Machine learning-based fault diagnosis and classification of three-phase transmission lines with RFE and domain knowledge
Baicun Guo, Bowen Yang, Shuhong Wang, Weizhan Shi, Fengye Yang, Dong Wang
Abstract
• A three-phase transmission line classification framework combining domain knowledge and machine learning is proposed. • Twenty-one new features based on domain knowledge is constructed. • Recursive Feature Elimination is used to extract new features. • Domain knowledge can improve the classification accuracy of various machine learning classifiers for three-phase transmission lines. • The best classification performance is XGBoost, with an accuracy of 94.25 %. In recent years, impressive achievements have been made in machine learning-based power system fault diagnosis. However, most approaches are purely data-driven and lack integration with power system domain knowledge, limiting their interpretability and robustness. This study proposes a hybrid diagnostic framework that combines features derived from domain knowledge with machine learning algorithms to improve fault classification performance in three-phase transmission lines. Based on power systems domain knowledge, 21 novel diagnostic features are designed, covering sequence components, phase angle dynamics, power characteristics, impedance indicators, and asymmetry metrics, to capture the multidimensional nature of fault behaviors. Recursive Feature Elimination is then applied to find the top 10 most discriminative features in order to focus on core variables and reduce redundancy. Experiments involving various mainstream machine learning algorithms show that the XGBoost classifier achieves the best classification performance among all tested models by reaching an accuracy of 94.25 % across 11 fault categories after incorporating selected features derived from domain knowledge. Further experiments show that the proposed features consistently improve classification performance across various machine learning methods, confirming their general applicability and effectiveness in fault diagnosis tasks. Moreover, under the same data and evaluation conditions, the proposed method is systematically compared with advanced approaches reported in the literature, showing advantages in accuracy.