Non-intrusive load curve disaggregation using sparse decomposition with a translation-invariant boxcar dictionary

We present a non-intrusive load monitoring (NILM) method based on sparse decomposition techniques in order to extract the individual appliance signals from the aggregated load curve of an household. The proposed method is generic and does not need to be adapted for each home. It is based on a translation-invariant boxcar dictionary of atoms, each of them modeling a complete on-off appliance activity event. We evaluated our algorithm on synthetic and real load curve dataset. Experiments showed that we can extract the individual appliance signals with more than 20 dB of signal-to-distortion ratio (SDR), and estimate the energy consumption of the main consumers with a relative error less than 1%. We think that this ability to automatically provide accurate feedback information on the user appliances consumption through a single point of measurement open new perspective in demand-side management.