<p>Accurate traffic flow forecasting plays a pivotal role in alleviating traffic congestion and enhancing the operational efficiency of transportation systems. Owing to the robust ability of graph neural networks (GNNs) to extract topological features, they are extensively utilized in traffic forecasting to capture the spatial characteristics of road networks. Despite the promising results achieved by existing GNN-based traffic forecasting methods, they still suffer from the following two issues: (1) Road networks exhibit multiple high-order correlations among nodes, yet traditional graph structures represent node relationships merely through pairwise connections, leading to the neglect of these high-order dependencies; (2) They typically consider only short-term or long-term temporal patterns, overlooking the multi-scale temporal characteristics in traffic data. However, traffic data exhibit distinct dynamic patterns across different time scales. To address these challenges, we propose Hypergraph-based Multi-scale Spatio-Temporal Graph Convolution Network for traffic forecasting (HMSTGCN). Specifically, HMSTGCN converts the edges in traditional graph structures into hypergraph nodes and dynamically updates the hypergraph information using hyperedge embeddings, thereby better representing the complex dynamic high-order relationships between nodes. Additionally, HMSTGCN incorporates a temporal convolution module consisting of multiple convolutional kernels of varying sizes to learn correlations across different time scales. Experimental results on six real-world traffic datasets demonstrate that HMSTGCN outperforms state-of-the-art traffic forecasting methods.</p>