Gaussian Process Regression for a PMV Prediction Model using Environmental Monitoring Data

Fanger's PMV index has been studied by many researchers as a control parameter for maintaining comfortable indoor environment using HVAC systems. Recently, couple of data-driven modelling techniques have been used for the prediction of thermal comfort. Machine learning based data-driven models for thermal comfort can predict the PMV with indoor parameters and human factors. However, current models still have two limitations. The first is that the input variables used for thermal comfort prediction still utilize Fanger's six variables, which are difficult to measure in real buildings. Hence, it is necessary to develop a data-driven model for predicting the PMV using data that can be easily monitored in real buildings. The second limitation is uncertainty associated to the predicted value of thermal comfort values from a data-driven model. In particular, uncertainty must be considered in real buildings, because the monitoring data from various sensors or measuring instruments are affected by noise or other sources, and lead to errors in the results. Therefore, this study aims to develop a Gaussian process regression (GPR) model for the prediction of thermal comfort, including uncertainty information. As a results, we identified the influences of the environmental factors on the thermal comfort of an occupant in a room and employed them to develop a data driven model using the GPR. The results of the thermal comfort models clearly show that the relative humidity and ambient temperature are major variables in enhancing the accuracy of the prediction. In addition, the 95% confidence interval was also significantly narrowed, indicating that the model uncertainty was reduced.