Ship track prediction based on Bayesian optimization in temporal convolutional networks

ObjectiveAs the traditional ship trajectory prediction method is prone to gradient explosion and long calculation time, this paper seeks to improve its accuracy and calculation efficiency by proposing a ship trajectory prediction model based on an improved Bayesian optimization algorithm (IBOA) and temporal convolution network (TCN). MethodA temporal pattern attention (TPA) mechanism is introduced to extract the weights of each input feature and ensure the timing of the historical flight track data. At the same time, a reversible residual network （RevNet） is introduced to reduce the memory occupied by TCN model training. The IBOA is then used to find the optimality of the hyperparameters in the TCN (size of kernel K, expansion coefficient d). The model is finally validated using a five-fold cross-validation method, and trajectory prediction is carried out after obtaining the optimal model. ResultThe trajectory data is collected by automatic identification system (AIS) and verified. The root mean square error (RMSE) is found to be 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.ConclusionThe proposed network has good adaptability to complex trajectories and higher accuracy than the traditional model and long short-term memory (LSTM) model, while maintaining high prediction accuracy for trajectories with strong coupling.