GL-LoiterDNet: A Hybrid Model for Ship Trajectory Prediction in Loitering Activity Scenarios

In loitering activity scenarios, vessels frequently execute course changes within localized maritime spaces, often exhibiting extreme turning maneuvers that generate ultralong, dense, and highly nonlinear spatiotemporal trajectories. Traditional prediction models demonstrate limitations in processing dynamically changing trajectory features, leading to insufficient prediction accuracy under such loitering conditions. To address this challenge, this study proposes a GL-LoiterDNet, a hybrid deep learning–based vessel trajectory prediction model. The model incorporates multidimensional trajectory characterization features including speed fluctuations, navigational positions, and course differentials. It integrates 1D convolutional neural networks (1D-CNN), bidirectional gated recurrent units (BiGRU), and bidirectional long short-term memory (BiLSTM) networks to capture both localized abrupt variations and long-term evolutionary patterns in loitering trajectories, thereby mitigating feature degradation phenomena. Experimental validation using trajectory data from vessels in Sagami Bay, Japan, demonstrates that the GL-LoiterDNet model outperforms 14 baseline models in prediction accuracy and robustness. The model exhibits rolling multistep trajectory prediction capability for loitering scenarios, achieving an average positioning error below 0.7 km within 10-min prediction windows. This research can provide reliable theoretical and data-driven support for continuous vessel positioning and monitoring in complex maritime operation scenarios.