A multi-perspective fusion model for operating speed prediction on highways using knowledge-enhanced graph neural networks

Abstract

This study proposes a multi-perspective fusion model for operating speed prediction based on knowledge-enhanced graph neural networks, named RoadGNN-S. By utilizing message passing and multi-head self-attention mechanisms, RoadGNN-S can effectively capture the coupling impacts of multi-perspective alignment elements (i.e., two-dimensional design, 2.5-dimensional driving, and three-dimensional spatial perspectives). The results of driving simulation data show that root mean squared error, mean absolute error, mean absolute percentage error, and R-squared values of RoadGNN-S are superior to those of other classic deep learning algorithms. Then, prior knowledge (i.e., highway geometry supply, driver expectations, and vehicle dynamics) is introduced into RoadGNN-S, and the models’ prediction accuracy and transferability are verified by field observation experiments. Compared to the above data-driven models, knowledge-enhanced RoadGNN-S effectively avoids the fundamental errors, improving the R-squared value in predicting passenger cars’ and trucks’ operating speed by 7.9% and 10.7%, respectively. The findings of this study facilitate the intelligent highway geometric design with multi-perspective fusion and knowledge enhancement techniques.