Spatial-temporal graph neural network based on gated convolution and topological attention for traffic flow prediction

Accurate traffic flow prediction is essential for developing intelligent transportation systems (ITS) and providing real-time traffic applications. This study proposes a novel Spatial-Temporal Graph Neural Network based on Gated Convolution and Topological Attention (STGNN-GCTA) to accurately model complex spatiotemporal traffic flow correlations. In the temporal dimension, we design a novel Gated-Memory Convolutional Neural Network (GMCNN) to capture the non-linear temporal dependencies by controlling the output based on the timing information position. In the spatial dimension, we develop a Multilayer Graph Topological Attention Network (MGTAN) to capture the dynamic spatial dependencies by identifying high-impact neighborhood segments in each time step. In particular, we improve the model’s prediction robustness in a noisy environment using the Network Smoothing Training (NST) method. Experimental results on two public traffic datasets demonstrate that STGNN-GCTA has higher prediction accuracy and execution efficiency than baseline methods and exhibits excellent robustness.