Fault Diagnosis of a Stand-Alone Wind System Using Supervised Learning Methods

Abstract

Fault detection is an essential component of an electrical system that speeds up fault source detection and prolongs the operation time. When a system operates in an islanded condition, fault diagnosis becomes highly significant. The system studied comprises a wind turbine, battery, inverter, and local load. There are several techniques to detect and classify faults. Here, we compare and apply machine learning, feature transformers, and feature reduction methods to find the best approach for fault detection and classification in an islanded doubly-fed induction generator (DFIG) system. The proposed method can classify sixteen conditions, including AC line faults in both stator and rotor sides, short and open circuit faults in the battery side, load changes, a control system fault during an increase in wind speed, and normal operation. Gaussian Naive Bayes (GNB), Support Vector Machine (SVM), Decision Tree (DT), K-Nearest Neighbor (KNN), and Random Forest (RF) methods, along with Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), quantile, and standardization transformers, are employed for the classification purpose. Also, the comparison is conducted using the cross-validation technique. The goal of using the fewest electrical measurements has also been achieved, and the results demonstrate that some mixtures can reach over 99% accuracy and amazingly distinguish between conditions. MATLAB/Simulink is used for data acquisition, and Python is utilized for feature engineering and model selection.