Conference Agenda

Machine learning approaches are increasingly discussed in port logistics, but practical value depends on whether they can improve real operational decisions. This paper presents an applied machine learning use case for a port terminal operator: forecasting container flow volumes in order to support resource planning, equipment readiness, and operational preparation. The main goal is to investigate the impact of machine learning based forecasts for container volumes under real-world operational conditions and to show that forecasting of container volumes is feasible in a real terminal environment, and to study under which data conditions and forecast horizons the results become useful for practice.

The work focuses on shift-level forecasts of inbound and outbound full and empty container volumes—an area where planning still relies on manual estimates, spreadsheets, and simple historical averages. In practice, workloads are shaped by multiple concurrent drivers (e.g., internal moves, planning changes, slot deviations, pre-announcements, weather, holidays, traffic, and disruptions), making the task both highly relevant for machine learning and methodologically challenging.

The paper follows a classical machine learning workflow. First, raw operational movement logs and related terminal data are transformed into analytical datasets. Second, extensive feature engineering is applied to represent the operational state of a shift. This includes lag variables, rolling aggregates, calendar and weekday effects, yard-related indicators, planning signals, and selected external variables. A key research question is the extent to which forecast quality improves when internal terminal movement data are combined with external data and planning-related information, instead of relying on movement data alone.

Several machine learning approaches are compared, including statistical baseline models, autoregressive learning setups, and tree-based ensemble methods. The paper discusses which model class performs best under volatile port conditions and how feature importance can be used to better understand the drivers of predicted volumes. Special attention is given to model validation in a real operational setting. Historical rolling backtests and holdout-based black-box validation are used to assess not only general forecast error, but also the practically useful forecast horizon. In other words, the paper examines whether the forecast quality is sufficient only a few hours ahead, for the next 24 hours, or even for several days.

The expected benefits are both operational and economic. More accurate machine learning-based forecasts can improve staffing and equipment planning, reduce under- and over-planning, stabilize terminal operations, and lower coordination effort with operators and service units. This, in turn, can enhance service reliability for shipping lines and hinterland partners and, over time, strengthen overall port performance and competitiveness.