Investigation of the surrogate model in an ANN-Meanline Hybrid model for Radial Turbine Performance Prediction

Abstract

The ability to efficiently optimise the charging system as part of the complete powertrain for a given duty is attracting significant research interest. A hybrid meanline model integrating Artificial Neural Networks as surrogate models for loss and blockage prediction has shown great potential in wide-range radial turbine performance prediction, demonstrating enhanced accuracy compared to traditional approaches. However, the configuration of surrogate models employed in the hybrid meanline modelling approach has not been studied thoroughly considering the wide range of geometrical variables and the dimensionality of the problem. This paper presents an investigation into a hybrid meanline model with regard to the choice of the surrogate model algorithm and the corresponding impact of the training database size. By optimizing the surrogate model hyperparameters via Bayesian Optimization, the effect of the hyperparameters on the performance of the surrogate models has been isolated. It was found that the mean square error level of the surrogate model was reduced significantly by 29% when the training database size increased from 60% to 100%. The two surrogate models investigated were ANN and SVR. Both were sensitive to changes in the size of the training data set. The overall results showed that ANN performed better in this turbomachinery loss modelling application.