<p>Accurately predicting traffic flow becomes more and more important to intelligent transportation systems in the age of autonomous driving. Some existing methods extract spatial and temporal features of road networks respectively extract temporal and spatial features via recurrent neural network and graph convolutional network from road networks for traffic flow prediction. However, there are still threefold challenge of spatiotemporal coupling interrelations, dynamic spatial correlations, and the impact of periodic patterns and external factors. Addressing at the issues, we propose a novel approach of stacking multi-timescale graph attention network (GAT) with bidirectional gated recurrent units (BiGRU) for traffic flow prediction, called SMGAT-BiGRU. Firstly, a GAT feature extraction module is explored to capture the dynamic spatial correlation of different road segments, and then a multi-timescale feature fusion module is constructed to mine periodic patterns and to capture the impact of external factors by extracting fine-grained spatiotemporal features and external dynamic time-varying features. Secondly, a BiGRU module with attention is proposed to capture temporal dependencies. Furthermore, a GAT-BiGRU module is presented to be stacked to fully capture the spatiotemporal coupling interrelations. Experimental results on two PeMS datasets show that SMGAT-BiGRU is able to remarkably promote the performance of traffic flow prediction and outperforms many state-of-the-art methods (such as T-GCN, STGCN, STSGCN, and DMSTGCN) for predicting traffic flow.</p>