Posters

Abstract

Floating Lidars Systems (FLS) play a fundamental role in providing affordable high-quality data within the off-shore wind industry. While the averaged 10-minutes value of the Horizontal Wind Speed (HWS) measured by FLS has proven to be reliable when compared to the output of a traditional cup anemometer, the Turbulence Intensity (TI), defined as the ratio between standard deviation and average of HWS over a 10-minute window, shows significant differences when FLS data is compared to data coming from a fixed Meteorological Mast (MM). This well-known disagreement stems from several aspects that characterize FLS measures, such as different sampling rates, spatial variability and other factors originating from the current state of the surrounding environment. In this work we propose a Machine Learning framework thanks to which it is possible to correct the FLS-measured TI values and thus perform a direct comparison with the TI values coming from a fixed MM. The model has been extensively trained on a high volume of data gathered by EOLOS FLS200 buoys on four Stage-3 validation campaigns at 3 different validations sites (LEG, IJJM, NOAH), for a total of over 542 full days (18 months) of data. The data gathered during every validation campaign is not restricted to wind speed and direction only. Moreover, the values of the same wind variables coming from MM are also fed into the model during the training phase. The trained model therefore takes as input several features on top of wind variables, which is used as a continuous multi-dimensional data stream to correct the FLS TI values and outputs a corrected 10-minute data stream for the FLS TI, which is then compared with the reference data stream coming from a fixed MM. The improvement in the Mean Absolute Percentage Error (MAPE) between FLS and MM TI values before and after the correction is significant, and can range between 30% and 50%, depending on the level of fine-tuning of the model. The model has also the flexibility to be applied from general cases where less data is available to more fined-tuned instances depending on a richer data source as well as different geographical locations.