Article: Structural optimisation and interpretable analysis of graph neural network for fault diagnosis of ecological power distribution system Journal: International Journal of Environmental Technology and Management (IJETM) 2026 Vol.29 No.7 pp.23 - 41 Abstract: Aiming at the challenges faced by the fault diagnosis of distribution system in the new power system, such as dynamic topology change, heterogeneous multi-source data and insufficient interpretability of the model, this paper proposes a topology-aware dynamic graph neural network (TDGNN). In this method, the impedance weighted dynamic adjacency matrix is constructed by electrical-physical joint coding, realising the co-evolution modelling of topology change and fault characteristics. The spatio-temporal dual attention module is designed to precisely capture the spatial correlations and temporal patterns of fault propagation, and the jump connection is introduced to alleviate the long-term dependence problem. This method provides a high-precision, robust and interpretable solution for intelligent fault diagnosis of distribution system, supporting proactive O&M decision-making and new employee training in distribution network. Inderscience Publishers - linking academia, business and industry through research

Title: Structural optimisation and interpretable analysis of graph neural network for fault diagnosis of ecological power distribution system

Authors: Yimin Ding; Huiling Liang; Rui Chen; Xiaobo Yang

Addresses: State Grid Wuzhong Electric Power Company, Wuzhong, 751100, China ' State Grid Wuzhong Electric Power Company, Wuzhong, 751100, China ' State Grid Wuzhong Electric Power Company, Wuzhong, 751100, China ' State Grid Wuzhong Electric Power Company, Wuzhong, 751100, China

Abstract: Aiming at the challenges faced by the fault diagnosis of distribution system in the new power system, such as dynamic topology change, heterogeneous multi-source data and insufficient interpretability of the model, this paper proposes a topology-aware dynamic graph neural network (TDGNN). In this method, the impedance weighted dynamic adjacency matrix is constructed by electrical-physical joint coding, realising the co-evolution modelling of topology change and fault characteristics. The spatio-temporal dual attention module is designed to precisely capture the spatial correlations and temporal patterns of fault propagation, and the jump connection is introduced to alleviate the long-term dependence problem. This method provides a high-precision, robust and interpretable solution for intelligent fault diagnosis of distribution system, supporting proactive O&M decision-making and new employee training in distribution network.

Keywords: interpretable analysis; graph neural network; GNN; fault diagnosis; power distribution system; topology-aware dynamic graph neural network.

DOI: 10.1504/IJETM.2026.155729

International Journal of Environmental Technology and Management, 2026 Vol.29 No.7, pp.23 - 41

Received: 07 Apr 2026
Accepted: 29 Jun 2026
Published online: 11 Aug 2026 *

Title: Structural optimisation and interpretable analysis of graph neural network for fault diagnosis of ecological power distribution system

Keep up-to-date