Fault-Type Identification in Power Systems Based on Cross-Validating Machine Learning Algorithms

The identification of the fault-types in power lines is the prerequisite for avoiding large-scale blackouts and restoring an abnormal or a faulty power system to its normal operation. Given an increasingly steady foundation for the application of artificial intelligence in power systems, machine learning has become one of the major directions applied to studying the fault problems in the systems. Performances of different machine learning algorithms may not be exactly the same in fault-type identification. In this paper the fault data of voltage and current obtained from a fault simulation model are collected and formed into sample sets. Five models of machine learning algorithms, i.e., one-dimensional convolutional neural network, principal component analysis & support vector machine, support vector machine, random forest, and K-nearest neighbor, are constructed and programmed in Python to identify various fault-types and then cross-validate their identification performances. After these models are trained in the training set, they are tested in the test set to obtain the values of such performance indexes as accuracy, precision, recall, and F1 score of the fault-type identification, and the confusion matrixes for the quantitative analysis of the misidentified and misclassified samples. The results show that the above five models built by the five machine learning algorithms perform significantly in identifying short-circuit fault-types, and all the values of the performance indexes exceed 98.81%. Especially, among them the model of the one-dimensional convolutional neural network outperforms the other four ones owing to its impressive overall performance.