A RL-based human behavior oriented optimal ventilation strategy for better energy efficiency and indoor air quality

In biological cleanrooms for pharmaceutical and biosafety laboratories, escalating cleanliness standards have made the high energy consumption associated with increased air change rates in ventilation systems untenable. Moreover, stringent pressure differential requirements are crucial in biological cleanrooms to ensure the physical isolation of pathogenic microorganisms. This study developed a multi-zone ventilation system model for a biopharmaceutical cleanroom using the Modelica language. A deep reinforcement learning (DRL) model was implemented in Python based on the actor–critic and proximal policy optimization (PPO) algorithms, utilising the Modelica model as the training environment. To maintain particulate matter (PM) concentrations and conserve energy, the DRL control model was trained to adjust air damper positions by identifying patterns in occupancy changes and pollutant concentration dynamics across various times and workspaces within the cleanroom. Results indicated that, relative to the conventional baseline control strategy, the developed reinforcement learning control approach achieved a 14.7 % reduction in energy consumption while maintaining pollutant concentrations within regulatory limits for cleanrooms, culminating in annual energy savings of 11,212.8 kWh. Additionally, pressure fluctuation ranges in the three controlled work zones of the cleanroom were diminished by 59.16 %, 9.58 %, and 29.32 %, respectively. SHapley Additive explanation (SHAP) analysis was employed to elucidate the contributing factors influencing the outputs of the developed DRL control model. Furthermore, the generalisation of the DRL control model was discussed by altering the control period and inner source of the PM.