船舶轨迹预测对于智能船舶理解复杂遭遇场景并做出正确决策至关重要。然而，由于船舶固有的不确定性和多船之间复杂交互影响，预测未来轨迹是一个非常具有挑战性的问题。为此，提出基于多关系加权图Transformer (MG-Transformer）的船舶轨迹预测模型。首先从AIS数据中提取轨迹中相似船舶行为模式，捕获不同移动特征。在此基础上，通过对不同船舶的历史行为模式进行学习，以提高模型预测精度和效率。其次，构造多关系加权图来说明多船之间复杂的空间关系，通过Transformer学习与周围船舶的交互影响以细化轨迹，并预测合理的轨迹。采用宁波舟山港AIS数据开展实验验证，结果表明：在对不同时间步长的轨迹进行预测时，对比LSTM、BiLSTM、Seq2seq、Social-SGCNN, MG-Transformer模型在平均位移误差、最终位移误差指标中均有大幅下降，各项指标平均降低27.54%，所提船舶轨迹预测模型的精度有显著提升，对于海上交通安全和效率至关重要。

Ship trajectory prediction is essential for intelligent ships to understand complex encounter scenarios and make wise decisions. However, due to inherent uncertainty and complex interactions between different ships, predicting future trajectories is a very challenging problem. Therefore, a ship trajectory prediction model based on Multi-relational weighted graph Transformer (MG-Transformer) is proposed. First, the motion patterns of ships with similar trajectories are extracted from AIS data to capture different movement features. On this basis, the historical motion patterns of different ships are learned to improve the prediction accuracy and efficiency of the model. Secondly, the multi-relational weighted graphs is constructed to illustrate the complex spatial relationship between multiple ships. The interaction with surrounding ships is learned through Transformer to refine the trajectory and predict a reasonable trajectory. The AIS data of Ningbo-Zhoushan Port is used for experimental verification. The results show that when predicting trajectories of different time steps, compared with LSTM, BiLSTM, Seq2seq, and Social-SGCNN, the MG-Transformer model has a significant decrease in the average displacement error and final displacement error indicators. The average reduction of each indicator is 27.54%. The accuracy of the proposed ship trajectory prediction model has been significantly improved, which is crucial for maritime traffic safety and efficiency.