Faulty Feeders Identification for Single-Phase-to-Ground Fault Based on Multi-Features and Machine Learning

The identification of single-phase-to-ground (SPG) faults in power distribution networks is crucial for ensuring the reliability of power supply. However, the traditional identification methods based on a single feature lack accuracy and robustness in complex fault scenarios. To address this issue, this article proposes a novel method that combines multiple features and machine learning techniques. Firstly, the article analyzes the differences in electrical signals between pre-fault and post-fault power distribution networks from multiple perspectives to construct a feature engineering reflecting the global fault characteristics. Secondly, random forest algorithm and feature influence curve are used to screen the features via measuring the importance of each feature. To deal with the problem of high feature dimensions and class imbalance, a data preprocessing framework that includes principal component analysis (PCA) technology and synthetic minority over-sampling technique (SMOTE) are proposed. Based on the above pretreatment, the model based on Bayesian optimization and support vector machine (BO-SVM) for identifying faulty feeders is finally established, which realizes automatic optimization of hyperparameters. The performance and generalization ability of the proposed model is evaluated through the learning curve and receiver operating characteristic (ROC) curve. Finally, the experiment data extracted from actual fault waveforms verify the model's strong fault identification ability in various complex project scenarios.