An approach to flow field prediction over the horizontal axis wind turbine by passive flow control device using CNN-Autoencoder model.

Abstract

The objective of this research is to predict the flow pattern on the horizontal axis wind turbine (HAWT) by introducing spherical dimples on the blades' surfaces. For this purpose, 600 sections are defined for the blades of both dimpled HAWT and the original turbine. The airfoil shape images from each section undergo processing and conversion into distance functions (DFs) using the Signed-Distance Field (SDF) method. Following the preparation of the input, which includes the airfoil profiles, and the output, consisting of pressure (P) and velocity (U) contours, four Convolutional Neural Network (CNN) models in auto-encoders form (CNN-AEs) have been designed, two of which include with skip-connection (SC). The mean square error (MSE) and Similarity Index Measure (SSIM) of the CNN-AE without SC for P and U test data has been reported as [Formula: see text], [Formula: see text], 7.970[Formula: see text] ± 1.381[Formula: see text], and 7.256[Formula: see text] ± 1.169[Formula: see text], respectively. The results show that the models can accurately estimate P and U from the airfoil section images. In addition, the hyperparameters of the designed CNN-AE models have been optimized to reduce the number of trainable parameters, which can overcome the limitation of the small amount of data.