Deep Learning-based Electric Appliances Identification from their Switching-On Current Waveforms

Yassine Chemingui, Adel Gastli, Mahdi Houchati

Abstract:

The field of non-intrusive load monitoring offers a multitude of methods for investigating and diagnosing energy demand
per appliance. Thus, energy-aware strategies can be derived and implemented. With the widespread of smart meters, the rich information
of the main current variation is within reach for many households. Through continuous analysis of the main current waveform, switchingon loads can be identified, and energy-saving practices can be devised. This paper proposes a deep learning model, a Convolutional
Siamese neural network for appliance classification based on the WHITED raw high-frequency current dataset. The model is trained on
pairs of appliance, measuring their similarity. Based on that, the appliance is identified. With minimal data preprocessing, an F1 macro
measure of 0.95 was achieved on the training appliances, and a 0.79 score on previously unseen devices.