Comparative analysis of data-driven models on detection and classification of electrical faults in transmission systems: Explainability, applicability and industrial implications

Most data-driven fault detection methods often face challenges in accuracy, adaptability, and real-time implementation, particularly in complex transmission networks. To address these issues, this study presents an in-depth comparative analysis of data-driven models, including machine learning, neural networks, and deep learning techniques, for detecting and classifying electrical faults in transmission lines. A few key gaps were identified to persist particularly in terms of the accuracy, computational complexity, explainability, applicability, and industrial implications of these models. Hence, a Quadratic Discriminant Analysis by Projection (QDA-P) model which effectively captures non-linear relationships between electrical fault features is proposed, enhancing classification accuracy, less computationally complex and explainable for both binary and multi-class scenarios. Using a publicly available simulated dataset generated through MATLAB Simulink to represent various fault types, the QDA-P model achieved binary classification scores of 0.988, 0.980, 0.982, 0.995, and 0.987 while recording a multi-class classification score of 0.982, 0.979, 0.982, 0.982, and 0.980 for accuracy, precision, specificity, recall, and F1-score, respectively. Feature importance analysis showed that voltage at the sending or receiving end of the transmission line for phase A was the most influential while the length of the transmission line in phase A had the lowest importance. In contrast, techniques such as SHAP, LIME, and PDP reveal that the length of the transmission line in Phases A & C significantly influenced Class 1 and Class 5 predictions, with the length of the transmission line in Phase C contributing positively and the length of the transmission line in phase A showing varied effects. An industrial applicability analysis using the Mahalanobis distance plot confirmed that the QDA-P model effectively captured underlying data patterns without significant outliers. These findings highlight the model's potential to improve fault detection accuracy and reliability, contributing to more efficient and secure energy transmission infrastructures.