Vessel destination prediction: A stacking approach

This paper proposes a stacking model for ship destination prediction that incorporates both a static model of historical information, based on a Bayesian neural network (BNN) and a dynamic model of current trajectory information. We propose a modified DBSCAN clustering method for trajectory clustering and core trajectory identification. By identifying core (representative) trajectories, we do not need to follow previous studies that compare a current trajectory with each historical trajectory, which significantly reduces the computation time and increases the accuracy of our similarity calculation. In order to obtain the optimal weights for the static and dynamic models, we develop a multi-response constrained linear regression model. This method has high interpretability, as the weights directly indicate the different roles of static and dynamic models at different stages after departure. Our results show that the stacking model is highly accurate after three days from departure, indicating the indispensable role of both information types.