Reference implementation accompanying the paper.
Heatwave deaths happen indoors, and they concentrate: in the simulated cities, the top quarter of households accounts for 70-82% of mortality. ThermoGuard v0.1 ships everything needed to reproduce that finding and act on it:
- 2R-2C building thermal twins for five residential archetypes.
- Physics, ML, and hybrid indoor-temperature forecasters (hybrid wins where humidity nonlinearity bites: Chicago 1995 care home MAE 2.07 -> 1.69 °C).
- Per-resident heat-mortality risk from heat index, WBGT, and six vulnerability factors.
- Equity-aware triage of limited cooling budgets, with the greedy-vs-floor tradeoff quantified rather than assumed.
- Synthetic heatwaves for PNW 2021, EU 2003, Chicago 1995;
make reproduceruns all three experiments in about 3 minutes on CPU.
Known limitations are listed honestly in the README (synthetic ground truth, linear residual learner, no external mortality validation yet).