Long-Term Turbulence Intensity Characterization and Evaluation of Statistical, Machine Learning, and Remodelling Calibration Methods

Abstract

An accurate long-term characterization of turbulence intensity is crucial for assessing wind conditions at future wind farm sites. Yet, no standardized method exists for extrapolating shortterm measurements into reliable long-term predictions. This thesis investigates how well selected statistical, machine-learning, and remodelling methods can predict long-term turbulence intensity using short-term measurements and long-term model data. Two contrasting sites are analyzed: an offshore site with stable wind conditions and an onshore site with a more irregular wind speed range. Eight calibration methods are tested and evaluated using three metrics - MRBE, RRMSE and DEL - and further refined through a three-tier post-processing correction framework. Results show that the Vortex Times Remodelling and Extreme Gradient Boosting perform most consistently across MRBE and RRMSE, particularly at the offshore site. The three-tier correction framework can partially improve variance representation and extreme behavior across both sites. Regarding the DEL, some methods produce errors below the sites' natural year-toyear variability, meaning they outperform the current industry standard of relying on a single measurement year. In conclusion, none of the tested methods consistently stay within industry-defined thresholds across all conditions, meaning full reliability remains an open challenge. Nevertheless, all calibration approaches bring the long-term model data closer to measured turbulence behavior. This study offers a clear comparison of the methods’ strengths and limitations, providing practical guidance for method selection and improvement in future turbulence intensity predictions.