Scholarlink

<p>In the era of industrial digitalization, high-quality data is critical for optimizing operations and decision-making, yet inconsistent, fragmented datasets from heterogeneous sources often undermine its utility. Traditional data governance approaches, reliant on manual rule creation and rigid workflows, struggle to address scalability and adaptability challenges in dynamic environments like power grid systems. This study proposes a new framework leveraging generative AI to automate SQL rule generation for anomaly detection, integrating domain-specific constraints and reinforcement learning to iteratively refine rules. By training Qwen2 on preprocessed grid data, the methodology transforms raw sensor logs and metadata into standardized, actionable rules, while embedding compliance checks to align with industry standards. Experimental validation demonstrates that the proposed model outperforms mainstream techniques (e.g., LSTM, GRU, CNN) by 2.3–3.6% in accuracy and F1-score, achieving 88.5% precision in detecting voltage anomalies. Ablation studies are conducted to highlight the importance of incorporating physical features in the model performance.</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

A generative artificial intelligence framework for automated data quality rule generation in power grid systems

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月12日

查看全文
  • Wenxiang Yang,
  • Li Tang,
  • Guanyu Zhang

摘要

In the era of industrial digitalization, high-quality data is critical for optimizing operations and decision-making, yet inconsistent, fragmented datasets from heterogeneous sources often undermine its utility. Traditional data governance approaches, reliant on manual rule creation and rigid workflows, struggle to address scalability and adaptability challenges in dynamic environments like power grid systems. This study proposes a new framework leveraging generative AI to automate SQL rule generation for anomaly detection, integrating domain-specific constraints and reinforcement learning to iteratively refine rules. By training Qwen2 on preprocessed grid data, the methodology transforms raw sensor logs and metadata into standardized, actionable rules, while embedding compliance checks to align with industry standards. Experimental validation demonstrates that the proposed model outperforms mainstream techniques (e.g., LSTM, GRU, CNN) by 2.3–3.6% in accuracy and F1-score, achieving 88.5% precision in detecting voltage anomalies. Ablation studies are conducted to highlight the importance of incorporating physical features in the model performance.