Prescriptive analytics in maritime transportation

Abstract

The maritime transportation sector is increasingly relying on optimization and analytics to maximize the potential of the extensive data generated in the industry. This thesis leverages real-world data to optimize ship maintenance planning, port state control officer (PSCO) routing, and container ship bunkering decisions by using advanced prescriptive analytics methodologies. The first study introduces a smart predict-then-optimize (SPO) framework, using an ensemble of SPO trees (SPOTs) for ship maintenance planning. This method significantly enhances the traditional predict-then-optimize (PO) approach by integrating operational, repair, and risk costs into the machine learning (ML) model. It effectively reduces total operating expenses of ship inspections by approximately 1% over the PO-based scheme and at least 3% over non-ML-based schemes. The SPO-based plans also contribute to more efficient port operations by reducing the necessity for intensive PSC inspections, thus alleviating port congestion. The second study explores the PSCO routing problem in the face of uncertain ship conditions. By embedding this optimization problem into the ML model training process through a decision-focused learning framework, this study addresses the limitations of the traditional two-stage framework that separates prediction and optimization. A compact model using undominated inspection templates and a surrogate decision loss function based on noise-contrastive estimation enhances solution efficiency and decision accuracy, underscoring the importance of integrating decision making directly into the learning process. The third study optimizes container ship bunkering decisions under uncertain fuel prices. It employs a two-channel long short-term memory model to capture spatiotemporal correlations among multi-port fuel prices, significantly outperforming traditional predictive models. Two prescriptive analytics frameworks are compared: a two-stage contextual deterministic programming model and a multistage contextual stochastic programming model. The suitability of these frameworks varies based on the variance of inter-port fuel prices and the number of ports on each shipping service, offering crucial insights for maritime operators. This thesis collectively demonstrates the efficacy of integrating ML with mathematical optimization to address critical operational challenges in maritime transportation, leading to substantial economic benefits and operational efficiencies. The approaches applied in this thesis pave the way for innovative practices in maritime logistics, emphasizing the transformative potential of data-driven decision making in complex operational contexts.