Exploiting Breadth in Energy Datasets for Automated Device Identification

Abstract

The recent explosion of interest in smart building energy-efficiency has led to a proliferation of public energy datasets. Most of these datasets focus on depth (i.e., many devices in a few buildings) as opposed to breadth (e.g., a few devices in many buildings), and thus most smart building algorithms are evaluated on depth-oriented datasets. We argue that increasing data breadth conveys important benefits that are not easily achieved by even a large quantity of deep data. As an illustrative case study, we consider the problem of classifying previously unseen appliances using an off-the-shelf classifier trained on known instances of other devices. Our experiments on multiple real-world datasets (both depth- and breadth-oriented) demonstrate significant and sustained benefits from increased data breadth, and point to the importance of incorporating greater breadth into similar techniques that rely on generalized electrical load models.