<p>Predicting traffic flow on urban road networks is extremely challenging due to complex spatio-temporal correlations. Currently, the most popular method is to stack multiple layers of spatio-temporal graph convolutions, which generally combines the graph convolutional network (GCN) with the temporal convolutional network (TCN) or the recurrent neural network (RNN) for joint spatio-temporal prediction. However, the existing methods for constructing the predefined adjacency matrices of the GCN fail to accurately capture the real world situation. In addition, these methods restrict the potential of GCN since the GCN module is limited to aggregating the features within a relatively limited local spatio-temporal neighborhood. In this paper, we propose a neural network called Gformer for predicting traffic flow based on the time lagged and cross correlation (TLCC) algorithm. Gformer captures the temporal and spatial features using the linear self-attention and TLCC-based GCN. The combination of these two models expands the range of the temporal and spatial neighborhoods where feature aggregation is effective, further unleashing the potential of GCN. Additionally, a novel parametrized skip-connection is used to merge the outputs of multiple layers of the encoder and feed them into the decoder for final prediction. To validate the effectiveness of Gformer, we conduct experiments on the two large-scale urban traffic flow datasets and find that our proposed model exhibits superior overall performance compared to all baseline models. The experiments demonstrate the validity and robustness of the proposed model.</p>