An Efficient Method for Learning Personalized Thermal Preference Profiles in Office Spaces

Conference Year

July 2018

Authors

Seungjae Lee, School of Civil Engineering, Purdue University, United States of America; Center for High Performance Buildings, Ray W. Herrick Laboratories, Purdue University, United States of AmericaFollow
Panagiota Karava, School of Civil Engineering, Purdue University, United States of America; Center for High Performance Buildings, Ray W. Herrick Laboratories, Purdue University, United States of AmericaFollow
Athanasios Tzempelikos, School of Civil Engineering, Purdue University, United States of America; Center for High Performance Buildings, Ray W. Herrick Laboratories, Purdue University, United States of AmericaFollow
Ilias Bilionis, School of Mechanical Engineering, Purdue University, United States of AmericaFollow

Keywords

Bayesian modeling, Clustering, Learning thermal preference, Personalized preference profile

Abstract

Incorporating occupant preferences in sensing and control operations of thermal systems has the potential to maximize thermal satisfaction and to supply energy when and where it is needed. Learning personalized thermal preferences is an essential part in this process. The latest studies have shown that it is possible to learn personalized thermal preference profiles using machine-learning and data-driven modeling algorithms. However, adequate data is required for each occupant, which is challenging in real buildings. This study presents: (i) a data-efficient method for learning personalized thermal preference profiles, based on Bayesian formalism, which shows good performance even with unobserved variables (i.e., air speed, metabolic rate, and clothing insulation level) and (ii) an efficient occupant feedback collection algorithm that minimizes disturbance of occupants. An experiment was conducted in identical perimeter offices, where human subjects were exposed to various thermal conditions and reported their respective thermal preference votes. Air temperature, globe temperature, and relative humidity were monitored continuously during the experiment. The collected data were used to infer occupants’ personalized thermal preference profiles, which showed better prediction performance and faster learning speed compared to existing methods. The method presented in this paper enables integration of occupant thermal preferences in smart environmental control systems of office buildings.