LES-based investigation of wake dynamics: from single wind turbine to wind farm flow interaction

Predictive modelling of utility-scale wind turbines demands numerical frameworks capable of resolving coupled physical processes across a wide range of scales, from blade-level aeroelastic deformations to farm-wide wake evolution within a turbulent Atmospheric Boundary Layer (ABL). This thesis develops and applies a Computational Fluid Dynamics - Computational Structural Dynamics (CFD--CSD) framework in which the fluid domain is resolved through Large Eddy Simulation (LES), rotor aerodynamics is represented via Actuator Line and Disc Models (ALM/ADM), and complex solid boundaries are handled through an Immersed Boundary Method (IBM). Structural dynamics is evaluated with a modal beam solver, yielding fully coupled aeroelastic simulations at tractable computational cost. The framework underpins five distinct but interconnected investigations. The first examines wake coherent structures of the NREL 5MW turbine under two distinct inflow regimes: an idealised pressure-driven boundary layer and a fully developed, Coriolis-driven Ekman layer. Proper Orthogonal Decomposition (POD), Dynamic Mode Decomposition (DMD), and Sparsity-Promoting DMD (SPDMD) are applied to the LES flow fields to extract spatially and dynamically dominant modes. The analysis quantifies the role of atmospheric wind veer in wake skewing, lateral recovery asymmetry, providing the modal basis for physics-informed Reduced-Order Models (ROMs). The second study performs a fully coupled aeroelastic comparison between the NREL 5MW and IEA 15MW reference turbines under identical sheared inflow conditions, isolating the effect of rotor scaling on torsional deformation, blade load distribution, and wake-tower interaction. The results show that torsion-bending coupling becomes increasingly pronounced at larger rotor diameters, introducing non-negligible feedback on sectional aerodynamic loading that linearised structural models fail to capture. In the third study, these atmospheric and structural insights are extended to a two-turbine in-line array of IEA 15MW machines under Coriolis-driven inflow. Yaw misalignment is imposed for active wake steering, and its interaction with wind veer is systematically characterised in terms of farm power output, wake deflection morphology, and downstream blade fatigue loading, which is evaluated through the Rainflow cycle counting method. The fourth study investigates Diffuser-Augmented Wind Turbines (DAWTs) using LES coupled with IBM. Multiple duct geometries are assessed under partial-load Maximum Power Point Tracking (MPPT) operation, quantifying the effect of area ratio and diffuser half-angle on rotor torque augmentation, internal adverse pressure gradient, tip-vortex suppression, and near-wake recovery length. Finally, a Multi-Agent Reinforcement Learning (MARL) framework is developed for farm-level yaw control, benchmarking nine algorithms across three families — PPO, SAC, and TD3 — under independent, CTDE, and fully centralised coordination paradigms, on a canonical three-turbine array and the irregular Ablaincourt Energies farm. Training is conducted in the FLORIS engineering surrogate and policies are transferred to LES to assess the sim-to-real fidelity gap. PPO and entropy-regularised SAC achieve 100\% convergence reliability and power gains of approximately 25\% on the three-turbine array and 10\% on the seven-turbine Ablaincourt farm over the uncontrolled baseline, while deterministic TD3 policies exhibit severe convergence degradation and persistent yaw oscillations — a structural limitation traced to the absence of a stable fixed point in the deterministic policy map.