An eXplainable Multi-Stage Stochastic Optimisation for Bunker Procurement Planning

Abstract

Bunker procurement decisions significantly impact shipping industry operating expenses. However, several challenges hinder obtaining optimal decisions, including the need to forecast future fuel prices, incorporate uncertainties and capacity constraints, and understand feature importance for transparent solutions. To address these difficulties and foster confidence in the conservative business sector, we propose an eXplainable multi-stage bunker procurement planning (X-BPP) framework for the maritime industry. Our study develops a machine learning forecaster for accurate and robust price predictions from short to long term, considering multiple bunker ports on the route. We demonstrate coherent forecast integration into operation optimization through tractable reformulation. Kernel Shapley is employed to reveal feature importance in non-linear and multi-stage stochastic planning. In a real-world implementation, we evaluate the unified planning framework’s practical knowledge and awareness. Results show an average operating cost reduction of $257,541.51 for a fleet of six vessels during a 42-day Asia-North America trip, based on data from July 2020 to December 2021. During the Russia-Ukraine war, the framework still achieved $351,247.71 in savings. Additionally, we identify varying important features between short- and long-term forecasting.

Keywords: Bunker procurement planning, Bunker price forecasting, Machine learning, eXplainable AI, OR practice

JEL Classification: R42, G13, G53