Vessel ETA Prediction Integrating BiLSTM with Attention Mechanism Using AIS Data

Abstract

Maritime transportation carries more than 80% of global cargo, making efficient port operations essential for international trade. Accurate prediction of ship arrival time is important for berth allocation, resource scheduling, and operational management. However, prediction accuracy is often affected by complex sea conditions, delayed vessel information, and reliance on human experience. To address these challenges, this study proposes a deep learning method that combines unidirectional (UniLSTM) and bidirectional long short-term memory (BiLSTM) networks with an attention mechanism to predict ship estimated time of arrival (ETA). The proposed models are evaluated using Automatic Identification System (AIS) data from the Port of New York, USA. Shapley additive explanations (SHAP) are also employed to analyze the contribution of different variables to the prediction results. The results show that BiLSTM performs better than the UniLSTM, and the attention mechanism further improves prediction accuracy. In particular, the root mean square error (RMSE) of the attention-based BiLSTM is reduced by an average of 5.7 compared with the traditional recurrent neural network (RNN), while the deviation between predicted and actual arrival times remains below 5%. These findings can support port operators and shipping companies in berth allocation, resource scheduling, and operational decision-making.