基于时频域多维特征提取的用户异常用电行为检测

A Detection Method for Abnormal Electricity Consumption Behavior of Users Based on Time-Frequency Analysis and Multidimensional Feature Extraction

中文关键词:窃电检测  模态选择  时频分析  深度学习  多尺度特征提取

keywords:electricity theft detection  mode selection  time-frequency analysis  deep learning  multiscale feature extraction

中文摘要:传统的窃电检测方法大多直接在原始功率序列上构建模型，且没有同时考虑长时序列中的周期依赖关系以及周期之间的局部关联特征，限制了电力用户行为规律的深层挖掘。提出了一种综合采用时频模态融合和多尺度特征提取的高精度窃电检测模型。采用经验模态分解方法，将原始信号分解为多个本征模态信号和一个残差信号，依据模糊熵值与皮尔逊相关系数找到同时包含局部信息与原始信号信息较多的模态，并将选择的模态信号与原信号进行拼接，这样既可以提升模型的维度，又能放大窃电用户与正常用户的局部差异；将拼接好的数据先输入卷积神经网络进行局部特征提取，并从提取到的特征输入多头自注意力机制神经网络模型中提取全局特征，从而实现多尺度特征提取，以增强模型提取特征的适应性。在公开数据集上的实验结果表明，所提模型的F1值达到了76.71%、召回率达到了87.99%、曲线下面积（area under the curve，AUC）值达到了93.11%，相比于现有方法均取得了明显提升。

Abstract:Traditional detection methods for electricity theft mostly construct models directly on original power sequences without simultaneously considering periodic dependencies in long time sequence and local correlation features between periods， which limits the in-depth mining of electricity user behavior patterns. To this end， this paper proposes a high-precision electricity theft detection model that comprehensively employs time-frequency modal fusion and multi-scale feature extraction. Firstly， the empirical mode decomposition method is employed to decompose the original signal into multiple intrinsic mode functions （IMFs） and a residual signal. Secondly， based on fuzzy entropy values and Pearson correlation coefficients， the modes containing both local information and more original signal information are identified， and the selected modal signals are concatenated with the original signal. This approach not only enhances the model's dimensionality but also amplifies the local differences between electricity theft users and normal users. Furthermore， the concatenated data is first input into a convolutional neural network for local feature extraction， and the extracted features are then fed into a multi-head self-attention neural network model to extract global features， thereby achieving multi-scale feature extraction to improve the model's adaptability in feature extraction. Experimental results on public datasets show that the proposed model achieves an F1-score of 76.71%， a recall rate of 87.99%， and an area under the curve （AUC）value of 93.11%， demonstrating significant improvements compared to existing methods.