Predicting individual thermal preferences in an office: Assessing the performance of mixed-effects models

Multiple studies suggested that existing thermal comfort models inadequately predict occupants' individual thermal preferences. Personalised comfort models offer an alternative to conventional comfort models aiming to forecast individual's thermal preference. Implementation of these personalised models in occupant-centric control of heating, ventilation, and air-conditioning (HVAC) systems can enhance their performance. A promising technique for personalised comfort modelling is mixed-effects (ME) modelling. A ME model accounts for fixed effects, representing the trends in the general sample, and for random effects, representing variations of specific clusters in the data. In contrast to fixed-effects (FE) models, ME models can capture individual differences. However, its effectiveness in predicting occupants' thermal preferences based on field measurement data, as well as the influence of variations in ME models on prediction accuracy, remains to be thoroughly investigated. This study aims to assess the prediction accuracy of ME models in contrast to FE models using field measurement data, including thermal preference votes from 30 unique occupants. The prediction performance was evaluated across three testing scenarios, each representing a different application of the models. Furthermore, two random effect structures were tested for the ME model: an intercept-only model and an intercept and slope model. The results show that ME models, in comparison to FE models, achieve an improved prediction performance of 8.0 % on average and up to 28.4 % for individual occupants. Moreover, the addition of a random slope to the ME resulted in deteriorated predictions. Finally, occupants' individual variations were determined with an uncertainty of 6 % after 20 observations.