A new unsupervised event detector for non-intrusive load monitoring

Abstract

Event detection plays an important role in today's Non-Intrusive Load Monitoring (NILM) systems faced more and more with nonlinear and variable loads. For this purpose, the paper presents an unsupervised NILM event detector based on kernel Fisher discriminant analysis (KFDA) which provides accurate start and end times of so-called active sections. Active sections are an extension of classical NILM events which are introduced to include pulses, variable load intervals and noisy signals which makes the event classification more flexible. The detector achieves good segmentation into steady states and active sections. When applied to the BLUED dataset, the detector yields a recall / precision of 98.78 % / 99.66 % for phase A and 92.17 % / 86.32 % for phase B, respectively.