Real-time hyperlocal indoor heat-risk forecasting and equity-aware triage for vulnerable housing during extreme heat events.

---

Heatwaves kill more people than any other natural hazard:

• European 2003: ~70,000 deaths
• Chicago 1995: ~700 deaths in 5 days
• Pacific Northwest 2021: ~1,400 deaths in 3 days

Almost all of these deaths happen indoors, in homes whose temperature lags the outdoors by 12–24 hours and depends on building physics, AC access, and occupant vulnerability. Existing heat-health warning systems issue outdoor forecasts. ThermoGuard closes the gap.

The paper is at paper/thermoguard.pdf.

Quick start

pip install -e .
pytest -q tests/         # 9 unit tests, ~4s
make reproduce           # all 3 experiments, ~3 min on CPU
make paper               # build paper/thermoguard.pdf

What it does

1. Building thermal digital twin: a 2R-2C lumped-capacitance model parametrised by five residential archetypes spanning the documented vulnerability stratification.
2. Hybrid physics+ML forecaster: physics RC backbone + closed-form ridge residual learner; runs on a single CPU, deterministic latency.
3. Per-resident heat-mortality risk model: Steadman heat index + Stull WBGT + age, chronic conditions, mobility, AC access, social isolation, income quintile.
4. Equity-aware triage optimiser: allocates limited intervention budgets (cooling visits, hydration kits) with a per-stratum equity floor.

What's in the box

Path	Purpose
thermoguard/building.py	2R-2C physics + 5 archetypes
thermoguard/weather.py	Synthetic heatwave templates (PNW 2021, EU 2003, Chicago 1995)
thermoguard/data.py	Synthetic city generator with vulnerability mix
thermoguard/forecast.py	Physics / ML / Hybrid forecasters
thermoguard/risk.py	Heat-index, WBGT, vulnerability, mortality
thermoguard/triage.py	Random / greedy / equity-aware allocators
tests/	9 unit tests, all green
experiments/forecast_accuracy.py	Three forecasters across three events
experiments/risk_stratification.py	Mortality concentration analysis
experiments/triage.py	Lives-saved vs coverage tradeoff
scripts/make_figures.py	Build figures from results
paper/thermoguard.tex	Paper source
paper/figures/*.pdf	Real figures from real data
results/*.json	Raw experiment records

Headline results

Experiment	Result
Indoor temp under PNW 2021 (low-income no-AC)	peaks at 44 °C (real epidemiology)
Indoor temp under PNW 2021 (care home top floor)	peaks at 52 °C (lethal)
Forecast MAE physics vs ML vs hybrid (PNW, no-AC)	0.40 / 1.94 / 0.54 °C
Forecast MAE on Chicago 1995 (care home)	physics 2.07 → hybrid 1.69 (humidity nonlinearity)
Mortality per 100k residents	PNW 484 / EU 1,007 / Chicago 1,823
Top 25% of households / total deaths	70–82% (long tail)
Greedy vs equity-aware lives saved (PNW, budget 80/160)	5.43 vs 3.46
Min stratum coverage greedy vs equity-aware	0.00 vs 0.11

Honest limitations

• Synthetic ground truth. Forecast accuracy uses the 2R-2C simulator as ground truth; real-world validation on Ecobee Donate-Your-Data or Quebec smart-thermostat dataset is the next step.
• Linear residual learner. Closed-form ridge; swappable for a tree ensemble or LSTM.
• Risk-model calibration. Hazard constant tuned to reproduce documented mortality; real deployment needs local coroner data.
• No external mortality validation. PNW 2021 / Chicago 1995 zip-code mortality data exists; we did not retrieve it for this v0.1.
• Equity floor is a policy choice. ThermoGuard quantifies the efficiency-vs-coverage curve but does not impose a value.

License

MIT.

Citing

@misc{debes2026thermoguard,
  title  = {ThermoGuard: Real-Time Indoor Heat Risk Forecasting
            for Vulnerable Housing under Extreme Heat Events},
  author = {Debes, Anwar},
  year   = {2026},
  note   = {Reference implementation v0.1, May 2026}
}