<p>Traffic flow prediction is critical to the collaborative decision support function of traffic management. However, it remains a challenging problem because of the inherent nonlinear dynamics, stochasticity, and spatial-temporal correlation of traffic flows. In this paper, a traffic flow prediction model called Spatial-Temporal Attention based ChebNet and BiGRU model (STA-CN-BiGRU), which consists of several layers of ST-Residual Block, is proposed. Firstly, ChebNet is used to learn the spatial dependence of traffic flow caused by the topology of the road network. BiGRU is combined with CNN for capturing the spatial and multi-scale temporal correlation of traffic flows. Then, in order to incorporate the different importance levels of temporal and spatial dependencies for each node, a hybrid spatio-temporal attention module is incorporated into the traffic flow prediction model and then a ST-Residual Block is constructed. Further, the outputs of multiple component traffic flows are weighted fused to obtain the final predictions. Finally, the performance of the proposed model has been extensively evaluated on real-world datasets in terms of one-step ahead prediction, multi-step ahead prediction, and ablation study. The results show that the proposed model can improve the prediction accuracy over the baselines, especially outperforming other models in long-term prediction.</p>