A Wavelet Scattering Approach for Load Identification with Limited Amount of Training Data

The performance of Non-Intrusive Load Monitoring is strongly depending on the choice of features and their ability to uniquely describe the signatures of different appliances. Especially for transfer-learning applications, the invariance of the features with respect to changes in the time domain is crucial. In this article, a two-dimensional wavelet scattering approach for load identification is presented. In detail, time-domain, frequency-domain, and scattering-domain based features are investigated and both 1D and 2D input features are considered, which are evaluated by five different classifiers. It was shown that wavelet scattering outperforms other load identification by 3.7%, while showing robustness to the amount of training data and to the selected sampling frequency.