Online Comfort-Constrained HVAC Control via Feature Transfer

Policy transfer can reduce energy consumption of the Heating, Ventilation, and Air Conditioning (HVAC) system in a target building without requiring extensive offline data or costly online interaction.However, the energy savings often come at the expense of violating constraints, such as thermal comfort constraints, in the target building.In this paper, we propose a novel approach to augment black-box policy transfer methods for simultaneous minimization of energy consumption and thermal comfort constraint violation.To ensure compliance with these constraints, the algorithm leverages nonlinear feature functions learned from historical data of source buildings and adapt these features to approximate the target building dynamics via online learning in a sample efficient manner.We present an upper bound on the number of time steps during which thermal comfort violations may occur under worst-case disturbances, and also establish a finite stopping time beyond which all constraints are consistently met, assuming certain stochastic properties of disturbances.Additionally, we evaluate the algorithm using the EnergyPlus simulator across 19 climate locations in January and July.OCFT reduces the total number of constraint violation of the black-box policy transfer algorithm by an average of 81.28% while resulting in 11.23% more energy consumption on average.