Unsupervised disaggregation of appliances using aggregated consumption data

Non-Intrusive Load Monitoring (NILM) is a technique that determines the electrical load composition of a household through a single point of measurement at the main power feed. In contrast with the majority of the existing approaches to solve this problem which require training, here we explore an unsupervised approach to determine the number of appliances in the household, their power consumption and state, at any given moment. We attempt to achieve this without using any a priori information on the number and type of appliances. Our approach is to first create clusters of steady-state changes and then employ a matching pursuit algorithm to reconstruct the original power signals using the clusters that were found as the sources in a linear blind source separation strategy. Changes in steady-state, sometimes referred to as events, are characterized by their change in real and reactive power (P and Q). Ultimately, the results may be applied to other features in an attempt to improve the separation between clusters. The preliminary results point toward a mixed scenario: large appliances (roughly above 400W) were easily identified, but the small appliances typically clustered together and were difficult to separate. We conclude that the errors occur during clustering which indicates that, in order to increase the purity of the clusters, perhaps other features could be used.