Appliance Identification in NILM Through Image or Signal Features: a Performance Comparison

To implement a NILM technique, feature extraction, and appliance identification are two essential steps. Many state-of-the-art approaches are based on image-based features and deep neural networks. However, they require high model complexity and large training costs. This limits the practical application of such approaches to NILM. In addition, the need for complex image features to achieve accurate appliance identification is debatable. To better understand this issue, this paper compares the performance of a Convolutional Neural Network (CNN) classifier based on various image features with the performance of a simpler and more classic appliance identification approach based on electrical signal features. Specifically, three kinds of image features are considered, i.e., the V-I trajectory, the Weighted Recurrence Graph (WRG), and the Gramian Angular Field (GAF). Multiple tests were performed using the PLAID2017, COOLL, and WHITED public datasets. Our results show that the use of image features has no significant advantage over more classical approaches based on electrical signal features. In addition, increasing image resolution leads to negligible performance improvements, while greatly increasing computational burden and training time.