A grey-box deep learning modelling strategy for fuel oil consumption prediction: A case study of tuna purse seiner

Operational measures and policies by International Maritime Organization have been proposed to reduce ship fuel consumption and associated emissions. These initiatives emphasise the need for precise predictive models that account for the impact of real operational conditions on Fuel Oil Consumption (FOC). FOC is an essential financial concern for fishing vessels, often comprising over 50% of operating costs. Accurate prediction of FOC during fishing trips and events is essential but challenging. The challenge lies in the distinct characteristics of fuel consumption across the operation modes, and the various factors that influence it. This study introduces a novel grey-box modelling approach for the main engine FOC for a tuna purse seiner, combining a Multi-head Bi-directional Long Short-Term Memory (Bi-LSTM) network with a domain-knowledge based operating mode detection methodology. This approach leverages the sequential data processing capabilities of attention-enhanced Bi-LSTM along with domain-specific insights. As a result, the FOC model developed using the proposed approach demonstrated a mean accuracy of at least 97.66% across 10 unseen cruising trips and at least 90.93% across eight fishing events, respectively, which suggests that this methodology has the potential to offer accurate predictions for decision support systems aimed at optimising the operations of tuna purse seiners. • Data from on-board sensors and Copernicus service for data-driven FOC modelling. • Domain knowledge to discern between operating modes for potential model improvement. • A novel grey-box deep learning FOC model informed with operating modes. • High accuracy yielded during both unseen ship voyages and fishing events.