TransES-ETA: A Novel Transformer-based Explainable and Efficient Structure for Predicting Estimated Time of Vessel Arrival

ICLR 2026 Conference Submission15550 Authors

Keywords: maritime AI, AI application, temporal model

Abstract:

The estimated time of arrival (ETA) prediction is crucial in maritime AI for improving maritime shipping operation efficiency and resilience; however, they currently still face significant challenges. Traditional machine learning-based methods struggle with extracting accurate representations of Automatic Identification System (AIS) data, while neural network-based methods rely heavily on data quality and lack of explainability. In this paper, we propose TransES-ETA, a transformer-based, multi-task framework that jointly handles port call scheduling and ETA prediction in an end-to-end and explainable manner, leveraging the strong correlation between the two tasks.
We first design a novel AIS data extractor composed of multiple modules, each responsible for processing different semantic data attribute categories, enabling the transformer to capture accurate representations. Then, the port call schedule module of our proposed pipeline predicts vessel trajectories and generates features that serve as inputs for the subsequent ETA prediction module. Finally, the ETA module aggregates both AIS-derived features and port call schedule features to produce the final ETA estimate. The entire multi-task pipeline is trained in an end-to-end manner, allowing for the simultaneous optimization of both tasks with the shared feature and thereby improving training efficiency. Furthermore, by incorporating port scheduling information, the ETA predictions become more interpretable. Comprehensive evaluations demonstrate that our proposed model achieves mean absolute error (MAE) 2.46h in ETA prediction. Additionally, the pipeline accelerates training and inference by 80.93\% and 8.84% respectively, making the framework more efficient.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 15550