Comparison of the Performance between Three Doppler wind Lidars and a Novel Wind Speed Correction Algorithm

Doppler wind Lidars (DWLs) have been widely used to detect wind vector variations, based on ground monitoring of atmospheric boundary layer and wind shear. This study evaluates the performance between three DWLs and in situ balloon radiosonde. Lidars data comparison focuses on low altitudes (height < 2 km) from July to September 2021 from three producers: MSD (Minshida), CUIT (homemade), and WP (windprofile) Lidars. Within the research height range, comparisons show the root mean square errors (RMSE) for wind speed were 1.11 m s^-1, 4.45 m s^-1, and 5.15 m s^-1, while wind direction RMSE were shown at 49.83°, 82.89°, and 84.87°, respectively. The measurement accuracy decreases with the altitude increase (up to 2km). The Lidar performance requires a certain amount of aerosol backscattering, when PM_2.5 ranges within 35-50 µg m^-³, MSD Lidar exhibited the highest wind speed correlation (R² = 0.82) with radiosonde, and the wind direction accuracy observed with the three Lidars is enhanced with the increase of aerosol concentration, indicating that particle loading is the critical factor affecting the wind profile. Lidar performance varied significantly with planetary boundary layer heights (PBLH), particularly, the Lidar performance is relatively optimal when the PBLH within 500-750 m, with the Pearson correlation coefficients (PCCs) of wind speed are 0.97, 0.92, and 0.72, while the wind direction is shown at 0.98, 0.75, and 0.70, respectively. The vertical relationship between cloud base height (CBH) and PBLH had also varied influences on the Lidar measurements. Machine learning was used to remove anomalies and complement missing values, the random forest (RF) demonstrated superior performance, with the Area Under the Curve (AUC) of 0.93(CUIT) and 0.90(WP) in the Receiver Operating Characteristic (ROC) curves. RF-based correction of CUIT data enhanced the R² from 0.42 to 0.65. The R² between the RF-based CUIT and Aeolus satellite data was 0.83, indicating that the method effectively improved data, even in circumstances of anomalies. We proposed a new correction algorithm combined with the isolation forest (IF) and RF to handle high-dimensional and incomplete datasets. Our procedure could increase the Lidar measurement quality of wind.