A mechanism-embedded neural network model for predicting performance parameters of centrifugal pump

A novel mechanism-embedded neural network (MENN) model is proposed to ensure the prediction accuracy of centrifugal pump performance under limited data conditions. The MENN model, embedded into the neural network module through a ‘Transition layer’, offers a versatile solution applicable to various fluid-mechanical performance prediction problems. Structural optimization using a genetic algorithm (GA) enhances the model's reliability and robustness across different datasets and application scenarios. Comparative analysis with traditional models in predicting centrifugal pump performance demonstrates superior performance, with a significantly smaller average relative error and higher determination coefficient. Specifically, the mean relative errors for predicting head and efficiency are only 0.796 % and 1.240 %, respectively, showcasing a marked improvement in prediction reliability compared to single-model approaches. The method's practicality is validated through successful predictions on a dataset of 63 commercial centrifugal pumps, achieving a determination coefficient greater than 0.997 in the performance prediction model. In summary, this study represents a notable advancement in fluid mechanics, providing a reliable and applicable solution for accurate performance predictions in scenarios with limited sensor and data resources, with broad engineering utility.