Presentations

Abstract

Floating LiDAR System (FLS) turbulence intensity (TI) measurements are well known for being strongly influenced by buoy motion and volumetric averaging effects. Both phenomena significantly affect the standard deviation of 10-minute wind speed measurements and, consequently, the resulting TI estimates. In response to these challenges, several mitigation strategies have been developed within the wind industry in recent years, ranging from deterministic motion compensation (MC) methods to data-driven machine learning (ML) approaches and hybrid solutions combining both paradigms. Among these, EOLOS has developed a dedicated ML-based approach for the correction of FLS TI measurements. One of the key advantages of the EOLOS ML model is its inherent capability to address both motion and volumetric effects simultaneously. This is achieved by training the model using pre- and post-deployment verification campaigns, during which the FLS is benchmarked against a meteorological mast over extended periods. An additional benefit is that the model operates directly on standard 10-minute statistics, enabling significantly faster evaluation of commercial campaign data, both retrospectively and potentially in near real time. However, as a purely data-driven approach, particular care must be taken to ensure the robustness and applicability of the model across a wide range of operating conditions. To address this, EOLOS has previously demonstrated the model’s capability for both site extrapolation and height extrapolation. These studies showed that the methodology can be safely applied to sites not included in the training dataset, as well as to heights above those seen during training. Nevertheless, these developments were constrained to applications involving the EOLOS FLS200 buoy, which constituted the sole source of training data. In this work, we present a significant step forward toward removing this limitation. A new, general statistical pre-processing stage has been developed and integrated with the existing ML framework. This combined transfer-learning approach enables the application of an ML model trained on EOLOS FLS200 data to other floating LiDAR systems that employ the same LiDAR technology. The methodology has recently undergone third-party validation by DNV. The proposed universal, buoy-agnostic approach has been tested across different verification campaigns with different FLS, demonstrating consistent and significant improvements in all relevant TI metrics. In particular, the model generally achieves mean bias error (MBE) values within best-practice limits, typically between −1% and 1%, alongside overall improvements in complementary TI performance indicators. In conclusion, the proposed transfer-learning framework substantially extends the applicability of EOLOS’ ML-based TI correction beyond a single FLS. By enabling robust, buoy-agnostic performance while maintaining industry-accepted accuracy levels, this approach represents a meaningful advancement toward standardized and scalable TI correction for floating LiDAR measurements in commercial wind energy applications.