A Comparative Study on Performance Improvement of Main Engine Fuel Consumption Prediction Models Based on Noon Report Data

Growing demands for maritime decarbonization have increased the need for data-driven approaches to improve fuel efficiency. This study develops a main engine fuel-consumption prediction model using Noon Report data from an operating bulk carrier. A systematic framework was constructed, including preprocessing, voyage-based cross-validation, hyperparameter tuning, and hybrid modeling. Ridge regression was selected as the baseline and further optimized using Optuna. To address residual errors, a Gradient Boosting Regressor (GBR) was introduced as a secondary learner. The hybrid Ridge–GBR model achieved the best performance, improving RMSE from 2.9164 to 2.8137 and R² from 0.9263 to 0.9315. Residual and cumulative-error analyses showed that the hybrid approach preserved the overall characteristics of the Ridge model while providing modest reductions in localized errors, particularly in segments with higher fuel-consumption variability. Feature analysis indicated that propulsion-related variables—engine distance and shaft revolution rate—were dominant in both the Ridge estimator and the GBR residual corrector, while course, mid and aft drafts contributed to correcting remaining nonlinear patterns. Weather-related variables exhibited limited influence in residual modeling. These results demonstrate that the proposed hybrid framework effectively combines linear and localized nonlinear patterns, offering a reliable methodology for fuel-efficiency assessment using routine Noon Report data.