Data-driven prediction of oil spill risk in ship-to-ship bunkering by reinforcement learning and Bayesian belief networks

Ship-to-ship (STS) bunkering is a sensitive shipboard operation which might pose acute environmental risks due to the potential for oil spill pollution. It requires conceptual modelling of complex interactions between operational, human, technical, and environmental variables addressed by uncertainty and non-linearity. The aim of this article is to introduce a data-driven tool integrating reinforcement learning (RL) and Bayesian belief networks (BBN) in order to predict oil spill occurrence probability during STS bunkering operations. In the model, whilst RL is used to model and optimize the nonlinear interactions between the most significant predictors, the BBN is capable of both causal reasoning and probabilistic inference under uncertainty. The RL-BBN modelling considerably improves predictive accuracy and interpretability when compared to single-model baselines. The findings of this research will provide contributions for marine environmental protection agency, rule makers, ship crew, superintendents, ship owners and Health, Safety, Environment and Quality (HSEQ) managers about oil spill risk reduction, operational planning, and regulatory compliance in STS operations.