Gulf humid-heat research
Research on forecast-driven heat safety for outdoor work in the Gulf.
Qatar's Ministerial Decision 17/2021 is the enforceable baseline: a WBGT standard and a fixed midday rest window. This project studies what a daily weather forecast adds on top of it, and reports what it does not. Harara is the assistant built from that work.
Background
Why humid heat needs its own instruments.
Humid heat is harder on the body than the number suggests.
The body cools by sweating, and sweat only helps when it evaporates. In humid Gulf air it barely does, so the real physiological load is far higher than air temperature alone implies. WBGT is the index that captures this, and it is what the regulation is written against.
Qatar's fixed window is a strong baseline.
Ministerial Decision 17/2021 was the first GCC regulation to adopt WBGT as its legal standard, with a WBGT stop-work line and a fixed midday rest window. It is simple and enforceable, and it is the baseline this research measures against.
The question is what a daily forecast adds.
Heat risk shifts day to day with the weather. This project tests, on 16 years of data, where a forecast-driven layer genuinely helps and where it does not.
What the data shows
Six findings, including one about our own instruments.
Qatar's Ministerial Decision 17/2021 is the enforceable baseline every result is measured against. Some of the strongest findings are that a proposed improvement did not hold up.
- 01
The fixed midday window is a floor, not a ceiling.
It covers about a quarter of daylight warm-season hours. Roughly 60 percent of the hours that are physiologically unsafe for heavy unacclimatised work fall outside it, in the mornings, evenings and shoulder months.
Technical report, section 7 - 02
Modern public forecasts already predict Doha WBGT well.
Raw 24 to 72 hour WBGT forecasts run 1.0 to 1.1 C mean error at a 12 to 14 percent miss rate. A learned bias-correction, trained walk-forward, lowers the mean error but more than doubles the miss rate: it regresses toward climatology and under-predicts the hot tail, which is the one failure mode that matters for a stop-work call.
Technical report, section 5.1 - 03
AI weather models are not the risk here. NOAA GFS is.
On identical hours, ECMWF IFS and AIFS place the 32.1 C stop-work hours within 12 to 18 percent. NOAA GFS runs 0.3 to 0.8 C cold on WBGT and misses 40 to 45 percent of those hours at every lead, and the bias worsens in the days before a heat wave.
Results ledger, rows 23 and 26 - 04
A daily schedule does not beat the fixed rule once worker time is counted.
Held to the same hours on site, the forecast-driven optimiser runs hotter than the calendar rule on peak retained heat load: the earlier advantage came from spreading work across a longer on-site day. A plain earlier start is better on heat but cannot deliver full output on about half of peak-season days.
Results ledger, row 14a - 05
A physics filter beats heart rate alone on real bodies, but is not field-ready.
With a skin-temperature channel it estimates core temperature at 0.41 C RMSE against a rectal reference, below the heart-rate-only baseline, with the bias removed. Its stated uncertainty is not calibrated out of domain and it has not been tested on Gulf outdoor workers.
Technical report, section 9.1 - 06
The rising-heat trend was understated, not overstated, by our own weather archive.
Cross-checking the working forecast archive against a second reanalysis and the measured airport station found a second, previously undocumented defect: its Doha humidity quietly drifted dry from 2018 onward. Correcting it with measured data turns a reported flat trend into a clearly rising one, confirmed by both independent sources. We are publishing the defect we found in our own primary dataset, not just the ones we found in others'.
Technical report, section 5.8
Full method and every intermediate result in the technical report and the results ledger.
Method
How the research was done.
All of it on public data, reproducible from the repository.
- 01
Sixteen years of Doha weather.
Hourly temperature, humidity, wind and radiation for 2010 to 2026, from the Open-Meteo archive, with a measured-wind patch from November 2024 where the archive drifted.
- 02
WBGT by the Liljegren energy-balance method.
The physical model through ECMWF thermofeel, not an algebraic approximation. An early approximation biased the index 1 to 2 C low and was replaced.
- 03
Walk-forward validation.
Every model is scored only on days after the ones it was fitted on. Hand-picked case days that seemed to support a hypothesis turned out to be selection bias, so the whole record is used.
- 04
The miss rate is the metric.
The share of true stop-work hours a forecast places below 32.1 C, not average error. A model can cut mean error and still miss more dangerous hours.
- 05
Negative results are reported.
Four forecasting hypotheses were tested and all four rejected. The scheduler result was revised down when worker time was counted honestly. Those are in the ledger with the rest.
The technical report has the full method, and the results ledger has every question tested, including the rejected ones.
The language layer
Where the language model sits, and where it does not.
Reads rules
It turns outdoor-work regulations from their text into machine constraints, with a character-offset citation for every value, checked against the source before anything uses it.
Talks plainly
It gives a plain-language interface. Describe a shift, or ask a heat, WBGT, forecast or rules question for the Gulf, and it routes to a deterministic tool and shows the source.
Never counts
It never computes or estimates a number. Every figure comes from the physics, the forecast, or a cited rule. It runs on K2-Horizon from IFM at MBZUAI.
The daily plan
One working day in Doha, from the forecast to the plan.
Built up in four steps: the forecast, the regulatory line, the fixed-window baseline, and the plan the solver returns inside it.
Worked example
A real past week, hour by hour.
What it is used for
Three applications, honestly scoped.
Plan a shift.
Give a site, a day, the work type, the hours and whether the crew is used to the heat. Get an hour-by-hour work and rest plan inside Decision 17/2021, with the time on site, the heat dose and the peak load of every choice shown next to each other.
Answer heat questions for the Gulf.
WBGT and how it is computed, heat illness and first response, acclimatisation, the forecast for a site, whether it is safe to work outside right now, and the outdoor-work rules for Qatar, the UAE and Saudi Arabia. Every answer cites its source, and it does nothing outside that scope.
Model rule coverage for a regulator.
Feed in a draft rule or an existing one and see which physiologically unsafe hours it covers and which it leaves out, against the 16-year record, by workload class and acclimatisation state.
Open the assistant.
Describe a shift the way you would to a colleague, or ask a heat, WBGT, forecast or rules question for the Gulf. Or read the full method first.