A Feature and Classifier Study for Appliance Event Classification

The shift towards advanced electricity metering infrastructure gained traction because of several smart meter roll-outs during the last decade. This increased the interest in Non-Intrusive Load Monitoring. Nevertheless, adoption is low, not least because the algorithms cannot simply be integrated into the existing smart meters due to the resource constraints of the embedded systems. We evaluated 27 features and four classifiers regarding their suitability for event-based NILM in a standalone and combined feature analysis. Active power was found to be the best scalar and WaveForm Approximation the best multidimensional feature. We propose the feature set [P, cos Phi ,TRI, WFA] in combination with a Random Forest classifier. Together, these lead to F-1-scores of up to 0.98 on average across four publicly available datasets. Still, feature extraction and classification remains computationally lightweight and allows processing on resource constrained embedded systems.