A cross-construction method is proposed to establish a wind turbine wake dataset with significantly reduced computational fluid dynamics (CFD) costs. This method involves adjusting one operating parameter, such as the tip speed ratio (TSR), while maintaining the others at their optimal values. This procedure is repeated across another parameter (inflow velocity) to generate a sparse but informative dataset. CFD simulations were performed using large eddy simulation (LES) coupled with an actuator line model (ALM) to generate data. A pre-training and fine-tuning network based on error classification (PFNEC) was developed, achieving high prediction accuracy with coefficients of determination of 0.9750 and 0.9851 for two validation conditions. Two models based on a softmax function and a residual block were designed, and they achieved the best performance, with coefficients of determination of 0.9921 and 0.9891 under different conditions. The Fourier embedding was applied to enhance input features of neural networks. Four samples added to the original dataset improved the prediction accuracy for extreme operating conditions, from coefficient of determination values of 0.7143 and 0.7034 to 0.9939 and 0.9886 with Fourier embedding. This cross-construction method can significantly reduce the cost of dataset establishment. The models exhibited reliable generalization and prediction accuracy.
Tang et al. (Mon,) studied this question.