An algorithm for denoising and gap filling of sodar wind data based on the statistical learning theory

Wind energy industries around the world have witnessed a significant surge in recent years. However, to make wind a cost-effective and reliable alternative energy source, the industry will need to make significant scientific advancements on wind power meteorology by reducing the wind power production uncertainties. Modern-day utility-scale wind turbines have blade tips extending up to 150 m above ground level; new turbines are expected to extend even above 200 m. As the future wind turbines are designed in increasingly higher hub-heights and larger rotor diameters, accurate measurements of lower boundary layer wind fields will become more important. Given the high construction, operation, and maintenance costs associated with tall-tower-based wind measurements, the wind energy community is exploring standalone, and inexpensive alternatives. Active ground-based remote sensing instruments like sonic detection and ranging (sodar), radio detection and ranging (radar), and laser imaging detection and ranging (lidar) are gaining popularity for various wind resource assessment, inflow characterization, and short-term forecasting. Modern-day sodars are very portable and can measure vertical profiles of three-dimensional velocity components and turbulence characteristics with high spatial and temporal resolutions.

Susceptibility to ambient noise and data-loss with increasing altitude are two major weaknesses of contemporary sodars. In this presentation, we will introduce an algorithm for sodar data denoising and gap-filling. The proposed algorithm based on the Vapnik-Chevonenkis Statistical Learning Theory is designed to be fully automated. We will test the effectiveness of this algorithm by using sodar wind observations (Scintec Flat Array Sodar system) along with a 915 MHz wind-profiler from an ongoing field campaign in Texas, USA. In-situ wind observations from a collocated 200 m meteorological tower will be utilized for validation. The proposed algorithm will complement the existing sodar/profiler noise reduction schemes by reducing the uncertainty associated with the missing data.