Decomposition Techniques for Non-intrusive Home Appliance Load Monitoring

Abstract

Energy-saving is a key element of Smart Grid. By encouraging consumers to moderate their energy demands, utilities can make more efficient use of their generation assets, and reduce total fuel consumption. For this purpose, we must provide homeowners with appliance energy consumption data, without requiring sensors on each appliance. This means that energy consumption from the house main feeder must be disaggregated into individual appliances. In this thesis, two novel methodologies for disaggregating household power consumption are evaluated. The first method is multi-label classification, which is used to predict appliance participation in the power signal. The second method is a new signature-based sequence matching algorithm. Two sets of features have been used. In the time domain, a delay embedding of the observed power signal is constructed. The second feature set is a wavelet decomposition of the power signal, using Haar wavelet. We evaluate our techniques and features on two synthetic datasets, and two households from REDD.