With the continuous development of the wind power industry toward high power and large capacity, the soaring unit capacity and expanding blade radius of large-scale wind turbines have resulted in increasingly complex spatial distributions of the inflow wind field in front of the turbine and significantly enhanced vertical wind shear effects. The conventional method of characterizing wind conditions using single-point wind speed at hub height can no longer fully reflect the wind speed distribution differences and dynamic patterns within the ultra-large rotor swept area, which is prone to causing issues such as wind power prediction deviations and inadequate adaptability of operation control strategies. To address these challenges, this study proposes a multi-point feature wind speed selection and estimation method for large-scale wind turbines, which can accurately capture key wind speed information in the rotor swept area, overcome the limitations of single-point feature wind speed, and provide data support for the optimal operation of wind turbines.

To achieve the aforementioned research objective, this study adopts a step-by-step technical approach for systematic investigation. Firstly, based on high-precision grid data of the inflow wind field in front of the turbine, a two-stage stepwise feature selection algorithm is proposed, which first carries out preliminary selection with random forest and then implements refined selection via Boruta (RF-Boruta). The Boruta algorithm is employed to conduct significance tests on the feature importance scores output by the random forest model, thereby eliminating redundant and irrelevant wind speed grid points and realizing stable and accurate selection of feature wind speed points within the ultra-large rotor swept area. Secondly, for the selected feature wind speed points, the extended long short-term memory neural network (xLSTM-Mixer) algorithm is introduced, combined with an embedded feature engineering strategy that accounts for input-output delay orders. This strategy fully exploits the temporal correlation and spatial correlation of wind speed sequences, and constructs a unit dynamics-driven ultra-short-term multi-step dynamic estimation model for multi-point feature wind speed points. Finally, to verify the effectiveness and superiority of the proposed method, large eddy simulation (LES) of a 10 MW wind turbine is performed on the SOWFA platform. Meanwhile, 7 typical wind conditions covering the full wind speed range specified in the IEC standards (including complex wind conditions such as shear wind and turbulent wind) are configured for numerical simulation and flow field data collection. The feature selection performance and speed estimation accuracy of the proposed method are comprehensively validated based on the collected high-fidelity data.

The numerical simulation and verification results demonstrate that four representative feature wind speed points, including the hub center, are identified via the RF-Boruta stepwise algorithm. These feature points are arranged at a radius of 50~60 m with an angular interval of 120°, which can effectively cover the key regions of the rotor swept area and fully characterize the spatial distribution features of the inflow wind field. The constructed xLSTM-Mixer model exhibits excellent performance in the multi-point feature wind speed estimation task: the relative error of 80-step-ahead (second-level) prediction for multi-point wind speeds is ≤2.8%, achieving second-level high-precision estimation. Statistical characteristic analysis shows that the Kolmogorov-Smirnov (KS) statistic between the model estimation results and the actual wind speed data is ≤0.2, and the structural similarity index (SSIM) is ≥0.96, indicating a high degree of consistency in both distribution characteristics and structural features between the two datasets. Comparative experiments with mainstream time-series prediction models such as the conventional LSTM and Transformer reveal that the estimation accuracy of the xLSTM-Mixer model is improved by approximately 10%, with distinct advantages in wind speed distribution matching and spatial structure capture capabilities.

The multi-point feature wind speed selection and estimation method proposed in this study effectively breaks through the limitations of conventional single-point feature wind speed, realizing accurate selection and efficient estimation of key wind speed information within the ultra-large rotor swept area. The high-precision multi-point wind speed data provided by this method can reliably support wind power prediction, operation control optimization, and power generation evaluation of wind turbines, helping to enhance the operational stability and energy utilization efficiency of wind turbines. It holds important theoretical significance and engineering application value for promoting the high-quality development of the wind power industry.