Trajectories prediction in multi-ship encounters: Utilizing graph convolutional neural networks with GRU and Self-Attention Mechanism

Multi-ship encounters are the essential prerequisite for collision accidents at sea, with the result that the study of trajectories prediction in multi-ship encounters is of great significance for reducing collision accidents and avoiding collision hazards in time. Based on the massive ship trajectories data recorded by the automatic identification system (AIS), we propose the spatiotemporal GCN with GRU and Self-Attention Mechanism model (ST-GRUA) for synchronized high-dimensional trajectories prediction in multi-ship encounters. In this study, GCN is employed to capture spatial dependencies while utilizing gated recurrent units of incorporation of attention mechanism to improve the interactive capability between time series. The model transforms the features of multi-ship encounters into topological relationship graphs in a graph-theoretic manner and converts the encounter process into the collection of relational topological graphs with timestamps by time slicing as the training data for the model, which is used to acquire the avoidance decisions and routine encounters processes in multi-ship encounters embedded in the actual AIS data to predict the trajectories of real multi-ship encounters. Numerous experiments have shown that the method improves navigational safety during multi-ship encounters while ensuring regulatory compliance. The method has important reference value for ship operators to formulate and implement effective collision avoidance strategies.