Optimization of submersible LNG centrifugal pump blades design based on support vector regression and the non-dominated sorting genetic algorithm Ⅱ

Optimizing the impeller blade design of submersible LNG centrifugal pump significantly enhances mechanical efficiency, reducing energy consumption and carbon emissions during LNG storage and transportation. This study focuses on optimizing blades of the centrifugal pump, combining Bezier blade design, CFD simulations, optimal Latin Hypercube sampling, Support Vector Regression (SVR), and Non-Dominated Sorting Genetic Algorithm II(NSGA-II) multi-objective optimization into a systematic method. During the development of this optimization method, it was found that directly applying sampled CFD simulation data for machine learning led to poor overall predictive performance of the models. To address this, a data augmentation method based on Gaussian noise was proposed. Through Bayesian optimization of the machine learning model's hyperparameters, the R2 of the predictive model was successfully increased from below 0.8 to above 0.99. The optimized design improved the prototype pump's efficiency from 70 % to 77.8 % under rated flow and head conditions. This efficiency gain is due to the significant reduction in flow separation vortices between blades and decreased turbulent kinetic energy between the impeller and diffuser vanes. Additionally, sensitivity and linear relationship analyses using the Sobol method and Pearson correlation provided valuable insights for blade optimization design.