DUET: Towards a Portable Thermal Comfort Model

Thermal comfort, achieved by estimating the thermal sensation of occupants, has long been an important research topic. Numerous models and systems have been developed to improve the estimates of the accuracy of thermal comfort. Many either require extra devices to be installed; or require occupants to provide frequent feedback hindering the large scale deployability of the system. Data-driven models separate the process of collecting data used to establish the thermal comfort model from the process of deploying the model, making these models portable in deployment. Recent studies on data-driven thermal comfort models often make use of a single model. A single model can introduce large errors in practice, as thermal comfort is highly dependent on a variety of contextual factors, such as building type, location, and so on. In this paper, we for the first time study the contextual adaptation involved in predicting the thermal comfort of individuals by training multi-task models. We develop a Dynamic MUlti-task PrEdiction on Thermal Comfort (DUET) model. A key idea of our model is to use metadata to automatically define multi-task. Fortunately, there are ongoing efforts in metadata development in buildings, e.g., Brick. We extract metadata from Brick and evaluate our model using the public ASHRAE dataset. We demonstrate that in terms of error rate, DUET outperforms PMV model by 39% and STL by 31%.