A novel dual-stage grey-box stacking method for significantly improving the extrapolation performance of ship fuel consumption prediction models

Ship Fuel Consumption Prediction (SFCP) is the foundation of ship energy efficiency assessment and optimization. However, existing research neglects to examine the model extrapolation performance, leading to significant degradation in predictive accuracy when models face dataset shift. To address this, a novel dual-stage grey-box stacking (DSGBS) model is proposed. First, based on the traditional grey-box model (GBM), a light grey-box model (LGBM) is proposed to enhance the extrapolation ability by incorporating more prior knowledge. Then, an improved stacking framework is used to fuse multiple GBMs to build the DSGBS model. Finally, a physics-based white-box model (WBM) is established, along with black-box model (BBM), traditional GBM, and LGBM based on nine machine learning algorithms . The extrapolation performance of these models is compared using data from three independent voyages. Results show that DSGBS model has a significant advantage in extrapolation performance, reducing its RMSE by about 63.51 %, 10.91 %, and 52.52 %, respectively, compared to the best model in BBMs, the best model among GBMs and LGBMs, and WBM. Therefore, the DSGBS model mitigates prediction accuracy loss from dataset shift, significantly improve the extrapolation performance, and support the practical application of ship energy efficiency management, with great significance for reducing operation cost and emission. • A DSGBS model with high extrapolation performance is proposed for ship fuel consumption prediction. • Reveals that BBM's predicted performance significantly degrades in extrapolation. • An LGBM is proposed that relies only on the physical model to generate input features. • First full revelation of the differences in extrapolation performance across models. • DSGBS model improves extrapolation performance by 63.51 %, 52.52 % and 10.91 % compared to BBM, WBM and GBM.