Innovative approaches to addressing the tradeoff between interpretability and accuracy in ship fuel consumption prediction

Ship fuel consumption is a major component of maritime transport costs and most of its emissions are harmful to the environment. Hence, it is essential to build an accurate ship fuel consumption prediction model, thereby providing reference to the navigation operations. However, maritime industry experts are wary of advanced black-box models since they cannot interpret the outcomes of these models. The application of advanced black-box models in the shipping industry remains limited and it is necessary to develop both accurate and interpretable ship fuel consumption prediction models. This study uses domain knowledge to develop two innovative methods for predicting ship fuel consumption—the first is a physics-informed neural network (PI-NN) model that improves the interpretability of the black-box model while maintaining accuracy and the second is a mixed-integer quadratic optimization (MIO) model that considers more forms of feature variable expressions in an additive white-box model. The proposed approaches address the tradeoff between model interpretability and model accuracy in ship fuel consumption prediction. The experiment results demonstrate that the PI-NN model improves the interpretability of the black-box model while preserving accuracy. The MIO model considers alternative variable expressions, leading to the flexibility of the white-box model. Finally, SHapley Additive exPlanations (SHAP) is used to explain how each feature value contributes to the predictions of the black-box model, thereby providing insights into how each value of feature variables affects fuel consumption. This study provides a solution to the tradeoff between model interpretability and model accuracy and can promote the application of data-driven models in ship fuel consumption prediction. Moreover, this study gives implications for the application of explainable machine learning models in practice.