Ship Trajectory Prediction Method of Gated Recurrent Unit Based on Spatial-temporal Attention Mechanism

The accuracy of ship trajectory prediction is crucial for the intelligence level of ship's navigation. Addressing the insufficient capability of the gated recurrent unit (GRU) in capturing spatial-temporal information from ship data, which leads to poor accuracy in trajectory prediction, a method of GRU ship trajectory prediction based on spatial-temporal attention mechanism (STA-GRU) is investigated. The traditional activation function in GRU is improved by a weighted activation function set to retain more comprehensive ship trajectory data. A spatial attention mechanism module is introduced to extract spatial location features of ships using latitude, longitude, relative latitude, and relative longitude as input sequences. This module computes spatial-temporal weight attention factors to obtain spatial feature vectors. The resulting vectors serve as the training dataset for the STA-GRU model used for ship trajectory prediction. Experimental validation is conducted using AIS data from Qingdao Port, with an input duration of 20 minutes and a sampling frequency of 2 min. A ship navigation trajectory dataset is constructed under these conditions. Results indicate that, compared to LSTM, AT-GRU, and Bi-GRU algorithms, the STA-GRU model not only converges faster during training but also significantly reduces the root mean square error, mean absolute error, and final displacement error. The average reductions of the aforementioned indexes for trajectory prediction are 50.2%, 38.7%, and 48.3%, respectively. For longitude prediction, the average reductions are 43.8%, 50.5%, and 49.5%, respectively. For latitude prediction, the average reductions are 52.4%, 48.4%, and 50.5%, respectively. Therefore, the proposed STA-GRU model exhibits significantly improved accuracy in ship trajectory prediction and meets the real-time requirements for trajectory prediction.