Wind Field Estimation by Machine Learning regression
report
In preparation, the offline LAWINE dataset has been used. A reduced dataset of a few hours of measurement from both a two-beam LiDAR and a Meteorological mast was selected. SCADA data was also used to compute (offline) an estimated rotor speed as a pseudo-signal. This estimated rotor wind speed has been used as the objective signal to be mapped by a number of regression algorithms. The algorithms that have been tested in the scope of this project were two: (1) a Gradient Boost ensemble method and (2) a Feed Forward Neural Network. Using these two approaches, it was possible to obtain both live and (time) predictive estimates of the wind speed at the turbine rotor location using only far-field LiDAR data (measurement range: 80-440m from the turbine). It was observed that the prediction accuracy was best for 5s ahead in time predictions and would then decrease until reaching a minimum for any attempt to predict 30s in advance, or further ahead in time. Furthermore, the second and third measurement ranges of the LiDAR (120-160m, or 1.5D-2D) were observed to have the highest correlation with the objective wind speed estimate. Therefore the range, or more generally the location, of the input measurements has an impact on the accuracy of the prediction and should be smartly chosen.
Topics
TNO Identifier
961412
Publisher
TNO
Collation
17 p.
Place of publication
Petten