Long-term traffic flow prediction by integrating multi-scale temporal features and spatially constrained attention mechanisms

Abstract Accurate traffic flow prediction is important for intelligent transportation system scheduling and planning. The main challenge in traffic flow prediction lies in the non-stationary characteristics of multi-band superimposed traffic flow data and its dynamic spatiotemporal coupling evolution. Additionally, traditional models are limited by the inability to decouple the orthogonality deficiency of long-term stability, medium-term periodicity, and short-term volatility in the temporal dimension, as well as the difficulty in adapting to the dynamic correlation between local topological constraints and global diffusion. This paper proposes a novel prediction framework that integrates multi-scale temporal features and spatial constraint attention mechanisms (MTF-SCA), adding a multi-level dynamic embedding module to capture the intrinsic coupled spatiotemporal features of the data, and constructing a spatial constraint attention mechanism based on road network topology priors to capture the dynamic correlations of the data at both local and global levels. Experiments were conducted on the METR-LA, PEMS-BAY, and PEMS series datasets, and the results show that it outperforms existing traditional models.