PaperlayerView exact revision on arXiv

Source-linked AI summary

Noise Steering: Exploiting Wake Steering and Acoustic Directivity for Noise-Constrained Wind Farm Control via Reinforcement Learning

Martín de Frutos, Baptiste Corban, Claire Bizon Monroc, Ana Bušić, Donatien Dubuc, Jiamin Zhud, Esteban Ferrer

arXiv:2610.36888v1cs.AIcs.SYeess.SYphysics.ao-phphysics.flu-dyn

TL;DR

Wind-farm noise mitigation commonly reduces turbine power without exploiting wake interactions or noise directivity. This paper proposes Noise Steering, a reinforcement-learning controller that jointly adjusts yaw and power using a low-fidelity acoustic model. It reports lower power losses than the NRO baseline, with constraint performance also evaluated under turbulent and higher-fidelity simulation conditions.

  • Problem

    Existing wind-farm noise mitigation reduces turbine power without accounting for wake interactions and noise directivity, while learning-based wind-farm control had not considered noise constraints.

    Source: §1 · p. 2

  • Method

    Noise Steering uses reinforcement learning to jointly set turbine yaw and power under a local noise constraint, with acoustic estimates from a low-fidelity model coupled to steady-state FLORIS.

    Source: §1 · p. 2, §2.1 · p. 3, §5 · p. 12

  • Results

    The policy meets the noise constraint across 8–15 m s−1 with a 2% average power deficit relative to greedy control versus 17% for baseline NRO, and under turbulent inflow retains 92% of greedy power with only transient violations.

    Source: §5 · p. 12

  • Takeaways & Limitations

    The results support the potential of combining wake interactions and acoustic directivity in wind-farm control to address power lost to acoustic curtailment.

    Source: §5 · p. 12

    Evaluation uses a single wind direction and fixed control region, and the acoustic model neglects ground reflection, refraction, and atmospheric absorption.

    Source: §5 · p. 12

Abstract

from arXiv · show

Existing noise mitigation strategies for wind farms usually consist in reducing turbines power output, without accounting for the wake interactions between them and the noise emission directivity. This leads to over-conservative policies and unexploited wind power. Yet, turbine yaw can simultaneously affect wake interactions and the direction of noise emission, offering an additional degree of freedom for farm-level noise control. We therefore propose Noise Steering, a model-free reinforcement learning (RL) strategy that jointly controls turbine yaw angles and power setpoints to maximise farm power while satisfying a local noise constraint in a defined region. To enable training at a reduced computational cost, a low-fidelity acoustic model is coupled with the steady-state FLORIS simulator. The RL agent effectively learns how to redirect wakes and noise directivity away from the protected region while redistributing power across the farm. On a six-turbine test farm, the policy satisfies the 40 dB noise constraint throughout the full range of turbulent winds, with an average power deficit of only 2% relative to greedy unimpeded control, and up to 7.7% at high wind speeds. In comparison, the industry-standard noise-reduced operation (NRO) baseline here generates an average power loss of 17%. When deployed in the higher-fidelity FAST.Farm simulator, the policy largely retains its performance under steady and turbulent inflow, despite differences in simulation fidelity. These results demonstrate the potential of combining yaw control and power derating through RL to improve the power-noise trade-off in wind farm operation.

1. Introduction

Existing wind-farm noise controls often curtail power without using wake interactions or noise directivity, while learning-based farm control had not addressed noise constraints. Noise Steering combines yaw and power control under a local noise limit, using reinforcement learning and a computationally efficient acoustic model.

1. Introduction

  • Existing farm-level noise mitigation through layout optimisation and NRO reduces power generation, with median French AEP loss estimated at 2.8% from 2020 to 2022.Prior approaches include layout optimisation and optimisation of individual NRO levels.Source: §1 · p. 2
  • Yaw control can redirect wakes and change dipole-like noise directivity, motivating Noise Steering: jointly controlling yaw and NRO levels to maximise farm power under a local noise constraint.Existing noise-reduction strategies relied on power adjustments rather than turbine yaw control.Source: §1 · p. 2
  • Noise constraints had not been considered in prior learning-based wind-farm control approaches [25].Reinforcement learning offers a data-driven alternative to model-based control, whose optimisation generally must be solved for each operating condition.Source: §1 · p. 2
  • The cooperative RL controller uses a low-fidelity acoustic model coupled to steady-state FLORIS for training, then deploys without retraining in FAST.Farm under steady and unsteady inflow.The acoustic model combines Hagg’s analytical model with enforced dipole directivity and is validated against the FAST.Farm BPM acoustic model.Source: §1 · p. 2

2. Methodology

The method frames wind-farm operation as maximizing power under a local noise limit, using coordinated yaw and power control. It trains an RL policy with FLORIS and a low-fidelity acoustic model, then calibrates and evaluates the approach against FAST.Farm.

2.1. Wind farm control problem and set-up

  • The six-turbine farm is controlled to maximize power while keeping maximum sound level in a nearby control region below its limit.The studied setting uses a 40 dB limit; inflow direction is aligned with the two rows, creating strong wake interactions.Source: §2.1 · pp. 2–3

2.2. Reinforcement Learning for wind farm control

  • At each step, the RL agent observes wind speed, turbine powers, and yaw angles, then updates turbine power and yaw setpoints to optimize the farm objective.PPO is trained in FLORIS to reduce computational cost, and the resulting policy is evaluated in higher-fidelity FAST.Farm.Source: §2.2 · p. 4

2.3. FLORIS to FAST.Farm calibration

  • After calibration, the reported value for the complete data set is 7%; overall, calibration strongly reduces the simulation gap between FAST.Farm and FLORIS.Calibration uses 16 FAST.Farm simulations over 8–15 m s−1, first fitting wake-deficit parameters on aligned cases and then wake-deflection parameters on the complete data set.Source: §2.3 · pp. 4–5

2.4. Aeroacoustic modelling

  • The FLORIS acoustic model combines Hagg’s analytical model with an enforced dipole directivity, logarithmically sums turbine contributions, and uses maximum noise in the control region as its metric.The model assumes spherical sound propagation and is validated against FAST.Farm’s acoustic model with a mismatch in the 1 dB range.Source: §2.4 · p. 5

3. Training and analysis of the Noise Steering policy in FLORIS

The FLORIS-trained policy keeps noise below the constraint across the wind-speed sweep while retaining power close to greedy control. It combines yaw steering with selective turbine derating to manage noise and wake interactions.

3. Training and analysis of the Noise Steering policy in FLORIS

  • Training used 300k FLORIS steps in 500-step episodes at random 8–15 m s−1 steady winds aligned with the turbine rows; residual penalties reflect exploration and action changes.The reward and penalty curves are shown in Figure 2.Source: §3 · p. 6, Figure 2 · p. 6
  • Across wind speeds, Noise Steering keeps noise below the threshold with a 2% average power deficit relative to greedy control; NRO's average power deficit is 17%.Figure 3 compares power and noise emissions for the control strategies across the wind-speed range.Source: §3 · pp. 6–7, Figure 3 · p. 7
  • Up to 11 m s−1, wake steering raised power by approximately 400 kW relative to greedy control; at higher speeds, yaw redirected noise through dipole directivity.At 15 m s−1, the two last turbines yaw to 34° and 18°, respectively.Source: Figure 4 · p. 7, §3 · p. 8
  • Above 11.5 m s−1, only row 1's middle turbine is derated, reducing its noise and enabling better wake recovery for the last turbine in that row.Yaw control alone is insufficient in some cases, so the policy also uses turbine power setpoints.Source: §3 · p. 8
  • At 15 m s−1, turbines orient their rotor planes toward the acoustic region; upstream yaw also steers wakes for higher downstream power, while row 1's curtailed middle turbine emits less noise.Figure 5 shows the wind-speed and acoustic fields, with the control region marked.Source: §3 · pp. 8–9, Figure 5 · p. 8

4. Evaluation in FAST.Farm: robustness to model mismatch and unsteady inflow

The policy trained in FLORIS was deployed directly in FAST.Farm, with its yaw and power commands adapted to that simulator’s controllers. It retained much of its performance under steady and turbulent inflow, though model mismatch and turbulence affected results.

4.1. FAST.Farm integration

  • FAST.Farm integration translated RL yaw and power commands into controller setpoints and limited yaw rate to 0.3 °s−1.Power setpoints modified rated operating conditions to implement turbine derating; noise was estimated on an 8 × 2 control-region grid with exponential smoothing.Source: §4.1 · p. 9

4.2. Steady wind speed comparison

  • In FAST.Farm, Noise Steering produced similar power to FLORIS, but its noise constraint was met at all wind speeds except the highest considered.FAST.Farm predicted noise up to 2.5 dB lower than FLORIS at 11 m s−1, suggesting model mismatch may leave the policy sub-optimal there.Source: §4.2 · p. 10, Figure 6 · p. 10
  • Under steady inflow, NRO exceeded the noise threshold above 14 m s−1, while Noise Steering met it except at the highest wind speed.NRO’s transferred uniform derating levels were optimized on FLORIS; Figure 6 compares the three strategies across the wind-speed range in both simulators.Source: §4.2 · p. 10, Figure 6 · p. 10

4.3. Turbulent wind assessment

  • The turbulent inflow combined a varying wind-speed profile, a shear exponent of 0.2, and TurbSim fluctuations at 6% turbulence intensity.The prescribed profile rose from 10 to 14 m s−1 and then fell to 8 m s−1; Figure 7 shows hub-height wind at three first-row positions, with bold curves filtered over 60 seconds.Source: §4.3 · pp. 10–11, Figure 7 · p. 11
  • The turbulent-inflow assessment included a 200-second wake-development period before the three control strategies were activated.Figure 8 compares their farm power and noise over the unsteady inflow.Source: §4.3 · p. 11, Figure 8 · p. 11
  • 91.7% of the turbulent-inflow evaluation stayed below the noise limit with 7.4% less power than greedy control, versus 18.1% less power for NRO.Greedy exceeded the limit for 53.0% of the time and NRO for 30.5%; Noise Steering briefly reached 41.0 dB, then fell to 39.4 dB within 6 seconds.Source: §4.3 · p. 11
  • Ambient turbulence increased model mismatch and wake recovery, reducing the added benefit of wake steering.At the 10 m s−1 plateau, Noise Steering produced less power than greedy and NRO, unlike under steady inflow.Source: §4.3 · p. 12

5. Conclusion

Noise Steering jointly uses yaw misalignment and power derating to maximize farm power under a local noise constraint. The policy performs across steady and turbulent inflows, while its evaluation is bounded by a single wind direction, fixed control region, and simplified acoustic model.

5. Conclusion

  • The policy meets the noise constraint across 8–15 m s−1 with a 2% average power deficit versus greedy control, compared with 17% for baseline NRO.Its deficit reaches 7.7% at the highest wind speed.Source: §5 · p. 12
  • At low wind speeds the agent uses power-maximizing wake steering; at higher speeds it turns rotor planes toward the control region so noise directivity radiates away.This combines wake redirection with control of dipole-like noise directivity.Source: §5 · p. 12
  • Without retraining, transfer to FAST.Farm under steady inflow retains similar power performance and stays below the noise limit except at the highest wind speed.The transferred policy appears too conservative, and the authors identify simulator mismatch as a cause; online learning is proposed for future work.Source: §5 · p. 12
  • Under turbulent inflow, the policy has only transient noise violations, retains 92% of greedy power, and improves conventional NRO power generation by 10%.It was trained exclusively under steady inflow.Source: §5 · p. 12
  • The evaluation uses one wind direction and a fixed control region, while the acoustic model omits ground reflection, refraction, and atmospheric absorption.The authors say the impact of relaxing these assumptions should be assessed in future work.Source: §5 · p. 12
  • The authors identify wake interactions and acoustic directivity as a promising direction for wind farm control, given energy production lost to acoustic curtailment..Source: §5 · p. 12

CRediT authorship contribution statement

The contribution statement assigns conceptualization, investigation, methodology, software, validation, visualization, supervision, project administration, writing, and funding acquisition across the named authors.

CRediT authorship contribution statement

  • Martín de Frutos and Baptiste Corban are credited with conceptualization, investigation, methodology, software, validation, visualization, and original-draft writing; the other contributors’ roles are also specified.Claire Bizon Monroc, Ana Bušić, Donatien Dubuc, Jiamin Zhu, and Esteban Ferrer have varying methodology, supervision, project administration, review, validation, conceptualization, and funding roles.Source: CRediT authorship contribution statement · p. 13

A. RL hyperparameters

The appendix reports that reward-function and PPO parameters were selected by trial and error, with action penalties kept small to limit policy oscillations.

A. RL hyperparameters

  • Reward-function and PPO hyperparameters were chosen by trial and error.Table 2 is identified as listing the parameters used for RL training, but its values are not supplied here.Source: A. RL hyperparameters · p. 15, Table 2 · p. 15
  • Action penalties were kept as small as possible while limiting policy oscillations; higher penalties can make operational actions excessively slow.Source: A. RL hyperparameters · p. 15

B. Noise model validation

The validation compares three noise-modelling approaches across operating conditions and reports that the lower-fidelity FLORIS + Hagg model captures directivity and amplitude levels, with noise levels within 1 dB of FAST.Farm across the evaluated cases.

B. Noise model validation

  • The comparison uses FAST.Farm + BPM as ground truth, FAST.Farm + Hagg to isolate the aeroacoustic model change, and FLORIS + Hagg as the training approach.Source: B. Noise model validation · p. 15
  • The lower-fidelity approach captures wind-turbine noise directivity and amplitude levels across the evaluated operating conditions.The cases vary wind speeds, yaw configurations, and power setpoints, and the validation considers only the first turbine row of the Figure 1 farm.Source: B. Noise model validation · p. 15
  • FLORIS + Hagg remains within 1 dB of FAST.Farm for the reported maximum control-region noise level across all evaluated conditions.Table 3 reports the higher- versus lower-fidelity difference; its caption identifies FAST.Farm + BPM as the higher-fidelity approach and FLORIS + Hagg as the lower-fidelity approach.Source: Table 3 · p. 15, B. Noise model validation · p. 16
  • Figure 9 compares three modelling approaches for three operational conditions, with time-averaged OASPL to reduce amplitude-modulation oscillations.The control region is marked for reference.Source: Figure 9 · p. 16

C. FAST.Farm and FLORIS policy comparison

At a steady free wind speed of 13.2 m s−1, the Noise Steering policy’s yaw actions and turbine power outputs converge in FAST.Farm toward those obtained in FLORIS after FAST.Farm wake dynamics stabilize. The comparison indicates the policy can operate under FAST.Farm’s transient wake dynamics after calibration reduces the static modelling discrepancy.

C. FAST.Farm and FLORIS policy comparison

Loading 2610.36888v1…