Presentations

Abstract

Turbulence Intensity (TI) measurements from lidars are one of the unresolved challenges of wind resource assessment. This is why several methods, both physics- and data-driven, have been developed to “calibrate” lidar TI measurements so that they are as close as possible to what a mast-mounted cup anemometer would have measured at the same location. Several models had accurately calibrated Lidar TI by utilizing co-located met masts. However, the wind resource community is still lacking a generalized method for the calibration of stand-alone lidar TI measurements.   This study benefits from a dataset of 42 co-located lidar-mast pairs spread across 4 continents and covering a variety of site characteristics. The dataset includes WindCube 2.0 and 2.1 from Vaisala. We used this dataset to train a gradient-boosted decision tree model through a leave-one-out approach, i.e. the model is recursively trained at 41 locations and tested at the 1 location left out of the training, so that each of the 42 locations act as a stand-alone lidar through the testing. This provides the largest known example of such dataset used for the evaluation of turbulence intensity (TI) lidar measurements through co-located met masts. The dataset includes co-located measurements from several heights, ranging between 40m and 140m, leading to 101 co-located measurements in total.  We evaluated the accuracy of the model by computing the mean relative bias error (MRBE) and the relative root mean square error (RRMSE) between the 10-min timeseries of calibrated lidar TI measurements and the synchronized TI measurements from mast-mounted cup anemometers. We evaluated MRBE and RRMSE using modified DNV JIP thresholds: we weight errors by the frequency of observations in each wind speed bin to avoid flagging sites as failing when minor violations occur in one or few wind speed bins, which are not meaningful when considering the site’s full wind speed distribution.   Our results show that 33 out of 42 sites (79%) meet the thresholds for the frequency weighted MRBE and RRMSE at all heights, while 37 out of 42 (88%) meet the criteria for at least one height. When computing MRBE and RRMSE between cup and raw lidar measurements, only 4 out of 42 sites (10%) meet the criteria at all available heights. Additionally, it is interesting to notice that no pattern was observed across the failing sites in terms of orography, topography, height, stability. This is a new finding compared to what is generally assumed in the wind resource community, where stability, terrain complexity and distance from the ground have been regarded as some of the main parameters driving the difference between lidar and cup turbulence measurements.   Our study shows that it is possible to calibrate TI measurements from stand-alone lidars through a data-driven calibration model, which is trained across a large and diverse dataset. Additionally, across the sites where the model provides a less reliable calibration, there is no consistent pattern in terms of terrain complexity, stability and distance from the ground.