High-accuracy prediction of vessels’ estimated time of arrival in seaports: A hybrid machine learning approach

Abstract

Optimizing the Estimated Time of Arrival (ETA) for seaport-bound vessels is crucial to maritime operations since inaccurate ETA predictions can have a ripple effect, causing vessel schedule disruptions, congestion, and decreased port operational effectiveness. To address these challenges and fill substantial deficiencies in existing prediction models, we have introduced a novel hybrid tree-based stacking machine learning framework integrating Extra Trees, AutoGluon Tabular, and LightGBM, with Random Forest Regressor (RFR) as the meta-learner. Utilizing Automatic Identification System (AIS) data from vessels in the Baltic Sea, our model significantly improves ETA predictions, achieving a mean absolute percentage error (MAPE) of 0.25 %. Compared to existing machine learning algorithms, our stacking model exhibits superior prediction performance. Our study's feature importance analysis highlights the crucial role of variables like speed, distance, course, and vessel type in ETA forecasts. Cross-validation further confirms the robustness of our ensemble model. In conclusion, this study improves predictive analytics in marine logistics by giving useful information about real-time ETA estimates. This helps port authorities make the best use of their resources, reduces vessel idle time and congestion, and increases overall efficiency and sustainability. This way, this study can significantly contribute towards attaining operational excellence and provide a strong foundation for future predictive models, advancing smart port management and maritime logistics.