Series arc fault detection method based on a residual deformable convolutional network for complex branch circuit

Abstract

When a series arc fault (SAF) occurs in one branch of a low-voltage alternating current power distribution system with complex connections and many types of loads, the load branches interact with one another, and thus, detection becomes more difficult. To avoid the occurrence of electrical fire, an SAF detection method based on a residual deformable convolutional network (RDCN) is proposed. First, an arc fault data acquisition experimental platform for low-voltage complex branches is built to measure trunk current signals during normal operation and SAF occurrence. Second, 1D current signals are mapped to 2D space as input to the model, more comprehensively responding to the amplitude and variation of the signals. Deformable convolutional networks are used to extract spatial distribution information from feature maps to avoid the limitation posed by the fixed shape of convolutional kernels and improve spatial differentiation among different data. Considering the ability to focus better on fault information, channel attention is introduced to strengthen the correlation among feature channels. Then, the experimental platform data verify that the method can effectively distinguish whether SAF occurs and the type of load where the fault occurs, with the highest accuracy of up to 98.98% and 98.84%, respectively, and an average test time of 1.8 s in determining whether a fault occurs in a six-branch circuit. Finally, compared with existing network models, RDCN has a shorter running time while obtaining a higher accuracy rate.