Ship Trajectory Prediction Based on CNN-MTABiGRU Model

Abstract

To address the limitations of traditional neural networks in extracting ship trajectory sequence features and achieving the required accuracy for ShenZhong Link navigation, this paper proposes an improved model, CNN-MTABiGRU, which integrates a temporal attention mechanism and the Mish activation function. Firstly, convolutional neural networks (CNN) are used to extract key features from ship trajectory data. Secondly, a BiGRU network is combined with CNN to form a CNN-BiGRU model, capturing temporal dependencies in trajectory features to predict ship trajectories. To further enhance model training efficiency, the Mish activation function is introduced at the CNN output to achieve smoother nonlinear mappings, effectively preventing gradient vanishing and overfitting. Additionally, the temporal attention mechanism is incorporated to highlight feature data items that significantly impact prediction accuracy. Together, these elements form the optimized CNN-MTABiGRU ship trajectory prediction model. Practical results indicate that the CNN-BiGRU model yields the smallest prediction error and highest accuracy compared to CNN, BiGRU, CNN-LSTM, and CNN-GRU models for predicting ship trajectories in the ShenZhong Link. Moreover, the optimized CNN-MTABiGRU model, which incorporates both the temporal attention mechanism and the Mish activation function, reduces the mean squared error (MSE) by 17.03% in latitude and 35.10% in longitude, and the mean absolute error (MAE) by 14.42% in latitude and 21.47% in longitude, compared to the CNN-BiGRU model. These findings demonstrate that the optimized CNN-MTABiGRU model has significant advantages for predicting ship trajectories in the ShenZhong Link.