A time-delay aggregation transformer model for traffic flow forecasting

Forecasting traffic flow is vital for the efficient operation of modern transportation systems and plays a pivotal role in advanced intelligent traffic management. Enhancing forecasting accuracy hinges on precisely grasping dynamic changes and spatial-temporal relationships of traffic flow. This study mainly focuses on such patterns as trends, periodicity, and spatial heterogeneity and proposes a time-delay aggregation transformer model (short for TDATF) to forecast traffic flow. The model leverages a time-delay aggregation and a self-attention mechanism to extract the trends and periodicities in traffic flow effectively, and a dynamic graph convolution module dynamically captures spatial dependency relationships. Meanwhile, a spatial position embedding matrix is utilized, which is constructed by a Laplacian-smoothed embedding vector derived from the Graph Convolutional Network (GCN) layer, to effectively model spatial heterogeneity. Experimental results on datasets PEMS03/04/07/08 demonstrate that the proposed TDATF model outperforms the baseline models in terms of forecasting accuracy.