Application concept of artificial neural networks for turbomachinery design

Abstract

This paper presents the results of an extensive investigation evaluating and improving the development of artificial neural network (ANN) models for turbomachinery design purposes. A set of 1100 differing axial compressor geometries based on 5 single-stage compressor rigs was prepared. Computations with the mean line analysis tool AXIAL™ took place to determine the according compressor maps defined by 15 operating points each. The challenge of ANN model development in terms of dimensionality reduction (feature selection), data normalization, defining the networks necessary plasticity, and network training is discussed using the example of three different models. As a result, the first model is able to predict the total pressure loss of the rotor blade row with a mean magnitude of the relative error (MMRE) of 3.6%. The second model predicts the total pressure ratio with an average accuracy of 0.8%. The third and last model was trained to predict basic geometrical parameters by presenting the load level and the performance data as an input. The achieved MMRE varied between 2.4% and 5.6% in respect of the particular output variable. The results show that ANNs are applicable to develop efficient models for turbomachinery design and analysis purposes, respectively.