Enhanced Fault Detection in Power Electronics-Dominated Microgrids: A Machine Learning Framework with Residual-Based Features and Class Imbalance Mitigation

Abstract

 This study presents a machine learning and deep learning framework for fault detection in power electronics-dominated microgrids using residual-based features and class-imbalance mitigation. The framework compares the performance of Support Vector Machine, K-Nearest Neighbors, Logistic Regression, Gradient Boosting, Convolutional Neural Network, and Long Short-Term Memory models, with data created from a simulated environment for a microgrid. To enhance the representation of fault related deviations, residual based and temporal features are created, and class imbalance is compensated for by applying Synthetic Minority Over-sampling Technique (SMOTE) to the training data. The metrics employed to evaluate the models include accuracy, precision, recall, F1 score and other metrics. The experimental results demonstrate that the Gradient Boosting model with 99.80% accuracy is best in the evaluated dataset and the models based on CNN are providing a competitive performance in the fault detection task. The results show the feasibility of a machine learning method based on residuals to enable automatic microgrid fault detection. But it is still required to validate with measured hardware or data from the field before use in practice.