Abstract This paper addresses the simulation of turboprop engine performance through Artificial Neural Network (ANN)-based modelling. It arises from the need for developing surrogate models that explore different engine designs and their entire flight envelope operations, while being both fast, accurate, and convenient for preliminary design. First, a database is generated using a commercially available thermodynamic cycle software. Design parameters of interest include performance, thermodynamic and geometric attributes of the engine and its propeller, while operating conditions cover an envelope of ambient temperatures, altitudes, flight speeds and aircraft offtakes. Their variability is captured through a Latin-Hypercube sampled Design of Experiment (DoE). The produced datasets are used to train a Multi-Layer Perceptron (MLP) to predict the engines resulting power and fuel consumption, while results are evaluated based on the Mean Absolute Percentage Error (MAPE). The conventional – random – split of data for training, validation and testing according to set ratios is found to yield inconsistent results, when testing for unseen engines that are not within the training dataset. Thus, a new method for structuring the subsets in terms of whole engine designs is proposed. While both methods result in equal ANN performance on unseen engine designs, the new method overcomes the inconsistency identified for the conventional practice with the validation/testing subsets. The final ANN-based surrogate model yields a 3.3% accuracy while predicting absolute or non-dimensional power. Finally, predictions for fuel flow and Specific Fuel Consumption (SFC) yield a 7% MAPE.
Psaropoulos et al. (Mon,) studied this question.