FDST-GCN: A Fundamental Diagram based Spatiotemporal Graph Convolutional Network for expressway traffic forecasting

Abstract

Expressway Traffic forecasting is a crucial issue in Intelligent Transportation System (ITS), and extensive studies have been conducted in this field. However, most existing models do not distinguish between expressway and urban road networks in traffic forecasting, which limits the performance improvement of traffic forecasting in expressways. Moreover, how auxiliary features contribute to prediction models are also hardly discussed in current studies. To deal with the above shortages, we put forward a Fundamental Diagram based Spatiotemporal Graph Convolutional Network (FDST-GCN) in this paper, which incorporates the prior knowledge in Fundamental Diagrams (FD) of the expressway and is dedicated to expressway traffic forecasting. FDST-GCN takes the Sequence to Sequence (Seq2Seq) structure as the backbone and assigns tasks to the encoder and decoder. The encoder aims to extract historical information; note the relations between speed, density, and capacity in FD; FDST-GCN uses two GCNs in the encoder to extract the spatial dependencies of speed and volume simultaneously. Decoder devotes to predicting future traffic status, which integrates the Time Feature, Weather Feature, and FD Feature into the model. In experiments, we validate the performance of FDST-GCN by comparing the prediction accuracy with baselines on two real-world traffic datasets; the results show that FDST-GCN improves 8.23% and 7.49% in RMSE compared to the best baselines in the above two datasets, respectively; 2.94% and 10.41% in MAE, respectively. Besides, we also illustrate the characteristics of auxiliary features and how those features boost the prediction accuracy.