Simulation of the Droplet Movement in the Interblade Channel of a Turbine Nozzle Cascade Using Neural Networks

Abstract

The issue of application of neural networks for analyzing the regularities of droplet motion in the interblade channel of turbomachines is examined. The droplet flow in a nozzle cascade was numerically investigated in a wide range of steam flow regimes and liquid phase conditions. The calculations were performed using an experimentally verified model of the liquid phase flow. The theoretical Mach number behind the cascade varied from 0.4 to 0.9, the relative density of the liquid phase from 1800 to 5100, the droplet diameter from 5 to 205 µm, the initial slip coefficient of the droplets from 0.1 to 0.9, and the initial angle between the velocity vectors of steam and droplets from ‒15° to +15°. The effect of various parameters on the characteristics of droplet movement through the interblade channel and droplets deposition on the blade surface was revealed. The numerical simulations yielded an array of approximately 1 million droplets, which was used to train neural networks. Based on the analysis of these data, an algorithm for using neural networks to predict the behavior of primary droplets in a turbine cascade was developed. The algorithm includes two neural networks: the first solves the problem of binary classification to determine the probability of a droplet collision with a blade, and the second predicts the features of droplet interaction with the blade surface. This algorithm was tested against a set of data that had not been engaged in the training but were in the same range of parameters. The test set consisted of three flow patterns with four different droplet diameters. The root mean square error determined for the test data set was 5.2% for the relative coordinate of the droplet deposition point and 1.5% for the dimensionless coefficient of collision energy. Estimation of the calculation time for the simulation has revealed that the algorithm using neural networks runs more than 100 times faster than its closest analogue.