Abstract: With the continuous enhancement of global shipping's requirements for energy conservation and emission reduction,accurate prediction of ship fuel consumption has become crucial for optimizing shipping efficiency and mitigating environmental impacts. However,the issues of feature redundancy and noise interference in multi-source data restrict the accuracy and stability of fuel consumption prediction models. To address these problems,a comprehensive framework for precise fuel consumption prediction is proposed. First,ship noon report data,ERA5 reanalysis data,global ocean physical analysis and prediction data,and Automatic Identification System(AIS) data are integrated,and a multi-dimensional feature space dataset is constructed. Second,to address feature redundancy and noise interference,a genetic algorithm is introduced for feature selection,and the original 32 features are reduced to 15. Third,nine machine learning regression models are systematically evaluated,and the results indicate that the Random Forest model exhibits significant advantages in capturing the nonlinear characteristics of fuel consumption. Finally,feature importance analysis based on the RF model reveals that time-related variables(such as daily operating hours and total operating hours) play a dominant role in fuel consumption prediction,providing quantitative decision support for energy efficiency management. The results show that the genetic algorithm-based feature selection effectively reduces computational complexity by approximately 53.1%,and the Random Forest model performs best in capturing the nonlinear characteristics of fuel consumption,significantly improving prediction accuracy. The research findings provide theoretical support for fuel optimization and emission reduction strategies in the shipping industry,and are of great significance in promoting the green transformation of maritime transport.