STFAENet: Multiattention Enhanced Spatiotemporal Fusion Network for Traffic Flow Prediction

Traffic flow prediction is a fundamental task in intelligent transportation systems (ITS). Due to the influence of urban functional zones and their neighboring regions, traffic flow data exhibit complex spatiotemporal correlations, making it challenging to effectively capture temporal dependencies and spatial structures for accurate prediction. To address this issue, this article proposes a multiattention enhanced spatiotemporal fusion network (STFAENet) based on the TransUNet architecture. STFAENet consists of an encoder, a skip-connection mechanism, and a decoder, and is designed to jointly learn fine-grained local features and global spatiotemporal dependencies in dynamic traffic scenarios. Specifically, an instance pyramid spatial attention (IPSA) module is introduced in the encoder to enhance multiscale spatial feature representation through hybrid normalization and pyramid attention, enabling the extraction of high-resolution fine-grained features. In the skip-connection stage, a spatiotemporal self-attention (STSA) module is embedded to jointly model spatial and temporal dependencies and reduce the semantic gap between the encoder and decoder. The decoder further fuses local details with global contextual information to generate high-resolution prediction results. Experiments on the TaxiBJ and TaxiCQ datasets demonstrate that STFAENet achieves superior prediction performance compared with several baseline methods.