Application of a modified MOPSO algorithm and multi-layer artificial neural network in centrifugal pump optimization

Abstract

Centrifugal pump optimization problems usually have strong nonlinear characteristics and are sometimes non-differentiable. The traditional multi-objective particle swarm optimization (MOPSO) algorithm was modified to solve this situation, and performed better with respect to both accuracy and search speed in validation experiments. Based on the modified algorithm and multi-layer artificial neural networks, the shape of the impeller blades of an industrial inline pump was optimized to improve the comprehensive performance under multiple operating conditions. The non-uniform rational B-spline was applied in the parametric design of the blade geometry, and 14 design variables of the spline were finally utilized in the iteration. With constraint of the computational head, the efficiencies of the part-load condition, the nominal condition, and the overload condition were selected as the objective functions. After optimization, a dramatic efficiency rise was obtained in all the three specified operating conditions, and correlation between the inflow conditions before the impeller and the performance of the inline pump was indicated.