A Data-driven Fault Classification Method for Microgrids

Abstract

This paper proposes a method that combines wavelet decomposition and deep learning for fault classification in microgrids (MGs). Random Search optimization is used to find the best wavelet function for matching traces to extract deep features of data. And these extracted deep features are used as complementary inputs to a deep learning model. One-dimension convolutional neural network (1-D CNN) is used to complete the secondary conversion of data features and high-dimensional feature mapping. The data are transformed by dimensionality and bidirectional long and short-term memory network. The data information mining is completed by using the multi-branch structure of the network to filter and merge the multidimensional effective information learned from each network branch. We also compare and analyze five different classification techniques (i.e., k-nearest neighbor(KNN), decision tree(DT), support vector machine(SVM), random forest(RF), and XGBoost(XGB) $)$ and compare their performance statistically. After modeling the MGs system in the simulation software, the Consortium for Electrical Reliability Technology Solutions (CERTS) MG effectively illustrates the validity of our proposed method.