ABSTRACT
In order to address the issue of low precision in ship trajectory prediction within complex navigable waters, a ship trajectory prediction model named GRU-ABiLSTM, which is based on the encoder-decoder architecture, is proposed. The prediction model employs gated recurrent units (GRU) in the encoder part of the model to capture the temporal features within the trajectory sequences, and utilizes a bi-directional long and short-term memory network (BiLSTM) in the decoder, and incorporates an attention mechanism to adjust the weights of the data features. The inputs of the model take the longitude, latitude, speed and heading of the ship at historical moment as basic features, while the density of ships in the waters, which has undergone a median filter smoothing process, is introduced as an additional feature. The AIS data collected from the core port area of Ningbo Zhoushan Port in March 2024 is selected for model training and validation, and compared quantitatively and qualitatively with GRU, LSTM, Seq2Seq-LSTM, and Attention-BiLSTM models, which provide better prediction results under different prediction duration and sailing scenarios.
