Managing vessel arrival uncertainty in container terminals: a machine learning approach

A container terminal is a complex system where a broad range of operations are carried out involving a wide array of resources that need to interact over a 24 hour operating cycle. Since the various activities are mutually related to each other, there is a need not only to maximise the efficiency of each one, but also to ensure proper coordination, hence to solve integrated decision-making problems. Several factors can affect the quality of the services provided and the overall efficiency. Vessel arrival uncertainty further complicates the task of the planners and, as a result, of the effectiveness of the planning itself, in particular at the operational level. Each arrival produces high peak loads for other terminal activities, as well as for the supporting arrival activities (pilotage, towage, etc.) and hinterland transportation (waiting, congestion etc.). Deviating arrivals only worsen this peak load. On a daily level, the actual time of arrival of the vessels often deviates from the scheduled time. Despite contractual obligations to notify the Estimated Time of Arrival (ETA) at least 24 hours before the arrival, ship operators often have to adapt and update the latest ETA due to unexpected circumstances. This aspect results in a last-minute change of plans in terminal operations resulting in higher costs. In fact, the ability to predict the actual time of a vessel’s arrival in a port 24 hours in advance is fundamental for the related planning activities for which the decision-making processes need to be constantly adapted and updated. Moreover, disruptions in container flows and operations caused by vessel arrival uncertainty can have cascade effects within the overall supply chain and network within which the port is part. Although vessel arrival uncertainty in ports is a well-known problem for the scientific community, the literature review highlights that in the maritime sector the specific instruments for dealing with this problem are extremely limited. The absence of a reference model that specifies the relationship between vessel arrival uncertainty and the involved variables resulted in the application of a specific machine learning approach within the Knowledge Discovery in Database process. This V approach, that abandons all prior assumptions about data distribution shape, is based on the self-learning concept according to which the relation between an outcome variable Y and the set of predictors X is directly identified from the historical collected data. The approach has been validated thanks to two different case studies: the container terminal of Cagliari, located in the Mediterranean basin, and one of the main container terminals of Antwerp, located at the North Sea. Depending on the framework and planning purposes several estimates can provide useful information on vessel arrivals. Sometimes, it can be useful for planners to infer a quantitative estimate of the delay/advance in minutes, sometimes it may be useful to have a qualitative estimate, even only knowing whether or not an incoming vessel is likely to arrive before or after the scheduled ETA. For this reason a two-step instrument is proposed is made up of two different modules. The fitted algorithmic models used to obtain predictions are Logistic Regression, CART (Classification and Regression Trees) and Random Forest. All the proposed models are able to learn from experience, following the well-known Data Mining paradigm “learning from data”. From a practical point of view, the probability, associated to the continuous estimation, of specifically identifying the work-shift of the incoming vessel is calculated. In all predictions Random Forest algorithms still show the best performance. This aspect can help planners, in the daily strategy decision making process, in order to improve the use of the human, mechanical and spatial resources required for handling operations. This could maximise terminal efficiency and minimise terminal costs, hence improving terminal competitiveness. Moreover, the interpretation of the discovered knowledge, made it possible to evaluate the most discriminating variables of the analysis, even thanks to graphical visualisation of the Importance-plots.