Comprehensive feature-based machine learning for fast prediction of marine propeller's open-water performance

Abstract

To address the strong demand for the efficient iterative design of marine propellers, this study researched a rapid prediction surrogate model for propeller open-water performance using a dataset comprising 1980 propeller open-water performance test results. A dimension reduction method based on comprehensive features is proposed, and eight input parameters were determined through correlation and importance analysis. Five machine learning algorithms were utilized to construct the prediction surrogate model employing the Grid Search combined with K-fold Cross-Validation. The validation results indicate that the SVR model performed the best on the validation set, with errors in predicting K_T,10K_Q, and η within 2 %. Further validation was conducted on three unseen propellers in the test set and two new design propeller schemes. It is found that the SVR model, based on comprehensive features, demonstrated good accuracy for the open-water performance prediction of unseen propeller schemes, with errors within 4 %. Compared with the CFD method, the computational performance of the SVR model is approximately 1000 times faster. Additionally, it effectively identifies load variations resulting from overall or local adjustments in pitch distribution, providing new means for rapid performance prediction and optimization design of marine propellers.