基于CNN⁃LSTM混合模型的多类别窃电行为检测

Multi‑class electricity theft detection based on the CNN‑LSTM hybrid model

针对复杂电网环境下窃电行为难以准确检测的问题,提出一种基于CNN‐LSTM混合模型的多类别窃电行为检测方法。首先基于卷积神经网络(CNN)良好的特征抽象能力提取一维用电数据的非周期性的局部特征,通过长短时记忆网络(LSTM)捕捉每日电能消耗数据间的相关性,提取周期性的用电特征建立特征融合层网络,再将CNN与LSTM提取的特征向量横向拼接获得新的融合向量,据此实现多类别窃电行为的准确检测。实验结果表明,本文提出方法能准确识别多类别窃电行为,相比现有检测方法检测结果更加全面准确。

ion ability of convolutional neural networks(CNN)is utilized to extract the non‑periodic local features of one‑dimensional electricity consumption data.Then,the long short‑term memory(LSTM)is adopted to capture the correlation between daily power consumption data and extract periodic power consumption features to establish feature fusion layer network.After that,the feature vectors extracted by CNN and LSTM are horizontally splicing to obtain a new fusion vector.Based on this,the accurate detection of multiple types of electric theft behavior are realized.Experimental results show that the proposed method can accurately identify multiple types of electric theft behavior,and the detection results are more comprehensive and accurate than the existing detection methods.