Traffic flow prediction is a cornerstone of intelligent transportation systems and is pivotal for the advancement of smart cities. However, current approaches face two major challenges: (i) methods relying on predefined graph structures fail to capture the global semantics and dynamic nature of traffic networks, limiting their ability to extract meaningful spatial features, which in turn degrades prediction accuracy, and (ii) many models overlook the spatio-temporal heterogeneity of traffic flow, failing to account for varying traffic distributions and flow patterns across regions and over time. To address these challenges, we propose a Spatio-Temporal Multi-Graph Fusion Network (STMGFN) for traffic flow prediction. The model leverages a convolutional multi-head self-attention mechanism to capture spatio-temporal features from both local and global perspectives. Additionally, it enhances traffic flow graph data by incorporating dynamic flow patterns and evolving graph topologies. By generating heterogeneous correlation graphs for each time step via spatio-temporal embeddings and associations, and then performing convolutional fusion with topological and semantic graphs, the model captures multi-dimensional correlations between nodes. Finally, the multi-view graph convolutional module is integrated with a recurrent neural network to model spatio-temporal dependencies and predict future traffic conditions. The experimental results demonstrate the effectiveness and generalizability of the model and significantly improve the accuracy of traffic flow prediction.