AI-driven ventilation control policy proximal optimization coupled with dynamic-informed real-time model calibration for healthy and sustainable indoor PM2.5 management

Abstract

Indoor air quality (IAQ) is an important factor for determining quality of life and urban sustainability. In underground subway stations, improving IAQ through ventilation systems remains challenging due to the complexity and nonstationary nature of IAQ resulting from diverse influential factors such as subway schedules, passenger volume, and outdoor air quality (OAQ). Therefore, this study aimed to develop a novel artificial intelligence (AI)-driven ventilation system for healthy and sustainable IAQ management in subway stations. First, an IAQ mechanistic model coupled with genetic algorithm (GA)-driven rolling-horizon calibration was developed from the collected IAQ big dataset, and global sensitivity analysis was then employed to identify the dominant variables in IAQ dynamics. Subsequently, proximal policy optimization (PPO), one of the reinforcement learning (RL) algorithms, was employed to control the ventilation system in both the lobby and platform areas of a subway station. The results demonstrated that the IAQ mechanistic model can capture IAQ dynamics with acceptable modeling performance, achieving around 19 % of mean absolute percentage error (MAPE). Furthermore, the PPO-driven ventilation control system can reduce energy consumption by around 22 % while maintaining IAQ at an acceptable level.