Regularized Long Short-Term Memory Neural Network for Vessel Trajectory Prediction

In this article, we propose a regularized Long Shortterm Memory Neural Network (RLSTM) architecture for vessel trajectory prediction. The input layer is utilized to perform input encoding of the AIS data, while a regularization factor is incorporated to impose constraints on the weight matrix in the network layer. This weight matrix reduces redundant connections among the network nodes, enhancing the model’s efficiency. The RLSTM exhibits a relatively rapid convergence rate and offers a wide range of regularization options. Notably, even after pruning the weight matrix, the proposed method maintains relatively high precision in predicting ship trajectories. Compared to existing approaches, the model resulting from our pruning process demonstrates superior precision.