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

The objective of this work is to develop and demons trate an efficient Bayesian inference algorithm to learn individual occupants’ thermal preferences in office buildings. We present an experimental study to col le t data representative of thermal comfort delivery conditio ns for which the algorithm would be implemented in actual buildings. Subsequently, we demonstrate the efficie ncy of our algorithm by showing the evolution of pe rsonalized thermal preference profiles as the training data si ze increases and by evaluating profiles inferred wi th limited data. The results show more reliable personal profiles wi th our approach when the training data are limited compared to typical learning approaches, training a model from scratch with maximum likelihood estimation.