Identification of Single Line‐Ground Faults in Active Distribution Networks Based on R‐KAN

ABSTRACT In active distribution networks (ADNs), power electronic devices like inverters introduce noise interference into fault monitoring signals, compromising identification accuracy. This challenge is particularly pronounced in high‐impedance grounding faults, where weak fault signatures degrade conventional detection performance. This study proposes an improved residual CNN architecture, ResNet‐Kolmogorov‐Arnold‐network (R‐KAN), for accurate single line‐ground fault (SLGF) identification. The method leverages the rich fault features contained in transient zero‐sequence current (ZSC) and zero‐sequence voltage (ZSV) waveforms following SLGFs in DG‐integrated systems. The model employs superimposed ZSC‐ZSV images as inputs and replaces standard ReLU activation with KAN functions, reducing linear components and computational burden. A comprehensive dataset generated through PSCAD simulations trains the R‐KAN alongside conventional neural networks. Comparative evaluations demonstrate R‐KAN's superior classification performance across multiple metrics. Rigorous testing, including high‐resistance fault scenarios, noise interference conditions, and missing data cases confirms the model's enhanced generalization capability. Field validation using actual recorded waveforms further verifies the model's practical effectiveness in real‐world SLGF identification. The proposed approach addresses critical challenges in modern ADNs by combining advanced network architecture with optimized feature extraction from transient zero‐sequence components.