LCTA-Net: a deep fusion model for multi-scenario vessel trajectory prediction

Abstract Vessel trajectory prediction tasks currently face several challenges, including the diversity of behavioral patterns, interference from anomalous data, and limited generalization capabilities of existing models. These challenges are particularly pronounced when dealing with multi-pattern trajectory characteristics, where achieving both high prediction accuracy and robust cross-scenario adaptability remains difficult. To address these issues, this paper proposes a high-precision vessel trajectory prediction framework. Firstly, the local outlier factor algorithm is applied to eliminate anomalies from raw automatic identification system data, thereby improving data quality. The Douglas–Peucker simplification and the Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) trajectory clustering algorithm are then used to construct datasets characterized by different turning behaviors. Subsequently, a prediction model called local-causal temporal attention network (LCTA-Net) is developed, based on a pruned Transformer encoder. LCTA-Net incorporates a temporal residual encoding module and a causal feature-augmented attention mechanism. These components jointly improve the model’s capacity to model temporal dependencies within trajectory sequences. In addition, a fixed-window local neighborhood search strategy is introduced to improve the spatial continuity and physical feasibility of the predicted trajectories. Finally, experimental results on two constructed trajectory datasets demonstrate that the proposed method significantly outperforms several state-of-the-art models across multiple error metrics, confirming its superior prediction accuracy and strong cross-scenario generalization capability.