Machine learning based modelling and control of wind turbine structures and wind farm wakes

With the fast development of wind energy, new technological challenges emerge, which calls for new research efforts to further reduce the cost of wind power. A lot of efforts have been spent to tackle the modelling and control of wind turbines and wind farms. However, big research gaps still exist due to the complexity and strong nonlinearity of the underlying structural and fluid systems. On the other hand, machine learning (ML), which is very powerful in handling complex and nonlinear systems, is developing very fast in the past years. Therefore, this thesis aims to tackle the modelling and control issues arising from the fast-developing wind industry, based on both traditional methods (including structural mechanics, control engineering, fluid dynamics, and scientific computing) and ML (including reinforcement learning, supervised ML, dimensionality reduction, generative adversarial network, and physics-informed deep learning). First, at the turbine level, mitigation of dynamic response of a floating wind turbine using active tuned mass dampers is investigated, where a reinforcement learning algorithm is employed and a neural network structure is designed to realize the employed algorithm. Second, at the farm level, novel static and dynamic wind farm wake models are developed by proposing novel ML-based surrogate modelling methods for distributed fluid systems and then training the model based on highfidelity CFD database generated by large eddy simulations. Third, the prediction of the spatiotemporal wind field in the whole domain in front of a wind turbine is investigated by combining data (i.e. LIDAR measurements at sparse locations) and physics (i.e. Navier-Stokes equations) in a unified manner via physics-informed deep learning. The results presented in this thesis fully demonstrate the great performance of the proposed structural controllers, the great accuracy, efficiency & robustness of the developed wind farm models, and the great accuracy of the full spatiotemporal wind field predictions. xi