Heating, ventilation, and air conditioning (HVAC) systems are the primary means of regulating indoor air quality (IAQ) and thermal environments. However, inappropriate HVAC control methods and equipment damage consume more energy and increase discomfort for indoor occupants. Traditional methods, such as field measurements and numerical simulation, are inefficient to optimize the HVAC control parameters. Artificial intelligence (AI)-based methods offer a promising alternative solution for efficient building control. However, due to the complexities of HVAC systems and indoor occupant activities, it is challenging to develop control methods for different building HVAC systems. To tackle these problems, this research develops an integrated framework that enhances the reliability and flexi...[ Read more ]

Heating, ventilation, and air conditioning (HVAC) systems are the primary means of regulating indoor air quality (IAQ) and thermal environments. However, inappropriate HVAC control methods and equipment damage consume more energy and increase discomfort for indoor occupants. Traditional methods, such as field measurements and numerical simulation, are inefficient to optimize the HVAC control parameters. Artificial intelligence (AI)-based methods offer a promising alternative solution for efficient building control. However, due to the complexities of HVAC systems and indoor occupant activities, it is challenging to develop control methods for different building HVAC systems. To tackle these problems, this research develops an integrated framework that enhances the reliability and flexibility of HVAC system control operations from data reconstruction and indoor occupancy states and optimizes HVAC control methods from a room level to a building level.

Firstly, a neural method consisting of the variational autoencoder, and long short-term memory (VAE-LSTM) network is proposed for early fault detection and fault data reconstruction to enhance the sensor data reliability in HVAC system control. Based on the periodicity and stable fluctuation of IAQ data, the analysis of variance and average change rate are conducted to detect abnormal variations before failure occurs. The IAQ dataset is corrupted by introducing complete failure, bias failure and precision degradation fault and removing data intervals with different missing rates to verify the feasibility of the VAE-LSTM model.

Secondly, we develop a novel occupancy-driven (OD-LSTM-DQN) HAVC control framework, which optimizes HVAC control operations by considering the indoor occupancy state. By analyzing the relationship between indoor carbon dioxide (CO₂) levels and occupant behavior, an occupancy state estimation method is developed using the change point analysis method (B-G algorithm). Based on the estimation analysis of indoor occupancy states, we develop the OD-LSTM-DQN control method that integrates Long Short-Term Memory (LSTM) networks with Deep Q Networks (DQN) to reduce unnecessary operations of HVAC systems.

Thirdly, we develop a rapid prediction and optimization framework of IAQ, occupant comfort, and energy consumption. Building Information Modeling (BIM) technology and Computational Fluid Dynamics (CFD) simulations are used to create the database containing indoor air velocity, temperature, CO₂ concentration and ventilation parameters under different distributions of occupants. Based on the database, the extreme learning machine (ELM) model optimized by the grey wolf optimizer (GWO) algorithm is developed to predict the thermal comfort level and CO₂ concentration. The input parameters of the prediction models are interpolated to generate more cases. The optimal air supply parameters for various optimization objectives are determined by combining the predicted CO₂ concentration and PMV value with the energy consumption analysis.

Last but not least, a multi-source transfer learning and deep reinforcement learning (MTL-DRL) integrated framework is developed for efficient HVAC system control through utilizing control experiences from different source domains. In order to select appropriate source domains, the contribution of various source domains to the target task is quantified, followed by a comprehensive evaluation of the transfer performance of each source domain in terms of average energy consumption and average temperature deviation. The well-pretrained DRL parameters from the optimal multi-source transfer set are then sequentially transferred to the target DRL controller. We also discuss the impact of the source domain transfer sequence on the DRL-based HVAC control performance.

All the developed approaches are validated with relevant experiments and scenarios, demonstrating superior performance to conventional methods. It is expected that this research will improve the energy efficiency of HVAC system control while ensuring the indoor air quality and thermal comfort of indoor occupants.