MA-STNet: A multi-scale adaptive spatio-temporal fusion algorithm for dynamic link weight prediction in traffic flow networks

With the increasing complexity of urban systems, link weight prediction in transportation has become a key issue for understanding system behavior and optimizing resource allocation. As a typical dynamic complex network, transportation systems exhibit highly nonlinear and multi-scale characteristics, where link strengths dynamically evolve over time and space. Traditional methods based on static graph structures or fixed time windows are insufficient for effectively modeling the dynamic evolution of link weights, particularly when it comes to distinguishing between node inherent patterns (non-diffusive) and state propagation (diffusive). To tackle these challenges, this paper presents a Multi-Scale Adaptive Spatio-Temporal Network (MA-STNet) for traffic flow prediction, which systematically models the complex spatio-temporal dependence of link weights in dynamic networks. In the temporal dimension, the model incorporates periodic temporal embeddings and a multi-scale causal temporal attention mechanism, effectively capturing hierarchical temporal dependencies ranging from local fluctuations to long-term trends. For spatial modeling, an adaptive dynamic graph spatial fusion encoder is designed to explicitly distinguish and integrate non-diffusive structural features and dynamically diffusive propagation information, enabling dynamic modeling and prediction of link strength. Experimental results on four real-world large-scale traffic datasets (PEMS03, PEMS04, PEMS07 and PEMS08) show that MA-STNet achieves lower MAE, RMSE and MAPE compared with mainstream methods, effectively capturing both short-term fluctuations and long-term trends in traffic flow, and demonstrating consistency and reliability in link weight prediction tasks within dynamic complex networks.