LiDAR-Referenced Inflow Wind Condition Estimation from SCADA Data Using a Deep Learning Model

Accurate inflow wind conditions are essential for operational wind farms. However, wind conditions from the Supervisory Control and Data Acquisition (SCADA) system are significantly affected by rotor-induced disturbances and thus cannot reliably represent the true inflow. Although LiDAR can directly measure inflow wind conditions, its data availability is highly sensitive to environmental conditions, frequently leading to insufficient valid samples. Existing studies generally apply the Nacelle Transfer Function (NTF) to empirically correct SCADA wind speed, yet its accuracy remains limited. Consequently, this study proposes a deep learning model for LiDAR-referenced inflow wind condition estimation from SCADA data. First, variations in LiDAR data availability and their influencing factors are systematically analyzed. The deviations and correlations between SCADA data and LiDAR measurements are quantitatively characterized. Subsequently, a deep learning model is developed, employing a time–frequency dual-branch residual network to extract features from SCADA data, while incorporating the Gram matrix as an additional input to provide auxiliary information. Finally, the proposed method is validated using measurements from two offshore turbines with different rated capacities. The results demonstrate that the proposed approach outperforms comparative methods, enabling more accurate estimation of inflow wind speed and direction.