Effective non-intrusive load monitoring of buildings based on a novel multi-descriptor fusion with dimensionality reduction

Recently, a growing interest has been dedicated towards developing and implementing low-cost energy efficiency solutions in buildings. Accordingly, non-intrusive load monitoring has been investigated in various academic and industrial projects for capturing device-specific consumption footprints without any additional hardware installation. However, its performance should be improved further to enable an accurate appliance identification from the aggregated load. This paper presents an efficient non-intrusive load monitoring framework that consists of the following main components: (i) a novel fusion of multiple time-domain features is proposed to extract appliance fingerprints; (ii) a dimensionality reduction scheme is introduced to be applied to the fused time-domain features, which relies on fuzzy-neighbors preserving analysis based QR-decomposition. The latter can not only reduce feature dimensionality, but it can also effectively decrease the intra-class distances and increase the extra-class distances of appliance features; and (iii) a powerful decision bagging tree classifier is implemented to accurately classify electrical devices using the reduced features. Empirical evaluations performed on three real datasets, namely ACS-F2, REDD and WHITED collected at different sampling rates have shown a promising performance, according to the accuracy and F1 score achieved using the proposed non-intrusive load monitoring system. Reported accuracy and F1 score have reached both 100% for the WHITED dataset, 99.79% and 99.76% for the REDD dataset, and up to 99.41% and 98.93% for the ACS-f2 dataset, respectively. The outstanding performance achieved using the proposed solution determines its effectiveness in collecting individual-appliance consumption data and in promoting energy saving behaviors.