Wind power generation is a representative of high-quality new energy. Real-time monitoring and accurate prediction of wind turbines are critical to ensure their stable operation. Due to sensor failures, network congestion, and communication errors, wind turbine monitoring data are often accompanied by data losses which affects the performance of the wind power prediction model. To address the challenge, we propose a two-stage method for inferring missing values in wind power data. First, the missing value supplement and selection of variables with high similarity in changes are applied, and the top-k nearest neighbors are employed to construct coarse-grained estimation. Second, we proposed a multi-view graph learning framework to capture the latent representation of wind power data from three views. The missing values will be inferred based on these latent representations. Finally, experiments with real world data demonstrate that our method has better inference accuracy than traditional and deep learning inference methods.