G-Trans: A hierarchical approach to vessel trajectory prediction with GRU-based transformer

Current trajectory prediction methods are relatively blind regarding trajectory selection and lack dynamic motion characteristics of ships. Additionally, many models achieve high prediction accuracy through complex network structures, increasing computational load. To tackle this issue, we introduce G-Trans, a ship trajectory prediction technique that relies on Transformer architecture and feature clustering. First, we smooth the trajectories continuously using moving averages and sliding windows, ensuring the trajectory data's time sequence and integrity. Second, we perform trajectory motion feature clustering through latent space to identify trajectory segments with apparent motion features for further processing. Finally, we modify the Transformer by adding a G-block and additional positional encoding to the encoder and decoder. The GRU structure extracts spatiotemporal features of vessel trajectory with a more straightforward structure than LSTM. The self-attention mechanism of Transformer solves the problems associated with analyzing long-term trajectory sequences. G-Trans combines both, complementing each other in long-term prediction, improving the trajectory prediction's fidelity. The algorithm's effectiveness is verified using trajectory data collected by AIS, with G-Trans improving by 30.8%, 20.2%, and 7.1% over LSTM, GRU, and Transformer, respectively, improving predictive precision and computational efficiency.