STTransformer: A Physics-Informed Spatial-Temporal Transformer for Ship Trajectory Prediction

Maritime traffic safety has become increasingly critical with the growing vessel density and complexity of shipping routes, making accurate ship trajectory prediction an urgent challenge for maritime traffic management. Traditional methods based on statistical models and simple deep learning architectures struggle to capture complex spatio-temporal dependencies and often fail to maintain physical consistency in their predictions. Recent advances in transformer architectures have shown promising results in sequence prediction tasks, yet their direct application to maritime scenarios remains unexplored. This paper proposes STTransformer, a novel spatial-temporal transformer architecture specifically designed for ship trajectory prediction that integrates domain-specific physical constraints and multi-scale temporal features. The key innovations of our approach include: (1) a hierarchical attention mechanism that captures both local and global trajectory patterns, (2) a physics-informed loss function that ensures predictions comply with maritime navigation constraints, and (3) an adaptive feature fusion module that dynamically weights different input features based on their temporal relevance. Extensive experiments on real-world AIS data from New York Harbor demonstrate that STTransformer significantly outperforms state-of-the-art baselines in terms of prediction accuracy and physical consistency.