Scholarlink

<p>With the increasing complexity of modern power systems, traditional approaches like Long Short-Term Memory (LSTM) and statistical models often act as black boxes, making fault interpretation challenging. Additionally, a few abnormal events can significantly reduce detection accuracy. This study proposes an LSTM-SHAP model combined with TimeGAN to enhance time series data, enabling improved information fusion and anomaly detection in power systems. TimeGAN is first used to augment small-sample data, which is then analyzed using LSTM for anomaly detection on both original and expanded datasets. The SHAP model is applied to interpret the detection outcomes. Experiments utilized data from a large power company’s monitoring system (March–June 2024), with performance evaluated in terms of data generation quality and detection accuracy. Results demonstrate that the LSTM model achieved an accuracy of 0.95 on the expanded 3260 × 6 dataset, which is a 0.07 improvement over the original dataset. Additionally, performance metrics such as recall, precision, and F1 score also improved with the expanded data, with the F1 score increasing from 0.80 on the original data to 0.83 on the expanded data of 3260 × 6. These improvements in detection performance confirm the effectiveness of the proposed LSTM-SHAP model combined with TimeGAN for anomaly detection in power systems.</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

Interpretable information fusion and small sample dataset expansion in power system anomaly detection based on artificial intelligence algorithms

  • 期刊论文

  • 发表日期：

    2025年11月8日

查看全文
  • Junhao Zhang,
  • Junwei Kuang,
  • Yufei Teng,
  • Siyu Xiang,
  • Shengjie Li

摘要

With the increasing complexity of modern power systems, traditional approaches like Long Short-Term Memory (LSTM) and statistical models often act as black boxes, making fault interpretation challenging. Additionally, a few abnormal events can significantly reduce detection accuracy. This study proposes an LSTM-SHAP model combined with TimeGAN to enhance time series data, enabling improved information fusion and anomaly detection in power systems. TimeGAN is first used to augment small-sample data, which is then analyzed using LSTM for anomaly detection on both original and expanded datasets. The SHAP model is applied to interpret the detection outcomes. Experiments utilized data from a large power company’s monitoring system (March–June 2024), with performance evaluated in terms of data generation quality and detection accuracy. Results demonstrate that the LSTM model achieved an accuracy of 0.95 on the expanded 3260 × 6 dataset, which is a 0.07 improvement over the original dataset. Additionally, performance metrics such as recall, precision, and F1 score also improved with the expanded data, with the F1 score increasing from 0.80 on the original data to 0.83 on the expanded data of 3260 × 6. These improvements in detection performance confirm the effectiveness of the proposed LSTM-SHAP model combined with TimeGAN for anomaly detection in power systems.