A More General Lightweight Network for Urban Regional Traffic Prediction

Abstract

With the increasing complexity of metropolitan management, the importance of urban regional traffic prediction has become increasingly prominent. However, existing traffic prediction models are often limited to specific regional traffic scenarios, lacking sufficient generalization ability. To address this issue, we propose a more generalizable lightweight spatiotemporal neural network (GL-STNN) for urban regional traffic prediction. By leveraging a fully convolutional design, GL-STNN circumvents the scalability issues of graph convolutional networks and the quadratic complexity of Transformers in dense spatiotemporal modeling. This network designs a novel deep convolutional architecture that adaptively integrates and captures global and local spatiotemporal correlations and heterogeneities by combining dilated convolutions and gating mechanisms. Furthermore, a multilevel grouped fusion convolutional architecture is designed to effectively fuse and align feature dimensions. To further enhance prediction accuracy, we also design a global masking mechanism to mark blank areas in the city. Experimental evaluations on a comprehensive benchmark of eight public datasets, spanning international cities and diverse traffic domains, demonstrate that GL-STNN outperforms existing baseline methods in terms of both performance and computational efficiency.