A domain adaptation technique for fine-grained occupancy estimation in commercial buildings

Abstract

Fine-grained occupancy information is essential to improve human experience and operational efficiency of buildings, yet it is quite challenging to obtain this information due to the lack of special-purpose sensors for occupancy monitoring, and insufficient training data for developing accurate data-driven models. This paper addresses this challenge by (a) utilizing recurrent neural network models to uncover latent occupancy patterns in individual rooms from trend data available through the building management system, and (b) applying a domain adaptation technique to transfer existing occupancy models trained in a controlled environment (i.e., the source domain) to another environment (i.e., the target domain) where labelled data is sparse or non-existent. We adjust the model parameters based on the apparent differences between the two environments and apply the adapted model to estimate the number of occupants in the target domain. Our results from two test commercial buildings in two continents indicate that the adapted model yields only slightly lower accuracy than a model that is originally built on the target domain given a large amount of labelled data. Furthermore, we study how much labelled data is required from the target domain for the semi-supervised domain adaptation technique to achieve promising results.