Scholarlink

<p>The yaw control system of large-scale wind turbines is required to rapidly track wind direction changes while exhibiting hysteresis characteristics. This design aims to reduce yaw actions and switches between yaw states, thereby minimizing mechanical loads on the turbine. Existing reinforcement learning (RL)-based yaw control methods either rely on complex post-processing or adopt a fixed yaw action duration, which prevents them from achieving optimal control performance. Furthermore, all these methods lack theoretical analysis regarding their stability. To address these issues, this paper proposes the HysteresisPPO method, which integrates the hysteresis-aware dual-head architecture of HysteresisQNet into the Proximal Policy Optimization (PPO) framework. A corresponding policy gradient estimator is derived to ensure that the training process of HysteresisPPO remains consistent with that of the classic PPO algorithm. This paper also designs an adaptive weight adjustment strategy to balance power generation efficiency and yaw actuation frequency, thereby avoiding the need for post-optimization procedures. Additionally, sufficient conditions are presented to guarantee that HysteresisPPO satisfies practical stability in expectation, resolving the critical issue of controller stability analysis that has long been absent in RL-based yaw control systems. Three sets of simulations are conducted to evaluate the performance of HysteresisPPO. In the first set, simulations based on wind data from five different wind turbines demonstrate that HysteresisPPO outperforms existing RL-based yaw controllers in both energy capture and mechanical load reduction, with no requirement for post-processing; these results further empirically validate the derived sufficient conditions for practical stability in expectation. The second set verifies HysteresisPPO’s strong generalization capability across diverse wind conditions on four synthetic datasets. In the third set, we evaluate HysteresisPPO under two representative disturbances: sector-wise occlusion and wind vane offset. Results show that HysteresisPPO learns an effective feedback-based policy when each disturbance is applied individually, but degrades under their combination—revealing its reliance on feedback-driven control rather than internal environmental modeling.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Yaw Control for Large-Scale Wind Turbine Based on HysteresisPPO

  • 期刊论文

  • 发表日期：

    2026年5月28日

查看全文
  • Shuai Dong,
  • Kang Deng,
  • Kun Zou

摘要

The yaw control system of large-scale wind turbines is required to rapidly track wind direction changes while exhibiting hysteresis characteristics. This design aims to reduce yaw actions and switches between yaw states, thereby minimizing mechanical loads on the turbine. Existing reinforcement learning (RL)-based yaw control methods either rely on complex post-processing or adopt a fixed yaw action duration, which prevents them from achieving optimal control performance. Furthermore, all these methods lack theoretical analysis regarding their stability. To address these issues, this paper proposes the HysteresisPPO method, which integrates the hysteresis-aware dual-head architecture of HysteresisQNet into the Proximal Policy Optimization (PPO) framework. A corresponding policy gradient estimator is derived to ensure that the training process of HysteresisPPO remains consistent with that of the classic PPO algorithm. This paper also designs an adaptive weight adjustment strategy to balance power generation efficiency and yaw actuation frequency, thereby avoiding the need for post-optimization procedures. Additionally, sufficient conditions are presented to guarantee that HysteresisPPO satisfies practical stability in expectation, resolving the critical issue of controller stability analysis that has long been absent in RL-based yaw control systems. Three sets of simulations are conducted to evaluate the performance of HysteresisPPO. In the first set, simulations based on wind data from five different wind turbines demonstrate that HysteresisPPO outperforms existing RL-based yaw controllers in both energy capture and mechanical load reduction, with no requirement for post-processing; these results further empirically validate the derived sufficient conditions for practical stability in expectation. The second set verifies HysteresisPPO’s strong generalization capability across diverse wind conditions on four synthetic datasets. In the third set, we evaluate HysteresisPPO under two representative disturbances: sector-wise occlusion and wind vane offset. Results show that HysteresisPPO learns an effective feedback-based policy when each disturbance is applied individually, but degrades under their combination—revealing its reliance on feedback-driven control rather than internal environmental modeling.