Effectiveness of data-driven wind turbine wake models developed by machine/deep learning with spatial-segmentation technique

Abstract

In this paper, the effectiveness of three machine/deep learning algorithms, namely, the artificial neural network (ANN), convolutional neural network (CNN) and U-shape neural network (Unet), in constructing wind turbine wake modeling is investigated. In order to enhance the performance of different neural networks, the spatial-segmentation technique for wake flow field is adopted which aims to divide the original wake field configuration (4D × 50D, D is the rotor diameter) into several small pieces (each with 4D × 6.25D). This is followed by separately training the subdivided small piece of wake flow fields and the resultant sub-models are consolidated to predict the whole wake flow field. Both wake velocity field and turbulence intensity field are predicted by the wake model to facilitate its applications to alleviate both wind turbine power losses and fatigue loads caused by wake interactions. Through comparative study, it is found that by using the spatial-segmentation technique it can significantly reduce the prediction error of the wake velocity but not for the prediction of turbulence intensity. Among the three selected network structures, ANN has the best prediction performance yielding the wake model with the maximum error of 11.6 % near to the rotor place, while for other regions it is generally below 8 %. By dividing the wake flow field into pieces, the maximum error located right behind the rotor reduces to 7.2 % with others less than 6 %. Through further repetitive training analysis, it proves a better and robust wake model can be achieved by ANN with the spatial segmentation. In comparison, the prediction error of turbulence intensity field is higher, but still fairly accurate for the far wake prediction with the error less than 5 %.