A CNNGRU-MHA method for ship trajectory prediction based on marine fusion data

• Published in: Ocean engineering 2024-10, Vol.310, p.118701, Article 118701
• Main Authors: Bi, Jinqiang, Gao, Miao, Bao, Kexin, Zhang, Wenjia, Zhang, Xuefeng, Cheng, Hongen
• Format: Article
• Language: English
• Subjects: AIS data; Convolutional neural network (CNN); Gated recurrent unit (GRU); Marine environmental data; Multi-head attention mechanism(MHA); Trajectory prediction
• ISSN: 0029-8018
• DOI: 10.1016/j.oceaneng.2024.118701

Predicting ship trajectories is significant for early warning of potential collision and reducing the probability of marine accidents. Previously, multi-ship trajectories prediction methods have been proposed, of which deep learning methods received more attention because it can take into account correlated factors and adapt to complex scenarios. However, they are limited in practical application due to insufficient extraction of environmental factors and characteristic information. Herein, the influence of internal and external factors in the course of ship navigation is considered. The spatial and temporal matrix of ship navigation geography including wind direction, speed, wave height, direction and velocity was constructed. A hybrid model, convolutional Neural Network and Gated Recurrent Unit based on Multi-Head Attention mechanism (CNNGRU-MHA) is established, combining the ability of one-dimensional convolutional units to extract high-dimensional features and learn time series features. The MHA mechanism was used to capture key factors affecting ship trajectory prediction, reducing the influence of interference information while preserving important features. It enables the model to process longer timing information with better robustness and accuracy. The proposed model was tested and validated in different situations and compared other models. The results demonstrate that the CNNGRU-MHA model is readily deployable, exhibits high accuracy and dependability. •A ship trajectory prediction method based on hybrid neural network is proposed, namely CNNGRU-MHA model.•The spatial and temporal matrix of ship navigation geography including wind direction, speed, wave height, direction and velocity is constructed.•The Hybrid neural network (CNNGRU-MHA) shows the ability of one-dimensional convolutional units to extract high-dimensional features and learn time series features.