Graphs for Scalable Building Decarbonisation: A Transferable Approach to HVAC Control (Proposals Track)
Anaïs Berkes (University of Cambridge); Donna Vakalis (Mila); David Rolnick (Mila); Yoshua Bengio (Mila)
Abstract
Direct building CO2 emissions need to halve by 2030 to get on track for net zero carbon building stock by 2050. Buildings consume 40% of global energy, with HVAC systems responsible for up to half of that demand. Limiting global warming to 1.5°C requires immediate deployment of scalable building efficiency solutions. However, current approaches fail to scale. We introduce HVAC-GRACE (Graph Reinforcement Adaptive Control Engine), the first graph-based RL framework for building control that enables zero-shot transfer by modeling buildings as heterogeneous graphs and integrating spatial message passing directly into temporal GRU gates. Our architecture supports zero-shot transfer by learning topology-agnostic functions. Our framework is the first to meet the fundamental requirement for scalable, transferable building control that could enable rapid climate impact across the global building stock.