Deep Spatio-Temporal 3D-Transformer Network for Traffic Flow Prediction

Accurate traffic flow prediction is increasingly vital for urban road planning and management. Recently, the existing models have made some progress in learning spatio-temporal features. However, they only focus on learning the dynamic spatio-temporal features from either local perspective or global perspective, while ignoring the comprehensive learning of correlations from both local and global perspectives, which limits their predictive performance. To this end, this paper proposes a Deep Spatio-Temporal Three-Dimentional Transformer Network (ST3DT-Net) for traffic flow prediction. Specifically, we first extract local dynamic spatio-temporal features via three-dimentional spatio-temporal decomposed convolutional blocks (3DSTDCB). Then, the global dynamic spatio-temporal features are learned based on three-dimentional vision Transformer (3D-Vit). The experiments are conducted on two typical traffic flow datasets. The results demonstrate the superiority of ST3DT-Net compared with the eight baselines.