Adaptive control for building energy management using reinforcement learning

Efficient energy management of building operation shall consider the individual and time variant characteristics of the building and its systems to maximize the potential energy savings without compromising the comfort level of occupants. Model-free control approaches, such as Reinforcement Learning, process building operation data to find control actions to operate the building systems while integrating seamlessly into their decisions changes in the building dynamics. These methods, however, do not scale well to complex problems due to the curse of dimensionality, which limits their practical applicability. To address the state explosion problem we propose a Reinforcement Learning controller for a two zone building model that gets state approximation inputs from an Artificial Neural Network. The results show that the system is able to maintain comfort levels while achieving significant energy gains by finding untapped potential for energy performance improvements.