Predicting Ship Waiting Times Using Machine Learning for Enhanced Port Operations

Abstract

Port congestion and prolonged ship waiting times pose challenges for global trade and increase operational costs and inefficiencies. In this study, a novel machine learning-based predictive approach was proposed to improve port operations by accurately forecasting vessel waiting times. By using a dataset of 121,401 voyage records, we evaluated nine regression models, including conventional, ensemble-based, and deep learning models. Shapley additive explanation (SHAP)-based feature selection is typically applied to enhance interpretability, and its effect is compared with principal component analysis-based dimensionality reduction and nonselection methods. The XGBoost Regressor (XGBR) is optimized using genetic-algorithm-based hyperparameter tuning, reducing mean squared error (RMSE) from 20.9531 to 19.6387, mean absolute error (MAE) from 13.6821 to 12.6753, and improving coefficient of determination (R2) from 0.2791 to 0.2949. A stacking ensemble model, integrating random forest regressor, XGBR, LightGBM regressor, and CatBoost regressor, improves performance, achieving an RMSE of 18.9023, MAE of 12.3287, and an R2 of 0.3265. ANOVA tests confirm numerous differences in model performance and computational complexity. The results demonstrated that tree-based ensemble models outperform deep learning models in this setting. The proposed approach enables proactive scheduling, reduces congestion, and cost savings. The scalability of the model renders it suitable for broad maritime logistics and intelligent transportation systems.