TFDTSTGCN: Time-frequency domain transformation space-time graph convolution

Traffic forecasting is a crucial element of Intelligent Transportation Systems (ITS) and plays a significant role in traffic planning, management, and control. Graph Convolutional Network (GCN)-based models have become the standard approach for traffic prediction due to their capability to model spatial correlations through the mechanism of message passing to transmit traffic information. However, existing models often fall short in effectively capturing the frequency information of traffic flow and delivering accurate long-term predictions. To address the challenge of current model frameworks that struggle to learn the periodic nature of traffic information, this paper introduces a novel architecture that integrates bidirectional time-frequency transformations. This design utilizes the Fourier Transform to perform graph convolution in the frequency domain, followed by the Inverse Fourier Transform for time-domain prediction. Additionally, to enhance the model’s learning capacity, we incorporate an adaptive learning unit and a multi-scale recurrent neural network structure. The proposed model has been validated using two real-world datasets and compared with six other typical forecasting models, consistently outperforming them in predictive accuracy.