Real-world deployment of deep reinforcement learning-based control for cost-efficient PM_2.5 mitigation in a naturally...

Residential PM_2.5 pollution, particularly from cooking-related emissions, needs to be controlled to reduce adverse impacts on human health. However, existing indoor PM_2.5 control studies based on deep Q-networks (DQN) remain limited to simulations or single-room experiments, with limited real-world validation under high-emission conditions. This study developed a DQN-based controller to intelligently manage windows, air purifiers, and range hoods in a full-scale, multi-room residence to mitigate indoor PM_2.5 while minimizing energy consumption. Baseline and automated control strategies were designed based on typical Chinese household behaviors and default automated mode settings, respectively. Compared with the baseline, the automated strategy achieved PM_2.5 removal efficiencies of 4.48 ± 5.30 %, 49.06 ± 6.54 %, and 43.58 ± 11.25 % in the kitchen, living room, and bedroom, respectively, while the DQN-based strategy improved these to 20.61 ± 16.37 %, 80.61 ± 2.87 %, and 69.43 ± 7.17 % (p < 0.01). The baseline, automated and DQN-based control strategy prolonged the duration of indoor PM_2.5 levels below the WHO 15 μg/m³ guideline to over 38 %, 52 % and 80 %, respectively. The effectiveness-to-energy ratios were 135.26 ± 23.31 μg/(m³·kWh) and 137.67 ± 116.28 μg/(m³·kWh) for the automated and DQN-based control strategies, respectively. The effectiveness-to-cost ratio increased from 19.07 ± 4.46 to 44.51 ± 15.65 μg/(m³·CNY) when only operational cost was considered, and from 12.15 ± 1.34 to 23.79 ± 5.25 μg/(m³·CNY) when total cost was considered. Overall, the proposed DQN-based control strategy effectively balances PM_2.5 mitigation with energy and cost efficiency, providing a scalable, data-driven solution for intelligent indoor air quality management in real residences.