Scholarlink

<p>Traditional control strategies such as proportional–integral–derivative (PID) and linear quadratic regulator (LQR) controllers are widely used for DC–DC converters but face limitations when dealing with nonlinear system behavior, dynamic load variations, and input disturbances. PID controllers offer simplicity but struggle with robustness in varying operating conditions, while LQR controllers, though optimal for linearized models, lack adaptability to real-world uncertainties. To overcome these challenges, this paper presents a mixture of experts (MoE) control framework, which adaptively blends PID and LQR controllers through a reinforcement learning-based gating network. The MoE controller dynamically selects the most effective control action based on real-time system conditions, ensuring improved transient response, steady-state accuracy, and disturbance rejection. Simulation and experimental results demonstrate that the proposed MoE framework achieves a 68.3% reduction in settling time, 45.5% lower overshoot, and a 5.1% increase in efficiency compared to stand-alone PID and LQR controllers under varying load conditions. These improvements validate the MoE approach as a robust and adaptive solution for real-time DC–DC converter control.</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-enhanced expert mixture of LQR and PID for optimized control in DC–DC boost converters

  • 期刊论文

  • 发表日期：

    2025年5月14日

查看全文
  • Rongmei Zhao,
  • Ahmad Alkhayyat,
  • Mohammad Ahmar Khan

摘要

Traditional control strategies such as proportional–integral–derivative (PID) and linear quadratic regulator (LQR) controllers are widely used for DC–DC converters but face limitations when dealing with nonlinear system behavior, dynamic load variations, and input disturbances. PID controllers offer simplicity but struggle with robustness in varying operating conditions, while LQR controllers, though optimal for linearized models, lack adaptability to real-world uncertainties. To overcome these challenges, this paper presents a mixture of experts (MoE) control framework, which adaptively blends PID and LQR controllers through a reinforcement learning-based gating network. The MoE controller dynamically selects the most effective control action based on real-time system conditions, ensuring improved transient response, steady-state accuracy, and disturbance rejection. Simulation and experimental results demonstrate that the proposed MoE framework achieves a 68.3% reduction in settling time, 45.5% lower overshoot, and a 5.1% increase in efficiency compared to stand-alone PID and LQR controllers under varying load conditions. These improvements validate the MoE approach as a robust and adaptive solution for real-time DC–DC converter control.