Determining the Profile Loss of Flow Kinetic Energy in Turbine Cascades with the Use of Neural Networks

The article addresses matters concerned with the use of neural networks for predicting the gas dynamic characteristics of turbine machinery cascades. The results of elaborating the architecture of deep machine learning models for determining the profile kinetic energy loss downstream of plane nozzle vane and rotor blade (impulse type) turbine cascades are presented. A procedure for preparing the training dataset with using numerical simulation of viscous flows is described. The dataset generated is analyzed; its shortcomings, which should be removed for improving the quality of trainable neural networks are identified. Work on selecting the architecture of neural networks for rotor and nozzle vane cascades was carried out. The studies have shown that the same structure of models is efficient for both nozzle vane and rotor blade cascades. The use of prepared models yielded good agreement between the predicted results and the data available for all types of the cascades considered. It is pointed out that the neural networks yield incorrect predictions in transonic and supersonic operation conditions near the theoretical Mach number downstream of the cascade equal to unity. This stems from the lack of information on such operation conditions in the training dataset. After the models had been additionally trained under supersonic operation conditions, it became possible to “trace” the influence of the flow wave structure on the power performance characteristics downstream of the cascade. The data obtained served as a basis for stating the importance of representing curvilinear blade passages in parametric form and the necessity to prepare the training data in a wide variation range of the majority of independent parameters. The neural networks have demonstrated high-efficient performance in solving the stated problem, which made it possible to formulate a number of algorithmic concepts for applying them in solving turbine stage design problems.