Vessel Trajectory Prediction Based on EDA_GRU Network

The Automatic Identification System (AIS) broadcasts and receives vessel-based information. However, due to the undetermined nature of AIS broadcasts and the flexibility of vessel navigation routes, AIS data often contains numerous noisy trajectories, which lowers the accuracy of predicted trajectories. To alleviate the influence of existing trajectories on predicted trajectories and hence improve the prediction accuracy, our study explores a deep learning approach utilizing AIS data, proposing Encoder_Decoder Attention_GRU (EDA_GRU) architecture. In particular, the EDA_GRU implementation incorporates an attention mechanism that concentrates on key trajectory details, adjusts relevant information weights, and enhances prediction accuracy. Additionally, GRU is utilized in the encoder to process input AIS trajectory sequences, extracting the spatial and temporal features to be combined with the GRU in the decoder, which utilizes attention operations to obtain attention weights. The resulting weight vector is then introduced to the GRU for decoding, ultimately leading to the final prediction outcome. Our proposed method is extensively experimented on public datasets. Compared with traditional GRU and its variants, the RMESE value of EDA_GRU model is reduced by 0.4371 and 0.1010 compared with GRU model and ED_GRU model.