Artificial neural network for the performance improvement of a centrifugal pump

Abstract

Performance improvement is very important to the energy saving of the pumps and industrial pumping systems. To increase the efficiency at design point, an artificial neural network is applied to construct a non-linear function with high accuracy between the optimization objective and design variables of the impeller, then particle swarm optimization is used to globally optimize the mathematical model. A database consists of 200 sets of impellers generated from Latin Hypercube Sampling method and corresponding efficiencies obtained from numerical simulation. A whole computational domain considering the leakage between the impeller and suction is calculated with SST k-ω turbulence model. Design variables are the distribution of blade angle is controlled by fourth-order Bézier curve with six points. The results show that the numerical performance curve has a faithful agreement with the experimental data. The approximate function can predict the optimization objective with high R-square 0.9311. The pump efficiency at design point is 0.12% higher than the original one. The velocity streamline distribution in the impeller illustrate the optimization eliminates the flow separation at the pressure side of impeller blade.