DS-SGCHN: dynamic-static synergetic graph convolutional hierarchical network for traffic forecasting

Abstract Accurate traffic flow prediction is critical for intelligent transportation systems. However, multivariate time series forecasting remains challenging due to dynamic spatio-temporal dependencies and complex latent association patterns. Existing graph neural network approaches typically rely on fixed or predefined graph structures, limiting their ability to model continuously evolving traffic patterns. To address these challenges, we propose a novel dynamic-static synergetic graph convolutional hierarchical network (DS-SGCHN) that jointly captures stable long-term patterns and dynamic short-term variations. Specifically, we introduce a dynamic-static dual-graph synergetic learning mechanism that adaptively constructs static and dynamic feature graphs from data without requiring prior topological assumptions. Furthermore, we design a dual-modal graph convolution method to aggregate spatial information from both topological and feature graphs, and employ a hierarchical architecture that integrates spatio-temporal features in stages. Within this framework, we incorporate two specialized modules: the Temporal Feature Interaction module, which enhances salient dynamic changes through a cross-time gating mechanism; and the spatial feature enhancement module, which reconstructs fine-grained spatial distributions via a cross-scale attention mechanism. Extensive experiments on four real-world traffic datasets demonstrate that DS-SGCHN consistently outperforms state-of-the-art baseline models in both short-term and long-term forecasting tasks. Our model achieves outstanding performance while maintaining efficiency, showing strong potential for practical deployment in dynamic traffic environments.