AI Controller Optimizes HVAC in Tropical Commercial Buildings

A new research paper introduces CQD-ERL (Contextual Quality-Diversity Evolutionary Reinforcement Learning), an advanced controller designed for the supervisory control of Heating, Ventilation, and Air Conditioning (HVAC) systems in tropical commercial buildings. Traditional control systems often struggle to adapt optimally to the dynamic and varied conditions found in such environments, which include fluctuating weather patterns and building loads. Unlike controllers that converge on a single optimal policy, CQD-ERL maintains a diverse archive of specialized policies. These policies are indexed by both data-driven operating contexts (like daily weather and load regimes) and context-invariant behavior descriptors. This allows the system to select the most appropriate policy for the current conditions, optimizing for energy efficiency and occupant comfort. The controller uses a combination of gradient-free evolutionary operators and a soft-actor-critic policy-gradient operator, sharing a single replay buffer for efficient learning. A crucial feature of CQD-ERL is its deterministic safety shield, which filters every action before execution to prevent unsafe operations. Trained on a reduced-order environment representing a Singapore commercial building, the controller was evaluated over a full annual backtest. It demonstrated superior performance compared to an ASHRAE Guideline 36 baseline, showcasing its potential for significant energy savings and improved environmental control in tropical climates.