Data-driven optimization for container ship bunkering management under fuel price uncertainty

Fuel prices are a crucial and volatile component of operational costs in maritime transportation. This paper optimizes container ship bunkering decisions under the uncertainty of multi-port fuel prices, using data-driven optimization frameworks that integrate machine learning and mathematical programming models. We address two primary challenges: (i) incorporating spatiotemporal correlations between multi-port fuel prices into predictive models, and (ii) determining the most effective data-driven modeling framework for this problem. To address the first challenge, we develop a two-channel long short-term memory model specifically designed to capture the spatiotemporal dependencies of multi-port fuel prices. For the second challenge, we construct two data-driven modeling frameworks for ship bunkering management: a two-stage contextual deterministic programming model with point predictions (TDP framework) and a multistage contextual stochastic programming model with distributional estimates (MSD framework). Through comprehensive computational experiments using both real-world and synthetic data, we obtain two crucial insights: (i) accounting for the spatiotemporal correlations among multi-port fuel prices significantly improves the accuracy of fuel price predictions; and (ii) the TDP framework is more suited to container shipping routes with fewer ports, while the MSD framework offers advantages in contexts with a higher number of ports.