Research on performance prediction of a small sample centrifugal pump based on a pre-processing approach for imbalanced regression and key hyperparameters optimization

This article proposes a method for optimizing key hyperparameters and its application in predicting the performance of centrifugal pumps. Centrifugal pumps, as essential equipment widely utilized in diverse engineering perspectives such as construction, industry, and water treatment, play a crucial role in both product design and engineering practice, and their performance prediction has been a widespread concern and should be critically looked at. Meanwhile, the traditional numerical simulation methods have limitations regarding computational complexities and time consumption, hence emanating to a growing interest in machine learning methods based on neural networks. By combining a genetic algorithm and a synthetic minority over scaling technique for regression with Gaussian noise and using a simple back propagation neural network, higher prediction accuracy and stability have been achieved in small sample sets, the average absolute percentage error of the head is 7.22%, and the efficiency is only 1.38%. To verify the adaptability of the method, we build a new dataset and compare it with experimental results, and the maximum relative error of the head is 5.06%, while the maximum error of the efficiency is only1.93%. Finally, the reasons for the higher accuracy of efficiency prediction were analyzed.