Abstract
Future compressors require the use of novel technologies to improve their efficiency, off-design performance, and thermal management. A frequently discussed option to achieve these goals is the use of active cooling methods in compressors. Maintaining a low temperature during the compression process improves the overall performance of the compressor such as its efficiency or mass flow capacity. It is thus essential to consider the effect of cooling during the early design stages. However, current preliminary design methods rely heavily on empirical data and experience and are, therefore, only partially applicable for the design of cooled compressors. In this work, we demonstrate how modern data-driven methods may be used to obtain a surrogate model for the fast and accurate prediction of performance parameters for cooled compressors. To do so, we train a feed-forward neural network to predict the total pressure and temperature ratio, as well as the mass flow rate of a 4.-stage axial compressor test case with arbitrarily cooled stator vanes. The performance predictions of our machine learning model extend over the full (numerical) operating range of the test case and deviate by less than 1% from computational fluid dynamics (CFD) simulations on the test data set. We further demonstrate and discuss the accuracy and generalisation capabilities of this approach by predicting entire performance maps for different cooling configurations.
