Abstract: Objectives To address the limited interpretability of conventional black-box models with weakly physical explainability for ship fuel-consumption prediction by machine-learning methods, a dual-driven modeling strategy for by combing the ship speed performance and the maritime data is proposed. Methods Based on preprocessed onboard monitoring data from an ocean-going bulk carrier, a direct black-box fuel-consumption model, a shaft-power transformation based grey-box model, and a wind-wave added-resistance separation based grey-box model are respectively developed. Accounting for the influences of wind and waves on propulsion efficiency, an iterative separation strategy cooperating propulsion efficiency with added resistance is introduced, while the supervised learning, ensemble learning, and deep learning algorithms are employed for fuel-consumption prediction. Results Compared with the direct black-box model, the shaft-power-transformation-based and wind-wave-added-resistance-separation-based grey-box models improve the coefficient of determination by 34.2% and 37.1%, respectively, while reducing the <italic>RMSE</italic> by 75.9% and 88.6%, respectively. Conclusions The proposed grey-box prediction model can more accurately characterize the variation of main-engine fuel consumption under complex sea states, while improving prediction accuracy and enhancing physical interpretability, thus providing technical support for onboard operational energy-efficiency assessment.