Forecasting efficiency parameters of an experimental turbojet engine by a hybrid deep learning approach

Abstract

Forecasting several parameters using machine learning algorithms for micro turbojet engines (MTEs) is of high importance for characterizing engine behavior at different run regimes, thereby facilitating research and design processes. In this study, the MTE is experimentally operated under different operating regimes. Since fuel consumption is the primary determinant of engine performance, the accuracy of prediction models is particularly evaluated in fuel flow-based efficiency parameters. Namely, it is attempted to predict fuel flow, thermal efficiency, exergy efficiency, and improvement potential rate regarding MTE with an artificial neural network (ANN) and hybrid deep learning algorithm called convolutional neural network-long short-term memory (CNN-LSTM). For this aim, five input variables such as RPM, exhaust gas temperature (EGT), compressor outlet pressure, exhaust speed, and air mass flow are preferred. Accordingly, fuel flow varies between 0.000969 kg/s and 0.00475 kg/s, whereas improvement potential rate changes between 43.37 kW and 186.11 kW due to varying RPM. As for modeling outcomes, the thermal efficiency of MTE is forecasted with R2 of 0.988848 by ANN, whereas it is enhanced to 0.996341 by CNN-LSTM. Similarly, R2 is measured for exergy efficiency as 0.968578 with ANN and 0.998723 with CNN-LSTM. Thanks to the hybrid model, the success of modeling of the parameters could be enhanced compared with the ANN algorithm. The results suggest that the efficiency prediction of an experimental turbojet engine with high accuracy via machine learning is seen as a cost-effective method by reducing the need for multiple experiments, and this acquisition could be applied to other gas turbine engines.