TR2021-149

Meta-Learned Bayesian Optimization for Building Model Calibration using Attentive Neural Processes


Abstract:

Currently, building model calibration algorithms do not leverage data archived from previous related calibration tasks. In this paper, we propose the use of Attentive Neural Processes (ANPs) to meta-learn a distribution across previously seen calibration tasks, which is used to accelerate Bayesian Optimization-based calibration of the unseen target task. Our proposed MetaBOAN algorithm is demonstrated on a library of residential buildings generated by the United States Department of Energy. The experiment results show the significantly improved data efficiency in model calibration.