Objective In order to improve the accuracy and calculation efficiency of ship track prediction, the traditional method is prone to gradient explosion and long calculation time.

Method A track prediction model based on improved Bayesian optimization algorithm (IBOA) and time convolution neural network (TCN) was proposed. The time pattern attention mechanism (TPA) is introduced to extract the weights of each input feature and ensure the timing of historical data of flight track. At the same time, the reversible residual network is introduced to reduce the memory occupied by TCN model training. Then the Bayesian optimization algorithm is used to find the optimality of the hyperparameters in the time-convolutional network (the size of the kernel K, the expansion coefficient d), and finally the model is validated by using a 5-fold cross-validation method, and the trajectory prediction is carried out after obtaining the optimal model.

Result The track data collected by AIS was verified, and the root mean square error (RMSE) was increased by 5.5×10⁻⁵, 3.5×10⁻⁴ and 6×10⁻⁴ in weak coupling, medium coupling and strong coupling track prediction, respectively.

Conclusion The proposed network has good adaptability to complex flight paths, its accuracy is better than the traditional model and the Long Short-Term Memory （LSTM） model, and it still maintains high prediction accuracy in the flight paths with strong coupling.