Reinforcement Learning Based Monitoring and Control of Indoor Carbon Dioxide Concentration Integrating Occupancy Presence

摘要

Carbon dioxide (CO_2) concentration has long been recognized as a most representative indicator for indoor air quality (IAQ). Real-time monitoring CO_2 concentration and further developing a proactive mechanical ventilation control policy contribute to the trade-off between indoor air quality and energy conservation for buildings. As the indoor CO_2 concentration exhibits strong randomized behavior due to incomplete modeling of exogenous environmental factors and occupancy, this paper proposes a model-based reinforcement learning approach to determine an optimal strategy to ventilate in buildings mechanically. Instead of using a simplified physical model, a Gaussian process modeling based probabilistic dynamics describing indoor CO_2 concentration evolution is learned from past on-site observations, including outdoor CO_2 concentration, window open and close status, and occupancy. Supported by the learned dynamics model, the optimal mechanical ventilation rate is intelligently adjusted by evaluating the expected consequences through a radial basis function (RBF) neural network based controller. Using a data efficient approach to simulate and control the indoor environment means that a controller can be learned with very few interactions with the real system, which avoids sacrificing the occupants comfort in the early stage of application. Finally, a single zone building is simulated to verify the effectiveness of the proposed method to maintain indoor air quality to a comfortable condition. The results demonstrate that the model-based reinforcement learning could prove a valuable tool to integrate stochastic occupancy whilst improving indoor air quality.