Cross-Pacific Vessel Estimated Time of Arrival and Next Destination Prediction with Automatic Identification System Data

• Published in: Transportation research record Vol. 2679; no. 3; pp. 67 - 80
• Main Authors: Lloret-Batlle, Roger, Lin, Sen, Guo, Jiequn
• Format: Journal Article
• Language: English
• Published: Los Angeles, CA SAGE Publications 01.03.2025
• Subjects: data and data science; marine automatic identification systems (AIS); marine transportation (water transportation); freight transportation data
• DOI: 10.1177/03611981241275551
• ISSN: 0361-1981

With the increase in global trade uncertainty and supply chain disruptions, accurately predicting the estimated time of arrival (ETA) of container vessels can effectively help carriers, terminals, and freight forwarders improve operational efficiency. The Asia-North America route has been recently under stress because of strikes and trade wars between the U.S. and China. Voyages are subject to multiple external factors leading to uncertainty in arrival times. This is especially true for cross-Pacific voyages, where long distances without intermediate port visits allow for a large feasible set of trajectories and vessel speed profiles. Large errors in ETA prediction not only hinder the effective planning and execution of other stakeholders but also lead to significant fluctuations in the types and quantities of goods arriving at the port, thereby hindering port competitiveness and efficient multimodal transportation. Existing literature focuses on estimating ETA and next positions for dense, compact areas at the vicinity of ports. We propose and evaluate model framework based on artificial neural networks (ANN) fed by automatic identification system (AIS) historical data to predict the next destination and ETA for cross-Pacific routes for cases where ETA from the captain is missing in the AIS data. Results show our model can effectively predict next destination and ETA of vessels, achieving a mean absolute error value of 4 h when the vessel is 1,500 nmi away from the port. For comparison, the ANN submodules are replaced with gradient boosted trees, providing similar results. We terminate by highlighting the challenges found to improve the model.