Experimental and machine learning optimization of centrifugal pump performance

Abstract

In this research, the Long Short-Term Memory (LSTM) technique coupled with response surface methodology (RSM) is used to optimize and predict the performance of a standard centrifugal pump for industrial applications. Model development is done through experimental data collection by varying aspiration pressure (9–95 kPa), discharge pressure (2–57 kPa), motor speed (2200–2600 rpm), and torque (0.31–0.84 Nm). The optimal pump efficiency was found to be 62.57 %. It is observed that the RSM model is capable of predicting the pump efficiency with an R² of 0.999 in comparison to the LSTM techniques, which exhibited an R² of 0.995. The Mean Absolute Error (MAE) is 5.179, the Mean Squared Error (MSE) is 64.38, while the Root Mean Squared Error (RMSE) is found to be 8.02 for the LSTM technique. Likewise, the MAE, MSE, and RMSE were found to be 0.23, 0.08, and 0.29, respectively, for the RSM model. The results of this study highlight that the pump efficiency might be effectively optimized and predicted with high accuracy by utilizing RSM along with the LSTM approach. This study presents a crucial data-driven strategy for optimization and performance prediction of a typical centrifugal pump. The proposed framework not only provides accurate performance predictions but also offers practical guidance for improving pump efficiency, reducing energy consumption, and supporting scalable implementation in industrial pumping systems.