Global maritime trade carries over 80% of world cargo, yet vessel arrival time (VAT) prediction remains highly inaccurate. Hong Kong Port experiences average ETA-ATA deviations of 13.8 hours, causing massive congestion costs and supply chain disruptions.
Current methods rely exclusively on either static ETA reports or dynamic AIS data, missing the complete picture. This fragmented approach ignores how ships actually navigate—constantly responding to weather conditions, sea states, and their own physical capabilities.
This study develops a multi-source data fusion framework that integrates four key streams: ETA baselines, real-time AIS movements, marine weather data (wave height, wind speed, swell patterns), and vessel physical parameters (VPP). OpenFE automatic feature engineering handles complex data interactions, while six machine learning models (XGBoost, Random Forest, LightGBM, LSTM, Transformer, TabPFN) are systematically compared.
Testing on Hong Kong Port data shows TabPFN achieves optimal performance with 2.88–3.42 hour prediction errors, which means 43%–47% improvement over ETA baselines. Weather factors occupy 3 of the top 15 important features, contributing 20% of predictive power. Surprisingly, traditional machine learning consistently outperforms deep learning on this structured maritime data. These advances enable optimised berth allocation, reduced port congestion, and more reliable logistics planning, supporting the maritime industry’s digital transformation.