Improved centrifugal compressor impeller optimization with a radial basis function network and principle component analysis

Abstract

The development of a fast and reliable computer-aided design and optimization procedure for centrifugal compressors has attracted a great deal of attention both in the industry and in academia. Artificial neural networks (ANNs) have been widely used to create an approximate performance map to substitute the direct application of flow solvers in the optimization procedure. Although ANNs greatly decrease the computational time for the optimization, their accuracies still limit their applications. Furthermore, ANNs also bring errors to the final results. In this study, principal component analysis (PCA) or independent component analysis (ICA) is applied to transform the training database and make a radial basis function network (RBFN), a type of ANN, trained in a new coordinate system. The present study compares the accuracies of three different trained ANNs: RBFN, RBFN with PCA, and RBFN with ICA. Furthermore, the total performances of the centrifugal compressor impeller optimization procedures using these three different trained ANNs are compared. Genetic algorithm (GA) is used as an optimization method in the optimization procedure and influences of GA parameters on the optimization procedure performances are also studied. All results demonstrate that the application of PCA significantly increases the accuracy of trained ANN as well as the total performance of the centrifugal compressor impeller optimization procedure.