Accurate awareness of power system topology can enhance the assessment of weak links and facilitates development of load regulation strategies. To address the real-time acquisition challenge of dynamic changes in network topology, a topology identification method for distribution networks based on graph attention network（GAT）and multi-layer perceptron（MLP）is proposed. Firstly, the active distribution network is abstracted into a graph model, and the GAT adaptively learnes the relationships between different nodes. Additionally,multi-head attention is employed to calculate the fusion features of each node in the graph. Subsequently, the fused features of nodes and edge sets in the topology are fed into the MLP to learn the relationship between node features and the state of edge connections. The topological graph-level identification results are obtained by integrating all edge states within the network. Finally, the effectiveness of the proposed method is verified in IEEE 33-node and IEEE 123-node distribution networks. The robustness of the proposed method under different noise levels is analyzed. Simultaneously, the proposed model is compared with traditional machine learning and deep learning algorithms to determine its superiority.