Forecasting worldwide empty container availability with machine learning techniques

Abstract

Due to imbalances in the global transport of containerised goods, liner shipping companies go to great lengths to match the regional supply and demand for empty containers by transporting equipment from surplus to deficit regions. Making accurate forecasts of regional empty container availability could support liner companies and other involved actors by making better relocation decisions, thus avoiding unnecessary transport costs of empty equipment. Previously proposed container availability prediction models are limited to the application in individual regions and typically characterized by a high degree of temporal aggregation. Against this background, this paper introduces two novel approaches based on machine learning and probabilistic techniques to predict the future weekly availability of empty containers for more than 280 locations worldwide. The machine learning and probabilistic prediction models are built by analysing a unique data set of more than 100 million events from past container journeys. These events represent different stages during the transport process of a container. Both models use a two-step forecast logic. First, the expected future location of a container is predicted. Second, the expected timestamp for arriving at that location is estimated. The machine learning model uses artificial neural networks and mixture density networks to forecast the movements of containers. The models are quantitatively assessed and compared to the actual availability of containers and two more conventional forecasting approaches. The results indicate that the probabilistic prediction approach can keep up with conventional approaches while the neural network approach significantly outperforms the other approaches concerning every evaluation metric.

Optimizing Shipping Operations through Real-Time Monitoring and Control

“…To effectively manage its container assets and optimize turnaround time, a shipping company requires a tool or a system that can monitor and control container stripping processes in a practical and realtime manner. Examples can be derived from Martius et al [5] who used machine learning to forecast the worldwide empty container availability, and Gençer and Demir [6] who used mixed-integer linear programming and scenario-based stochastic programming to optimize the empty containers. In addition, Budipriyanto et al [7] used a simulation approach to solve the empty container problem.…”