An Improved Traffic Forecasting Model based on Efficient Spatiotemporal Graph Convolutional Network

Traditional approaches often ignore spatial and temporal connections, which cannot match the needs of forecasting assignments due to the extremely nonlinear and complicated nature of traffic flow. In this paper, a novel deep learning model, Efficient Spatiotemporal Graph Convolutional Network (EST-GCN), is proposed to address the time series prediction problem in the transportation domain. EST-GCN is able to jointly capture inter-sequence and temporal correlations through spectral transformation, which is combined with the graph convolutional network (GCN) and the gated linear unit (GLU). The design of the spectral transform enables the model to reduce the computational complexity by using an approximation method while maintaining the prediction accuracy. Furthermore, EST-GCN automatically extracts correlations between sequences from the data without the need of pre-defined prior knowledge. Results show that EST-GCN outperforms state-of-the-art baselines in prediction accuracy and training speed on real-world traffic dataset.