Scholarlink

<p>Accurate voltage prediction is crucial for power grid stability amid increasing renewable energy integration. However, existing prediction methods struggle to simultaneously adapt to sudden load changes and effectively capture both long-term steady-state trends and short-term transient fluctuations. To bridge this research gap, this paper proposes MSDAA-DAR (Multi-scale Dual-Agent Attention with Dynamic Adaptive Robustness). The primary contribution of this framework lies in integrating dynamic prompt engineering with a hierarchical attention mechanism. Unlike traditional static encoding, our model employs GRU-gated prompts to autonomously adjust to sudden voltage changes. Concurrently, a dual-agent structure is utilized to separately extract macroscopic trends and local transients, preventing feature interference and reducing computational overhead. Evaluated on a simulated dataset of 10,000 samples from the IEEE 33-bus system, MSDAA-DAR demonstrates robust performance, reducing prediction errors by up to 14.8% compared to state-of-the-art baselines in multi-step forecasting tasks. While the current evaluation primarily relies on simulated grid data, the framework’s proven ability to balance steady and dynamic fluctuations highlights its potential as an efficient solution for real-time grid stability management in modernized networks.</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

Multi-Scale Dual-agent Attention with Dynamic Adaptive Robustness for Voltage Prediction

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月14日

查看全文
  • Xu Wanting,
  • Kamsin Amirrudin,
  • Huang Jun,
  • Jia Huili,
  • Mei Yuan

摘要

Accurate voltage prediction is crucial for power grid stability amid increasing renewable energy integration. However, existing prediction methods struggle to simultaneously adapt to sudden load changes and effectively capture both long-term steady-state trends and short-term transient fluctuations. To bridge this research gap, this paper proposes MSDAA-DAR (Multi-scale Dual-Agent Attention with Dynamic Adaptive Robustness). The primary contribution of this framework lies in integrating dynamic prompt engineering with a hierarchical attention mechanism. Unlike traditional static encoding, our model employs GRU-gated prompts to autonomously adjust to sudden voltage changes. Concurrently, a dual-agent structure is utilized to separately extract macroscopic trends and local transients, preventing feature interference and reducing computational overhead. Evaluated on a simulated dataset of 10,000 samples from the IEEE 33-bus system, MSDAA-DAR demonstrates robust performance, reducing prediction errors by up to 14.8% compared to state-of-the-art baselines in multi-step forecasting tasks. While the current evaluation primarily relies on simulated grid data, the framework’s proven ability to balance steady and dynamic fluctuations highlights its potential as an efficient solution for real-time grid stability management in modernized networks.