ADCT-Net: Adaptive traffic forecasting neural network via dual-graphic cross-fused transformer

Abstract

The rapid development of road traffic networks has provided a wealth of research data for intelligent transportation systems. We are faced with vast high-dimensional traffic flow data, characterized by complex spatio-temporal dependencies, waiting for exploration of their internal relationships. Accurately representing these spatiotemporal relationships and improving the accuracy of spatiotemporal traffic prediction are critical challenges in current intelligent transportation forecasting. To tackle this issue, we propose an intelligent prediction framework for traffic flow based on the adaptive dual-graphic transformer with a cross-fusion strategy. Our aim is to uncover latent graphic feature representations that transcend temporal and spatial limitations. Furthermore, we establish a traffic spatiotemporal prediction model using a cross-fusion attention mechanism to capture dependency relationships represented by adaptive graphs. Extensive experiments demonstrate that our proposed model achieves superior prediction performance on practical urban traffic flow datasets compared to benchmarks, particularly for long-term predictions. Further analysis confirms its strength in balancing reliability and practicality, making it well-suited for applications in intelligent transportation systems.