In recent years, researchers have shown significant interest in various fields such as transportation, social networks, and recommendation algorithms regarding graph neural networks. However, achieving high-accuracy traffic flow prediction in transportation proves highly challenging due to its complex spatiotemporal dependencies and nonlinear traffic patterns. This paper proposes a method that utilizes a Hybrid-hop attention mechanism to establish spatiotemporal correlations through the combination of geographical and semantic information between different time steps. To reduce time complexity, we set the receptive field of attention to neighboring spatial nodes and introduce a dynamic spatiotemporal perception graph driven by data to capture hidden spatiotemporal dependencies. Additionally, we design a multi-scale gated convolution mechanism to extract dynamic temporal dependencies from multi-receptive field features of various scales. Experimental results on public transportation network datasets (METR-LA and PEMS-BAY) demonstrate the model’s excellent performance.