GATransformer: A vessel trajectory prediction method based on attention algorithm in complex navigable waters

Abstract

The prediction of vessel trajectories plays a crucial role in ensuring maritime safety and promoting the scientific and efficient management of waterway traffic. In complex navigable waters, where vessels frequently encounter avoidance maneuvers and face intricate route intersections, the accuracy of vessel trajectory prediction is challenged. This paper proposes a vessel trajectory prediction model based on attention mechanism: GATransformer. A spatial encoder module is used to capture the spatial interaction patterns among vessels, while a temporal encoder module is employed to extract the temporal sequence information of vessel trajectories. Subsequently, the encoded spatiotemporal features are concatenated to achieve the fusion of trajectories’ spatiotemporal information. Finally, the trajectories at future time steps are predicted through a decoder. The spatial information of vessels is enriched by introducing the artificial feature of "distance between vessels and intersection nodes of the waterway network". Extensive comparisons between GATransformer model and baselines were conducted in a typical complex navigable water, Ningbo-Zhoushan Port. The experiment results indicate that the GATransformer model exhibits superior predictive performance, and the introduction of the feature "distance between vessels and intersection nodes of the waterway network” contributes to the improvement of accuracy in predicting vessel trajectories in complex navigable waters.