Fault Diagnosis in HVDC Systems Using Machine Learning

Abstract

This thesis investigates machine learning (ML) for fault diagnosis in high voltage direct current (HVDC) systems, emphasizing generalization to unseen fault conditions. A line commutated converter (LCC) HVDC system was modeled in MATLAB/Simulink to simulate diverse faults, generating data from which statistical features were extracted. Seven ML models (logistic regression, SVM, KNN, decision tree, random forest, gradient boosting, MLP) were evaluated. Experiments included a standard random split and a crucial generalization test on unseen fault resistances. Random forest, neural network, and gradient boosting demonstrated superior robustness, with random forest achieving the highest accuracy in generalization. The study highlights the importance of generalization testing for reliable fault diagnosis.