Posters

Abstract

Today, ground-based lidars play a significant role in the wind energy industry, and they are widely accepted. They have many advantages over classic metmast instruments; however, it is still challenging to obtain accurate TI measurements from them. In our work, we aimed at finding a practical solution, oriented towards end-users of ground-based, continuous-wave lidars who have to rely solely on lidars in their measurement campaigns. As a basis for our work, we used measurements from the Østerild test center for large wind turbines. Data from a 250m metmast and two lidars installed at the foot of it was available to us - with a possibility to use measurements from four different heights - 37m, 103m, 175m and 241m. Our approach was to develop different types of neural networks and compare them based on their root-mean square error, mean average percentage error and a linear regression slope between the corrected lidar TI and the sonic TI measurements. The neural networks we developed could be divided into two general groups - physics-based neural networks (PBNN) and data-driven neural networks (DDNN). The difference between the two is the data used for the training. The PBNN is trained with data generated by physics-based models, whereas the DDNN is trained with actual measurements. We also experimented with different sets of inputs for the training to evaluate which would yield the best correction to the lidar TI. For 37m and 103m and wind speeds above 5m/s, we achieved a fairly good correction with the PBNN that uses as an input the variance of the longitudinal component of the wind speed as measured by the lidar. The corrected lidar TI compared very well to the sonic TI, with a slope of slightly above one. Further work is required to improve the performance of both PBNN and DDNN at higher heights.