Dynamic time delay methods for wind resources in complex terrain

Wind Energy is the process of generating electricity via wind turbines, with wind resources therefore being one of the most crucial factors to consider for its profitability and development. However, not only the stochasticity and turbulence in wind speed makes it a difficult resource to predict but the complex terrain where wind farms are pretended to be installed provides a high challenge for wind planners to rely on wind resources. A better and deeper understanding of atmospheric flow with more robust methodologies are demanded for wind resource assessment. In that case, data-driven science is a methodology proper to such challenge as it focuses on extracting knowledge from data despite the complexity of the system in question. In this thesis, wind speed from a complex terrain was subject to a study from a perspective of dynamical systems and in pure data-driven approach. Pointedly, the Hankel Alternative View of the Koopman Analysis (HAVOK) was used to make a dynamical model out of time-delay coordinates. HAVOK is a universal and data-driven decomposition of chaotic time series into an intermittently forced linear dynamical model. Recently, such approach has been improved with the name of Structured HAVOK (sHAVOK) to make a more accurate model. To do so, a unique and single time series was assessed from averaging hourly-taken measurements from met-masts installed in Mesa Sur, thus making a system to make the dynamical model. Fourier filtering via Fast Fourier Transformation was then used to denoise the time series in four different stages. The first study of this thesis consisted of comparing HAVOK and sHAVOK with different training set proportions at four different levels of filtering considering that the forcing term of the dynamical system is known. Conversely, in the second part of this thesis, the forcing term is not assumed to be known and such term is fitted through Fourier series to make future wind speed measurements, considering the highest level of denoising. This strategy resulted in a good prediction for the first trajectory and wind speed only at first stage of denoising, while fell behind on the remaining stages.