A Novel Method for Multiple Power Quality Disturbances Classification Using a Multi-Task Convolution Neural Network

Abstract

In this paper, we proposed a multi-task convolution neural network (MT-CNN) to realize the multi-label classification of multiple power quality disturbances (MPQDs). According to the characteristics of PQD signals, the multiple labels of MPQDs are assigned to three groups corresponding three learning tasks in MT-CNN. In the process of training, these tasks can help each other and the label correlations among various PQDs are utilized in the joint learning of interrelated tasks. Further, since each group can be designed different network structure, the MT-CNN can extract more discriminative features and obtain better recognition rate compared with traditional CNN. In addition, due to the special network structure, the MT-CNN has very strong ability to resist over-fitting. Extensive experiments have demonstrated that our network had better performance and it can greatly improve the accuracy rate for identifying MPQDs under different SNR conditions.