Accurately predicting the fuel consumption of a ship and associated carbon emissions is of great importance in terms of improving operational costs and satisfying the ever-stringent environmental regulations. This study proposes a new, physics-guided, unified fleet model that fuses proprietary voyage report data with high-resolution weather and ocean data to forecast bunker fuel consumption. The methodology extends prior work by training a single generalizable model on a heterogeneous fleet and engineering a superior physics-informed feature set comprised of direct engine load indicators, dynamic operational features, and performance-based proxies. This study undertook a comprehensive assessment of eleven machine learning algorithms. The results indicated that tree-based ensembles achieve state-of-the-art accuracy, with the top-performing Random Forest and Extremely Randomized Tree models achieving an overall R 2 Test score of 0.93. A granular, per-ship analysis discovered that while bagging ensembles (RF, ET) excel on vessels with challenging data, the neural network generalizes best to the largest ship class. Importantly, a SHAP analysis proves that the vessel’s fundamental geometry, viz., DWT, Breadth, among others, is the overwhelming primary driver of fuel consumption in a multi-vessel context. Given that, the unified, physics-guided approach offers a robust way to predict fuel use and its corresponding CO 2 footprint, yielding key actionable insights for fleet optimization and decarbonization strategies.
Ghosh et al. (Sun,) studied this question.