Scholarlink

<p>Electricity theft seriously affects the stability of the power system, causing damage to both legitimate users and power companies. Traditional data-driven electricity theft detection methods rely on manually designing the feature representation of electricity data, and often ignore the time-series nature, which are difficult to be applied to well extract the features of high-dimensional data. To solve these problems, this paper proposes a method, combining white noise test and Seasonal and Trend Decomposition using LOESS, to adaptively obtain the potential period range of time-series, and then determines the period to be decomposed by using prior tools. In the classifier construction part, convolutional neural networks with different kernel sizes are used to extract medium and short-term features, and residual networks are used to extract long-term features. Additionally, the proposed method here is comprehensively compared with other general classifiers on the public electricity theft dataset, and the results, that the multiple metrics of the proposed method are significantly improved compared with the baseline models, demonstrate superiority of the proposed method.</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

Temporal residual three-dimensional convolution neural networks for electricity theft detection

  • 期刊论文

  • 发表日期：

    2025年2月20日

查看全文
  • Lichao Feng,
  • Xiangyang Xia,
  • Chunyan Zhang,
  • Liping Du,
  • Nan Ji

摘要

Electricity theft seriously affects the stability of the power system, causing damage to both legitimate users and power companies. Traditional data-driven electricity theft detection methods rely on manually designing the feature representation of electricity data, and often ignore the time-series nature, which are difficult to be applied to well extract the features of high-dimensional data. To solve these problems, this paper proposes a method, combining white noise test and Seasonal and Trend Decomposition using LOESS, to adaptively obtain the potential period range of time-series, and then determines the period to be decomposed by using prior tools. In the classifier construction part, convolutional neural networks with different kernel sizes are used to extract medium and short-term features, and residual networks are used to extract long-term features. Additionally, the proposed method here is comprehensively compared with other general classifiers on the public electricity theft dataset, and the results, that the multiple metrics of the proposed method are significantly improved compared with the baseline models, demonstrate superiority of the proposed method.