A stacking ensemble learning for ship fuel consumption prediction under cross-training

Accurate ship fuel consumption prediction is vital for the shipping industry. In this study, a stacking ensemble learning is developed to predict tanker fuel consumption precisely, built on cross-training of the first-level learner. Among comparative experiments, stacking with Bayesian regression (BR) as the meta-learner and extremely randomized trees (ET), gradient boosting decision tree (GBDT) and light gradient boosting machine (LGBM) as firstlevel learners achieves superior performance, yielding the best results. The root mean square error (RMSE) on the test dataset is 0.2679, and on the training dataset is 0.1327. Ensemble model-based feature importance analysis reveals that ship attributes (speed, draught, trim) contribute around 80 %, while meteorological features contribute about 20 %.