Delay prediction in container terminals: A comparison of machine learning methods

One of the most important issues in Container Terminal (CT) management is to manage adequately late arrivals. In fact, despite contractual obligations to notify the Estimated Time of Arrival (ETA) 24 hours before arrival, often ship operators have to revise it due to unexpected events such as weather conditions, delay in previous port and so on. In a daily planning scenario, this causes a series of inconveniences directly associated with the resource allocation problem. Terminal operators need to increase the accuracy of incoming demand in order to allocate more efficiently human resources, equipment and spatial resources required to satisfy the predicted demand. For planners the decision-making processes related with demand uncertainty may sometimes be highly complex without the support of suitable methodological tools. Specific models should be adopted, in a daily planning scenario, to provide a useful support tool in CTs and to help mitigating the consequences of late arrivals. In this study, using data collected in a Mediterranean Transhipment Container Terminal, we illustrate a data mining approach for predicting the level of daily alarm related to late arrivals. First, we defined three levels of daily alarm ranking the delay of arrivals. Then we obtained an estimate of the alarm level using three different Machine Learning models (Naive Bayes, Decision Trees and Random Forests) and we compared their predictive power on a test data set.