Fault Diagnosis of Microgrids Using Branch Convolution Neural Network and Majority Voting

Abstract

Fault diagnosis is an important guarantee for the stable and safe operation of microgrids, which consists of fault detection and fault localization. However, most current researches separately deal with these two issues, which cannot obtain completed fault diagnosis results. This paper proposes a solution based on deep learning, namely branch convolution neural network (CNN) with a majority voting (B-CNN-MV) model, to simultaneously realize fault detection and fault localization through two branches. One of the branches realizes fault detection and the other performs fault localization. Firstly, in each branch, the CNN module extracts the two-dimensional image features of each sample in the spatial dimension and outputs primary classification results. Then, the classification results from the CNN module within one period of data constitute the temporal dimension input for the following majority voting module. Finally, the majority voting modules after each branch employ these temporal dimension inputs to calculate the final fault type and location results. Through this new design, the information on accurate fault type and fault location can be obtained simultaneously. Moreover, the test results show the proposed B-CNN-MV model can also achieve a high accuracy even in the case of insufficient data.