Reinforcement Learning for Model-Free Linear Quadratic Control of Building HVAC Systems

A novel model-free approach to the control of variable-air-volume (VAV)-HVAC systems for the regulation of thermal comfort in rooms of a non-residential building where the number of occupants may vary considerably during the day is proposed. Specifically, a reinforcement learning control algorithm based on model-free optimal linear quadratic control is developed. The quality function <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$Q$</tex>, derived from Bellman dynamic programming, is used to develop a learning control algorithm based solely on the time series generated by the building system, including its dynamics and occupants. Simulations are carried out on a new HVAC-VAV system installed in a university building, demonstrating the potential of the proposed method for maintaining the comfort of room occupants while optimizing the airflow demand of VAV boxes, which is correlated with the energy consumed per room.