From one building to many: transferability of a deep reinforcement learning agent for optimizing pollutant exposure and energy consumption

Minimizing indoor pollutant exposure while conserving energy is essential for protecting human health and the environment. Deep reinforcement learning (DRL) has emerged as a promising approach for optimizing residential ventilation and air conditioning systems. While DRL deployment is simpler than fully physics-driven strategies like dynamic optimization (DynOpt), its generalizability across diverse buildings and ambient conditions remains challenging. Although researchers have studied transfer and imitation learning techniques to address these challenges, they still require house characteristics and field measurements to adaptively train an agent. Therefore, the large-scale deployment of DRL agents can still be potentially challenging. This study assesses the performance of a trained DRL agent against the DynOpt (benchmark) when transferred to houses with varying characteristics and environmental conditions using digital twins. When varying house characteristics one at a time, the agent's performance remained comparable to DynOpt, with particulate matter (PM) exposure and energy ratios near unity (1.05 ± 0.03). Similarly, under simultaneous variations in house characteristics, the exposure (1.03 ± 0.07) and energy (1.09 ± 0.06) ratios remained close to one. However, the agent's performance declines in houses with high PM infiltration under high ambient parameters. The results indicate that the agent can still be integrated into different houses under varying ambient conditions by restricting the infiltration of PM, as evident by lower exposure and energy ratios in houses with lower infiltration. Moving forward, uncertainty quantification and benchmarking of the agent's performance are critical for enhancing confidence in predictions.