A lifelong meta-learning approach for learning deep grey-box representative thermal dynamics models for residential buildings

Thermal dynamics models of residential buildings are crucial for managing energy use and maintaining desirable indoor environment quality, in the context of decarbonizing the electric grid. Despite recent developments in data -driven models, the implementation with a massive number of buildings in the real world remains infeasible because the computation is expensive and the thermostat data of buildings can be online and sparse. A representative model for aggregated study and acceleration of personalizing models is needed for many downstream tasks. In addition, most existing data -driven models lack physical interpretability questioning the representativeness of any aggregated model. In this paper, we present an online framework for learning meta-models of thermal dynamics across various residential buildings. The meta-models are physically explainable according to the building types and can accelerate fine-tuning personalized models for any building saving data and computation. Experiments suggest that our method performs well with a real -world online dataset that produces representative models of different climate zones and effective initialization for training individual models.