Multistage hybrid model for performance prediction of centrifugal pump. (December 2022)

• Record Type: Journal Article
• Authors: Deng, Hongying; Xia, Zhaoshun; Sun, Zenan; Zheng, Shuihua; Liu, Yi
• Is Part Of: Advances in engineering software. Volume 174(2022)
• Publication Date: 2022-12
• Subjects: Centrifugal pump -- Performance prediction -- Multistage hybrid model -- Process knowledge -- Gaussian process regression; Computer-aided engineering -- Periodicals; Engineering -- Computer programs -- Periodicals; Engineering -- Software -- Periodicals; Periodicals; 620.0028553
• DOI: 10.1016/j.advengsoft.2022.103302
• Languages: English
• ISSNs: 0965-9978

Abstract

Highlights: A multistage hybrid model is constructed for the performance prediction of centrifugal pumps only using the measured speed and valve opening values. The available process knowledge of centrifugal pumps is effectively integrated into the just-in-time data-driven modeling process for practical use. By sequentially introducing the predicted process variables into the input vector, the data-and-knowledge hybrid model can utilize process information to enhance the prediction performance. The method can effectively reduce the dependence on the field sensors to facilitate practical applications. With limited input variables and modeling data, the proposed method can substantially reduce the prediction errors to satisfy the engineering applications. Abstract: To ensure reliable working and reduce energy consumption of centrifugal pumps, it is necessary to describe the relationship between performance indices and operation conditions. However, accurate modeling often encounters several challenges in practice, including some unknown hydrodynamic phenomena, limited test data in complicated conditions, and time-consuming test process. A multistage hybrid modeling method is proposed for practical use in this work. First, a just-in-time model is constructed with similar data collected from an experiment system. The useful process knowledge of centrifugal pumps is then integrated into the modeling process. Moreover, the prediction of the process variables is divided into four stages to reduce the dependence on the sensors and improve the modeling quality. Finally, the performance indices are calculated using the mechanism model. Consequently, the proposed data-and-knowledge hybrid modeling method can replace some field sensors to obtain process variables online and enhance the performance prediction efficiency. Experimental results show its superiority and simplicity for practical use.