Surrogate models for twin-VAWT performance based on Kriging and artificial neural networks. (1st April 2023)

• Record Type: Journal Article
• Authors: Chen, Yaoran; Zhang, Dan; Li, Xiaowei; Peng, Yan; Zhang, Xiangyu; Han, Zhaolong; Cao, Yong; Dong, Zhikun
• Is Part Of: Ocean engineering. Volume 273(2023)
• Publication Date: 2023-04-01
• Subjects: Twin vertical axis wind turbines -- Computational fluid dynamics -- Surrogate models -- Kriging -- Artificial neural network; Ocean engineering -- Periodicals; Ocean engineering; Periodicals; 620.4162
• DOI: 10.1016/j.oceaneng.2023.113947
• Languages: English
• ISSNs: 0029-8018

Abstract

Abstract: Though twin vertical axis wind turbines (VAWTs) have great potential in the application in oceanic energy harvest, their aerodynamic characteristics is quite complex, especially in the case of sufficient wake-blade interactions. At design stages, the prediction of their power performance usually relies on high-fidelity unsteady simulations based on computational fluid dynamics, whose time budget is high. In this paper, two surrogate models, i.e., Kriging and artificial neural networks (ANN), were adopted for the performance prediction of a twin-VAWT with a close staggered arrangement. Turbines' pitch angles and their averaged torques at the best tip speed ratio were taken as the input and output, respectively. The numerical study shows that both Kriging and ANN models can provide satisfactory predictions using only 22.45% of CFD observations as training set, and the R 2 values for both upstream and downstream turbine models reach more than 0.99 and 0.98, respectively. Among them, the Kriging-based models appear to be more time-efficient and stable than those based on ANN under the moderate dataset in hand. In addition, the current sampling strategy was tested to be modest and robust through sensitivity analysis. Highlights: Kriging and ANN surrogate models for twin-VAWT prediction were built. 77.55% computational cost were saved compared with pure CFD calculation. The R 2 performance of surrogate model predictions reached more than 0.98 Sensitivity to sampling plans were investigated and compared between models.