Graph attention temporal convolutional network for traffic speed forecasting on road networks

Traffic speed forecasting plays an increasingly essential role in successful intelligent transportation systems. However, this still remains a challenging task when the accuracy requirement is demanding. To improve the prediction accuracy and achieve a timely performance, the capture of the intrinsically spatio-temporal dependencies and the creation of a parallel model architecture are required. Accordingly, we propose a novel end-to-end deep learning framework named Graph Attention Temporal Convolutional Network (GATCN). The proposed model employs the graph attention network to mine the complex spatial correlations within the traffic network and temporal convolution operation to capture temporal dependencies. In addition, the multi-head self-attention mechanism is incorporated into the model to extract the spatio-temporal coupling effects. Experiments show that the proposed model consistently outperforms other state-of-the-art baselines for various prediction intervals on two real-world datasets. Moreover, we reveal that the proposed model can effectively distinguish the sophisticated traffic patterns of ramps on expressways by analyzing the graph attention heatmap.

Automated summary

Traffic speed forecasting is crucial in intelligent transportation systems, and improving prediction accuracy and achieving timely performance require capturing spatio-temporal dependencies and creating parallel model architecture. The proposed GATCN framework utilizes graph attention network and temporal convolution operation to predict traffic speed, outperforming other state-of-the-art baselines in experiments.