Learning Optimization Friendly Comfort Model for HVAC Model Predictive Control

Heating, Ventilation and Air Conditioning (HVAC) systems perform environmental regulations to provide thermal comfort and acceptable indoor air quality. Recently optimization based Model Predictive Control (MPC) has shown promising results to improve energy efficiency of HVAC system in smart buildings. However rigorous studies on incorporating data driving comfort requirement into the MPC framework are lacking. Previous research on comfort learning usually ignores the restrictions of the downstream control and merely focuses on utilizing existing machine learning tools, which induce undesirable non-linear coupling in decision variables. In this work, we adopt a novel "learning for application" scheme. The idea is to describe user comfort zone by a Convex Piecewise Linear Classifier (CPLC), which is directly pluggable for the optimization in MPC. We analyze the theoretical generalization performance of the classifier and propose a cost sensitive large margin learning formulation. The learning problem is then solved by online stochastic gradient descent with Mixed Integer Quadratic Program (MIQP) initialization. Experimental results on publicly available comfort data set validates the performance of CPLC and the training algorithm. HVAC MPC case studies show that the proposed method enables much better exploitation and seamless integration of individual comfort requirement in the MPC framework.