A Dynamic Context-Aware Approach for Vessel Trajectory Prediction Based on Multi-Stage Deep Learning

Accurate and reliable vessel trajectory prediction is challenging due to the highly non-linear, intricate and stochastic features of maritime transport networks, but its solutions are vital for ensuring maritime safety, intelligence and efficiency. In this study, a dynamic context-aware (DCA) approach based on multi-stage deep learning, termed DCA-MSDL, is proposed to address this challenge and improve prediction accuracy by considering a broader range of influencing factors. An inverted Transformer (iTransformer)-based deep learning architecture is first constructed for vessel turning status prediction. A deep generative framework is then presented for vessel trajectory augmentation. Following that, multiple potential trajectories are predicted using a novel data-driven algorithm based on multiple steps of nearest neighbor selection. Finally, trajectory enhancement considering dynamic traffic context is proposed to further improve prediction accuracy. With the separate steps of vessel turning status prediction, trajectory augmentation, trajectory prediction and trajectory enhancement, our approach allows us to explicitly explain the factors affecting the prediction accuracy and enable targeted improvements correspondingly. Extensive tests on real-world trajectories of vessels in the Singapore Strait have been conducted and the following encouraging results have been obtained: 1) the iTransformer-based turning status prediction achieves high accuracy at 93.37%, beating other state-of-the-art machine learning-based time-series models; 2) the deep generative model-based trajectory augmentation reduces error by 8.26%, and it significantly outperforms other oversampling techniques; 3) DCA-MSDL increases the prediction accuracy by at least 33.75% in comparison to existing benchmarking methods; 4) the devised dynamic context-aware trajectory enhancement in DCA-MSDL significantly improves the accuracy by 3.53%.