Wind Turbine Aeroelastic Simulation and Wake Control: A Verifiable Chain from OpenFAST to Reinforcement Learning

New
License: Apache-2.0 for OpenFAST, FAST.Farm, FLORIS and the IEA Wind Task 55 reference models · Updated 2026-11-11
My rating

Overview

Wind-energy discussions of physical AI routinely compress three layers into one story: high-fidelity whole-turbine aeroelastic simulation, engineering-scale wake modelling for control, and learned control policies. The layers differ sharply in how well they can be verified, and merging them produces claims that cannot be traced.

OpenFAST and FAST.Farm ship complete open-source code, modular documentation and a regression test suite, so their capability boundaries can be checked line by line against the official documentation. FLORIS, the engineering wake model, has been validated against historical SCADA data from three North Sea offshore wind farms and has supported a commercial-scale field trial of wake steering. On the learning side, one line of work has moved a policy from simulation into a wind tunnel, and another has trained a policy entirely in FLORIS and then deployed it, without retraining, into the higher-fidelity FAST.Farm dynamic simulator.

This guide assembles those layers into one chain and marks where verification stops. It also performs a correction. An earlier version of this entry cited an "IEA Wind Task 36 open dataset" and carried a table of prediction errors bucketed by wind speed. Task 36 was Forecasting for Wind Energy, running from 2019 to 2021, and its public outputs are forecast evaluation practice and an information portal, not open multi-machine simulation datasets. The open reference machine models come from the Task 37 and Task 55 series hosted under the IEAWindSystems GitHub organization. The error table has no traceable primary source, so it is withdrawn here and replaced with results that can be checked one by one.

The payoff is twofold: a concrete engineering chain with named repositories, documentation versions and paper identifiers, and a way to tell which wind-energy AI numbers are self-consistent inside a simulator and which have survived physical or high-fidelity dynamic testing.

How It Works

The backbone is the modular coupling framework inside OpenFAST. The official documentation describes OpenFAST as a multi-physics, multi-fidelity tool for simulating the coupled dynamic response of wind turbines. Its body is a layer of glue code that couples aerodynamic, hydrodynamic, control-and-electrical and structural-dynamics modules in the time domain, producing a nonlinear aero-hydro-servo-elastic simulation. The glue code and most modules are written in Fortran following the 2003 standard, while modules may also be written in C or C++. The framework covers two- and three-bladed horizontal-axis rotors, pitch or stall regulation, rigid or teetering hubs, upwind or downwind rotors, and land, fixed-bottom offshore and floating installations.

The module list is itself a statement of coupling scope. AeroDyn carries blade-element momentum theory and rotor wakes; OLAF provides a free-vortex wake option; BeamDyn models composite blades with geometrically exact beam theory; ElastoDyn handles tower, drivetrain and blade dynamics through modal and multibody formulations; HydroDyn computes hydrodynamics for fixed-bottom and floating foundations; SubDyn covers multi-member fixed-bottom substructures; SeaState describes wave kinematics and marine conditions; MoorDyn solves mooring dynamics; ServoDyn implements blade pitch, generator torque and yaw control; InflowWind supplies inflow; TurbSim synthesises turbulent wind; and FAST.Farm extends the single-turbine solver to multi-turbine wake fields with wake advection, sway and merging. OpenFAST and its modules are published under the Apache License 2.0 and are maintained by a team at the National Lab of the Rockies.

This coupling is what makes the pipeline meaningful for physical AI. The surrogate model is meant to stand in for a time-domain solver that already binds aerodynamics, structure, control and hydrodynamics together. Fitting aerodynamic loads and structural responses separately lets coupling error accumulate, and it cannot reproduce the intertwined relationship among flapwise, edgewise and torsional root loads the way a whole-turbine model does.

Two levels exist for wakes and inflow. FLORIS is an engineering wind-farm model built on a Gaussian wake formulation, with velocity deficit, wake deflection, wake combination and turbulence intensity models made pluggable. It targets steady inflow and completes a full-farm computation in seconds to minutes, which is what makes it usable inside a reinforcement-learning training loop. FAST.Farm is the dynamic high-fidelity multi-turbine simulator, layering wake advection, sway and merging on top of the OpenFAST time-domain turbine model so that unsteady inflow and wake evolution are resolved, at far greater computational cost.

Learned policies generally bridge those two levels: trained in FLORIS for sample efficiency, then checked in FAST.Farm for whether they survive time-domain coupling and turbulent inflow. This division of labour, low-fidelity training against high-fidelity verification, is one of the clearer engineering patterns in wind-energy reinforcement learning, and it means a policy's usefulness need not rest entirely on the simulator's own fidelity.

Scope and Assumptions

Several assumptions must be settled before results from this chain can be extrapolated, and they are not formalities.

Fidelity range comes first. FLORIS's Gaussian wake model has been checked against historical SCADA data from three North Sea offshore wind farms spanning 35 to 111 turbines and 4.9 D to 7.2 D spacing. The validation concludes that wake depth and width are described well overall, yet that the model consistently underpredicts wake depth in deep-array situations, and that a further divergence could not be pinned to a specific mechanism and may relate to inflow heterogeneity. The same work notes that FLORIS's inflow-heterogeneity submodel is insufficiently mature, which is why it presents results under both homogeneous and heterogeneous inflow, and it states explicitly that the effect of wake deflection was not explicitly tested, because the turbines involved all operated in conventional commercial mode and always aligned with the inflow. The claim that FLORIS can evaluate wake-steering benefit therefore holds at the model level and lacks validation evidence.

Verification tier is the second assumption. Recent results occupy three escalating tiers: evaluation inside the FAST.Farm dynamic simulator, physical measurement in a wind tunnel, and zero-adjustment transfer of a FLORIS-trained policy into FAST.Farm. These carry very different evidentiary weight, and promoting an in-simulator gain to a field gain is the most common overreach on this chain.

Third, the objective of a wind system is not only energy capture. Farm-level power gain interacts with structural loads, component lifetime, curtailment obligations and grid-integration constraints. Coordinated control can target power, cumulative damage or lifetime extension, and a policy built for one objective does not substitute for another. A power-maximising policy applied to a load-sensitive scenario needs additional constraint modelling and revalidation.

Fourth, the data boundary. Public reference machine models are produced by the IEA Wind task series, positioned as openly available data representative of current turbine technology for aerodynamic, structural, aeroelastic and farm-flow modelling across fidelities. That is a model asset, distinct from a dataset of measured operating records, and treating it as a large-scale machine-learning corpus requires a separate assessment of whether sample counts suffice.

Fifth, controller realism. Actuators have rate and magnitude limits, yaw motion induces additional loading, and a policy with unconstrained action space can emit sequences no real machine can execute. Policies aimed at physical experiments typically add an action-consistency term to suppress chattering, which costs part of the theoretical optimum.

Getting Started

Begin by establishing what the reference assets actually are. The IEAWindSystems GitHub organization hosts the output of the IEA Wind task series, and the OpenFAST-relevant parts are directly usable: Task 37 produced the 3.4 MW land-based, 10 MW offshore and 15 MW offshore reference machines, while Task 55 REFWIND released the 22 MW offshore reference turbine in March 2024 and its first offshore reference wind plant of 74 ten-megawatt machines in April 2024. The Task 37 fifteen-megawatt repository ships both OpenFAST and HAWC2 aeroelastic input files, WISDEM optimisation files, the turbine ontology yaml files, plus SolidWorks CAD and an Ansys finite-element blade model, and states that NREL's Reference OpenSource Controller is required.

Next, pick a fidelity tier and assemble the controller. Credible OpenFAST results need controller, pitch and yaw settings, aerodynamic airfoil polar tables, and structural modal and damping data, and none of these can be omitted. The twenty-two-megawatt repository provides OpenFAST, HAWC2 and QBlade inputs, so cross-code comparison requires aligning controller versions first.

Then establish the wake layer. FLORIS is far cheaper to start with: pick a reference machine and run a single farm case. Its real value, though, is as a reference frame rather than a destination, generating data and training policies that are then transferred into FAST.Farm to see whether they hold.

Introduce reinforcement learning only after that. Implement an interpretable baseline first, such as a wake-steering lookup table or model predictive control, evaluate the learned policy under the same protocol, and check whether gains sit clearly outside baseline variation.

The surrogate-model and data pipeline belongs last. Feasibility depends far more on how inputs, outputs and training data are defined than on network architecture, because coverage of load extremes and operating boundaries usually dominates. If the target is real-time assessment, inference cost and online data quality have to be assessed alongside accuracy.

Comparison with Alternatives

The tools usually listed side by side in wind-energy AI sit at different levels of abstraction and are not substitutes.

OpenFAST is a whole-turbine high-fidelity time-domain solver whose core capability is the coupled response of aerodynamics, structure, control and hydrodynamics; FAST.Farm is its multi-turbine extension, responsible for dynamic wake transfer between machines. FLORIS sits one level down as an engineering wake model and a steady-state whole-farm calculator, resolving neither root bending moment histories nor controller dynamics. The relationship is division of labour: the high-fidelity tier supplies trustworthy loads and dynamic response, the low-fidelity tier supplies field distributions at a sampling rate that training loops can absorb.

Reinforcement learning acts as a policy generator within this chain and contributes no physics of its own, so its results inherit whatever fidelity the training environment provided and whatever verification environment was used for evaluation. That is why a policy trained in FLORIS that largely holds up inside FAST.Farm is a substantive result, while training and reporting inside FLORIS alone is not equivalent evidence.

IEA Wind reference machines and the windIO data format occupy a different dimension again. The reference machines are parametric whole-turbine models defined collaboratively across institutions and aimed at particular installation vintages; windIO is a data format for inputs and outputs of wind energy system models, providing schema definitions, a Python library that validates files against the schemas and loads them into a dictionary, and input files for test-case turbine and plant models. The Task 55 team is developing a 2.0 release of windIO. Toolchain comparisons are a separate group of frequently confused contrasts: ROSCO is NREL's reference open-source controller handling pitch and torque logic and is listed as a requirement in several reference machine repositories; HAWC2 is DTU's aeroelastic code, offered alongside OpenFAST inputs and used for cross-code load comparison; QBlade is the Berlin University of Technology implementation; and Bladed is DNV's commercial code, which also appears in the same comparison study.

The relationship between surrogate routes and physical modelling deserves equal care. A surrogate replaces repeated high-fidelity calls with offline training and online inference, at the price of being trustworthy only inside the covered operating envelope, with an error distribution determined by how data was generated rather than by model structure. Treating a surrogate as a replacement for the physical model is the most common misuse.

Benchmarks and Evidence

This section lists only results traceable to a specific paper, dataset or validation study, each tagged with its verification tier so the reader can judge how far it travels.

Field evidence for wake steering comes from NREL's campaign at the Peetz commercial wind farm in Germany, instrumented with a ground-based lidar, a meteorological tower and two sodars, with nacelle lidars added on selected turbines; results were published in two parts in Wind Energy Science in 2019. The reported order of magnitude is that wake steering can deliver roughly one to two percent additional annual energy production depending on plant size and layout, and that on a typical 300 MW plant two percent corresponds on the order of a million dollars per year. This figure is frequently cited as the practical ceiling for wake steering.

Three verifiable reinforcement-learning results follow. The first trains DDPG to control multiple wind directions within plus or minus 45 degrees across a nine-turbine array, achieving on average about seven percent more farm power than greedy control on the training conditions and about seventeen percent on a smaller subset containing unseen directions. That study also notes an existing wake-management review estimating real-farm wake-steering gains in the order of two to three percent based on multiple experiments and simulations, a reference that both corroborates the field measurement and frames the discount from purely simulated gains.

The second is an offline reinforcement-learning wind-tunnel study using MTD3-BC for yaw control under dynamic wind directions, with three scaled G1 turbines on a turntable in the atmospheric test section at Politecnico di Milano. The policy learns from a pre-collected offline dataset without interacting with a wind-farm simulator, and an action-consistency term constrains yaw smoothness. It delivers roughly ten percent farm-level power gain over the greedy strategy, on par with a data-calibrated model-based wake-steering benchmark. The experimental details are equally instructive: the offline dataset came from roughly 80 GB of results collected under greedy and lookup-table methods, and the inflow direction time series came from five days of 1 Hz records at an onshore site in northern Germany, compressed to 90 minutes and split into nine 610-second wind profiles with direction variation within plus or minus 15 degrees.

The third constrains wake steering by noise: the policy adjusts both yaw angles and power setpoints, satisfying a 40 dB noise constraint across the full turbulent-wind range on a six-turbine test farm, with an average power deficit of only two percent relative to unrestricted greedy control and up to 7.7 percent at high wind speeds, while the industry-standard noise-reduced-operation baseline loses 17 percent on average. The policy was trained exclusively under steady inflow in FLORIS and then deployed without retraining into the higher-fidelity FAST.Farm simulator, largely retaining its performance under both steady and unsteady turbulent inflow. The acoustic component couples a low-fidelity analytical acoustic model with enforced dipole directivity into steady-state FLORIS and validates it against the FAST.Farm acoustic model.

Read together, these three results cover in-simulator, cross-fidelity transfer and physical measurement. The trajectory of evidence in wind-energy reinforcement learning is moving from in-simulation gains toward physical validation, though scaled tunnel conditions and prescribed operating points still require care before mapping onto real farms.

Role in Physics-AI Workflows

The most direct use of this chain inside a physical AI workflow is to build the high-fidelity data supply that learning algorithms consume.

The first use is data generation. Taking reference machines and plant models as baselines, running OpenFAST and FAST.Farm across an operating matrix of wind speed, direction, turbulence intensity, shear and gusts, and exporting root bending moment histories, edgewise and torsional root loads, thrust, rotor speed, yaw angle, pitch angle and power yields the dataset for supervised learning and offline reinforcement learning. What matters here is not the sheer number of runs but the design of the operating matrix, since load extremes and failure modes tend to sit at matrix boundaries.

The second use is two-tier training with a low-fidelity and a high-fidelity layer. FLORIS provides the sampling density needed for policy training while FAST.Farm provides the time-domain coupled check. Aligning outputs between the two tiers is the implementation crux: state inputs and action spaces must stay consistent across layers, otherwise transfer silently degrades into retraining.

The third use places surrogates inside the online part of a digital twin. Real-time farm-level assessment cannot run whole-turbine simulation on every decision, so the surrogate maps state to loads or power quickly, with its valid range defined by the training data and a fallback to the physical model or an alert when operating conditions leave that range.

The fourth use connects reinforcement learning to coordinated wake control. The policy must then handle constraints explicitly: yaw rate and magnitude, induced loading, noise limits, and possible conflicts with grid frequency support or curtailment orders. Constraint handling is the decisive step for deployment and is usually harder than the algorithm itself.

The fifth use is cross-fidelity calibration. FLORIS parameters, especially turbulence intensity and wind-direction variability, suffer a waterbed-type non-uniqueness: different parameter combinations can produce nearly identical energy-ratio curves while behaving completely differently under wake steering. Policies for wake steering therefore require parameter identification rather than inherited default values.

Limitations and Failure Modes

The first failure mode comes from the low-fidelity model's envelope. The Gaussian wake model underpredicts wake depth in deep-array situations, its inflow-heterogeneity submodel is insufficiently mature, and wake deflection has no measured counterpart in that validation. A policy trained under uniform inflow and applied to complex terrain or a site with heavy wake overlap can therefore overstate its benefit.

The second failure mode is the inherent fragility of reinforcement learning. Policies depend on the operating conditions covered in training and behave unpredictably under wind directions, turbulence or machine availability outside that distribution. Offline methods avoid live interaction with the environment but relocate the risk to the representativeness of the offline dataset. Without physical constraints on the action space, a policy may also emit yaw sequences that real equipment cannot execute.

The third failure mode is objective mismatch. A policy trained to maximise power can deliver negative consequences in load- or lifetime-sensitive operation, and adding noise limits, curtailment or grid-support constraints raises both the dimensionality and the difficulty of the control problem, requiring retraining and revalidation.

The fourth failure mode is the gap between simulator and hardware. Scaled tunnel sections differ in Reynolds number effects, turbulent structure, sensor noise and actuator dynamics, so treating tunnel gains as commercial expectations overstates the outcome. The one-to-two percent range measured in the field provides a more conservative reference.

The fifth failure mode is the reference machine itself. IEA reference machines target particular installation vintages and serve modelling comparison, and their parameter sets reflect specific design assumptions rather than replicating any particular machine. Applying their simulation conclusions directly to a commercial turbine requires first checking sensitivity to the parameter differences.

The sixth failure mode is mixing data conventions. Machine power curves, SCADA records and simulator power output can carry different definitions, and load channels can differ in units and reference points. These details are easy to miss when stitching data from several sources, yet they are enough to make a surrogate's error metric meaningless.

The seventh failure mode is version and interface drift. OpenFAST modules and input formats change between versions, reference machine repositories get updated to correct design oversights, and windIO is evolving toward 2.0. Different machine repositories state different OpenFAST version requirements and in several places note that the models have not been tested for backward compatibility. Ignoring version constraints lets a model run while its results stop being trustworthy.

Version and Source Notes

This section records the verification results for this entry item by item, with the basis for each.

First, the identification of an "IEA Wind Task 36 open dataset" containing public multi-machine simulation data does not hold. Task 36 was Forecasting for Wind Energy, started in 2019 and concluded in 2021, operated by Gregor Giebel of DTU and Caroline Draxl of NREL. Its three work packages cover numerical weather prediction physics and information on accessible datasets, an international pre-standard for benchmarking and comparing wind power forecasts including probabilistic forecasts with benchmarks run alongside IEA Wind Task WakeBench, and dissemination of best practice to forecast users. The public outputs are a Recommended Practice guideline set published open access by Elsevier in November 2022 with ISBN 9780443186813, and an information portal listing meteorological masts, meteorological experiments, publicly available forecasting benchmarks, research projects and open weather data. No multi-machine aeroelastic simulation dataset appears among them. Task 51 is the successor task. The resource reference is therefore withdrawn and replaced by the Task 37 and Task 55 reference machine repositories and the windIO data format.

Second, the original wind-speed-bucketed error table, covering random-forest and LSTM error at three ranges from 3 to 12 metres per second, 12 to 25 metres per second and extreme turbulence, has no corresponding traceable primary source, and its test-set construction, error definition and operating conditions cannot be confirmed. The table is withdrawn and the Benchmarks and Evidence section carries locatable published results instead.

Third, the statement that traditional FAST or OpenFAST simulation takes minutes per run and cannot support real-time operational decisions has no traceable source, and it sits awkwardly against published work in which policies were evaluated in FAST.Farm under turbulent inflow, since the cost gap between FLORIS and FAST.Farm is precisely what makes cross-fidelity training possible. That phrasing is not adopted; the claim is restated as a division of labour between a high-fidelity tier and a low-fidelity tier.

Fourth, three source layers must be separated. OpenFAST and FAST.Farm are maintained by an NREL team, with source and documentation in the OpenFAST GitHub organization under the Apache-2.0 licence and maintenance now carried by the National Lab of the Rockies. FLORIS is also an NREL product. The IEA Wind Task 37 and Task 55 reference machines and windIO are coordinated by IEA Wind TCP, produced with NREL, DTU and other institutions, released under Apache-2.0 and hosted under IEAWindSystems. The original entry gave only the GitHub organization link without distinguishing these, so they are labelled separately here.

Fifth, the NREL 5 MW reference turbine is a usable open model that pairs well with OpenFAST, but it is published by NREL and is a separate asset from the IEA reference machines; the two must not be cited as one and the same.

Sixth, the publication status of the MTD3-BC wind-tunnel work needs correcting. Several secondary reports state it appeared at ACC 2024, whereas it is in fact an arXiv preprint submitted on 11 September 2026 with identifier 2609.12905, authored by the ICSE research group at the University of Warwick and the Wind Energy Institute at the Technical University of Munich, with the tunnel campaign conducted at Politecnico di Milano, and marked as submitted to IEEE. The arXiv identifier should be used when citing.

Seventh, the Noise Steering work carries arXiv identifier 2610.36888, first submitted on 27 October 2026 and updated the following day, with authors from a research group spanning Geneva and Barcelona, noted as 17 pages and 11 figures and classified under artificial intelligence, systems and control, atmospheric and oceanic physics, and fluid dynamics.

Eighth, the nine-turbine DDPG study is Expert Systems with Applications volume 282 (2025), titled Investigations into deep Reinforcement Learning for wind farm set-point optimisation, by Sheehan, Poole, Silva Filho, Bossanyi and Landberg, with DOI 10.1016/j.eswa.2025.127627. Its seven percent and seventeen percent gains correspond to two different framings, training conditions and a subset containing unseen directions, and the two must not be used interchangeably.

Ninth, the scale parameters of the IEA reference machines should be taken from the reports: the twenty-two-megawatt offshore reference machine has a 284 metre rotor diameter and 170 metre hub height, is a class 1-B machine with specific power approaching 350 W per square metre, mounts on fixed-bottom or semi-submersible floating platforms and targets installation in the 2025 to 2030 timeframe; the fifteen-megawatt offshore reference machine exists in monopile and semi-submersible UMaine VolturnUS-S variants, documented in NREL/TP-5000-75698 and NREL/TP-5000-76773 respectively; and the 3.4 MW land-based reference machine is documented in NREL/TP-5000-73492.

Tenth, the twenty-two-megawatt release used a cross-solver comparison involving OpenFAST from NREL, Bladed from DNV, HAWC2 from DTU and QBlade from the Technical University of Berlin, recorded in the Task 55 2024 annual report. The existence of that comparison is a reminder that the correctness of a reference machine is established in the sense of multi-code consensus, and that a single code's self-consistent output does not represent the machine.

Eleventh, expectations for wind-energy physical AI should be anchored to field measurement rather than pure simulation. Wake steering measures in the order of two to three percent in field trials and reviews, while reinforcement learning commonly reports gains above ten percent inside simulators. The gap comes from farm size, operating distributions, how the greedy baseline is implemented and which constraints apply, and quoting a reinforcement-learning result should carry that reference range alongside it.

Twelfth, licensing information follows each project's official repository: OpenFAST and its modules are under the Apache License 2.0, and the IEA Wind Task 37 and Task 55 reference machine repositories and windIO are released under Apache-2.0, with windIO additionally archived on Zenodo under DOI 10.5281/zenodo.15191297.

Thirteenth, related entries on this site include OpenFOAM industrial multiphysics pipeline, surrogate modelling and reduced-order methods, the cost boundary of classical solvers, WEC-Sim wave energy converter hydrodynamics and multiphysics coupling. The division of labour is that this guide focuses on whole-turbine aeroelasticity and wake control, while the others cover general solvers, general surrogate methodology and marine-energy coupling.

References

  1. National Lab of the Rockies, OpenFAST documentation, version 4.2.0, dated 12 February 2026, https://openfast.readthedocs.io/en/dev-tc/
  2. National Lab of the Rockies, OpenFAST documentation, section 7 on licensing, Apache License Version 2.0, https://openfast.readthedocs.io/en/dev-tc/source/license.html
  3. Wilson R., Buhl M., Jonkman J., Sprague M., OpenFAST, SWR-08-12, DOI 10.11578/dc.20210521.94
  4. IEA Wind TCP, Task 37 on systems engineering in wind energy, 2016 to 2022, https://iea-wind.org/?p=4069/
  5. IEA Wind TCP, Task 55 REFWIND on reference wind turbines and plants, https://iea-wind.org?p=14198/
  6. Bortolotti P., Bay C., Barter G., Gaertner E., Dykes K., McWilliam M., Friis-Moller M., Molgaard Pedersen M., Zahle F., System Modeling Frameworks for Wind Turbines and Plants: Review and Requirements Specifications, 2022, DOI 10.2172/1868328
  7. IEAWindSystems, windIO data format repository and documentation, Apache-2.0, Zenodo DOI 10.5281/zenodo.15191297
  8. Zahle F., Barlas A., Lonbaek K., Bortolotti P., Zalkind D., Wang L., Labuschagne C., Sethuraman L., Barter G., Definition of the IEA Wind 22-Megawatt Offshore Reference Wind Turbine, DTU Wind Report E-0243, 2024, DOI 10.11581/DTU.00000317
  9. Gaertner E., Rinker J., Sethuraman L., Zahle F., et al., Definition of the IEA Wind 15-Megawatt Offshore Reference Wind Turbine, NREL/TP-5000-75698, 2020
  10. Fleming P., et al., Initial results from a field campaign of wake steering applied at a commercial wind farm, Part 1, Wind Energy Science, volume 4, page 273, 2019, DOI 10.5194/wes-4-273-2019
  11. Su Y., Dong H., Tamaro S., Campagnolo F., Bottasso C. L., Zhao X., Offline Reinforcement Learning for Wind Farm Control: A Wind Tunnel Study under Dynamic Wind Directions, arXiv 2609.12905, submitted 11 September 2026
  12. De Frutos M., Corban B., Bizon Monroc C., Busic A., Dubuc D., Zhud J., Ferrer E., Noise Steering: Exploiting Wake Steering and Acoustic Directivity for Noise-Constrained Wind Farm Control via Reinforcement Learning, arXiv 2610.36888, 2026
  13. Sheehan H., Poole D., Silva Filho T., Bossanyi E., Landberg L., Investigations into deep Reinforcement Learning for wind farm set-point optimisation, Expert Systems with Applications, volume 282, 2025, DOI 10.1016/j.eswa.2025.127627
  14. Barlas A., et al., Comparison of the Gaussian Wind Farm Model with Historical Data of Three Offshore Wind Farms, Energies, volume 15, issue 6, article 1964, 2022, DOI 10.3390/en15061964
  15. IEA Wind TCP, Task 36 on forecasting for wind energy, 2019 to 2021, https://iea-wind.org/task36/
Tags: