Predicting vessel service time : a data-driven approach

Abstract

Vessel Service Time (VST) refers to the period from when a ship arrives at a berth until it departs. VST is a critical metric for port operational efficiency and service quality. Uncertainty in VST can undermine the operational efficiency in port management and lead to financial setbacks. To mitigate this uncertainty and lay the foundation for subsequent berth allocation, vessels typically provide an estimated departure time (EDT). However, substantial discrepancies often exist between the reported EDT and the actual departure time (ADT). These discrepancies mainly stem from unforeseen port handling inefficiencies and supply chain disruptions. This variability results in significant differences between the actual VST and its anticipated duration, thereby complicating port operations. To tackle this issue, our research represents the first study to predict VST from a data-driven perspective. We introduce an advanced tree-based stacking regression model for VST prediction, utilizing vessel port call records from 2020 to 2023. Our machine learning stacking approach achieves more accurate VST predictions than EDT reported by vessels, significantly reducing the mean absolute error (MAE) by 29.7% (from 4.54 to 3.19 h) and the root mean square error (RMSE) by 31.9% (from 6.58 to 4.48 h). The model also demonstrates reliable predictive power with an R-squared (R2) value of 0.8. These results underscore the significant scientific value of data-driven approaches in maritime studies. Our findings highlight the potential of the proposed tree-based models to surpass traditional models and originally reported data in predictive accuracy for VST. This advancement not only represents a notable improvement in predictive capabilities for VST but also lays the groundwork for further research into enhancing vessel scheduling efficiency through machine learning.