A Graph-Based Semi-Supervised Learning Approach Towards Household Energy Disaggregation

Non-Intrusive Appliance Load Monitoring has drawn increasing attention in the last few years. Many existing studies that use machine learning for this problem assume that the analyst has access to the actual appliances states at every sample instant, whereas in fact collecting this information exposes consumers to severe privacy risks. It may, however, be possible to persuade consumers to provide brief samples of the operation of their home appliances as part of a "registration" process for smart metering (if appropriate financial incentives are offered). This labeled data would then be supplemented by a large volume of unlabeled data. Hence, we propose the use of semi-supervised learning for non-intrusive appliance load monitoring. Furthermore, based on our previous work, we model the simultaneous operation of multiple appliances via multi-label classification. Thus, our proposed approach employs semi-supervised multi-label classifiers for the monitoring task. Experiments on publicly-available dataset demonstrate our proposed method.