Loss prediction of axial compressors using genetic algorithm–back propagation neural network in throughflow method

Abstract

While the consistent advance in computational power has enabled the Computational Fluid Dynamics (CFD) an effective tool for compressor performance characterization, the need for quick performance estimates at initial design phase of compressor still requires the use of low order models. Thus, the throughflow method remains the backbone of compressors design process. The accuracy of the throughflow calculation mainly depends on the adopted empirical correlations. However, the traditional empirical models are just accurate for the conventionally loaded compressor at normal working conditions. In this article, the mechanism of blade profile loss generation and the formation mode of existing empirical correlation are studied, and the reason why the traditional diffusion factor based empirical models are not applicable for modern high-loading compressors or conventional-loading blades at negative incidence is also discussed. Then, the Genetic Algorithm assisted Back Propagation Neural Network (GA-BPNN) is used to train the surrogate model for the design and off-design loss prediction along the blade span of the compressor. Based on the test data of four transonic compressor stages, a database containing 72 sets of blade element geometry and about 1400 sets of blade element performance data is established. Considering the different mechanisms of rotor and stator losses at different working conditions, the entire database and surrogate model are divided into four components according to the rotor and stator at positive incidence and negative incidence. Comparing the prediction results of the surrogate model with the traditional empirical correlations and experimental data, the results show that the GA-BPNN is an alternative solution for developing the empirical model.