A Comprehensive Feature Study for Appliance Recognition on High Frequency Energy Data

Awareness about the energy consumption of appliances can help to save energy in households. Non-intrusive Load Monitoring (NILM) is a feasible approach to provide consumption feedback at appliance level. In this paper, we evaluate a broad set of features for electrical appliance recognition, extracted from high frequency start-up events. These evaluations were applied on several existing high frequency energy datasets. To examine clean signatures, we ran all experiments on two datasets that are based on isolated appliance events; more realistic results were retrieved from two real household datasets. Our feature set consists of 36 signatures from related work including novel approaches, and from other research fields. The results of this work include a stand-alone feature ranking, promising feature combinations for appliance recognition in general and appliance-wise performances.