Fault Location Algorithm for Distribution Network With Distributed Generation Based on Domain-Adaptive TGATv2

Abstract

When a grounding fault occurs in a distribution network with distributed generation (DG), it poses significant challenges for fault localization, including difficulties in extracting fault features, complex changes in network topology, and limited fault samples. A distribution network fault localization method based on a domain-adaptive transfer dynamic graph attention network (TGATv2) is proposed to tackle these challenges. First, a feature selection (FS) module is embedded into the dynamic graph attention network (GATv2), which can automatically select zero-sequence current statistical features related to node faults, thereby reducing noise interference. Subsequently, the network topology is reconstructed to account for both fault conditions and economic efficiency. This enhances the model's generalization ability across various degrees of network topology changes and scenarios. Additionally, a domain-adaptive transfer learning method with an optimized dynamic loss function is employed to reduce the distribution differences between the source domain and target domain data, addressing the issue of fault location with limited samples. Finally, simulations and experimental tests are conducted using different distribution networks. The results demonstrate that compared to TGATv2 without FS embedding, the proposed method improves average accuracy by at least 8% and achieves approximately 3–22% higher accuracy compared to other methods, demonstrating strong robustness.