Semi-supervised Energy Disaggregation Framework using General Appliance Models

Abstract

Providing detailed appliance-level energy consumption information helps consumers to understand their usage behavior and encourages them to optimize their energy usage. Non-intrusive load monitoring (NILM) or energy disaggregation aims to estimate appliance-level energy consumption data from the aggregate consumption data of households. NILM algorithms can be broadly classified into supervised and unsupervised (or semi-supervised) techniques. The former requires a large amount of prior data for each appliance and the latter relies on manual tuning of models of appliances based on some metadata information. While there is a significant interest from academia and industry, NILM techniques are still not adopted widely across households. This is mainly because the techniques developed for one household cannot be generalized and applied in other households (applicability), require tremendous manual-tuning to apply across households (scalability), and cannot run in real-time. To overcome the above issues, we propose a novel semi-supervised energy disaggregation framework – UniversalNILM. The key idea of UniversalNILM is to model appliances in a few (3-10) training houses, which has detailed appliance-level data and transfer this learning to test houses (blind disaggregation), which has only aggregate house consumption data to derive fine-grained appliance energy consumption. To this end, we develop an automated appliance modeling technique that creates general appliance models across various appliance brands and models. The general appliance models are analytical models which describe power consumption of each appliance. These general appliance models are then fine-tuned automatically on test houses to accurately disaggregate the energy consumption in real-time. To test the robustness of UniversalNILM, we empirically evaluated it across three publicly available real-world datasets. We show that the general appliance models learnt on a few households is able to accurately disaggregate on unseen test houses in the same dataset, as well as unseen houses from different datasets. This is the first work in NILM which is able to perform disaggregation across datasets. Another improvement is that UniversalNILM outperforms the reported accuracy from both state-of-the-art supervised and unsupervised NILM techniques.