Multi-Step Spatial-Temporal Fusion Network for Traffic Flow Forecasting*

Traffic flow forecasting plays an important role in vehicle routing, traffic signal control and urban planning. However, this task is very challenging since the traffic flow shows high nonlinearities and has complex patterns. Most existing traffic flow prediction methods have limitations to model both spatial and temporal correlations of the traffic data and few of them consider this task in a multi-step prediction framework. To improve the prediction accuracy, we propose a novel attention-based multi-step spatiotemporal fusion network (called GCGRU att) to perform the traffic forecasting task. Based on a Seq2Seq structure, the model fuses the graph convolutions (GC) in the unit of GRU, applies attention mechanism (att) to capture the dynamic spatial-temporal correlations, thus realize the short-time multi-step traffic flow prediction at road network level. Experiments on real-world traffic datasets demonstrate that the proposed model outperforms other baselines and has interpretability in real scenarios.