A deep reinforcement learning-based autonomous ventilation control system for smart indoor air quality management in a subway station

•DRL was used to design an autonomous ventilation control system for a subway station.•A DQN algorithm was used to train an agent to regulate IAQ in a subway station.•A box model was calibrated to simulate the IAQ system and generate a training dataset.•State and reward functions targeting IAQ were designed for training a DQN agent.•The proposed method reduced ventilation energy consumption by 14.4% to within a healthy IAQ. Mechanical ventilation has been widely implemented to alleviate poor indoor air quality (IAQ) in confined underground public facilities. However, due to time-varying IAQ properties that are influenced by unpredictable factors, including outdoor air quality, subway schedules, and passenger volumes, real-time control that incorporates a trade-off between energy saving and IAQ is limited in conventional rule-based and model-based approaches. We propose a data-driven and intelligent approach for a smart ventilation control system based on a deep reinforcement learning (DeepRL) algorithm. This study utilized a deep Q-network (DQN) algorithm of DeepRL to design the ventilation system. The DQN agent was trained in a virtual environment defined by a gray-box model to simulate an IAQ system in a subway station. Performance of the proposed method over three weeks was evaluated by a comprehensive indoor air-quality index (CIAI) and energy consumption under different outdoor air quality scenarios. The results show that the proposed DeepRL-based ventilation control system reduced energy consumption by up to 14.4% for the validation dataset time interval and improved IAQ from unhealthy to acceptable. [Display omitted]