Accurate vessel trajectory prediction is essential for improving maritime safety, traffic coordination, and autonomous navigation. However, traditional machine learning and deep learning models often struggle to maintain predictive accuracy over multiple forecasting steps owing to their limited ability to capture complex temporal dependencies. To address these challenges, this study proposes a hybrid deep learning framework that integrates convolutional neural networks (CNNs), Transformer encoders, and multilayer perceptrons (MLPs). The CNN module extracts robust local motion patterns and mitigates noise effects in raw AIS data. The Transformer component models long-range temporal dependencies through self-attention mechanisms, enhancing the capability of the model to understand sequential vessel dynamics. The MLP decoder jointly processes feature-wise and temporal relationships to produce precise trajectory forecasts. K-means clustering was employed to generate static features that capture representative navigational behaviors to improve model generalization. Experimental results on AIS datasets demonstrate that the proposed CNN–Transformer–MLP architecture consistently outperforms baseline models across all evaluated prediction horizons, achieving a root mean square error of 1.5599 km and a mean absolute error of 1.0710 km at the 60-min horizon. • A novel hybrid deep learning framework for ship trajectory prediction is proposed. • CNN, Transformer, and MLP are combined to model complex spatio-temporal dynamics. • K-means clustering augments features to enhance model generalization. • The model outperforms baselines on both short-term and long-term prediction.
Lee et al. (Mon,) studied this question.