Scholarlink

Virtual Synchronous Generator (VSG) technology has become essential for renewable energy integration, but its susceptibility to low-frequency oscillations poses significant stability risks. Current oscillation suppression methods face two critical challenges: (1) Traditional fixed-parameter control strategies struggle to adapt to the dynamic operating conditions of high-penetration renewable grids, where oscillation characteristics exhibit time-varying nonlinearity. (2) Existing adaptive methods require complex system modeling and expert tuning, limiting their practical implementation. To overcome these limitations, this paper proposes a Twin Delayed Deep Deterministic Policy Gradient (TD3)-enabled intelligent control framework. The method develops a four-dimensional feature space combined with a dual-parameter adjustment mechanism that dynamically optimizes virtual inertia and damping coefficients through reinforcement learning. Experimental results demonstrate that (1) The TD3-based controller reduces frequency deviation peaks by 75% and power fluctuations by 80% compared to conventional PI control. (2) The adaptive parameter coordination achieves 85% energy attenuation in critical 0.5–2 Hz oscillation bands, with response latency reduced to 4 ms without additional hardware requirements. Field tests on industrial energy storage systems confirm the framework's capability to maintain synchronous stability while accommodating rapid renewable power fluctuations.

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 Control Strategy for Low-Frequency Oscillation Suppression in VSG

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Heran Zhou,
  • Yonggang Li,
  • Yichen Zhou,
  • Fanchao Wei

摘要

Virtual Synchronous Generator (VSG) technology has become essential for renewable energy integration, but its susceptibility to low-frequency oscillations poses significant stability risks. Current oscillation suppression methods face two critical challenges: (1) Traditional fixed-parameter control strategies struggle to adapt to the dynamic operating conditions of high-penetration renewable grids, where oscillation characteristics exhibit time-varying nonlinearity. (2) Existing adaptive methods require complex system modeling and expert tuning, limiting their practical implementation. To overcome these limitations, this paper proposes a Twin Delayed Deep Deterministic Policy Gradient (TD3)-enabled intelligent control framework. The method develops a four-dimensional feature space combined with a dual-parameter adjustment mechanism that dynamically optimizes virtual inertia and damping coefficients through reinforcement learning. Experimental results demonstrate that (1) The TD3-based controller reduces frequency deviation peaks by 75% and power fluctuations by 80% compared to conventional PI control. (2) The adaptive parameter coordination achieves 85% energy attenuation in critical 0.5–2 Hz oscillation bands, with response latency reduced to 4 ms without additional hardware requirements. Field tests on industrial energy storage systems confirm the framework's capability to maintain synchronous stability while accommodating rapid renewable power fluctuations.