Predicting out-terminals for imported containers at seaports using machine learning: Incorporating unstructured data and measuring operational costs due to misclassifications

Abstract

Persistent bottlenecks at container ports have significantly disrupted global supply chains, necessitating more efficient operations at seaports to address yard density and port congestion. An untapped but potentially critical approach to mitigating these challenges is to leverage container characteristics and machine learning to predict the out-terminals of containers upon their discharge from vessels. The predicted results can then guide the development of a more effective container storage strategy. To formulate such a strategy, this research developed a data-enabled methodological framework that integrates four key components: 1) Utilization of structured and unstructured data to enhance prediction accuracy. 2) Practice and knowledge-informed feature engineering to construct relevant features for the machine learning models. 3) Explanatory machine learning based classification models to understand the factors influencing terminal predictions. 4) Model-induced cost analysis to capture the monetary value of the prediction model including assessing the cost implications of misclassifications. An empirical study conducted at a seaport shows that our framework yields cost savings ranging from 14.90% to 30.45% compared to the Business-as-Usual scenario. Incorporating unstructured data as an additional feature in the machine learning models improves prediction performance by up to 6%. Moreover, integrating this framework into the existing operational system poses minimal risk and can be seamlessly executed. Additionally, the proposed methodological framework and its four components has broad applications beyond the shipping industry.