Pith. sign in

REVIEW 4 major objections 4 minor 55 references

Reinforcement learning-enhanced genetic algorithm for wind farm layout optimization

T0 review · 4 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper claims that a reinforcement learning agent choosing genetic algorithm parameters on the fly produces wind farm layouts as good as or better than a fixed-parameter genetic algorithm in about a third of the generations, with the…

desk verdict Sensible Q-learning wrapper around a genetic algorithm for WFLO, with a plausible speedup claim that needs a fair baseline and repeated runs before it can be taken at face value. read the letter →

arxiv 2412.06803 v1 pith:DHGJM33X submitted 2024-11-24 cs.NE

Guodan Dong , Jianhua Qin , Chutian Wu , Chang Xu , Xiaolei Yang This is my paper · ORCID
classification cs.NE
keywords reinforcementlearninggeneticalgorithmwindfarmlayoutoptimizationQ-learningJensenwakemodelparameteradaptationphysics-informedlayoutsmetaheuristics
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes RLGA, a genetic algorithm for wind farm turbine layout whose crossover, mutation, and mating choices are made on the fly by a Q-learning agent rather than fixed in advance. The claim is that this dynamic parameter selection converges to layouts as good as or better than a standard GA's, while taking roughly one third as many generations, and that the advantage widens as the search space grows from about 100 to 900 potential turbine positions. The practical stake is that large and densely gridded wind farms, where wake interactions make the layout problem highly non-convex, can be optimized with far less computational effort and without case-by-case parameter tuning. The paper tests this on aligned, staggered, sunflower, and unstructured potential-layout grids under unidirectional, omnidirectional, and non-uniform wind, and matches or improves on previous GA results.

What carries the argument

The load-bearing mechanism is a Q-learning agent wrapped around the genetic algorithm loop. At each generation, the state is a single bit: whether the best fitness improved since the previous generation. The actions are the GA parameter choices: number of mating parents (2 or 3), crossover type (single-point, uniform, two-point, or scattered), and mutation rate (1% to 4% of genes). The reward is the fitness difference between consecutive generations, and Q-values are updated with the Bellman equation $Q_{t+1}(S_t,A_t)=Q_t(S_t,A_t)+\alpha(R_t+\gamma\max_{a}Q(S_{t+1},a)-Q(S_t,A_t))$, with $\epsilon$-greedy action selection. This lets a small population ($N_p=5$) adapt its search operators to whether the last generation improved, which is what the paper credits for avoiding local optima and accelerating convergence.

What would settle it

A decisive check would be to run RLGA and a GA whose parameters are tuned separately for each case (for example, larger population, adaptive mutation schedule) over many independent random seeds, and compare the median number of generations to reach the same objective. If the tuned GA converges in the same or fewer generations, the threefold-efficiency claim would not survive.

Watch

Extended reading notes

Core claim

The central claim is that replacing fixed genetic algorithm hyperparameters with a Q-learning controller yields layouts that are as good as or better than those from a standard GA, while converging about three times faster, with the advantage growing as the problem becomes more complex. The Q-learning agent chooses among two mating counts, four crossover schemes, and four mutation rates at each generation, using whether the best fitness improved as the state and the fitness difference as the reward. On the small wind farm, RLGA reproduces the previous aligned and staggered results exactly and improves power output by about 0.4% for the sunflower layout and about 8% for the unstructured layout. On the larger and finer cases, with roughly 625 and 900 potential turbine positions, the convergence advantage over the fixed-parameter GA reaches approximately threefold.

Load-bearing premise

The claimed threefold speedup assumes the fixed-parameter GA used for comparison is a fair baseline, and that the shown convergence curves—apparently from single runs—represent typical behavior.

Editorial extensions

If this is right

  • For wind farm layout problems with 625 to 900 potential turbine positions, RLGA reaches a comparable or better objective roughly three times faster than the fixed-parameter GA in the cases tested.
  • The dynamic parameter selection removes the need to hand-tune mating, crossover, and mutation when the wind farm size, grid spacing, or wind conditions change.
  • On sunflower and unstructured potential-layout grids, RLGA improves power output by about 0.4% and 8% over the previous physics-informed GA results in the small-farm unidirectional case.
  • For the 6 km by 6 km farm under unidirectional wind, the sunflower layout yields the highest power, about 8.5% higher than the aligned layout, consistent with the small-farm trend.
  • The method is positioned as a general recipe that can be carried to other combinatorial optimization problems where genetic operator choice matters.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: if the same Q-learning wrapper were tried on other combinatorial search problems whose bottleneck is operator choice rather than evaluation cost, the speedup pattern could transfer, but the paper does not claim this.
  • Editorial inference: because the convergence comparison fixes the GA at one hyperparameter set, the 'three times' figure is best read as an upper bound on the benefit of online adaptation; a per-case-tuned GA baseline would probably narrow the gap.
  • Editorial inference: the binary improvement-state reward function discards information about the size of fitness gains, so a graded reward might give the Q-learning agent a sharper signal on harder landscapes; this is a testable variant, not a paper claim.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes RLGA, a Q-learning-enhanced genetic algorithm that dynamically selects mating, crossover, and mutation parameters during wind farm layout optimization. The method is evaluated on four physics-informed layouts under unidirectional, omnidirectional, and non-uniform wind conditions, in small and large wind farms, and is compared against a fixed-parameter GA. The accuracy is validated by reproducing the aligned and staggered results of Wu et al. [17] and marginally improving the sunflower and unstructured results. The central claim is that RLGA is about three times more efficient than GA, with the advantage growing with problem complexity.

Significance. If the efficiency claim holds, the paper makes a useful contribution: it demonstrates a practical way to eliminate manual GA parameter tuning in WFLO, and the validation against prior published layouts gives the wake-model and fitness evaluation a credible anchor. The authors also provide a clear algorithmic description of RLGA and test it across a wider range of cases (coarse/fine grids, small/large farms, three wind conditions) than is common. The main weakness is that the central efficiency claim rests on convergence curves that appear to come from single runs and on a GA baseline that is not fairly tuned; without repeated-seed statistics and a fair baseline, the claimed threefold gain is not yet established.

major comments (4)
  1. [Section 3.3, Table 3] The baseline GA is fixed at a single parameter configuration (Np=5, Pc=2, single-point crossover, Mc=4%), whereas RLGA may select Pc in {2,3}, four crossover types, and Mc in {1,2,3,4}% in each generation. Because the GA's parameters are a single arbitrary point inside the same action space, the reported "three times more efficient" gain could be an artifact of an unlucky fixed parameter choice rather than a benefit of dynamic selection. Please either tune the baseline GA (e.g., a grid search over the same ranges, reporting the best configuration) or show that the chosen fixed parameters are not inferior to other fixed choices in the action set.
  2. [Figs. 7-9 and Section 3.3] The efficiency comparison is supported only by convergence curves, but the manuscript does not state whether these are single runs or averages, how many independent runs were performed, or whether any error bars or confidence intervals apply. GAs are stochastic, and run-to-run variance can be large, especially for high-dimensional layouts with 625-900 potential positions. Please report the mean and spread over at least 10 independent seeds for both GA and RLGA, and provide a statistical comparison (e.g., Wilcoxon signed-rank test or a clear effect size) for the generations-to-target or fitness-at-generation metric.
  3. [Section 3.3] The phrase "about three times more efficient" is not quantitatively defined. The text does not specify the generation budget Ng, the target threshold used to measure convergence, or whether the comparison is in generations, fitness evaluations, or wall-clock time. Without an explicit definition of the efficiency metric and the termination criterion, the main claim is not falsifiable. Please define the metric (e.g., generations needed to reach 99% of the final fitness) and state the numerical values used to obtain the threefold factor.
  4. [Algorithm 3 and Section 2.2.3] The Q-learning hyperparameters alpha, gamma, epsilon, the initial Q-table values, and the reward scaling are never specified. Since the contribution of RLGA is precisely the Q-learning-driven parameter selection, the absence of these values prevents reproduction and leaves the sensitivity of the method unexamined. Please report the chosen values and, ideally, a small sensitivity study to show that the efficiency gain is robust to the RL settings.
minor comments (4)
  1. [Section 1] The introduction states that RLGA uses "a significantly smaller population size per generation" than GA, but Table 3 shows the same Np=5 for both algorithms; please clarify this inconsistency or correct the introductory claim.
  2. [Section 3.4] The text says the omnidirectional-uniform results are "presented in Table 6", but the table that follows is Table 5; the cross-reference should be corrected.
  3. [Throughout] There are several typos, including "minic" for "mimic", "dfferen-tial" for "differential", "futrure" for "future", and "flow chat" for "flowchart"; a careful proofread is needed.
  4. [Captions of Figs. 7-9] The captions do not specify the quantity plotted on the y-axis or the units; please state whether it is the objective value, the fitness, or the cost per unit power, and add a legend distinguishing GA and RLGA in each panel.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the RLGA efficiency claim rests on empirical convergence comparisons, and its wake, cost, and fitness models are imported from external or separately published sources.

full rationale

I walked the paper's claimed derivation chain and found no step in which a prediction or first-principles result is equivalent to its inputs by construction. The wake model, wake superposition rule, cost function, and turbine parameters are all taken from established literature (Jensen, Frandsen, Mosetti, Grady, and others), so the objective function is not derived from the RLGA itself. The RL reward in Eq. (11) is the change in fitness, and the Q-table update in Eq. (12) is the standard Bellman equation; using fitness improvement as a reward for operator selection is the intended mechanism of the method, not a circular redefinition of the outcome. The validation in Section 3.2 compares against a separately published prior study by the same group [17] and also against Grady et al. [39]; reproducing the aligned and staggered results exactly, and improving on the prior sunflower and unstructured results, is an empirical benchmark rather than a self-referential proof. The central efficiency claim in Section 3.3 is supported by convergence comparisons shown in Figs. 7-9 and Table 3. A possible concern is that the GA baseline is fixed at one parameter combination (Np=5, Pc=2, single-point crossover, Mc=4) while RLGA samples a broader action space, and the convergence curves appear to come from single runs; this is a fairness or reproducibility risk, not a circularity. The GA's poor performance could in principle be due to an unlucky fixed parameter choice, but the claim that the RLGA is about three times more efficient is an empirical assertion subject to that baseline, not a quantity forced by the paper's definitions. The layouts labeled sunflower and unstructured are imported from prior work [17], but they are used only as test cases, not as premises that guarantee the efficiency result. No self-citation is load-bearing in the sense of forbidding alternatives or providing the uniqueness of the proposed method. Overall, the derivation chain is self-contained relative to its externally sourced models and its internally standard RL update rules, so I find no significant circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on algorithm settings rather than new physics. No new entities are invented; the wake model, superposition rule, and cost function are taken from prior literature. The hand-chosen GA baseline and RL action ranges are the most consequential free choices, and the unreported RL hyperparameters are a reproducibility gap.

free parameters (3)
  • Baseline GA fixed parameters = Np=5, PC=2, CC='single point', MC=4
    The comparison fixes GA at one hand-chosen configuration (Table 3) without tuning or sensitivity analysis; the efficiency gain of RLGA is measured against this single baseline.
  • RLGA action ranges = PC in [2,3]; CC in {single point, uniform, two points, scattered}; MC in [1,2,3,4]
    These hand-chosen ranges define the Q-learning action space; different ranges could change the comparison outcome.
  • RL hyperparameters alpha, gamma, epsilon = Not reported
    Learning rate, discount factor, and exploration rate are never specified, yet they control the Q-learning dynamics and affect convergence behavior.
assumptions (4)
  • standard math Jensen wake model with momentum conservation (Eq. 2) and quadratic-sum superposition (Eq. 3)
    Standard low-fidelity wake model taken from prior literature [8,42]; the partial-overlap correction (Eqs. 4-5) is also standard.
  • domain assumption Discrete potential turbine positions on a grid with specified spacing
    The optimization restricts turbines to predefined grids (Section 3.1), which shapes the solution space and the achievable layouts.
  • domain assumption The four physics-informed layouts (aligned, staggered, sunflower, unstructured) are representative of realistic design spaces
    These layouts are taken from the authors' prior work [17] and may not cover all real-world constraints or site geometries.
  • standard math Objective function based on cost model and fobj_ideal as N to infinity limit
    The cost model comes from [16,39,17]; fobj_ideal is computed from the wind rose as an asymptotic best case.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reinforcement learning-enhanced genetic algorithm for wind farm layout optimization." pith.science (2026). https://pith.science/paper/DHGJM33X

@misc{pith2026241206803,
  author       = {Pith},
  title        = {Pith review of: Reinforcement learning-enhanced genetic algorithm for wind farm layout optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DHGJM33X}},
  note         = {Machine review of arXiv:2412.06803}
}
abstract

A reinforcement learning-enhanced genetic algorithm (RLGA) is proposed for wind farm layout optimization (WFLO) problems. While genetic algorithms (GAs) are among the most effective and accessible methods for WFLO, their performance and convergence are highly sensitive to parameter selections. To address the issue, reinforcement learning (RL) is introduced to dynamically select optimal parameters throughout the GA process. To illustrate the accuracy and efficiency of the proposed RLGA, we evaluate the WFLO problem for four layouts (aligned, staggered, sunflower, and unstructured) under unidirectional uniform wind, comparing the results with those from the GA. RLGA achieves similar results to GA for aligned and staggered layouts and outperforms GA for sunflower and unstructured layouts, demonstrating its efficiency. The sunflower and unstructured layouts' complexity highlights RLGA's robustness and efficiency in tackling complex problems. To further validate its capabilities, we investigate larger wind farms with varying turbine placements ($\Delta x = \Delta y = 5D$ and 2$D$, where $D$ is the wind turbine diameter) under three wind conditions: unidirectional, omnidirectional, and non-uniform, presenting greater challenges. The proposed RLGA is about three times more efficient than GA, especially for complex problems. This improvement stems from RL's ability to adjust parameters, avoiding local optima and accelerating convergence.

Figures

Figures reproduced from arXiv: 2412.06803 by the authors.

Figure 1
Figure 1. The schematic (left) and contour (right) of the Jensen single wake model. [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Overlapped area between a wind turbine rotor and a wake stream. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Block diagram schematic for RL. The reward function evaluates the success of actions based on fitness im￾provements, as shown in Eq. (11). The Q-values in the Q-table are updated based on the expected maximum reward from the next action, using the Bell￾man equation shown in Eq. (12). By dynamically adjusting operator selection, 11 [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: The flow chart of the RLGA method with GA and RL algorithms shown in pink [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]
Figure 5
Figure 5. Figure 5: The wind rose for cases the omnidirectional uniform wind (left) and spread non [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 6
Figure 6. Figure 6: Optimal layouts (the first row) and the corresponding contours (the second row) [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: The comparison of convergence efficiency between GA and RLGA in optimization [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]
Figure 8
Figure 8. Figure 8: The comparison of convergence efficiency between GA and RLGA in optimization [PITH_FULL_IMAGE:figures/full_fig_p020_8.png]
Figure 9
Figure 9. Figure 9: The comparison of convergence efficiency between GA and RLGA in optimization [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]
Figure 10
Figure 10. Figure 10: Optimal layouts and contours obtained from the RLGA for WFLO under unidi [PITH_FULL_IMAGE:figures/full_fig_p022_10.png]
Figure 11
Figure 11. Figure 11: Optimal layouts and contours obtained from the RLGA for WFLO under omnidi [PITH_FULL_IMAGE:figures/full_fig_p023_11.png]
Figure 12
Figure 12. Figure 12: Optimal layouts and contours obtained for WFLO under spread non-uniform wind [PITH_FULL_IMAGE:figures/full_fig_p024_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 53 canonical work pages

  1. [17]

    C. Wu, X. Yang, Y. Zhu, On the design of potential turbine positions for physics- informed optimization of wind farm layout, Renewable Energy 164 (2021) 1108– 1120

    work page 2021

  2. [1]

    Veers, K

    P. Veers, K. Dykes, E. Lantz, S. Barth, C. L. Bottasso, O. Carlson, A. Clifton, J. Green, P. Green, H. Holttinen, et al., Grand challenges in the science of wind energy, Science 366 (6464) (2019) eaau2027

    work page 2019

  3. [2]

    Meyers, C

    J. Meyers, C. Bottasso, K. Dykes, P. Fleming, P. Gebraad, G. Giebel, T. G¨ o¸ cmen, J.-W. Van Wingerden, Wind farm flow control: prospects and challenges, Wind Energy Science Discussions 2022 (2022) 1–56

    work page 2022

  4. [3]

    X. Yang, F. Sotiropoulos, A review on the meandering of wind turbine wakes, Energies 12 (24) (2019) 4725

    work page 2019

  5. [4]

    J. S. Gonz´ alez, M. B. Pay´ an, J. M. R. Santos, F. Gonz´ alez-Longatt, A review and recent developments in the optimal wind-turbine micro-siting problem, Renewable and Sustainable Energy Reviews 30 (2014) 133–144

    work page 2014

  6. [5]

    G. C. Larsen, H. Aagaard Madsen, F. Bing¨ ol, Dynamic wake meandering model- ing (2007)

    work page 2007

  7. [6]

    G. V. Iungo, F. Viola, U. Ciri, M. A. Rotea, S. Leonardi, Data-driven rans for simulations of large wind farms, in: Journal of Physics: Conference Series, Vol. 625, IOP Publishing, 2015, p. 012025

    work page 2015

  8. [7]

    Z. Li, G. Dong, X. Yang, Onset of wake meandering for a floating offshore wind turbine under side-to-side motion, Journal of Fluid Mechanics 934 (2022) A29

    work page 2022

Show all 55 references
  1. [8]

    N. O. Jensen, A note on wind generator interaction, Risø National Laboratory, 1983

    1983

  2. [9]

    Katic, J

    I. Katic, J. Højstrup, N. O. Jensen, A simple model for cluster efficiency, in: European wind energy association conference and exhibition, A. Raguzzi, 1987, pp. 407–410

    1987

  3. [10]

    Bastankhah, F

    M. Bastankhah, F. Port´ e-Agel, A new analytical model for wind-turbine wakes, Renewable energy 70 (2014) 116–123

    2014

  4. [11]

    X. Gao, H. Yang, L. Lu, Optimization of wind turbine layout position in a wind farm using a newly-developed two-dimensional wake model, Applied Energy 174 (2016) 192–200. 28

    2016

  5. [12]

    Keane, Advancement of an analytical double-gaussian full wind turbine wake model, Renewable Energy 171 (2021) 687–708

    A. Keane, Advancement of an analytical double-gaussian full wind turbine wake model, Renewable Energy 171 (2021) 687–708

    2021

  6. [13]

    H. Sun, H. Yang, Study on an innovative three-dimensional wind turbine wake model, Applied energy 226 (2018) 483–493

    2018

  7. [14]

    Bastankhah, F

    M. Bastankhah, F. Port´ e-Agel, Experimental and theoretical study of wind tur- bine wakes in yawed conditions, Journal of Fluid Mechanics 806 (2016) 506–541

    2016

  8. [15]

    Cheng, M

    Y. Cheng, M. Zhang, Z. Zhang, J. Xu, A new analytical model for wind turbine wakes based on monin-obukhov similarity theory, Applied Energy 239 (2019) 96– 106

    2019

  9. [16]

    Mosetti, C

    G. Mosetti, C. Poloni, B. Diviacco, Optimization of wind turbine positioning in large windfarms by means of a genetic algorithm, Journal of Wind Engineering and Industrial Aerodynamics 51 (1) (1994) 105–116

    1994

  10. [18]

    X. Yu, W. Zhang, A teaching-learning-based optimization algorithm with rein- forcement learning to address wind farm layout optimization problem, Applied Soft Computing 151 (2024) 111135

    2024

  11. [19]

    Rajper, I

    S. Rajper, I. J. Amin, Optimization of wind turbine micrositing: A comparative study, Renewable and Sustainable Energy Reviews 16 (8) (2012) 5485–5492

    2012

  12. [20]

    Parada, C

    L. Parada, C. Herrera, P. Flores, V. Parada, Wind farm layout optimization using a gaussian-based wake model, Renewable energy 107 (2017) 531–541

    2017

  13. [21]

    J. F. Herbert-Acero, O. Probst, P.-E. R´ ethor´ e, G. C. Larsen, K. K. Castillo-Villar, A review of methodological approaches for the design and optimization of wind farms, Energies 7 (11) (2014) 6930–7016

    2014

  14. [22]

    Wilson, S

    D. Wilson, S. Rodrigues, C. Segura, I. Loshchilov, F. Hutter, G. Buenfil, A. Kheiri, E. Keedwell, M. Ocampo-Pineda, E. Ozcan, S. Pena, B. Goldman, S. Rionda, A. Hernandez-Aguirre, K. Veeramachaneni, S. Cussat-Blanc, Evolu- tionary computation for wind farm layout optimization,...

    2018 doi

  15. [23]

    Barrera-Cardenas, M

    R. Barrera-Cardenas, M. Molinas, Optimal lqg controller for variable speed wind turbine based on genetic algorithms, Energy Procedia 20 (2012) 207–216

    2012

  16. [24]

    S. Gao, Y. Wang, J. Wang, J. Cheng, Understanding differential evolution: A poisson law derived from population interaction network, Journal of computa- tional science 21 (2017) 140–149

    2017

  17. [25]

    Z. Lei, S. Gao, Y. Wang, Y. Yu, L. Guo, An adaptive replacement strategy- incorporated particle swarm optimizer for wind farm layout optimization, Energy Conversion and Management 269 (2022) 116174

    2022

  18. [26]

    S. Gao, Y. Wang, J. Cheng, Y. Inazumi, Z. Tang, Ant colony optimization with clustering for solving the dynamic location routing problem, Applied Mathematics and Computation 285 (2016) 149–173

    2016

  19. [27]

    N. Li, Y. Zhou, Q. Luo, H. Huang, Discrete complex-valued code pathfinder algorithm for wind farm layout optimization problem, Energy Conversion and Management: X 16 (2022) 100307

    2022

  20. [28]

    Elkinton, J

    C. Elkinton, J. Manwell, J. McGowan, Optimization algorithms for offshore wind farm micrositing, in: Proceedings of the WINDPOWER 2007 Conference and Exhibition, Los Angeles, CA, USA, 2007, pp. 3–6

    2007

  21. [29]

    F. Bai, X. Ju, S. Wang, W. Zhou, F. Liu, Wind farm layout optimization us- ing adaptive evolutionary algorithm with monte carlo tree search reinforcement learning, Energy Conversion and Management 252 (2022) 115047

    2022

  22. [30]

    H. Dong, X. Zhao, Reinforcement learning-based wind farm control: Toward large farm applications via automatic grouping and transfer learning, IEEE Transac- tions on Industrial Informatics 19 (12) (2023) 11833–11845

    2023

  23. [31]

    Vyshnav, M

    T. Vyshnav, M. Lavanya, K. S. Thampatty, Reinforcement learning based wind farm layout optimization, in: 2022 International Conference on Electronics and Renewable Systems (ICEARS), IEEE, 2022, pp. 1393–1398. 30

    2022

  24. [32]

    Y. A. Li, Deep reinforcement learning on wind power optimization, in: 2022 Inter- national Conference on Networks, Communications and Information Technology (CNCIT), IEEE, 2022, pp. 45–51

    2022

  25. [33]

    H. Dong, J. Zhang, X. Zhao, Intelligent wind farm control via deep reinforcement learning and high-fidelity simulations, Applied Energy 292 (2021) 116928

    2021

  26. [34]

    Z. Deng, C. Xu, X. Han, Z. Cheng, F. Xue, Decentralized yaw optimization for maximizing wind farm production based on deep reinforcement learning, Energy Conversion and Management 286 (2023) 117031

    2023

  27. [35]

    H. Dong, X. Zhao, Composite experience replay-based deep reinforcement learn- ing with application in wind farm control, IEEE Transactions on Control Systems Technology 30 (3) (2021) 1281–1295

    2021

  28. [36]

    Stanfel, K

    P. Stanfel, K. Johnson, C. J. Bay, J. King, A distributed reinforcement learn- ing yaw control approach for wind farm energy capture maximization, in: 2020 american control conference (acc), IEEE, 2020, pp. 4065–4070

    2020

  29. [37]

    H. Zhao, J. Zhao, J. Qiu, G. Liang, Z. Y. Dong, Cooperative wind farm control with deep reinforcement learning and knowledge-assisted learning, IEEE Trans- actions on Industrial Informatics 16 (11) (2020) 6912–6921

    2020

  30. [38]

    H. Dong, X. Zhao, Data-driven wind farm control via multiplayer deep reinforce- ment learning, IEEE Transactions on Control Systems Technology 31 (3) (2022) 1468–1475

    2022

  31. [39]

    Grady, M

    S. Grady, M. Y. Hussaini, M. M. Abdullah, Placement of wind turbines using genetic algorithms, Renewable energy 30 (2) (2005) 259–270

    2005

  32. [40]

    Schepers, ENDOW: Validation and improvement of ECN’s wake model, Energy research Centre of the Netherlands ECN Sint Maartensvlotbrug, The

    J. Schepers, ENDOW: Validation and improvement of ECN’s wake model, Energy research Centre of the Netherlands ECN Sint Maartensvlotbrug, The . . . , 2003

    2003

  33. [41]

    Frandsen, On the wind speed reduction in the center of large clusters of wind turbines, Journal of Wind Engineering and Industrial Aerodynamics 39 (1-3) (1992) 251–265

    S. Frandsen, On the wind speed reduction in the center of large clusters of wind turbines, Journal of Wind Engineering and Industrial Aerodynamics 39 (1-3) (1992) 251–265

    1992

  34. [42]

    G¨ o¸ cmen, P

    T. G¨ o¸ cmen, P. Van der Laan, P.-E. R´ ethor´ e, A. P. Diaz, G. C. Larsen, S. Ott, Wind turbine wake models developed at the technical university of denmark: A review, Renewable and Sustainable Energy Reviews 60 (2016) 752–769. 31

    2016

  35. [43]

    R. J. Stevens, C. Meneveau, Flow structure and turbulence in wind farms, Annual review of fluid mechanics 49 (1) (2017) 311–339

    2017

  36. [44]

    Port´ e-Agel, M

    F. Port´ e-Agel, M. Bastankhah, S. Shamsoddin, Wind-turbine and wind-farm flows: a review, Boundary-layer meteorology 174 (1) (2020) 1–59

    2020

  37. [45]

    J. Feng, W. Z. Shen, Wind farm layout optimization in complex terrain: A pre- liminary study on a gaussian hill, in: Journal of Physics: Conference Series, Vol. 524, IOP Publishing, 2014, p. 012146

    2014

  38. [46]

    Abdulrahman, D

    M. Abdulrahman, D. Wood, Investigating the power-coe trade-off for wind farm layout optimization considering commercial turbine selection and hub height vari- ation, Renewable energy 102 (2017) 267–278

    2017

  39. [47]

    J. S. Gonz´ alez, A. G. G. Rodriguez, J. C. Mora, J. R. Santos, M. B. Payan, Op- timization of wind farm turbines layout using an evolutive algorithm, Renewable energy 35 (8) (2010) 1671–1681

    2010

  40. [48]

    Kumar, A

    A. Kumar, A. Zhou, G. Tucker, S. Levine, Conservative q-learning for offline reinforcement learning, Advances in Neural Information Processing Systems 33 (2020) 1179–1191

    2020

  41. [49]

    C. J. Watkins, P. Dayan, Q-learning, Machine learning 8 (1992) 279–292

    1992

  42. [50]

    Pookpunt, W

    S. Pookpunt, W. Ongsakul, Optimal placement of wind turbines within wind farm using binary particle swarm optimization with time-varying acceleration coefficients, Renewable energy 55 (2013) 266–276

    2013

  43. [51]

    A. M. Abdelsalam, M. El-Shorbagy, Optimization of wind turbines siting in a wind farm using genetic algorithm based local search, Renewable energy 123 (2018) 748–755

    2018

  44. [52]

    Guirguis, D

    D. Guirguis, D. A. Romero, C. H. Amon, Gradient-based multidisciplinary design of wind farms with continuous-variable formulations, Applied energy 197 (2017) 279–291

    2017

  45. [53]

    X. Yu, Y. Lu, Reinforcement learning-based multi-objective differential evolution for wind farm layout optimization, Energy 284 (2023) 129300. 32

    2023

  46. [54]

    H. Sun, H. Yang, X. Gao, Investigation into spacing restriction and layout opti- mization of wind farm with multiple types of wind turbines, Energy 168 (2019) 637–650

    2019

  47. [55]

    Hansen, Aerodynamics of wind turbines, Routledge, 2015

    M. Hansen, Aerodynamics of wind turbines, Routledge, 2015. 33

    2015

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.