Modeling and Prediction of Personalized Thermal Comfort and Control of HVAC System for Indoor Climate

Thermal comfort is a crucial factor of people’s happiness and work productivity. In today’s predominant indoor lifestyle, people’s thermal comfort is controlled by HVAC (heating, ventilation, air conditioning) system in buildings. HVAC systems also play a major role in energy consumption. In 2011, 50.1% of energy consumption in building operations is caused by HVAC systems. Under the background of global energy crisis, it has substantial potential in energy savings. However, the metric and measurement of thermal comfort in HVAC system is often over-simplified to limited parameters such as temperature. Conventional building energy management also has primitive control strategy, which fails to accurately control the thermal comfort of indoor climate, and leads to energy waste. Traditional comfort models also lack the adaptability to fit individuals’ demand and sensation. To address these problems, the focus of this research is to develop a data-driven thermal comfort model using machine learning algorithms, and corresponding control strategies that can improve the overall thermal comfort of occupants in buildings. A cyber-physical system (CPS) based architecture is used to achieve those goals. The proposed system is examined in several ways. First, a thermal comfort-driven building automation simulation platform is designed and built. A single-space prototype lab room was simulated in EnergyPlus with external control algorithm in MATLAB. Results suggest that overall, compared to a conventional temperature-driven control strategy, the proposed system can minimize thermal comfort violation and improve occupants’ thermal comfort by 22% on average, while energy consumption remains same or is reduced (up to 2% reduction). Secondly, a platform that can simulate occupants with different thermal sensations is constructed to examine the performance of support vector machine (SVM) algorithm and compare with several popular machine-learning algorithms on thermal comfort prediction. The significance of each predictors, influence of training dataset size and adaptability are evaluated. We also proposed a novel hybrid thermal comfort classifier based on support vector machine (SVM) and linear discriminant analysis (LDA) that can improve the efficiency of model training without sacrificing the accuracy. The proposed data-driven prediction model is then paired with a fuzzy-based real-time control system and evaluated in experiment. The proposed method shows improvement in both thermal comfort prediction accuracy, as well as control performance. The difference between experimental result and simulation, particularly the influences of predictors, is also discussed. Furthermore, the scope of this research extended to a large multi-zone, open-space scenario, with occupants generated using adaptive models, using the simulation platform built in first part of the work. The results show 43.41% to 69.93% reduction on thermal discomfort of occupants and energy reduction of 3.63% to 1.01% in four cities