Voltage Sag Classification Based on Multi-task Parallel Convolutional Neural Network

Abstract

This paper proposed a novel sag classification method that is a parallel convolutional neural network with multi-task learning (PCNN_MTL). Voltage sag events will not only cause a sharp drop in the amplitude of single-phase or multi-phase voltage but also bring about phase changes after propagation. In order to obtain distinguishing feature information, a one-dimensional convolution neural network is employed to extract the distortion characteristics of single-phase voltage, and a two-dimensional convolution neural network is utilized to capture the correlation characteristics between three-phase voltages. The extracted features of them will be fused in the full-connection layer. Finally, the multi-task learning is adopted to classify the sag signals with two classification modes which are the ACD classification and the A~G classification. The experimental results show that the proposed PCNN_MTL achieves good classification effects in both classification modes, and can realize the refined classification of 19 types of sags.