With the rapid development of AloT, accurate traffic flow prediction has become the core of intelligent transportation systems, which is important for dynamic allocation of urban resources and optimization of travel efficiency. Existing models, however, struggle to capture the spatio-temporal heterogeneity and long-term dependencies of traffic data, show limited generalization to unseen urban scenarios, and cannot effectively integrate multi-source heterogeneous data. To address these challenges, this paper proposes a Multi-modal Spatio-Temporal Dynamic Fusion Network (MSTDFN) for traffic flow prediction. MSTDFN is a dynamic spatio-temporal model that integrates multi-source heterogeneous data. It addresses the distribution-shift issue through instance normalization, employs a patch-based embedding method to segment the time series, and combines linear transformations with positional encoding to integrate weather data. Furthermore, the model uses a temporal Multi-head Attention Mechanism to capture multi-scale temporal dependencies and dynamically model spatial neighborhood interactions through Graph Attention Network (GAT). Experiments show that MSTDFN reduces the MAE by 13.69% on average compared to existing models on multiple urban traffic datasets and maintains a prediction accuracy 3.09% better than the baseline model in unseen urban scenes.