Feature Selection for Thermal Comfort Modeling based on Constrained LASSO Regression

Abstract

Thermal comfort is influenced by many factors and can vary significantly between different individuals. Therefore, modeling personal thermal comfort is a complex challenge and requires a detailed knowledge about environmental as well as physical or even mental conditions of an occupant. However, only limited data are available which is usually restricted to environmental measurements. Furthermore in the context of commercial buildings, the calculation effort must be kept as low as possible that scaling issues related to the large number of occupants are reduced. To cope with this problem, the presented paper analyzes thermal sensation voting data collected in an open-plan office in Singapore and uses LASSO regression techniques for the identification of the most important comfort features. Well known relations between thermal comfort and the corresponding vote are considered via suitable constraints. To define a common set of optimal features, the individual regression problem is extended to an arbitrary number of occupants. This leads to multiple LASSO optimization problems that are coupled by nonlinear if-statements. A reformulation method is presented which results in a mixed integer quadratic program by introducing binary activation variables. Eventually, the method is applied to comfort modeling and the resulting model structures are compared regarding their complexity, number of selected features and prediction accuracy.