Aerodynamics prediction of vertical-axis wind turbines based on meta learning under regional interactions

Abstract

Due to the difficulty in balancing efficiency and accuracy using traditional methods, an increasing number of deep learning models are being used to study the aerodynamic parameters of vertical-axis wind turbines. Nevertheless, interactions between the upwind and downwind regions of vertical-axis wind turbines were often ignored in this process, resulting in inaccurate outcomes. In a meta learning framework, this study proposes a novel deep learning model, called Meta-Double Long Short-Term Memory, to accurately predict aerodynamics of vertical-axis wind turbines. In this model, Double Long Short-Term Memory module utilizes the upwind region and downwind region models to characterize the interactions among regions, while model-agnostic meta-learning is used to capture knowledge across different datasets in a two-stage strategy. Experimental results indicate that this model outperforms other baselines with the lowest errors of 2.21 % and 1.85 % for the peak torque coefficient and the corresponding azimuthal angle, respectively. Global sensitivity analysis reveals that turbine rotational speed (ω) has the greatest impact on prediction results, while upwind aerodynamics and geometrical parameters also significantly affect downwind predictions. Additionally, the proposed model can effectively optimize turbine parameters and provides detailed time series of aerodynamic parameters for in-depth analysis. By using a meta-learning approach and considering regional interactions, the proposed model improves the accuracy of aerodynamic predictions for vertical-axis wind turbines, and its training methodology can be applied to other renewable energy systems.