Detection and Classification of Faults in a Power System with Inverter Penetration Based on a Real-Time Machine Learning Application

This study investigates the implementation of a machine learning (ML) algorithm for the classification and localization of faults in a power system with inverter penetration. The approach utilizes the Clarke transform and cosine distance prediction method to detect fault types and insertion distances. However, the results indicate that the general objective was not met, as the algorithm exhibited significant errors in fault classification, frequently predicting multiple fault types per observation, leading to indecisive fault identification. Additionally, the fault insertion distance detection showed poor performance, achieving only 36% accuracy with a tuned window of =1250. The study concludes that the decision criterion using L2 normal for characteristic eigenvalues and cosine distance for eigenvectors holds potential, but improvements in the regression model are necessary for reliable fault classification and localization in inverter-penetrated power systems.