Turbofan Performance Estimation Using Neural Network Component Maps and Genetic Algorithm-Least Squares Solvers

Abstract

Computational models of turbofans that are oriented to assist the design and testing of innovative components are of fundamental importance in order to reduce their environmental impact. In this paper, we present an effective method for developing numerical turbofan models that allows reliable steady-state turbofan performance calculations. The main difference between the proposed method and those used in various commercial algorithms, such as GasTurb, GSP 12 and NPSS, is the use of neural networks as a multidimensional interpolation method for rotational component maps instead of classical β parameter. An additional aspect of fundamental importance lies in the simplicity of implementing this method in Matlab and the high degree of customization of the turbofan components without performing any manipulation of variables for the purpose of reducing the dimensionality of the problem, which would normally lead to a high condition number of the Jacobian matrix associated with the nonlinear turbofan system (and, thus, to significant error). In the proposed methodology, the component behavior can be modeled by analytical relationships and through the use of neural networks trained from component bench test data or data obtained from CFD simulations. Generalization of rotational component maps by feedforward neural networks leads to an average interpolation error up to around 1%, for all variables. The resulting nonlinear system is solved by a combined genetic algorithm and least squares algorithm approach, instead of the standard Newton’s method. The turbofan numerical model turns out to be convergent, and results suggest that the trend in overall turbofan performance, as flight conditions change, is in agreement with the outputs of the GSP 12 software.