Predicting older people's thermal sensation in building environment through a machine learning approach: Modelling, interpretation, and application

Abstract

There is insufficient knowledge on how environmental and physiological factors affect older people's thermal perceptions. In this paper, we present two data-driven models (a field study model and a lab study model) using the algorithm of random forests to predict older people's thermal sensation. These two models were developed from a field study dataset and a lab study dataset separately. The field study dataset was collected from 1040 old subjects (70 + years) who lived in 19 aged-care homes, which contains multi-dimension factors such as environmental parameters, subjects' demographic information, health condition, acclimatization degrees, living habits and thermal perceptions' votes. The lab study dataset was collected from a lab study and contains 18 old subjects' (65 + years) eight local skin temperatures and thermal perceptions' votes under five thermal environments (21/23/26/29/32 °C). After the procedure of feature selection, the field study model was developed with four environmental variables (air temperature, velocity, CO2 concentration, illuminance) plus two human-related variables (health condition and living time in aged-care homes) as inputs. It produced an overall accuracy of 56.6%, which was 24.9% higher than that of the PMV model. The lab study model was built on five local skin temperatures including head, lower arm, upper leg, chest and back temperatures, which demonstrated an overall accuracy of 76.7%, 30.1% higher than UC Berkeley thermal sensation model's accuracy. We then interpreted how these inputs distinguish thermal sensations by applying a partial dependence analysis. Finally, we proposed two applications of the above models and present older people's seasonally neutral indoor temperature zones.