A Semi-Supervised Deep Transfer Learning Architecture for Energy Disaggregation

Abstract

Energy disaggregation is the problem of estimating individual load consumption profiles from an aggregated waveform. Most existing methods ignore the discrepancy of load profile distributions within different sources, which may lead to robustness issues. Furthermore, limit amount of labeled data is a bottleneck for all existing literature. This paper proposed a deep transfer learning based method for energy disaggregation. Each load has its own disaggregtor, which consists of a feature extractor and a regressor. Unlike existing methods, a semi-supervised learning using deep domain adaptation is proposed, which can align the energy data distribution to some extent by utilizing both the labeled and the unlabeled data. Tests were carried out on a publicly available dataset. It can be shown that the proposed architecture can effectively improve energy disaggregation performance and enhance the robustness.