Answers are grounded in this paper and cite their sources.
The coordinated control of Heating, Ventilation, and Air Conditioning (HVAC) systems and windows is widely studied to maintain thermal comfort while reducing energy consumption. However, traditional Reinforcement Learning (RL) methods typically require extensive training data and prolonged training, hindering fast adaptation to new climatic conditions. To address this, we propose a Meta-Reinforcement Learning (Meta-RL) framework for HVAC and window coordinated control that reduces dependence on data scale and training cost in new environments. An optimal source-domain data selection strategy based on Wasserstein distance is proposed to select the most climatically similar subset from cross-climate multi-city datasets as the optimal pre-training dataset. Offline meta-training on the selected dataset extracts generalizable initial policy parameters encoding shared control knowledge across climate zones; these meta-parameters are then transferred to unseen target cities for rapid online adaptation of the coordinated HVAC and window control policy. Evaluated on the ASHRAE Global Occupant Behavior Database across unseen target cities (Odense, Mumbai, Singapore), the proposed optimal-selection method reduces target-city adaptation episodes by 99% versus conventional RL trained from scratch, improves thermal comfort rate by 6.84 percentage points, and decreases energy consumption by 46.88%. Compared with Meta-RL trained on the full source dataset, optimal selection reduces required training data by up to 91%, while further improving comfort rate by 4.06 percentage points and reducing energy consumption by 18.36%.
No takes yet. Share an insight, caveat, or question.
Huang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: