Towards Lifelong Thermal Comfort Prediction with KubeEdge-Sedna: Online Multi-task Learning with Metaknowledge Base

Thermal comfort, achieved by estimating the thermal sensation of occupants, has long been an important research topic. Numerous data-driven models and systems have been developed to improve the estimates of the accuracy of thermal comfort. However, many of these approaches do not have memory: they either require a sufficient number of samples, or require detailed context analysis by data scientists, hindering lifelong maintenance of the system. In this paper, we for the first time study the online multi-task maintenance involved in predicting thermal comfort by training lifelong learning models. We develop a Lifelong Thermal Comfort Prediction (LEON) framework. We introduce metaknowledge and its operators to automatically discover what models are generated, when they should be used and how they are updated. Fortunately, there are ongoing efforts in metadata database development in buildings, e.g., Brick. We build the metaknowledge base with Brick to extract metadata and evaluate our framework using the public ATCII database. We demonstrate that in terms of SMAPE, LEON outperforms all state-of-art baselines by 30.02% on average. Partial code has been adopted and released to KubeEdge-Sedna.