With the rapid development of the energy internet and smart grids, distribution networks are exhibiting increasingly dynamic evolution and complex multi-source characteristics. Accurate identification of the operating topology of distribution networks is crucial for ensuring system security, optimizing scheduling, and enabling rapid fault response. However, traditional topology identification methods rely heavily on complete measurement systems and static structural assumptions, limiting their effectiveness under conditions of incomplete observations and frequent topology changes. In recent years, Graph Neural Networks (GNNs) have garnered extensive attention in power systems for their superior structural modeling capabilities. Nevertheless, mainstream GNN models such as GCN and GAT struggle to capture high-order dependencies among distant nodes and exhibit limited robustness in noisy environments. To address these issues, this paper proposes a distribution network topology identification method based on the spectral-domain graph convolution network ARMAConv. By constructing a line graph structure, the branch state identification task is transformed into a node classification problem, and multi-layer stacked auto-regressive moving average (ARMA) filters are employed to achieve deep modeling of complex topological dependencies. Furthermore, a Focal Loss-based optimization objective is introduced to mitigate the severe class imbalance of branch states. Extensive experiments conducted on the IEEE 33-bus system demonstrate that the proposed method significantly outperforms traditional approaches in terms of identification accuracy, generalization ability, validating its practicality and potential for intelligent grid applications.