Multi-Dimensional Attention Based Spatial-Temporal Networks for Traffic Forecasting

Traffic flow prediction is the key problem of intelligent transportation system. Accurate prediction results are indispensable for traffic management and road planning. However, due to the complex spatial-temporal correlation of traffic flow data, including the spatial correlation and temporal correlation of adjacency, periodicity, and trend that exist between different roads. The existing forecasting methods consider the spatial-temporal correlation but lack the dynamic modeling of spatial-temporal correlation. To deal with this dynamic feature, this paper proposes a multi-dimensional attention-based spatial-temporal network (MA-STN). It mainly contains three parts, the spatial-temporal attention unit, the spatial-temporal feature extraction unit based on Graph Convolutional Network (GCN) and the fusion prediction unit, and the residual connection is also added to the model to avoid the gradient disappearance problem. Meanwhile, this paper divides the dataset into three subsets to deal with the three features in the temporal dimension separately. To verify the effectiveness of the proposed model, two real-world road traffic flow data collected by PeMS system are used for validation. By comparing six different models, the proposed network in this paper has a 7% accuracy improvement compared to the baseline model. To verify the effectiveness of the attention mechanism, ablation experiments are used in this paper for validation, and the results show that the attention mechanism can achieve a 5% accuracy improvement.