Lower peaks, similar growth: Isolating the value and limits of artificial-intelligence-driven thermal sensation modeling for building-grid resilience under future climates

Evaluations of artificial intelligence (AI) in building energy control often conflate supervisory policy and occupant prediction. We present a control-isolated EnergyPlus benchmark in which three predictor-driven arms share the same supervisory logic: analytical predicted mean vote (PMV), gradient-boosted trees implemented with the Light Gradient Boosting Machine (LightGBM), and a physics-informed neural-network variational autoencoder (PINN–VAE); a fixed-setpoint controller serves as a separate operational reference. Experiments cover five American Society of Heating, Refrigerating and Air-Conditioning Engineers (ASHRAE) climate archetypes, three 2001–2100 horizons, and Representative Concentration Pathway 8.5 (RCP8.5). The grid metric is the 95th-percentile coincident peak (CP95) of 15-minute heating, ventilation, and air-conditioning electric demand. City/weather realization accounts for 90.6% of fitted main-effect R 2 for absolute CP95 and 90.8% for warming-driven amplification; strategy identity accounts for 5.0% and 1.0%. Against the same-city, same-horizon fixed-setpoint reference, LightGBM lowers CP95 by 18–22 kilowatts (kW) per 10,000 square metres (m 2 ) on average. This is not overall superiority: its PMV-violation rate reaches 84%–94%, PINN–VAE reaches 68%–90%, and a separate annual EnergyPlus summary reaches 2,789 discomfort hours. The pooled stress-index result ( p = 0.203 ) reflects fixed design terms and hence design-induced degeneracy rather than conceptual invalidity; archetype and pathway screens alter absolute magnitudes. This applied-AI benchmark separates predictor value from its engineering application. Model comparison must combine peak and tariff effects with comfort guardrails, while long-term planning remains conditional on building stock and climate uncertainty.