Optimizing cavitation performance of a centrifugal pump: A hybrid approach of combining Random Forest and feature selection

This study presents a cavitation optimization framework combining feature selection with a Random Forest model to address performance degradation by cavitation in centrifugal pumps. A single-stage centrifugal pump with a specific speed of 36.77 is selected as the research object. Entropy production by cavitation (EPC) is introduced as the optimization objective to represent cavitation performance, effectively replacing the computationally intensive critical net positive suction head (NPSH3%) and thereby improving optimization efficiency. Furthermore, to address the dual challenges posed by redundant weakly correlated variables—namely, the inflation of optimization space dimensionality and the degradation of surrogate model accuracy—a feature selection method was implemented to identify significant influential parameters for use as optimization variables. Four surrogate models—Decision Tree (DT), Back Propagation Neural Network (BPNN), Extreme Gradient Boosting (XGBoost), and Random Forest (RF)—were selected to assess the effectiveness of feature selection. Through the feature selection approach, the Adjusted R2 coefficient of these surrogate models achieved a maximum improvement of 49.15%. Among the surrogate models, the Random Forest model demonstrates superior fitting performance, achieving Adjusted R2 coefficient of 0.93 for head and 0.94 for EPC. Final optimization using the non-dominated sorting genetic algorithm (NSGA-II) yields a 2.80% head increase, 3.89% NPSH3% reduction and 2.54% improvement in efficiency. The optimized impeller exhibits more uniform pressure distribution, reduced vapor volume, and decreased EPC, indicating mitigated cavitation intensity. The weakened cavitation intensity improves local entropy production distribution in wake regions, ultimately enhancing energy conversion efficiency.