A wind speed vector-wind power curve modeling method based on data denoising algorithm and the improved Transformer

The complex nonlinear relationship between the wind speed and the wind power, and the singularity of wind speed information leads to the lack of accuracy of current wind power curve modeling. To address the problem, this paper presents a high-precision wind power curve modeling method based on the wind speed vector, including the wind speeds and wind directions at different heights of the wind measuring tower. First, considering the stochastic fluctuation of the wind speed vectors and wind power sequences, complementary ensemble empirical modal decomposition with adaptive noise (CEEMDAN) is used to decompose and reconstruct the highnoise data. Second, based on the reconstructed data, dynamic time warping (DTW) is adopted to analyze the lagged causality between current wind power and historical wind speed series. In order to better mine the rules, the improved Transformer network is proposed with two convolutional layers and multi-head attention mechanisms to develop the wind speed vector-wind power curve model. Finally, through comparative experiments with the mainstream methods, the advancement of the proposed wind power curve modeling method is verified from the perspective of modeling error and its distribution

Automated summary

This paper proposes a high-precision wind power curve modeling method based on the wind speed vector, addressing the lack of accuracy in current wind power curve modeling due to the complex nonlinear relationship between wind speed and wind power, and the singularity of wind speed information. The method includes decomposing and reconstructing high-noise data using complementary ensemble empirical modal decomposition with adaptive noise, analyzing the lagged causality between current wind power and historical wind speed series using dynamic time warping, and developing the wind speed vector-wind power curve model using an improved Transformer network with convolutional layers and multi-head attention mechanisms. Comparative experiments verify the advancement of the proposed method in terms of modeling error and its distribution.