<p>Traffic forecasting plays an essential role in urban planning and traffic management. Nevertheless, the intricate spatio-temporal connections in traffic data make traffic prediction particularly tough. Although many traffic prediction methods have been developed, most of them fail to effectively model the decomposition of multiple traffic modes in traffic conditions, and it is also difficult to dynamically capture spatio-temporal dependencies between roads without relying on prior knowledge. A framework called Multi-scale Fusion Dynamic Graph Convolutional Recurrent Network (MDGCRN) is developed to tackle these difficulties. The approach initially separates intricate traffic conditions into trend and detail components, which are subsequently predicted independently using the dynamic graph convolutional recurrent network. Within this network, we propose a technique for creating dynamic graphs that, when combined with GRU, effectively captures short-term and long-term temporal information as well as spatial information. Comprehensive experiments conducted on three large-scale datasets demonstrate that MDGCRN outperforms existing state-of-the-art baselines across all metrics.</p>