ADAPT diffusion world model cuts HVAC energy use by 7.3%

A diffusion-based world model trims HVAC energy use by 7.3% and cuts occupant discomfort by 30.2% against state-of-the-art baselines in simulated tests, according to a new arxiv paper by Xu Yang, Kailai Sun, Dianyu Zhong and Qianchuan Zhao. The system, called ADAPT, is pitched as a controller that generalizes across climates without hand-calibrated thermal parameters, in a domain where buildings 'account for roughly one-third of global energy consumption and CO₂ emissions.'

ADAPT plugs a conditional diffusion model into a reinforcement-learning loop, then constrains the generated thermal trajectories with what the authors call a 'learnable multi-zone heat-balance regularizer' that 'satisfy transferable building thermodynamics without requiring known building geometry or manually calibrated thermal parameters.' The 7.3% and 30.2% gains come from the SemibuildingSim and Sinergym simulators under in-distribution control. On out-of-distribution scenarios spanning 'unseen seasons and climate regions,' the paper reports the model 'maintains robust performance with only marginal degradation.'

Two researchers on our watchlist shared the link. The abstract names no specific baselines, reports no real-building deployment, and gives no per-simulator breakdown; every figure is from simulation.