A Series Arc Fault Diagnosis Method in DC Distribution Systems Based on Multiscale Features and Random Forests

Series arc faults (SAFs) are able to cause electrical fires, which threaten the security of dc distribution systems critically. The noise of SAFs can interfere with the adjacent lines, and the SAF detectors based on the information in a single line are prone to trip. Therefore, to correctly diagnose SAFs, it is of significant value to consider the centralized SAF diagnosis methods that analyze the information in the entire system. This article analyzes the mechanism of arc noise interfering with normal lines in detail and proposes an SAF diagnosis method that comprehensively considers the information in the entire system. In the proposed method, the time-domain features, frequency-domain features, and singular values obtained from different scales of current are utilized to construct high-dimensional feature vectors. Then, to reduce the redundant information in feature vectors, feature importance analysis and feature selection are realized using random forests (RFs). Finally, RFs are employed to fuse the information of feature vectors for SAF diagnosis. The offline experimental results indicate that the diagnostic accuracy of the proposed method can reach up to 98.93%, and the effectiveness and advancement of the proposed method are verified by comparing it with different state-of-the-art methods. Besides, the proposed method is implemented in the microcontroller for real-time operations. The results of the online experiment show that the proposed method is of outstanding reliability and robustness under transient conditions, and, at the same time, the diagnostic speed is less than 250 ms and can satisfy the requirements of the standard.