The scale mismatch between wind turbine hub heights and conventional meteorological masts introduces uncertainties in wind resource assessment. Vertical wind speed extrapolation serves as a critical technique to bridge this spatial gap. Current extrapolation paradigms struggle with two fundamental limitations. Physical models fail to capture non-stationary atmospheric stability, whereas purely data-driven methods depend heavily on unavailable hub-height ground truth. To bridge this gap, this paper proposes a Physically Guided Neural Network framework. By integrating physical boundary-layer principles with an adaptive residual correction mechanism, the model introduces an inductive bias that maps near-surface observations to dynamic wind shear evolutions. The network employs a “Near-Surface Learning and Hub-Height” Transfer strategy. This approach optimizes the model exclusively on multi-level observations from 10 to 70 m to eliminate the dependency on high-altitude target labels. Validation on a 100 MW wind farm dataset, utilizing a 70 m proxy variable evaluation, demonstrates that this framework reduces the wind speed extrapolation root mean square error by 56.48% compared to traditional power law models. Furthermore, downstream theoretical power estimation errors are reduced by 10.72%, effectively mitigating power curve lag phenomena. This hybrid approach establishes a robust and low-cost paradigm for refined wind energy assessment in engineering scenarios lacking tall meteorological monitoring.
Wu et al. (Thu,) studied this question.