Modern power systems incorporate ultra-high-voltage transmission lines with series compensation to enhance their reliability and stability. These configurations generate complex data patterns in electrical signals, which present challenges for accurate fault diagnosis. This complexity arises from various potential faults, such as high-impedance faults and inductive-capacitive interactions, compounded by data scarcity owing to privacy and cybersecurity constraints. This study employs a simulation-based analysis to facilitate fault detection, classification, and localization as unified Multi-Classification tasks. The methodology improves the time-frequency domain data through meta-feature engineering and utilizes consensus-based feature selection by combining gradient boosting, variance thresholding, and mutual information. Model learning is conducted using Bayesian optimization within a Stacking Ensemble, integrating Extra Trees, Light Gradient Boosting Machine, and Extreme Gradient Boosting with an optimized Random Forest meta-classifier. The evaluation considers the accuracy, computational time, inference speed, and memory usage under challenging conditions to address practical limitations. The results indicate that the Stacking Ensemble achieves the highest performance despite its greater computational demand. Extra Trees offers the fastest execution but shows decreased accuracy with increased classification complexity. The Light Gradient Boosting Machine and Extreme Gradient Boosting demonstrate optimized performance across tasks owing to enhanced data processing. These findings confirm the practical applicability of the proposed fault diagnosis framework to modern ultra-high-voltage power systems.
Nawaz et al. (Sun,) studied this question.