Balancing indoor air quality and ventilation efforts via deep reinforcement learning: An agent-based approach applied to an office building

• DRL used 3–6x lower ventilation rate in sporadic-use spaces compared to the ASHRAE. • ASHRAE kept mean CO 2 below 1000 ppm but over-ventilated small, highly utilized rooms. • DRL properly adapted ventilation to the stochasticity of occupant-building dynamics. • DRL matched iteratively calculated ventilation while using 4x less computation time. The COVID-19 pandemic highlighted the critical role of Heating, ventilation, and air conditioning (HVAC) systems in mitigating airborne disease transmission within indoor environments. While increasing active building ventilation improves Indoor Air Quality (IAQ), it also leads to higher energy consumption. With buildings projected to account for over 30% of global energy demand by 2050, balancing IAQ and energy efficiency has become a pressing challenge for building designers and operators. However, existing research has predominantly focused on optimizing heating and cooling aspects of HVAC systems for thermal comfort and cost reduction, often neglecting optimal ventilation control for IAQ and energy efficiency. This paper presents a novel approach that combines Deep Reinforcement Learning (DRL) with Agent-Based Modeling (ABM) to optimize an open-loop HVAC ventilation control in office buildings. Unlike standard driven methods that rely on static equations and do not adapt to realistic occupancy patterns, our DRL agent takes into account stochastic user behavior simulated within the ABM environment. Simulation results demonstrate that, at times, the applied ventilation standard exceeds common target carbon dioxide (CO 2 ) levels and tends to over-ventilate on less or sporadically used rooms, leaving potential for energy savings associated with ventilation efforts. A brute-force (BF) approach, optimizing ventilation rates iteratively using the ABM environment, avoided over-ventilation best while maintaining CO 2 levels closest to the desired targets. However, the DRL agent achieved similar performance while requiring substantially less computational efforts than the BF method. This work contributes to the development of intelligent HVAC control systems that can balance IAQ and ventilation efforts better than static ventilation guidelines and thus promotes post-pandemic, sustainable building design and operation.