Explainable Stratified Sampling Blending Meta Ensemble for Ship Energy Consumption Prediction with Monte Carlo Sensitivity Analysis

Maritime transport is energy-intensive and contributes substantially to global emissions, motivating accurate and interpretable prediction of ship fuel consumption for operational optimization. This study uses four operational predictors (speed, draft, trim, and wind) and observed fuel rate to create an explainable stratified sampling blended meta-ensemble for fuel consumption prediction. Fuel and speed (0.98), draft (0.98), and wind (0.93) have very significant positive correlations, whereas trim (0.62) has a moderate correlation, indicating their usefulness as control variables. Three foundational models are trained and assessed: extra trees (ET), decision trees (DT), and linear regression (LR). With a mean squared error (MSE) of 0.026 and 0.019 for training and testing, respectively, LR attains R2 values of 0.990 and 0.988. With training R² = 0.976, testing R² = 0.893, and a higher testing MSE of 0.171, DT performs less well. With training R² = 0.997, testing R² = 0.987, and MSE of 0.008 and 0.020, ET provides robust accuracy and resilience. LR, DT, and ET are combined in a stratified sampling blending model (BL) to achieve balanced performance (training R2 = 0.993, testing R2 = 0.992; MSE = 0.019 and 0.014). Draft and speed are the primary factors influencing fuel usage, with wind and trim ranking second and third, respectively, according to SHAP, LIME, and Monte Carlo sensitivity calculations. With its rich interpretability and excellent predictive accuracy, the proposed framework offers valuable guidance for data-driven ship energy management.