Date of Award

May 2019

Degree Type


Degree Name

Master of Science



First Advisor

Lingfeng Wang

Committee Members

Jun Zhang, Rohit Kate


With the development of high-power electronic technology, HVDC system is applied in the power system because of advantages in large-capacity and long-distance transmission, stability, and flexibility. Therefore, as the guarantee of reliable operating of HVDC system, fault diagnosis of the HVDC system is of great significance. In the current variety methods used in fault diagnosis, Machine Learning based methods have become a hotspot. To this end, the performance of several commonly used machine learning classifiers is compared in HVDC system. First of all, nine faults both in AC systems and DC systems of the HVDC system are set in the HVDC model in Simulink. Therefore, 10 operating states corresponding to the faults and normal operating are considered as the output classes of classifier. Seven parameters, such as DC voltage and DC current, are selected as fault feature parameters of each sample. By simulating the HVDC system in 10 operating states (including normal operating state) correspondingly, 20000 samples, each containing seven parameters, be obtained during the fault period. Then, the training sample set and the test sample set are established by 80% and 20% of the whole sample set. Subsequently, Decision Trees, the Support Vector Machine (SVM), K-Nearest Neighborhood Classifier (KNN), Ensemble classifiers, Discriminant Analysis, Backward Propagation Neural Network (BP-NN), long Short-Term Memory Neural Network (LSTM-NN), Extreme Learning Machine (ELM) was trained and tested. The accuracy of testing is used as the performance index of the model. In particular, for BP-NN, the impact of different transfer functions and learning rules combinations on the accuracy of the model was tested. For ELM, the impact of different activation functions on accuracy is tested. The results have shown that ELM and Bagged Trees have the best performance in HVDC fault diagnosis. The accuracy of these two methods are 92.23% and 96.5% respectively. However, in order to achieve better accuracy in ELM model, a large number of hidden layer nodes are set so that training time increases sharply.