Abstract: Aiming at the problem of low-voltage series arc fault detection, an extreme learning machine (ELM) model based on multi-feature fusion is proposed. A multi-feature matrix including time-domain features, wavelet energy features, AR model prediction error ratio and variational mode decomposition(VMD) energy entropy is constructed to improve the accuracy and robustness of fault identification. An optimized ELM model is adopted for fault classification, and its performance advantages are verified by comparison with support vector machine (SVM),Random Forest and BP neural network. The experimental results show that the proposed method can effectively reduce the misdetection rate in complex load environments, the accuracy is increased from 85% of single feature to 96%, and the misdetection rate is reduced from 12% to 4% in mixed load scenarios. Compared with the BP neural network, the ELM model has a 5-fold improvement in training speed and a 60% reduction in memory usage. This method still has good robustness under the conditions of noise interference and load mutation, providing technical support for the safe operation of smart grids.

    Key words: low-voltage series arc fault; multi-feature fusion; extreme learning machine; wavelet energy; variational mode decomposition energy entropy