Accurate prediction of ship energy consumption is a fundamental requirement for optimizing energy efficiency and reducing emissions in intelligent shipping. However, most existing studies focus on improving algorithmic performance while neglecting the influence of data quality and sample distribution on prediction accuracy. To address this issue, this study applies four data augmentation methods, including SMOTE, GAN, LSTM, and Diffusion Models, to improve the representativeness and diversity of training data. These enhanced datasets are then used with CNN and RF models to analyze prediction performance for ship energy consumption. Experimental results based on real data from a large bulk carrier indicate that the incorporation of data augmentation significantly reduces prediction errors. Among all models, the CNN combined with SMOTE performs best, reducing MAE by 42.1% and improving R2 by 116.5% compared with the original CNN model. The proposed method effectively improves model robustness and generalization under small-sample conditions, providing technical support for ship energy efficiency management and green shipping development.