Hierarchical Classification for Symmetrized VI Trajectory Based on Lightweight Swin Transformer

Non-intrusive LoadMonitoring (NILM) is an important means to realize household energy management, and appliance identification is a significant branch of NILM. However, type II appliances, also called multi-state appliances, make it hard to correctly identify the type of appliance. In this paper, hierarchical classification based on swin transformer is proposed to improve the accuracy of appliance identification and reduce the adverse impacts of intra-class variety (IACV) mainly caused by type II electrical loads. By using k-means to pre-cluster target categories into more abstract subclasses, artificial classification operations in hierarchical classification are reduced. Meanwhile, VI trajectories with comprehensive and highly differentiated characteristics are generated, specifically, we skillfully symmetrize VI trajectories and map the higher order harmonic feature into the empty pixels in the background of the VI images for the first time, which improves the network's potential mining for features, and the symmetrical trajectory is more conducive to swin transformer's feature positioning and fine-grained learning through the shifted windowing configuration, and in order to effectively cope with the negative impacts of inter-class variety (IECV) and insufficient feature information in the existing load signatures, we adopt RGB color encoding to fuse multiple features. Compared with the existing methods, the experimental results indicate that our proposed method is more effective on the PLAID and Whited v1.1 datasets. The code is available at: https://github.com/linfengYang/HC_LST_NILM.