Optimisation of energy performance of axial-flow pump based on conditional tabular generative adversarial network enhancement learning method

In the optimisation of axial-flow pumps, common challenges include the need for large training datasets, high numerical simulation costs, and limited accuracy in fast performance prediction. To overcome these issues, this work proposes an integrated optimisation framework combining a conditional tabular generative adversarial network (CTGAN), a backpropagation neural network (BPNN), the rime optimisation algorithm (RIME), and a standard genetic algorithm (SGA) to improve pump energy performance. Key geometric parameters of an axial-flow pump with a specific rotational speed of 1030 are extracted from numerical simulations to build the parametric model. An initial dataset of 4000 samples is generated through random sampling and evaluated via numerical simulation. CTGAN is then used for data augmentation, producing 1912 high-quality synthetic samples. The entire data-generation and filtering process is completed within seconds, reducing computational effort by approximately 3824 h on 36 CPU cores. The augmented dataset enables the construction of a more accurate energy-performance prediction model using BPNN coupled with RIME, reducing the mean squared error from 1.015 × 10−4 to 7.88 × 10−5. Finally, global optimisation using SGA further enhances pump performance. Compared with the preliminary optimised design, the final optimised axial-flow pump achieves a 0.88% increase in hydraulic efficiency, along with improved internal flow structures and a marked reduction in vortex intensity and turbulent kinetic energy near guide-vane walls. This framework demonstrates an efficient and reliable approach for intelligent design, performance prediction, and optimisation of axial-flow pumps, significantly reducing computational cost while improving accuracy and hydraulic performance. Highlights An integrated framework combining CTGAN, BPNN, RIME, and SGA is proposed. The integrated framework significantly reduces the usage of computing resources. The optimized pump achieves 0.88% higher efficiency with improved internal flow. The improvement mechanisms are revealed by streamlines, vorticity, and TKE.