Passenger flow prediction is one of the auxiliary methods of urban rail transit (URT) control, which can predict traffic conditions in advance and guide decision-makers to take measures. With the development of big data and artificial intelligence, the accuracy of passenger flow prediction has been greatly improved. However, there are still some problems not to be solved, such as insufficient relationship extraction of node features and topology structure. So, a model based on the graph convolutional network (GCN) is proposed to mine the connection between node features and the topology structure. Firstly, for station adjacencies in different dimensions, a GCN-based double-channel feature extraction module (DCGCN) is constructed based on parameter sharing and independent parameters to capture the heterogeneity and consistency of two spaces. And the model introduces the attention mechanism for the adaptive fusion of heterogeneous features, long short-term memory (LSTM) is used to capture temporal correlation. To verify the accuracy and rationality of this model, two datasets are introduced to test it, and the results show that DCGCN performs better than other baselines.