Comparison of Real-Time Multivariate Dynamic Modeling Methods of Gas Turbine

Abstract This paper presents methods for generating multivariate real-time models of jet engines in real-time by system identification method from experimental data. In this paper, multivariate real-time models of the engine have been built based on two main methods: transfer function and artificial neural network. The output variables that need to be predicted by the model include rotational speed, EGT, compressor temperature, turbine inlet temperature, and compressor pressure. To ensure high model accuracy over the entire operating frequency domain of the engine, a special test was designed, in which an engine fuel supply signal consisting of multiple sine pulses of different frequencies was applied to collect data for generation of models. The advantage of the transfer function-based model building method is that it is possible to examine the dynamic characteristics of the model in terms of amplitude response and phase response on the Bode diagram before selecting the model. In addition, the transfer function-based model building method requires less number of calculations to produce results. Meanwhile, the method of building a model based on neural network has the advantage of being able to simultaneously predict multiple output variables in real time, in which the variables may have nonlinear properties and high complexity that the linear transfer function model cannot simulate.