Scholarlink

<p>Power quality disturbances (PQDs) are caused by many power system malfunctions such as switching, transient, and component failures. This paper proposes a cross-recurrence plot (CRP)-based method for accurate PQD categorization. CRP identifies the dynamic developments and interactions that may be lost from single-signal approaches through comparison of disturbance signals and nominal waveforms. This is particularly effective when there are several disturbances simultaneously, a situation where traditional methods lack in, since CRP depicts predictive dynamics between two signals. The resulting patterns are informative features of a light four-layer deep learning model specifically designed for feature extraction and classification. Trained on data modeled analytically and tested on actual datasets, the approach obtains 97.73% accuracy in 18 classes and maintains 89–91% accuracy at noise levels of 25–40&#xa0;dB. The architecture’s performance was compared with pretrained networks, including AlexNet, Shufflenet, Darknet-53, and SqueezeNet.</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

Cross-Recurrence Plot-aided Deep Learning Framework for Accurate Identification of Power Quality Disturbances

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Prity Soni,
  • Pankaj Mishra,
  • Debasmita Mondal

摘要

Power quality disturbances (PQDs) are caused by many power system malfunctions such as switching, transient, and component failures. This paper proposes a cross-recurrence plot (CRP)-based method for accurate PQD categorization. CRP identifies the dynamic developments and interactions that may be lost from single-signal approaches through comparison of disturbance signals and nominal waveforms. This is particularly effective when there are several disturbances simultaneously, a situation where traditional methods lack in, since CRP depicts predictive dynamics between two signals. The resulting patterns are informative features of a light four-layer deep learning model specifically designed for feature extraction and classification. Trained on data modeled analytically and tested on actual datasets, the approach obtains 97.73% accuracy in 18 classes and maintains 89–91% accuracy at noise levels of 25–40 dB. The architecture’s performance was compared with pretrained networks, including AlexNet, Shufflenet, Darknet-53, and SqueezeNet.