Optimization of horizontal axis wind turbine performance with the dimpled blades by using CNN and MLP models

Abstract

In this study, four NNs models are designed by three types of data: CFD data of original (no-dimpled), dimpled blades, and experimental data. These data are used to estimate torque and thrust. Models of the designed networks are Multilayer Perceptron (MLP) and the Convolutional Neural Network (CNN) so that the input of MLP is a vector and the input of CNN is a matrix. The networks are trained with the data ten times. Each training process obtains estimated torque and thrust with small errors. According to the MLP, the average error of MSE for estimating the torque test data is 4.64e-3 ±5e-4, and this error for estimating the thrust test data is 1.34e-3 ±5e-4, while these errors for CNN are 1.4e-3 ±5e-4, and 1.2e-3 ±5e-4 respectively. As a result, CNN models have better performance in comparison to MLP. Finally, the optimal point for the two networks that estimated the torque is obtained using a Maxnet network. The MLP network correctly finds the optimal point five times out of ten, and the CNN correctly finds the optimal point eight times out of ten, which is consistent with the CFD data. MLP and CNN estimate the torque and thrust with high accuracy, and the optimum point of the turbine is correctly recognized. Therefore, these models can be implemented to reduce the computational costs of CFD simulation.