Data-driven thermal comfort model via support vector machine algorithms: Insights from ASHRAE RP-884 database

Abstract

Many models can predict building occupants’ thermal comfort, but their accuracies were not always perfect due to the limited self-learning and self-correction capability when varying the application contexts. Advances in machine learning algorithms allow us to reveal the “hidden insights” behind a large amount of data, offering a great opportunity to understand more nuanced aspects of thermal comfort in buildings. This study applied the support vector machine (SVM) algorithm to the RP-884 thermal comfort database and developed a new model with self-learning and self-correction ability. We identified its application range according to the features of the SVM algorithm and the sample distribution characteristics of RP-884. With variables of indoor air temperature, clothing insulation, metabolic rate, air velocity and so forth, the model can largely reduce the previous models’ inaccuracy. Compared to the PMV model, the new model's sum of squares for residuals (SSE) reduced by 96.4%, and the fitting degree (Rnew) increased by 83.7%. It can also quantify the effects of each input variable on building occupants’ thermal sensation. Instead of using two separate models, the data-driven model can automatically distinguish occupants’ thermal comfort responses in natural ventilation (NV) and air-conditioning (AC) buildings. Using the new model, we determined thermally comfortable zones on the psychrometric chart for NV and AC buildings. Moreover, an open-access platform has been developed to help apply machine learning algorithms in thermal comfort data analysis. The work introduced in this paper can be a reference for a more comprehensive comfort model development.