System Identification of an Advanced Geared Turbofan Model Using Neural Networks

Gas turbines exhibit nonlinearities in both steady-state and transient operations, which also vary throughout the flight envelope. This work addresses the challenge of obtaining a high-fidelity simulation model without requiring detailed knowledge about individual gas turbine components and assumes that only input–output data are available. Two approaches were employed to model an advanced geared turbofan engine using data from a high-fidelity simulation. First, in the system identification approach, a quasi-steady model was derived using multivariate orthogonal functions, followed by the estimation of dynamic parameters during throttle transients. In the second approach, different neural network architectures were tested: feed-forward layers, a long short-term memory architecture, and a neural differential architecture. Model performance was quantified through the prediction accuracy of five outputs (fuel flow, thrust, turbine temperature, and shaft rotational speeds [Formula: see text] and [Formula: see text]) using only altitude, Mach number, and throttle as inputs. The results demonstrate that the neural differential architecture significantly outperforms both the system identification method and other neural approaches, reducing the standard deviations of residuals by 80%–90% compared with the conventional system identification method across all model outputs. This work advances the understanding of classical and deep learning methodologies for gas turbine modeling in flight simulation applications.