Abstract

Accurate vessel trajectory prediction using Automatic Identification System (AIS) data is crucial for maritime traffic supervision, collision risk assessment, and search and rescue operations. However, long-horizon prediction remains challenging because AIS trajectories involve complex navigation patterns, continuous motion variations, and cumulative forecasting errors. To address these issues, this study proposes MAKDformer, a Transformer-based model for vessel trajectory prediction that integrates multi-attribute discrete state modeling with kinematic differential perception. MAKDformer uses latitude, longitude, speed over ground, and course over ground as input variables, and employs a Transformer backbone to model long-range temporal dependencies. To capture continuous vessel motion dynamics, a Kinematic Differential Perception Module (KDPM) is developed to extract sequential variations in displacement, speed, acceleration, and course. These kinematic differential features are then fused with the discrete Transformer representations. Furthermore, auxiliary regression loss and motion-consistency trend loss are incorporated to regularize motion pattern learning and enhance long-horizon forecasting stability. Experiments on real-world AIS datasets demonstrate that MAKDformer achieves superior spatial prediction accuracy compared with baseline models, including the standard Transformer, TrAISformer, MART, and GeoTrackNet. Specifically, in the 15 h forecasting task, MAKDformer reduces the Haversine error by approximately 40.1% compared with the second-best model, GeoTrackNet. Ablation experiments further verify that explicitly integrating KDPM, the feature fusion mechanism, and motion-constrained losses effectively mitigates error accumulation and improves the robustness of vessel trajectory prediction.

IPC Classification

Keywords