Modeling of Relative Exergy Destruction for Turboprop Engine Components Using Deep Learning Artificial Neural Networks

Abstract

Abstract This study illustrates an deep learning approach supported by a metaheuristic design targeting the foremost features and parameters of artificial neural network (ANN) framework used in predicting relative exergy destruction ( ( f r e l , d e s t * ) (f_{rel, dest}^*) ) of a turboprop components. The development of deep ANN comprising of three-hidden layers using data obtained considering multiple engine input parameters was accomplished. Once the deep learning ANN frameworks were hybridized with a metaheuristic approach, such as genetic algorithms (GAs). The analysis of errors revealed a close fit involving the predicted values of the model and references made on data in f r e l , d e s t ∗ f_{rel, dest}^* for the engine main components. The use of appropriately chosen values in preceding networks weights produced more accurate testing results (Linear correlation coefficient values (R) for the engine components are found to be between 0.996497 and 0.998986) in networks using three hidden layers compared to those using lower hidden layers. Furthermore, optimizing deep ANNs using GAs delivers not only further improved accuracy (R is calculated to be in the range of 0.998929–0.999966 for the engine components) but also an effective utilization of time in the resulting models.