A thermodynamics-based neural network modeling approach for turbofan engines

Subject to the inaccuracies of the component characteristic maps， the traditional thermodynamic model for on-wing turbofans exhibits a significant modeling error. Moreover，when the model is iterated near the boundary points of the maps， it is prone to non-convergence， rendering it unreliable. To address these issues， a neural network modeling method for turbofans based on thermodynamic is put forward. This method improves modeling accuracy by fully considering the optimization of thermodynamic constraints during the training process of neural network models. By constructing a component-level network structure，implementing a components-cooperating loss function，and applying a fusion training process，we transform the traditional iterative process of the thermodynamic model based on the component characteristic maps into a multi-objective optimization and training process of the component-level neural network. This approach improves the convergence and modeling accuracy of the model. We train and test the model using 26 970 actual engine flight data. The results demonstrate that the maximum error of our proposed modeling method is approximately 7%，even under loose quasi-steady-state data， which is about 5% lower than that of the thermodynamic model based on the characteristic maps. © 2023 BUAA Press. All rights reserved.