为更好地预测船舶密集水域多船会遇避碰场景中的船舶轨迹，基于生成式对抗网络思想，提出一种多船轨迹预测模型——Vessel-GAN。针对船舶轨迹特点，Vessel-GAN模型基于时序卷积网络的多船历史轨迹编解码、船舶交互特征提取、终点信息引导和避碰损失函数等技术，通过在历史轨迹上的对抗训练，实现了能够拟合船舶航行行为数据分布的多船预测轨迹生成。基于琼州海峡近5000万条由AIS数据构建的船舶交互数据集的实验表明，相较Social-GAN基准模型, Vessel-GAN在计算速度方面提升了36%，平均位移精度、终点位移精度分别提升28%、41%，生成的多船预测轨迹更加符合船舶真实行为特征，且具有更好的预测实时性和稳定性。

To better predict vessel trajectories in multi-vessel collision avoidance scenarios in congested waters areas, a multi-vessel trajectory prediction model called Vessel-GAN was proposed based on the generative adversarial networks. For the vessel trajectory characteristics, based on technologies of multi-vessel historical trajectory encoding and decoding of temporal convolutional network, vessel interaction feature extraction, terminal information guidance, and collision avoidance loss function, the Vessel GAN model realized the generation of multi-vessel prediction trajectories of fitting the distribution of vessel navigation behavior data through adversarial training on historical trajectories. The experiment on the vessel interactive data sets of nearly 50 million messages constructed from AIS data in the Qiongzhou Strait show that compared with the Social GAN benchmark model, the calculation speed of Vessel GAN is improved by 36%, the average displacement accuracy and the terminal displacement accuracy are improved by 28% and 41% respectively, and the generated multi-vessel prediction trajectory is more consistent with the real behavior characteristics of vessels, and has better real-time prediction and stability.