On the Performance of Data-Driven Reinforcement Learning for Commercial HVAC Control

Commercial heating, ventilation and air conditioning (HVAC ) system consumes large portion of the building energy use. With the abundance of the available data in the building automation systems (BAS) of commercial buildings, ample opportunities have emerged to help develop adequate data-driven control of HVAC systems. This paper proposes the use of data-driven reinforcement learning (RL) that can evaluate control policies and develop new ones. A Q- learning algorithm is used as a type of reinforcement learning to minimize the building energy consumption cost while maintaining the comfort level. The proposed Q-learning algorithm is trained using actual data where the data is first used to develop temperature and energy models. Four different machine learning methodologies are used to obtain these models which are linear regression, deep neural network, support vector machines and random forests. The performance of the Q- learning algorithm under each methodology is tested and compared with the others. The algorithm is validated using a decent physics-based model of a 3-floor office/classroom building. The results showed that though all the four methodologies yield satisfactory results, the Q-learning algorithm performed the best under support vector machine and random forest.