A Kernel Discriminant PCA and contrastive learning–enhanced FCNN framework for UHVDC line fault diagnosis

With the rapid development of large-scale clean energy bases in remote regions, ultra-high-voltage direct current (UHVDC) transmission plays a vital role in modern power grids. However, existing line protection schemes exhibit limitations in fault feature representation and utilization, often leading to misoperations under extreme conditions. To tackle these challenges, this study develops a high-precision UHVDC line fault diagnosis framework that synergistically integrates Kernel Discriminant Principal Component Analysis (KDPCA) for enhanced feature separability with a contrastive learning–enhanced Fully Connected Neural Network (CLE-FCNN) for improved robustness and generalization. Physically interpretable zero-mode and line-mode voltage components are first extracted as initial features. KDPCA is then applied to perform nonlinear discriminative feature extraction, thereby enhancing inter-class separability and discriminative capability. Subsequently, the CLE-FCNN is developed to further enhance robustness and generalization by pulling samples of the same class closer together while pushing samples of different classes farther apart in the feature space. Simulation results show that the proposed method can accurately distinguish internal from external faults within a 1 ms window, even under high transition resistances up to 500 Ω . Furthermore, validation on real UHVDC fault samples confirms that the method maintains stable performance under practical operating conditions. • Proposed a UHVDC fault diagnosis framework combining KDPCA with CLE-FCNN. • Developed KDPCA to improve nonlinear feature separability for UHVDC fault data. • Designed CLE-FCNN to boost generalization under diverse operating conditions. • Reached near-perfect accuracy within 1 ms post-fault, enabling ultra-fast protection. • Ensured robustness across fault locations, high resistances, and real UHVDC cases.