Parameter identification study of dynamic characteristics for micro turbojet engines based on AO-NARX

Abstract

To address the strong nonlinearity and time-delay characteristics of micro turbojet engines (MTEs) under dynamic operating conditions, this paper proposes an identification modeling approach that employs the Aquila Optimizer (AO) to jointly optimize the number of hidden layer nodes, learning rate, and input/feedback delay orders of a NARX neural network. Based on a self-developed test bench, five-dimensional input data (including fuel flow rate and inlet temperature) and three-dimensional output data (including rotational speed, thrust, and exhaust gas temperature (EGT)) covering the entire acceleration and deceleration processes were collected. Experimental results demonstrate that, during the relatively stable acceleration phase, the AO-NARX model achieves extremely high fitting accuracy, with mean squared errors (MSEs) for rotational speed, thrust, and EGT predictions as low as 0.689 × 10^-5, 1.054 × 10^-5, and 0.740 × 10^-5, respectively, and coefficients of determination (R²) exceeding 0.999 for all. Compared with PSO-NARX, SSA-NARX, and GA-NARX models, the AO-NARX reduces the MSE for thrust prediction by more than 60% and the MSE for EGT prediction by more than 80%. In the deceleration phase characterized by intense nonlinear fluctuations, the AO-NARX model exhibits remarkable residual suppression capability, maintaining the mean relative error (MRE) for thrust within 5.25% and significantly outperforming the overshoot phenomena observed in the comparison models; the EGT prediction MSE is reduced by 73.0% relative to GA-NARX, demonstrating its quantitative performance in suppressing transient errors and capturing thermal inertia delays.