Multi-data fusion and machine learning for predicting ballast water discharge in ports
Jihong Chen1, Xizhi Chen2, Jia Shi3
1College of Civil and Transportation Engineering, Shenzhen University, Shenzhen, 518060, China; College of Management, Shenzhen University, Shenzhen, 518071, China; Shenzhen International Maritime Institute, Shenzhen, 518083, China; The Institute for Sustainable Development, Macau University of Science and Technology, 999078, Macao.
Abstract:
Although ballast water is essential for vessel stability, its discharge is a major vector for invasive species spread. With stricter regulations, accurate port-level discharge prediction has become critical for governance. This study develops a machine learning framework integrating AIS trajectories, ship attributes, port characteristics, and NBIC records (2021-2023), yielding 4137 samples with 23 features. Five regression models-RF, SVM, LGBM, XGBoost, and EN-were trained with Bayesian hyperparameter tuning via Optuna, under weighted and unweighted schemes. Results show RF achieves the most stable overall performance, while XGBoost improves high-discharge fitting, LGBM captures overall trends but is weight-sensitive, SVM attains high correlation yet struggles with extremes, and EN underfits. Quartile analysis and SHAP interpretation highlight next-voyage distance, vessel-type proportions, and vessel scale (DWT, Draft) as key drivers. Despite limited vessel coverage and ecological variables, the framework demonstrates robust generalizability and supports applications in invasion risk assessment and pollutant migration.
