Traffic speed prediction is an essential part of urban transportation systems that contributes to minimizing the environmental pollution caused by vehicle emissions. The existing traffic speed prediction studies have achieved good results, but some challenges remain. Most previously developed methods only account for road network characteristics such as distance while ignoring road directions and time patterns, resulting in lower traffic speed prediction accuracy. To address this issue, we propose a novel model that utilizes multigraph and cross-attention fusion (MGCAF) mechanisms for traffic speed prediction. We construct three graphs for distances, position relationships, and temporal correlations to adequately capture road network properties. Furthermore, to adaptively aggregate multigraph features, a multigraph attention mechanism is embedded into the network framework, enabling it to better connect the traffic features between the temporal and spatial domains. Experiments are performed on real-world datasets, and the results demonstrate that our method achieves positive performance and outperforms other baselines.
Keywords: cross-attention; graph convolutional network; traffic speed prediction.