Feature Extraction for Non-intrusive Load Monitoring System

Abstract

Compared with traditional intrusive load monitoring system, the non-intrusive load monitoring system only needs to install a monitoring equipment at the entrance of the power supply. In this paper, we focus on the feature extraction for non-intrusive load monitoring system. By extracting features from the monitored voltage and current and training them in machine learning model, we can identify the types and operation states of different loads. Meanwhile, these features reflect the electrical characteristics (e.g., power) and can be used to detect the events (i.e., switch of operation states) for every load. Specifically, the features are obtained via time and frequency domain analysis. In time domain, energy ratio is calculated to detect the events during the operation of loads. Besides, active and reactive power, power factor and current peak coefficient are also calculated. To identify the load with current peak distortion, we innovatively propose a position ratio, defined as the ratio of the ramp-off time over the ramp-up time for a period of current waveform. In frequency domain, harmonic and total harmonic current distortion are extracted. Apart from this, the harmonic ratio such as the ratio of third harmonic to fundamental wave is proposed for load identification. It can be seen that, for resistive load such as heating kettle, the harmonic ratio is quite close to zero; for inductive load such as washing machine, the harmonic ratio is close to 1.