MSDA-DiffNet: traffic flow prediction via multi-scale feature fusion and dual adaptive graph convolution with conditional diffusion

Traffic flow prediction presents significant challenges in modeling complex spatiotemporal dependencies and quantifying uncertainties. We propose MSDA-DiffNet, a novel framework that integrates three key components: a Multi-Scale Feature Fusion Module that captures temporal dependencies across various scales; a Dual Adaptive Graph Convolution Network that dynamically models spatial correlations; and a Conditional Diffusion Module that generates probabilistic predictions with comprehensive uncertainty quantification. Our approach addresses the limitations of existing methods that rely on static graph structures and single-scale features. Extensive experiments conducted on four public datasets demonstrate that MSDA-DiffNet significantly outperforms state-of-the-art models, reducing Mean Absolute Error, Mean Absolute Percentage Error, and Root Mean Square Error by 8.9%, 7.5%, and 9.3% respectively, while providing robust uncertainty estimation.