As the energy transition accelerates, improving wind energy efficiency and forecasting becomes increasingly critical. One key challenge lies in reconstructing high-fidelity atmospheric boundary layer (ABL) flow fields from sparse measurements, especially in regions influenced by wind turbines. This thesis explores the use of Latent Diffusion Models (LDMs) to reconstruct physically plausible ABL flow fields from limited spatial data. Where previous work focused on homogeneous, neutral ABL states, this research extends the methodology to more diverse and realistic conditions by including both stable and neutral boundary layers with embedded wind farms. A conditional diffusion model is trained in the latent space of an autoencoder (AE), using both local measurements and global atmospheric labels. The model is evaluated using statistical, physical, and spectral metrics, and its ability to generalize across multiple ABL regimes is analyzed. Results show that the model can generate realistic reconstructions from extremely sparse inputs, including turbine wake structures, and can potentially serve as a tool for initializing Large Eddy Simulations (LES). These findings mark an important step toward integrating generative models into operational wind forecasting and control systems.

Floating offshore wind energy is a technology that has gained significant interest over the past few years due to its potential to unlock vast, high-quality wind resources far from shore and in deeper waters, areas that fixed-bottom turbines cannot reach. However, harnessing this resource presents unique challenges, among them are maximising annual energy production (AEP) while minimising blade fatigue damage. These objectives inherently conflict. Aggressive control settings that boost energy capture tend to increase loads, accelerating material fatigue. Crucially, both AEP and fatigue life depend on the turbine controller. By systematically varying a set of key features of a proportional–integral (PI) controller, this thesis sets up a workflow to investigate how controller parameters shape the trade-off between AEP and blade fatigue. To navigate this multi-objective landscape, Pareto optimisation is employed, generating a front of controller configurations that balance energy yield against blade durability. Economic viability is assessed through levelised cost of energy (LCOE) calculations, linking extended fatigue life to deferred maintenance costs and potential improved lifetime. Within the existing literature, where past studies have treated AEP, fatigue, and control in isolation, this work offers a unified framework, demonstrating how slight adjustments in controller settings can unlock significant performance gains. The core contribution of this thesis lies in the framework itself. An end-to-end roadmap is presented, combining metocean data analysis from Utsira Nord, aero-servo-hydrodynamic simulations using SIMA, AEP analysis and fatigue life estimation, Pareto front construction, and LCOE impact assessment. Key results show that blade life can be extended by over 15% with less than a 1% drop in AEP, translating to a meaningful reduction in LCOE of at least 3% under typical economic assumptions. These findings highlight the opportunity to explore the controller parameter space further with this system, paving the way for more finely tuned strategies that optimise the balance between the two objectives.

To achieve net-zero greenhouse gas emissions by 2050, the Global Wind Energy Council (GWEC) emphasizes the need for a significant expansion in global wind power capacity. A key factor of this growth is the upscaling of wind turbines, which increases the swept area and exposes the blades to higher wind speeds at higher altitudes, thereby increasing energy yield. However, this upscaling introduces significant control challenges. Larger wind turbines increase aeroelastic complexity, with stronger nonlinear dynamics arising from increased blade flexibility and more significant wind shear across the expanded rotor diameter. Additionally, the increased rotor inertia delays the system response, making rotor speed regulation more difficult under varying wind conditions.

This thesis proposes an adaptive closed-loop Subspace Predictive Control (SPC) framework designed in an attempt to handle the complexity of larger, nonlinear wind turbines. Closed-loop SPC is a direct Data-Driven Predictive Control (DDPC) method that does not rely on explicit state-space modeling, but instead uses measured input-output data to predict future outputs and compute optimal control action. For the optimal control action, it sets up a receding horizon optimization that regulates above-rated rotor speed. This thesis focuses on the above-rated region, where aeroelastic complexity becomes more pronounced due to higher wind speeds and greater wind speed variations, posing significant control challenges.

To capture time-varying and nonlinear behavior more effectively, the closed-loop SPC incorporates Recursive Least Squares (RLS). The controller adapts to time-varying conditions through online parameter estimation using RLS, which updates a locally linear model in real time. Three RLS variants are examined: standard RLS without forgetting, RLS with exponential forgetting, and RLS with directional forgetting. Standard RLS weighs all past data equally, which may be effective when the system dynamics remain stationary but limits adaptability to changing conditions. Exponential forgetting addresses this by placing more weight on recent data, improving adaptiveness, but at the potential cost of losing parameter estimation accuracy in less-excited directions. Directional forgetting refines this further by applying forgetting selectively along the directions of incoming data, preserving excitation in recently unexcited directions and enhancing estimation robustness.

To reduce the phase lag introduced by increased rotor inertia, wind preview information is incorporated into the closed-loop SPC as a feedforward signal. This wind preview is included in the receding horizon optimization problem, enabling the controller to anticipate and proactively respond to upcoming wind changes. Additionally, the wind preview is demonstrated using more realistic measurements obtained through a LIDAR simulator.

The adaptive closed-loop SPC is validated on the DTU 10 MW reference turbine using QBlade, a high-fidelity wind turbine simulator. Various wind scenarios, including gusts, ramps, and turbulent inflow, are evaluated with and without wind preview feedforward. Results demonstrate that the inclusion of wind preview significantly improves rotor speed tracking performance and reduces pitch activity. This improvement is also observed when more realistic LIDAR wind measurements are used in simulations with a turbulent wind field. In the conducted wind cases, among the RLS-based adaptive closed-loop SPC strategies, exponential forgetting combined with wind preview consistently outperformed the other RLS approaches across all scenarios evaluated in this thesis. These findings demonstrate that introducing adaptiveness through forgetting, together with feedforward wind information, can enhance closed-loop SPC performance in rated rotor speed tracking.

Data-Driven Predictive Control (DDPC) has emerged as a promising alternative to Model Based Control (MBC), enabling direct control using Input-Output (I/O) data without requiring explicit model identification. This thesis advances DDPC by bridging the gap between theoretical developments and real-world implementation. The study focuses on four algorithmic variants: Subspace Predictive Control (SPC), Closed Loop Subspace Predictive Control (CL SPC), Recursive Closed Loop Subspace Predictive Control (R-CL SPC), and Constrained Recursive Closed Loop Subspace Predictive Control (CR-CL SPC), to enhance adaptability and ensure real-time feasibility.

A key insight in computational efficiency is the R-CL SPC algorithm, which updates system parameters online using recursive least squares estimation combined with Givens rotations. This reduces computational complexity by avoiding large-scale matrix inversions at each time step. Additionally, the CR-CL SPC variant introduces constraint handling via a Quadratic Programming (QP) solver, enabling input and output constraints.

These algorithms’ performances were evaluated through simulation studies and real-time experiments on a piezo-actuated beam setup, where the control objective was to suppress the first two natural vibration modes. The R-CL SPC algorithm demonstrated a strong balance between computational efficiency and control performance, achieving execution times below 0.2 ms while maintaining effective vibration damping. Meanwhile, though at a higher computational cost, CR-CL SPC validated constraint enforcement capabilities.

This thesis demonstrates that adaptive SPC algorithms can be successfully implemented on real-world hardware. The results contribute to the growing knowledge on direct DDPC strategies and provide a foundation for their broader application in real-time, constrained control systems.

Monopiles are the most commonly used foundation type for offshore wind turbines. Traditionally, monopile foundations were installed using jack-up vessels, which can fix themselves to the seabed using extendable legs. However, as offshore wind farms expand into deeper waters and monopile sizes continue to grow, jack-up vessels are becoming less suitable. To overcome these challenges, SSCV have become viable alternatives. These vessels are equipped with a DP control system, allowing for precise positioning of the vessel without the need for anchors. This advancement significantly improves installation efficiency. However, because these floating vessels are not fixed to the seabed, their movement introduces a new challenge: maintaining the monopile in a vertical position during the installation process while experiencing vessel motions due to waves, currents, and wind. To mitigate this issue, a MCGF can be used.

A major challenge in monopile installation occurs when the tip of the pile contacts the seabed and establishes a connection with the lateral soil. This introduces an abrupt change in the overall system dynamics and increases the risk of DP instability problems of the vessel. These problems are not new to the offshore industry. However, with the introduction of the MCGF, the risk of DP instability increases, as there is now a much stronger dynamic coupling between the vessel, the monopile and the seabed. Therefore, it is important that the MCGF is properly controlled, as this directly influences the reaction forces applied to the vessel. However, properly tuning the MCGF controller is challenging, as it strongly depends on the lateral soil behavior. Currently, these controller gains are tuned using simulations in which the soil behavior is modeled using CPT data. Nevertheless, the estimated behavior based on this data contain large uncertainties and therefore the control settings might become suboptimal during the installation. Furthermore, since soil dynamics also changes during installation, it is beneficial to use a real-time estimation method that takes this into account. Therefore, this thesis investigates different identification methods to obtain real-time estimates of lateral soil behavior during the monopile installation process.

Two main approaches are explored; the augmented EKF approach and the GPLFM approach. The augmented EKF approach shows to be capable of estimating soil behavior by directly identifying the lateral and rotational soil stiffness values, given that the filter is carefully tuned. However, tuning the filter is computationally demanding due to the large number of tunable parameters, which makes this method impractical for site-specific tuning prior to installation using the first measurements obtained. This limitation is critical, as offshore installations face varying conditions at each site, and accurate estimation therefore requires site-specific tuning. To address this tuning challenge, an alternative method is introduced: the GPLFM approach. In this framework, the identification task is reformulated as a GP regression problem. An important advantage of this method is that the process covariance matrix is determined in a data-driven manner, providing a complete covariance structure governed by only a small number of tunable parameters. Consequently, the parameter space is greatly reduced, enabling efficient site-specific tuning. As a result, it is shown that it is possible to obtain accurate estimation results across varying conditions.

Therefore, it is found that the GPLFM method offers a promising solution for real-time estimation of lateral soil behavior during monopile installation. By providing real-time estimates, this study supports the development of more effective MCGF control strategie. This can in the future be used to improve the MCGF controller as it can now be adjusted automatically during the installation process rather than manually. Furthermore, it can help to prevent DP instability problems.

Probabilistic or Bayesian modeling plays a fundamental role in engineering and science, providing a framework for integrating noisy measurements with predictive models through probability distributions.
While probabilistic methods have many benefits, such as recursive estimation and uncertainty quantification, they often come with substantial memory and compute requirements.
Computational challenges are particularly pronounced in large-scale settings, where data sets contain a high number of measurements, and for high-dimensional problems, which require exponentially many parameters to describe probability distributions.
These scenarios can suffer from the curse of dimensionality, which requires exponentially growing computing resources, making conventional approaches computationally intractable.

This dissertation addresses computational challenges by leveraging tensor networks (TNs) to develop computationally efficient probabilistic algorithms.
TNs, also known as tensor decompositions, extend matrix decomposition to higher dimensions by representing large multidimensional arrays, i.e., tensors, in a compact, decomposed format, defined by TN components and TN ranks.
Under the assumption of low-rank structure, TNs enable efficient storage and computation, making large-scale and high-dimensional problems more tractable, even on resource-constrained hardware such as conventional laptops.
The focus of this work is on scalable solutions for Bayesian estimation problems involving Gaussian distributions and exact inference, including recursive filtering and Gaussian process (GP) regression.

Driven by more stable wind conditions and abundant space, the offshore wind market has grown significantly, leading to larger turbines being installed farther offshore in deeper waters. Floating crane-vessels, with their large capacity cranes, can take the next generation of wind farms to new depths. However, the mitigation of wind- and wave-induced motions still remains a challenge.

This thesis explores the use of frequency domain models, full-scale data and Control Moment Gyroscopes to enhance the floating installation of wind turbine towers. The resulting contributions highlight the challenges and relevance of offshore motion compensation, advancing the state-of-the-art and contributing to the future of the wind sector.

Floating wind turbines face persistent wind–wave disturbances that conventional feedback-only controllers—operating without environmental awareness—manage reactively. This thesis advances proactive control by incorporating real-time wave information into turbine control. Building on LiDAR-based wind preview, this thesis investigates the usage of wave preview (e.g., RADAR) and develops data-driven feedforward controllers that account for unmodeled dynamics. Two controllers are synthesized—one to reduce generator-speed fluctuations and another to damp platform pitch—and are integrated alongside the standard speed-regulation loop. Mid-fidelity simulations under realistic seas show attenuation of wave-induced speed variations and pitching within the linear-wave frequency band. Hybrid wave-tank experiments, combining physical hydrodynamics with real-time numerical aerodynamics and control, corroborate these benefits across varied conditions. A second contribution addresses the negative-damping instability and bandwidth limits arising from non-minimum-phase zeros in the blade-pitch–to–generator-speed path. A new control architecture is introduced that mitigates these constraints without additional sensors, preserving industry practice. Together, these results enhance energy capture, reduce fatigue, and improve stability, supporting cost-competitive floating wind.

Wind energy has the potential to accelerate the transition from a carbon-based to a carbonfree energy supply system. This transition is essential in the ongoing global effort to combat the growing impacts of climate change. Due to the availability, as well as the increasingly competitive cost, the wind energy industry has enjoyed rapid growth in terms of installed wind capacity. Where the first onshore wind farm was composed of nearly 5000 turbines in 1980, with a capacity of 576 MW the current largest offshore wind farm in development has a design capacity of 1400 MW with only 100 turbines. Almost all offshore wind farms currently in operation, under construction, or in the planning phase are designed with bottom-fixed turbines and are in relatively shallow water.

The total available wind energy capacity increases significantly when deeper waters can also be accessed by wind turbines and wind farms. For these areas, floating wind turbine technology will play an essential role. When they are deployed in similarly sized wind farms as bottom-fixed wind farms they will also encounter challenges currently faced by these bottom-fixed farms. One of these challenges is the wake interaction between turbines, a cause of significant efficiency losses for a wind farm. The field of wind farm flow control aims to develop a control solution that can alleviate the negative effects of the wake interaction between turbines...

In today's data-driven landscape, the capacity to efficiently process and analyze vast datasets is crucial across various domains, including healthcare, climate modeling, and finance. Despite the growing need for scalable and interpretable machine learning models, traditional approaches, particularly kernel machines, face significant challenges due to the curse of dimensionality. As data complexity increases, the computational and memory demands of kernel machines often become prohibitive, limiting their applicability in high-dimensional applications. This thesis addresses these challenges by investigating the integration of tensor networks (TNs) with kernel machines, aiming to enhance scalability, efficiency without sacrificing predictive power and interpretability.

We propose that TNs, with their ability to represent high-dimensional data through low-rank structures, can effectively alleviate the limitations of kernel machines. Our research is structured around three central inquiries: first, we examine how TNs can accelerate kernel machine scalability while accurately approximating kernel functions; second, we elucidate the theoretical links between TN-constrained kernel machines and Gaussian processes, providing insights into convergence and generalization; finally, we introduce a novel optimization framework characterizing a specific TN, the multi-linear singular value decomposition (MLSVD), in terms of primal and dual problems.

Wind turbines have been deployed worldwide to address the growing energy demand while also targeting ambitious plans for the energy transition to renewable sources. Although wind energy is a key renewable energy source, it still faces unsolved challenges and offers ample room for innovation. One major concern is its variability, with the maximum power available fluctuating over time. Importantly, wakes from upstream turbines also contribute to reducing the power availability for downstream turbines. These wake effects are magnified by the construction of large wind farms with densely spaced wind turbines, aiming to efficiently use the allocated space.

Wind farm control strategies can be implemented to mitigate these wake effects and optimize wind farm power generation. In scenarios requiring on-demand response, such as those explored in this thesis, wind turbines are leveraged to provide flexibility, constrained by their maximum power availability. The power delivery of wind power plants upon request is facilitated by a closed-loop wind farm controller, providing active power control at fast timescales. Active power control involves adjusting the resource's active power to assist power grid operators in balancing energy supply and demand, thereby improving energy security.
Our proposed closed-loop control solution provides superior response capabilities by
compensating for reduced power availability, ultimately enhancing the reliability of on-demand power generation.

The wind variability across turbines, intensified by wake effects, contributes not only to attaining fluctuations in power generation but also to fluctuations in structural loads on the turbines. Amplified by wake-induced turbulence, this structural load variability across turbines leads to uneven degradation of turbine components over the long term. In offshore scenarios, where accessibility is limited and maintenance operations must be minimized due to higher costs compared to onshore counterparts, controlling turbines to prolong their lifetime is of significant interest. In this thesis, this aspect is addressed at both the wind farm and wind turbine levels.

At the farm level, we propose that farms fulfilling grid energy demands must also balance the aerodynamic forces of their turbines to evenly distribute structural degradation among them.
This can be achieved without compromising the power generation when the turbines operate below their maximum energy extraction capacity. We have demonstrated that by implementing a real-time feedback loop, it is feasible to balance aerodynamic loads while meeting wind farm energy demands, albeit limited by wind availability. Moreover, we have demonstrated that balancing aerodynamic forces is advantageous for active power control in a wind farm affected by wake effects, compared to simply distributing power requests uniformly.

At the wind turbine level, we introduced two wind turbine controllers designed to individually restrict real-time aerodynamic loads as a surrogate of structural loads in turbine components. These controllers are referred to as load-limiting controllers. The first load-limiting controller employs an optimal control approach. The operator can impose structural load constraints, using a convex model predictive control for power tracking. The second controller, which is more practical, utilizes a switching mechanism with integral control that allows the operator to prioritize a structural load setpoint over a power demand setpoint. This prioritization aims to reinforce structural safety in situations where turbines are compromised from their design conditions. This could be a consequence of numerous factors, such as unpredictable degradation, installation issues, vessel collisions, and others.

As wind turbines prove to be a viable, reliable, and eco-friendly energy source, new wind farm projects are becoming more ambitious, incorporating a larger number of turbines than ever before. Additionally, there is a substantial growth in wind turbine installations within existing wind farms. This growth in the number of turbines poses an implementation challenge for wind farm control systems. Similar challenges have been encountered in controlling other large-scale systems with collective goals, where agents must instead make decisions based on partial information due to communication limitations in processing or transmission.

Anticipating this implementation challenge, we transition from a centralized to a distributed wind farm control solution. Taking advantage of the time scale inherent in typical wind farm controller implementations, we exchange information with neighboring turbines rather than a central workstation. Our aim, in particular, is not to gather partial information but to achieve consensus across the entire farm. However, our control methodology has a negative implication - the addition of delays - which is carefully examined by the derived stability condition for the design and is assessed through simulations. Notwithstanding these delays, the proposed solution is fully distributed and has been demonstrated to be both simple and effective, facilitating the application of our control solutions in large-scale wind farms.

Lastly, we validate our wind farm control solutions through experiments conducted with scaled wind turbines in full-wake conditions. In this way, we verify the benefits of our control solutions not only through high-fidelity simulations but also through real-world experimentation.

The work presented in this thesis emphasizes the importance of wind turbine controllers capable of offering demanded power to the grid while enhancing reliability in power delivery and addressing structural and maintenance concerns. We introduce closed-loop wind farm controllers designed to handle these challenges. Furthermore, we expand the implementation through a distributed approach on one front, while on the other front, we validate the solutions by means of experiments. The findings from this research contribute to the efficient operation of future wind farms by employing feedback control strategies across clusters of wind turbines.

Despite a large number of numerical and experimental tests and, as of late, linear stability analyses, incomplete knowledge of the physics of manipulated wakes prevents the community from embedding dynamic induction control into analytical control-oriented models. This thesis addresses this gap. We first focused on the framework: we considered the need of modeling blade flexibility in the simulations, then evaluated the turbine model to be used [2]. The actuator line model has been the go-to approach for wind turbines and (small) wind farm simulations in academic contexts for more than a decade but no consensus has been reached over important issues, such as evaluating the free-stream velocity and choosing the width of the smearing function used to project volume forces into the computational domain. The thesis discusses how these issues are connected and proposes an alternative approach to velocity sampling [3].
However confident we can be in our high-fidelity computational framework, it is clear that it cannot be directly used for the optimization of wind farm control strategies as this should, ideally, happen in real-time. However, the results of high-fidelity simulations can be used for reduced order modeling. We simulated turbines in both idealized [1] and realistic [4] atmospheric conditions and with both standard control and dynamic induction control. The data was organized into snapshot matrices and fed to a dynamic mode decomposition (DMD) algorithm. DMD splits the data into purely spatial modes, scalar amplitudes, and purely temporal signals. This makes it suitable for the identification of dominant frequencies. With this approach, a reduced order model is obtained, which, for the analysed cases, is able to reconstruct the full flow field with a maximum 9% relative root mean square error, with only two modes.

The devastating effects of climate change are becoming more evident each day. We are running out of time, and our best collective effort is needed to revert the situation and ensure a sustainable and healthy future. Renewable energies are a key enabler for such a future. Wind energy has attained remarkable progress in the last decades, and the growth required to meet future net-zero scenarios is extraordinary. Wind turbine technology is rapidly evolving towards larger rotors and power ratings, resulting in much higher torque density demands for the drivetrain in general and the gearbox in particular. Maintaining or even improving gearbox reliability with increasing torque density demands is proving to be challenging. Accurate knowledge of the mechanical loads of wind turbine gearboxes has become essential for modern highly loaded gearbox designs with significant dynamic interactions. The contribution of this dissertation can be summarized by its goal: to develop a method to measure dynamic mechanical torque in geared wind turbines. A key requirement was set to enable fleet-wide implementation to monitor torque throughout the complete service life of the wind turbines. This dissertation proposes a method based on strain measurements on the outer surface of the static first-stage ring gear that overcomes the main drawback of traditional methods. A series of experiments are presented, ranging from proof-of-concept tests conducted using gearbox test benches to an extensive field validation campaign. These experiments have advanced the technology readiness level and demonstrated the accuracy and robustness of the proposed method, which is now deemed ready for commercial implementation. Future research should explore improving drivetrain loading using dynamic mechanical torque measurements with novel data-driven control strategies. Additionally, recursively tracking operational deflection shapes over time is recommended for fault detection.

Our planet is warming up with potentially disastrous consequences. The main cause of this climate change is the increase of greenhouse gases in the atmosphere, which are mainly emitted by burning fossil fuels to generate energy. Therefore, fossil fuels need to be substituted to reduce emissions from the energy sector. Renewable energies offer an alternative with reduced emissions. Among these, wind and solar energy are growing the fastest. This thesis investigates how the wind energy supply can be increased by improving its operational efficiency.

There are several reasons why a wind turbine may not generate its maximum capacity, one of them being its placement. Turbines are often placed in farms, which allows the collective use of infrastructure and minimizes land usage. The downside is that the turbines influence one another: As a turbine extracts energy from the wind, an area with lowered wind speed develops downstream. This area is called wake, and other turbines affected by it will generate less energy.

The ways to address this problem can be split into pre- and post-construction measures. Pre-construction the wind farm layout can be optimized, and post-construction control strategies are needed to operate the wind farm optimally. These strategies fall under the term wind farm flow control and aim to manipulate the flow between the turbines to optimize the farm performance. A turbine’s wake can be altered by changing the turbine’s resistance to the flow or by misaligning the turbine with the wind direction. The former leads to a faster wake recovery, and the latter results in a redirection of the wake, also called wake-steering.

The current state-of-the-art of wind farm flow control is to utilize wake-steering in an open-loop control configuration. To this end, steady-state engineering models of the wake are used to optimize the farm set points offline. This is done for a selection of atmospheric conditions and the set points are stored in a lookup table (LuT). During operation, the flow conditions are used to look up the precomputed turbine set points. A problem with this approach arises as open-loop control assumes a perfect match between the model and the actual conditions in the field. There are reasons why this might not be the case: (i) There is an inevitable modeling error, which creates a mismatch between the model and the reality; (ii) conditions can arise that are offline not accounted for, e.g., time-varying atmospheric conditions or layout changes due to turbine downtime.

These problems can be addressed by closing the loop. In closed-loop control, measurements are used to continuously correct the model and to adapt it to the current state of the true wind farm. Optimal set points are then found based on the current model state. The control strategy can, therefore, react to new conditions. A challenge is that the optimization needs to happen online and requires a way to incorporate sensor data into the model. Previous work has designed closed-loop approaches using the same computationally cheap steady-state models that were previously used for open-loop control. This was achieved by adapting the parameters of the model based on the mismatch between the observed and predicted measurements, like power generated. A core assumption these models make is that the flow is in a never changing steady state. However, flow conditions do change, and the large spacing between turbines leads to minutes of delay between the control action the upstream turbine takes and the effect that the downstream turbine experiences. The question arises: What could be achieved using dynamic wake models instead of steady-state ones? These can incorporate wake dynamics, which could lead to better decision-making.

This thesis designs a closed-loop model-predictive wind farm flow control strategy based on a dynamic wake model to maximize the energy generated by a wind farm under time-varying conditions. The thesis is comprised of three building blocks: (i) The development of a dynamic wake model, (ii) the derivation of a sensor fusion strategy to identify the state of the flow field, (iii) the composition of a control strategy that uses the model to optimize the control set points. The building blocks are then connected to form the closed-loop control strategy.

The model building is based on the further development of an existing model, which utilizes a steady-state wake model and reintroduces flow dynamics. In the first step, the underlying wake model is substituted by a three-dimensional one, and the formulation is adapted to heterogeneous flow conditions. In the second step, the model is reformulated as a framework that links to an arbitrary wake model. This is done to profit from advancements in the steady-state model development and to significantly decrease the computational cost of the model. In the third step, the dynamic model is compared to a steady-state one in a set of high-fidelity wind farm simulations under time-varying conditions based on field measurements. The results show that the dynamic model does provide a better match with a simulated wind farm.

In the second part of the thesis, a state estimation methodology is introduced. To this end, an ensemble approach is adopted, where the multiple versions of the model are simulated in parallel. The correlation between the ensembles is then used to correct them based on the predicted and measured wind direction and power measurements of the turbines. A byproduct of the ensemble approach is that each estimated state also has an uncertainty based on how much the ensembles agree on its value.

The third part of this thesis investigates the control and optimization problem. This part focuses on the cost function formulation and the behavior it leads to. In a steady-state frame, the delays do not have to be taken into account, but in a dynamic formulation, they become a challenge. We, therefore, propose a cost-function formulation that synchronizes the control actions with their effect at the downstream turbines. This leads to a series of smaller optimization problems instead of one larger one.

The three building blocks of this thesis are then tested in a case study: The closed-loop controller is employed to maximize the energy of a ten-turbine wind farm under time-varying conditions. Both the farm layout and wind direction time series are based on field conditions. The controller generates an overall energy gain of up to 4% over the baseline using noise-free wind direction measurements.
This is on par with the steady-state approach. However, the closed-loop approach is found to be more robust to disturbed wind direction measurements - Where the performance of the steady-state approach decreases to 1.7% due to the sensor noise; the closed-loop approach still achieves a 2.5% gain.

The conclusion of the work presented in this thesis is thereby: Closed-loop wind farm flow control based on a dynamic engineering surrogate model leads to a more accurate and robust state estimation of the wind farm flow field but, given no preview, does not necessarily lead to a higher energy generation than what can be achieved with steady-state models.

The global demand for renewable energy is driving the rapid growth of the wind energy sector, with wind farms increasingly preferred over stand-alone turbines due to their operational and economic benefits, such as reduced deployment costs, operational efficiencies, and minimized land use. Effective wind farm operation depends on optimizing energy extraction, necessitating accurate energy yield predictions. Traditional wind turbine design focuses on maximizing individual performance, often overlooking interactions within wind farms that can lead to suboptimal overall performance due to wake effects. Wakes are characterized by reduced wind speeds and increased turbulence downstream of turbines, negatively impacting the performance of downstream turbines. Managing these wakes is crucial for improving overall wind farm efficiency.

Wind Farm Flow Control (WFFC) aims to enhance wind farm performance by managing these wake effects through control strategies. This thesis addresses the need for integrating advanced control techniques, such as Individual Pitch Control (Helix Approach), into engineering wake models to improve wind farm simulations and real-time control applications.

To achieve this, the research identifies the most suitable wake model by evaluating various models implemented in PyWake, selecting the Super Gaussian model for its superior accuracy in representing the wake behavior in the whole region. A robust calibration framework is developed, balancing precision and computation time. The thesis investigates the stability of the calibration process, addressing common optimization challenges such as local minima, and identifies strategies to ensure the global optimum configuration.

The sensitivity of model parameters to different inflow conditions is analyzed, allowing the calibration process to be generalized across a range of wind speeds within the partial load region of the power curve. This generalization reduces the need for frequent recalibration, streamlining the process. The research further explores the responsiveness of different control configurations, with a focus on the helix approach, to establish correlations between control signal amplitudes and model parameters.

The newly extended model incorporates helix approach information through parameter calibration for varying pitch angle amplitudes, demonstrating a polynomial relationship between model parameters and control signal amplitudes. This integration reduces the calibration requirement to a single procedure, enhancing model adaptability.

Finally, the extended model is tested for optimizing online wind farm performance across different configurations. The helix approach shows significant improvements in reducing wake losses and increasing overall wind farm efficiency, particularly in turbine clusters. Comparisons with traditional "greedy control" highlight the helix approach's advantage, with power output increasing by up to 25% in clusters of five turbines when upstream turbines utilize helix control.

This thesis contributes to the field by providing a comprehensive approach to wind farm modeling and control, facilitating the practical application of advanced control strategies. The integration of control data into engineering wake models not only enhances calibration processes but also broadens their practical applications, paving the way for more efficient and sustainable wind farm operations.

This thesis investigates the concept of turbine repositioning to enhance energy production in floating wind farms. Due to the dense deployment of floating turbines, downstream units could potentially experience reduced wind speeds caused by the wakes of upstream turbines, leading to decreased power output—an effect known as the wake effect. To address this, methods such as power de-rating and yaw-based wake redirection have been extensively studied. Notably, for floating wind farms, the ability of turbine bases to move within a certain range has prompted the proposal of turbine repositioning as a novel wake mitigation strategy.

This study delves into optimal control strategies for turbine repositioning, with a particular emphasis on manipulating rotor yaw angles. It introduces two primary repositioning strategies: static repositioning, suitable for farms with relatively slack mooring lines, and dynamic repositioning, for those with tighter lines. Alongside, the research proposes optimization methods to identify the optimal control sequences for each repositioning strategy. Lastly, by analyzing rotor yaw angle control sequences in the frequency domain, this study distinguishes the frequency component crucial for repositioning turbines from that steering the wakes. The findings provide significant insights into enhancing the cost-effectiveness of power production in floating wind farms through effective wake interaction management.

The offshore wind energy sector has experienced remarkable growth in recent years, driven by increasing demand for renewable energy sources and technological advancements. As wind farms expand and turbine sizes increase, challenges associated with wake effects and turbine interactions become more pronounced. Therefore, wind farm control strategies have been developed to mitigate the negative effects of wakes, consequently enhancing overall power production or reducing fatigue loads. The helix approach has emerged as a promising method for controlling wake dynamics, shaping the wake into a helical pattern to increase wake mixing further downstream.
In this context, phase synchronisation has emerged as an extension of the helix approach, simultaneously applying the helix on up- and downstream turbines. By leveraging the existing helical periodic components induced by the upstream turbine, phase synchronisation can initiate the helix at the downstream turbine with less pitch action. % while improving overall power performance.
However, power is still compromised at the upstream turbine due to the execution of the helix. Consequently, the potential for improvement arises with the integration of phase synchronisation onto the meandering effect—a phenomenon often already present at the downstream turbine. This effect yields benefits similar to those of the helical wake concerning wake recovery, offering the potential to improve overall power performance without power loss at the upstream turbine.
%
This thesis has developed a phase synchronisation method to track the time-varying excitation frequency inherent to the meandering effect. Enhancements were made to a SOGI-PLL, enabling the extraction of phase, amplitude, and bias from periodic loads, which forms the basis of our closed-loop control strategy for the downstream wind turbine. The improvements led to the development of the ESOGI- and MSTOGI-HTan-QT2PLL methodologies, which address internal disturbances caused by offsets and enhance system stability, while accurately tracking of the phase during phase jumps and frequency steps and ramps.
The methodologies were tested at different levels of fidelity. First, a low-order study validated the effectiveness of the proposed method during uncertain conditions, which showed better performance of the MSTOGI-HTan-QT2PLL methodology. Then, the results of the higher-fidelity experiments demonstrated that our method could successfully amplify the meandering effect without loss in power, while attenuation led to reduced fatigue loads with minimal power loss. These results show the potential of integrating phase synchronisation on the meandering effect.

Floating Offshore Wind Turbines (FOWTs) pave the way to accessing deep water regions with abundant wind resources that are unreachable to bottom-fixed turbines. This technology is not widely deployed due to the increased cost of producing energy. A suitable control strategy can improve the power quality and extend the lifespan of the FOWT, resulting in a reduction of the final cost of energy. However, FOWTs face a specific set of control challenges in the above-rated operational region, such as the negative damping issue, rough environmental conditions, and increased model complexity due to both the floating structure and the increase in wind turbine size. Advanced control strategies are the most suited to resolve the negative damping problem, but obtaining a control model that balances low complexity with good accuracy is an increasingly difficult task. To mitigate the effect of environmental disturbances, feedforward control using wind preview is most commonly employed. Although waves have been shown to increase rotor speed oscillations and turbine loads, wave preview-based methods remain scarce.

This thesis implements a model-free, feedforward controller based on wave preview for the above-rated region of a FOWT. The controller uses a preview of wave forces acting on the floating platform and aims for simultaneous rotor speed regulation and platform motion reduction using collective blade pitch control. As a model-free approach, a modified Data-enabled Predictive Control formulation that considers past and future information about measurable disturbances is proposed. The controller is implemented with a linear model of the NREL 5-MW wind turbine installed on the OC3-Hywind spar-buoy platform and tested in several cases. The effectiveness of the wave feedforward data-driven controller is evaluated in a high-fidelity environment using the QBlade simulator. A decrease in rotor speed variance of 67% and platform pitching motions of 71% is obtained, at the cost of a 7-fold increase in blade pitching effort compared to the baseline controller. In turbulent wind conditions, the wind proved to be the dominant disturbance, and including both wind and wave previews in the controller is recommended. This work demonstrated the feasibility of a model-free feedforward control strategy for wave effect mitigation in FOWTs. Further efforts are required to adapt this strategy to closed-loop operation and to validate its effectiveness across the entirety of the above-rated region.

Wake losses in wind farms, caused by closely spaced turbines, reduce downstream power production and overall efficiency. These losses can be mitigated through wind farm flow control strategies, specifically wake steering, which involves yaw control to misalign turbines with incoming wind, and redirecting wakes to enhance overall power output. This thesis aims to develop a real-time, model-based controller that addresses operational and environmental challenges in large-scale wind farms. The controller adapts to off-design conditions such as changing atmospheric factors and offline turbines, making it more effective than traditional controllers based on pre-optimized lookup tables. The research modifies engineering wake models to account for heterogeneous wind conditions, improves the computational efficiency of the Serial-Refine optimization strategy for real-time use, and incorporates tuning of wake model parameters to ensure adaptability. Additionally, the impact of offline turbines on optimal yaw misalignment angles is examined. Results indicate that adapting wake models using sensor data improves power predictions in non-uniform wind environments, while the distributed optimization framework for real-time yaw adjustments performs comparably to centralized methods. Incorporating offline turbines into the control strategy further boosts power output, leading to increased revenue for the wind farm. This research advances wind farm flow control by developing control policies that are tailored to dynamic, real-world site conditions.