Scholarlink

<p>This paper focuses on optimizing the speed regulation of a certain twin-shaft turbofan aero-engine and explores the reinforcement learning predictive control (RLPC) method based on the Identifier-Actor-Critic architecture. First, the linear parameter variation model of the turbofan engine is used to generate offline data, from which an Identifier agent model is constructed to produce predictive data, thereby expanding the learning dataset and supporting the Actor-Critic optimization process. Secondly, an input penalty term is added to the one-step cost to prevent the fuel flow rate from exceeding the limit, and a decay term is introduced into the Q function to achieve the boundary optimization of the rotational speed. Finally, the gradient descent method is applied to obtain the optimized controller through an alternating iterative update of Critic (updating the Q-function) and Actor (updating the control policy) to realize the rotor speed regulation. The simulation results verify the effectiveness of this RLPC method, indicating that it not only effectively accelerates the rotor speed of the engine in the starting stage, but also achieves a smooth transition to the steady state.</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

Reinforcement learning based speed control for aero-engine

  • 期刊论文

  • 发表日期：

    2026年1月7日

查看全文
  • Junping Yao,
  • Chengxi Zhang,
  • Jiwei Wen,
  • Ruichao Li

摘要

This paper focuses on optimizing the speed regulation of a certain twin-shaft turbofan aero-engine and explores the reinforcement learning predictive control (RLPC) method based on the Identifier-Actor-Critic architecture. First, the linear parameter variation model of the turbofan engine is used to generate offline data, from which an Identifier agent model is constructed to produce predictive data, thereby expanding the learning dataset and supporting the Actor-Critic optimization process. Secondly, an input penalty term is added to the one-step cost to prevent the fuel flow rate from exceeding the limit, and a decay term is introduced into the Q function to achieve the boundary optimization of the rotational speed. Finally, the gradient descent method is applied to obtain the optimized controller through an alternating iterative update of Critic (updating the Q-function) and Actor (updating the control policy) to realize the rotor speed regulation. The simulation results verify the effectiveness of this RLPC method, indicating that it not only effectively accelerates the rotor speed of the engine in the starting stage, but also achieves a smooth transition to the steady state.