Convolutional Neural Network Identification of Stall Flow Patterns in Pump–Turbine Runners

Abstract

Stall flow patterns occur frequently in pump turbines under off-design operating conditions. These flow patterns may cause intensive pressure pulsations, sudden increases in the hydraulic forces of the runner, or other adverse consequences, and are some of the most notable subjects in the study of pump turbines. Existing methods for identifying stall flow patterns are not, however, sufficiently objective and accurate. In this study, a convolutional neural network (CNN) is built to identify and analyze stall flow patterns. The CNN consists of input, convolutional, downsampling, fully connected, and output layers. The runner flow field data from a model pump–turbine are simulated with three-dimensional computational fluid dynamics and part of the classifiable data are used to train and test the CNN. The testing results show that the CNN can predict whether or not a blade channel is stalled with an accuracy of 100%. Finally, the CNN is used to predict the flow status of the unclassifiable part of the simulated data, and the correlation between the flow status and the relative flow rate in the runner blade channel is analyzed and discussed. The results show that the CNN is more reliable in identifying stall flow patterns than using the existing methods.