Load oscillation pattern detection for NILM based on scale space decomposition

Abstract

Load pattern detection is the fundamental step for non-intrusive load monitoring (NILM). Some appliances’ power profiles exhibit long-duration cyclic oscillation patterns when performing certain physical tasks, which contain load signatures with strong discrimination for state recognition, but also tend to interfere with the feature extraction of other appliances. However, detecting load oscillation pattern (LOP) is extremely challenging, and the existing NILM methods cannot effectively cope with the multi-scale characteristics of LOPs in unseen scenarios and the overlapping interference of multiple LOPs. To this end, we propose a LOP detection method based on the scale space decomposition theory. LOPs are mined at different time scales, and then are optimally selected based on the minimum description length (MDL) principle. Finally, experiments are carried out on publicly-accessible Pecan Street dataset and private dataset, showing the proposed method generally outperforms two benchmarks in various evaluation metrics.