This work addresses the problem of site-specific calibration of engineering wake models for offshore wind farms by combining supervised machine learning (ML) with the PyWake simulation framework and a verification based on dual-Doppler scanning lidar data. Motivated by the need for improved predictive accuracy on how the wake within a wind farm evolves, we study whether machine learning can reliably infer wake expansion parameters from prevailing inflow and additional meteorological features. The open-source software PyWake is integrated as the wake modelling engine and has been extended to account for curtailment by deriving a throttle value for each individual turbine from a comparison between theoretical and active power in the SCADA signals. High accuracy dual-Doppler scanning lidar PPI scans deliver a base of calibration for verification of the wake models.

Our primary research question is: To what extent can supervised ML models learn generalisable dependency of wake expansion parameters from free stream and atmospheric conditions that reduce wind speed prediction errors at turbine locations, relative to engineering wake models with standardised parametrisation? Two engineering wake models, Jensen and TurbOPark are investigated. Concretely, we calibrate parameters per timestamp and assess the value of incorporating meteorological conditions based on ERA5 data.

Free-stream wind speed, direction, and turbulence intensity are derived from turbines, with iterative averaging across free-stream sectors to refine direction. The input to the ML models is enriched using ERA5 reanalysis (1-hour resolution), adding boundary layer height, surface heat fluxes, sea surface temperature, and friction velocity to the feature space. Filtering enforces operational ranges (wind speed ≤32 m/s, direction 200°–270°, turbulence intensity ≤0.5) and removes invalid records and potential wake effects from other wind farm clusters. For each timestamp, we perform brute-force optimisation over a parameter range to minimise MSE against SCADA wind speeds, yielding per-timestamp optimal targets. Supervised ML models are trained on training data, evaluated on test timestamps and ultimately validated with dual-Doppler PPI measurements.

The baseline using the mean of brute-force optima establishes a stringent reference, underscoring the value of data-driven parameterisation over generic offshore values (e.g., k or A ≈ 0.04). Key findings show that learned parameters consistently reduce error relative to fixed baselines. For the Jensen model, Random Forest achieved the highest MSE reduction with ERA5 features, followed by performance without ERA5, while XGBoost showed substantial improvement without ERA5. Gradient Boosting benefitted from ERA5 inclusion. TurbOPark exhibited smaller but consistent gains. Sector-focused calibration and additional throttle filtering further improved MSE performance, with learned values substantially reducing MSE compared to fixed approaches for both Jensen and TurbOPark models.

We conclude that timestamp-level optimisation coupled with supervised learning yields measurable, site-specific accuracy gains. Incorporating ERA5 features as input for the ML model is beneficial, particularly for the Jensen model. The dual-Doppler PPI enables a timestamp-wise analysis of the flow fields and confirms the ML-based parametrisation.

Designing floating offshore wind turbines (FOWTs) that efficiently harness wind energy while withstanding dynamic ocean environments requires a precise understanding of the resultant forces that arise from the interaction between aerodynamics and platform motion. While many simulation studies have attemped to unravel the complexities of unsteady aerodynamic effects on moving wind turbine rotors, a key challenge remains: identifying the most impactful effects. Consequently, there is ongoing discussion about the degree to which unsteady aerodynamic phenomena impact the aerodynamic loads on FOWT. Recent studies indicate that the unsteady airfoil (Theodorsen), dynamic wake (dynamic inflow) and the returning wake effect may occur in typical environmental conditions. The results of a recent numerical study based on free vortex wake simulations show a strong correlation between the dynamic wake effect and the thrust force amplitude of a surging a large-scale rotor. However, this contradict the findings of the IEA Wind Task 30 (Phase III), where experiments and extensive simulations reported almost no influence of unsteady effects on the thrust force amplitude induced by pure surging.
This study aims at investigating this aspect by examining the behavior of the axial induced velocity of a large-scale rotor using a free vortex wake model and a blade element momentum theory method. The investigation reveals that the reduction in thrust force amplitude caused by surge motions at increasing frequencies, as predicted in a previous work, can be directly attributed to the dynamic wake effect. Moreover, it is demonstrated that the influence of the dynamic wake effect on rotor thrust is governed not only by the motion itself but also by the steady-state behavior of axial induction across varying tip speed ratios. Building on these new insights, it can be explained why the dynamic wake effect appears negligible for the model wind turbine studied in the IEA Wind Task 30, yet becomes significant in the case of the large-scale rotor. In fact, the analysis shows that FOWTs experiencing typical tower-top surge motions are generally affected by the dynamic wake effect regardless of the motion period.
Furthermore, considerable differences in the modelling of the induced velocity were found between the free vortex wake and the blade element momentum methods. These differences ultimately lead to deviations between the results of both methods when it comes to the modelling of the dynamic wake effect.

Wake interactions remain a primary source of energy loss in and between large offshore wind farms, where the weak ambient turbulence of the Atmospheric Boundary Layer (ABL) limits natural wake recovery. The present study investigates a novel flow-control concept in which a tethered parakite entrains high-momentum undisturbed flow from above to re-energise wind turbine wakes. A three-dimensional URANS framework is developed for a neutral-stratified ABL inflow and applied to the IEA 15-MW Reference Wind Turbine(RWT). The turbine is represented by an Actuator Disc (AD) whose streamwise momentum sink is prescribed from high-fidelity, blade-resolved reference data as a function of wind speed, thereby retaining realistic induction and radial loading without the cost of fully-resolving the rotor. Several URANS computations were conducted for the AD–parakite configuration to examine momentum flux enhancement in the rotor near wake under different parakite positions. Performance was evaluated using centreline velocity-deficit recovery, turbulent kinetic energy generation, and power density at varying downstream distances from the rotor. Results show that kite-induced entrainment can accelerate wake recovery and mitigate wake-induced power losses over farm-relevant distances. The proposed methodology offers a tractable framework for concept screening and design-space exploration, highlighting the potential of airborne above-rotor flow-enhancing structures to optimise wind-farm performance.

Steady state wake models are widely used in wind farm control and resource assessment studies. To reduce complexity and computational costs, such models do not consider dynamic wake behavior, like wake meandering or time lag when a wind speed or direction change propagates through a wind farm. Some codes, like the Dynamic Wake Meandering and FloriDyn models have been proposed in literature. In this work, we add the dynamic option to the open-source modeling suite FOXES (Farm Optimization and eXtended yield Evaluation Software).
One of the underlying principles of FOXES is the superposition of ambient background flow fields and wakes. Dynamic wakes are computed using the frozen-turbulence hypothesis, i.e., they are transported by the time-varying ambient flow. Because of the simulation’s finite time resolution, this is implemented with Lagrangian wake parcels of finite size, which are emitted by the wake-generating rotor at each time step and then advected through the domain. Technically, this concept requires cross-communication between the parallel computations of neighboring inflow-state chunks, which are enabled by adopting an iterative approach.
Wake meandering is included using an analytical description of the wake center position. It estimates the wake center position at a downstream location using an Ornstein-Uhlenbeck process, similar to a random walk but with a mean-reverting behavior to avoid runaway. This expression includes the standard deviation of the wake center position, which can be tuned to match observational or large-eddy simulation (LES) data. The time series of wake center positions is then used to modify the wind direction signal that serves as input to FOXES.
To analyze the impact of these modifications, we compare the FOXES results of three model settings (steady-state (StSt), dynamic without meandering (DwoM), dynamic with meandering (DwM) to LES data of the PArallelized Large-eddy simulation Model (PALM). Two NREL 5MW turbines with a spacing of six rotor diameters are simulated in a Neutral Boundary Layer with a hub height wind speed of approximately 8 m/s. Two large-scale wind direction changes were simulated: one small gradual change, and one large abrupt change. Upstream (2.5 D) of the turbine, the wind speed and direction signals are extracted and low-pass filtered. These signal are either passed directly to FOXES (StSt and DwoM), or the meandering signal is imposed first (DwM).
Results of the power spectral density of the wake center location show that DwM, as expected, more closely reproduces the LES data than SS and DwoM. Vital for the accuracy of this method is FOXES’ time step. With a time step similar to the LES, DwM reproduces the LES spectra well. When using larger time steps, much of the meandering information is lost, but still more accurate than SS or DwoM at a small time step.
The full paper will also present the impact of these methods on the power estimates for the two turbines, as well as the annual energy production of an offshore wind farm.

The Lattice Boltzmann Method (LBM) offers significant advantages for wind energy simulations through its potential for high-fidelity, computationally efficient flow predictions compared to conventional CFD approaches. This capability is critical for enhancing industry competitiveness through faster and more accurate prediction of wind farm flows, individual turbine performance, and structural loads. Recent advances, such as the cumulant collision operator [1, 2] and GPU acceleration [3], have extended LBM’s applicability to high Reynolds number flows. Previous studies have successfully applied LBM to large-scale atmospheric and wind farm simulations [4, 5], typically using the Actuator Line Method (ALM) to represent turbine blades. However, the explicit influence of secondary structures such as the tower and hub has received limited attention, despite their importance for near-wake turbulence and recovery. This validation study applies the GPU-accelerated LBM framework VirtualFluids to a model-scale wind turbine from the NTNU blind test campaign [6]. The turbine model is extended with body-force representations of the tower and hub in addition to the already implemented ALM blades. Turbulent inflow is generated through a precursor simulation
of the wind tunnel domain, and wake development is analysed at several downstream positions. Preliminary validation results show a distinct tower signature in the velocity deficit and turbulence fields, while the hub exerts a weaker but measurable effect. Comparing simulations with and without tower/hub modelling demonstrates their relevance for accurately reproducing near-wake features observed in the experimental data. Ongoing validation work focuses on calibrating resistance coefficients against wind tunnel measurements and validating wake turbulence characteristics. This model-scale validation provides essential benchmarking for LBM-based simulation tools, supporting the need for fast, high-fidelity industrial wind farm design and optimization.

References
[1] Martin Geier, Martin Sch¨onherr, Andrea Pasquali, and Manfred Krafczyk. The cumulant lattice Boltzmann equation in three dimensions: Theory and validation. Computers and Mathematics with Applications, 70(4):507–547, aug 2015.
[2] Martin Geier, Konstantin Kutscher, Martin Sch¨onherr, Anna Wellmann, S¨oren Peters, Hussein Alihussein, Jan Linxweiler, and Manfred Krafczyk. VirtualFluids – open source parallel LBM solver. Computer Physics Communications, 317:109810, 2025.
[3] Henrik Asmuth, Hugo Olivares-Espinosa, Karl Nilsson, and Stefan Ivanell. The Actuator Line Model in Lattice Boltzmann Frameworks: Numerical Sensitivity and Computational Performance. Journal of Physics: Conference Series, 1256(1):012022, jul 2019.
[4] Henrik Asmuth, Hugo Olivares-Espinosa, and Stefan Ivanell. Actuator line simulations of wind turbine wakes using the lattice Boltzmann method. Wind Energy Science, 5(2):623–645, 2020.
[5] Henry Korb, Henrik Asmuth, and Stefan Ivanell. Validation of a Lattice Boltzmann Solver Against Wind Turbine Response and Wake Measurements. Journal of Physics: Conference Series, 2505(1):012008, may 2023.
[6] Per ˚Age Krogstad and P˚al Egil Eriksen. “Blind test” calculations of the performance and wake development for a model wind turbine. Renewable Energy, 50:325–333, 2013.