A Multi-Source Traffic Data-Driven Traffic Flow Evolution Modeling Method for Intelligent Transportation Systems

Abstract

The fusion of multi-source traffic data and dynamic traffic-state prediction provides an important approach for Intelligent Transportation Systems (ITS) applications. To address spatiotemporal-scale inconsistency among heterogeneous traffic data and insufficient representation of road-network correlations, a multi-source traffic flow evolution model based on graph modeling and spatio-temporal attention is proposed. Different traffic modalities are first aligned and encoded into a unified representation, and adaptive modality weighting is introduced to enhance complementary information rather than directly concatenating heterogeneous features. Graph-based spatial correlation and spatio-temporal attention are then jointly employed to capture congestion propagation and dynamic dependencies. Experimental results show that the model achieves a Mean Absolute Error (MAE) of 14.87, a Root Mean Square Error (RMSE) of 24.96, and a Mean Absolute Percentage Error (MAPE) of 8.31%, with the MAE reduced by 8.27% compared with Graph WaveNet. For the 60-min forecast, the MAE reaches 17.83, corresponding to an 8.56% reduction. The results indicate that this model can improve the accuracy and stability of traffic flow evolution predictions in complex road network environments.