Abstract: To solve the inaccurate performance prediction of the centrifugal pumps at multi-operation points, we develop a new method based on a genetic algorithm and support vector regression, which is able to predict the energy performance. 673 sets of multi-operation data were extracted from 35 sets of centrifugal pump performance curves as the test samples, and were divided into the training data group of 538 samples and a test data group of 135 samples. We also selected five centrifugal pumps, specific speed 32.2, 47.2, 58.7, 92.8 and 128.2, to verify our method via predicting heads and efficiencies at six operation points and comparing them with test data. Results show that it can effectively predict the performance of the centrifugal pump under different operation conditions. For these five pumps, the average relative errors of the head and efficiency are 0.49% and 3.76% respectively, demonstrating a significant error reduction in comparison to the corresponding neural network model errors of 1.12% and 4.66%, and the average relative errors increased by 128.57% and 23.94%.

Key words: centrifugal pump, genetic algorithm, support vector regression, multi-operation points, performance prediction