ADAPT Model Optimizes HVAC Energy, Improves Comfort Across Climates

Buildings are major contributors to global energy consumption and CO2 emissions, making efficient HVAC control crucial for climate mitigation. Existing control methods struggle with the complex, delayed thermodynamic responses and partial observability within buildings, especially when needing to generalize across different environments or seasons with limited data. To address these challenges, researchers developed ADAPT, a novel HVAC control model based on physics-aware conditional diffusion. ADAPT incorporates a short-horizon thermal baseline to capture a building's latent thermal inertia and uses a diffusion backbone for robustness. A key innovation is a learnable multi-zone heat-balance regularizer that ensures generated trajectories adhere to building thermodynamics without requiring explicit geometry or manual parameter calibration. Extensive experiments demonstrated ADAPT's effectiveness, showing a 7.3% reduction in HVAC energy consumption and a 30.2% decrease in occupant discomfort compared to state-of-the-art methods in controlled environments. Crucially, ADAPT maintained strong performance with minimal degradation when applied to out-of-distribution scenarios, such as unseen seasons and climate regions, significantly outperforming other methods in transfer robustness.