, wildfire detections, wildlife and crew reports. A mild forecast can never average away dangerous heat or a reported hazard.
3. Official NWS warnings set minimum scores: at least 9 for a Hurricane Warning, 7 for other supported warnings, and 4 for watches or advisories. Floors only raise a lower score. A nearby downed-line report also sets today's minimum to 7; our flood-risk alert sets it to 4.
4. Lightning stop-work overrides today's index to at least 9.0. The stop-work engine runs independently; no model forecast can shorten or clear its wait.
5. Tomorrow uses tomorrow's forecast, peak heat, and watches or warnings in effect then. Today's crew-report floors, advisories and lightning stop-work do not carry over. If the model cannot score, safety rules still run.

Machine learning model

A model estimates the chance of a crew-relevant hazardous weather day: NOAA Storm Events logs a storm, flood or other hazardous weather near the site. It mainly learns storms and floods; it does not predict injuries or the time of a lightning strike.

It learns from archived Open-Meteo forecasts matched against NOAA Storm Events, using the same kind of forecast it receives live.

Trained on 2022–2023 across 55 sites (40,150 site-days, 1,152 reported hazardous days). Calibrated on 2024; tested on 2025, kept out of training and tuning.

Compared with two baselines

The model's same-day PR-AUC is 4.0× Climatology; 3.2× Simple rule. This measures how well rare hazardous days rise to the top, not percent accuracy.

2025 test results · same-day forecasts

Method	PR-AUC ↑	ROC-AUC ↑	Brier ↓
Hazard model (LightGBM)Calibrated probability from the weather forecast, season and site.	0.1285	0.8419	0.02219
ClimatologyThe training years' hazardous-day rate for that calendar month.	0.0318	0.6154	0.02319
Simple ruleFlag the day if forecast rain is at least 1 in or the heat index reaches 108°F.	0.0396	0.5994	0.02308

PR-AUC and ROC-AUC measure ranking; higher is better. Brier measures probability error; lower is better. Reported hazards occurred on 2.37% of test site-days (476 of 20,075).

Day-ahead results and label definition

2025 test results · forecasts issued a day earlier

Method	PR-AUC ↑	ROC-AUC ↑	Brier ↓
Hazard model (LightGBM)	0.1021	0.8253	0.02271
Climatology	0.0318	0.6154	0.02319
Simple rule	0.0338	0.5644	0.02313

— = not evaluated. A site-day is one site on one local calendar day.

Label: NOAA Storm Events crew-relevant event within 15 mi or in the site's NWS zone/county, by site-local date

Weather input: Open-Meteo gfs_seamless (HRRR over Florida); training on the Historical Forecast API

The diagonal is a perfect match between predicted and observed chances. The model is better at ranking days than estimating exact probabilities; the highest probabilities can be overconfident.

Model evaluation limits

• The label is "NOAA logged a hazardous-weather event", not "it happened": rural sites and nights are under-reported, and lightning is logged only when it causes damage or injury.
• Location is partly reporting bias: areas near people get more reports, so the model rates them riskier. That's why "Area with frequent past severe-weather reports" can be a top factor.
• Heat isn't learned (NOAA logged Florida heat only in 2022-2023), so heat comes from the OSHA rule.
• Wildfire is barely predicted by weather alone, so fires come from the satellite-detection rule.
• Two training years, one calibration year (2024, with three hurricanes) and one test year: very high probabilities (above ~20%) run over-confident in ordinary years.
• Much better at ranking days than at exact probabilities: 4% Brier skill over monthly averages today, not significant a day ahead.

• Rainfall is a model estimate.

  Open-Meteo rain totals come from weather models, not a rain gauge at the site. Real totals can be higher or lower.

• Wildfire depends on satellite passes.

  VIIRS satellites see a given spot only a few times a day, so a new fire can burn for hours before it shows up. Hot spots within 1 km of a plant are treated as plant heat and ignored, so a real fire inside that ring would be missed. Without a FIRMS key, wildfire data is off and the card says so.

• It predicts hazardous weather days, not injuries.

  A high index flags conditions to check; it does not mean someone will get hurt. A low index does not mean a site is safe. It supports a supervisor's judgment; it doesn't replace it.

• Lightning positions are approximate.

  Satellite pixels are about 8–14 km across, so a strike reported at 9.8 mi could be a little closer or farther. A strike reaches us 20–40 seconds after it happens.

• Hazardous days are rare, and exact probabilities are uncertain.

  There are few reported events to learn from. The model ranks risky days better than it estimates exact chances; the highest probabilities can be overconfident. Day-ahead probability gains over the baselines are not statistically conclusive.

• Reporting bias affects the model.

  NOAA records reported events, not every event that happened. Rural sites and nights are under-reported; lightning is usually logged only after damage or injury. Location partly reflects where reports come from, so a low score at a remote site is weaker evidence of a quiet day.

• Heat and wildfire cannot be learned reliably from these labels.

  Heat reports are too sparse to validate the model's heat predictions, and weather alone captures wildfire poorly without drought or fuel conditions. OSHA heat bands and satellite fire detections supply the safety rules for these hazards.

Lightning doesn't strike on cue for a demo. Replay takes an hour of real GOES-19 satellite files from a past Florida storm and feeds them, one by one, through the same code the live system uses. A simulated clock starts at the storm's first file and runs 30× faster, so an hour of lightning plays in about two minutes.

• Why the real pipeline: it proves the code that protects crews actually works, and it gives exactly the alerts crews would have seen that evening. The alert engine has no clock of its own; it runs on whatever time it's given.
• Replay alerts are stored separately and never mix with live ones. Real lightning keeps being tracked during a replay: real alerts stay on screen next to the replay's, which are tagged REPLAY.
• After the last file, replay runs 31 quiet simulated minutes so every countdown can finish.
• Heat, rain, fire and wildlife stay live during a replay; only lightning is replayed.

Try it: open the dashboard and press ⚡ Replay storm