Automatic Label Correction and Appliance Prioritization in Single Household Electricity Disaggregation

Electricity disaggregation focuses on classification of individual appliances by monitoring aggregate electrical signals. In this paper we present a novel algorithm to automatically correct labels, discard contaminated training samples, and boost signal to noise ratio through high frequency noise reduction. We also propose a method for prioritized classification which classifies appliances with the most intense signals first. When tested on four houses in Kaggles Belkin dataset, these methods automatically relabel over 77% of all training samples and decrease error rate by an average of 45% in both real power and high frequency noise classification.