There is a gap between the current contribution of wind energy to the global electricity generation mix and its potential capacity. This discrepancy underscores the necessity for addressing social, economic, and technical hurdles that are impeding the broader integration and acceptance of wind energy. The research focuses on tackling the modelling challenges in wind energy by employing Surrogate Model (SM) techniques, combining probabilistic methods, machine learning, and simulation technologies.

This dissertation aims to develop SMs capable of mapping wind time series to the power output as well as extreme and fatigue loads on wind turbines. In this dissertation, I try to answer a number of crucial questions:​ determining the most effective type of SM for this mapping, identifying the optimal sampling method for building these SMs, extending the applicability of the developed SMs with minimal effort, and leveraging publicly available simulation tools and wind turbine models for turbine health assessment. These objectives are essential for improving wind turbine design, operation, and maintenance, enhancing their efficiency and reliability.

Throughout the dissertation, there is an effort to bridge the gap between theoretical research and practical application. The surrogate models developed are presented as a contribution to the integration of theoretical concepts with practical applications in the field of wind turbine design and maintenance. Central to this research is the development of SMs for effectively mapping wind time series to the extreme and fatigue loads experienced by wind turbines. The aim is to find the optimal SM type that balances accuracy with computational feasibility. As the wind turbine faces diverse conditions, I propose adaptable methodologies to optimize the SM performance across various settings. Additionally, I investigate the potential of combining publicly available wind turbine models with probabilistic data-driven models to assess turbine health.

First, a non-intrusive Polynomial Chaos Expansion (PCE) is constructed based on the outputs from the NREL 5MW Blade Element Momentum (BEM) model, demonstrating the convergence of sectional statistics in the results. Subsequently, I utilize the SM to estimate thrust and torque on the rotor and perform a sensitivity analysis of the extreme loads to the number of Monte Carlo Simulations (MCS) in the SM. Transitioning from the PCE realm, I adopt a sequential Machine Learning (ML) method to map wind time series to the Damage Equivalent Load (DEL) of wind turbine loads. I demonstrate that the developed SM, based on a Temporal Convolutional Network (TCN)-Fully Connected Neural Network (FCNN) architecture, is capable of capturing the wind turbine structural dynamics. It demonstrates adaptability in digesting the upstream wakes and accurately estimating the DEL utilizing Transfer Learning (TL). Moving beyond purely synthetic data, I propose the development of a probabilistic data-driven model, integrating limited wind turbine measurements with synthetic data for wind turbine health assessment purposes. I illustrate that an Approximate Gaussian Process Regression (AGPR) trained on a year’s worth of Supervisory Control and Data Acquisition (SCADA) data, combined with simulation outputs from a publicly available wind turbine model, emerges as a promising probabilistic tool for wind turbine health assessment.