Bai Yu, Sun Haitong, Zhang Lili, Wu Haoqi
School of Electronical and Information Engineering, Shenyang Aerospace University, Shenyang 110136, China.
Sensors (Basel). 2023 Oct 12;23(20):8405. doi: 10.3390/s23208405.
Illicitly obtaining electricity, commonly referred to as electricity theft, is a prominent contributor to power loss. In recent years, there has been growing recognition of the significance of neural network models in electrical theft detection (ETD). Nevertheless, the existing approaches have a restricted capacity to acquire profound characteristics, posing a persistent challenge in reliably and effectively detecting anomalies in power consumption data. Hence, the present study puts forth a hybrid model that amalgamates a convolutional neural network (CNN) and a transformer network as a means to tackle this concern. The CNN model with a dual-scale dual-branch (DSDB) structure incorporates inter- and intra-periodic convolutional blocks to conduct shallow feature extraction of sequences from varying dimensions. This enables the model to capture multi-scale features in a local-to-global fashion. The transformer module with Gaussian weighting (GWT) effectively captures the overall temporal dependencies present in the electricity consumption data, enabling the extraction of sequence features at a deep level. Numerous studies have demonstrated that the proposed method exhibits enhanced efficiency in feature extraction, yielding high F1 scores and AUC values, while also exhibiting notable robustness.
非法获取电力，通常称为窃电，是电力损耗的一个重要原因。近年来，神经网络模型在窃电检测（ETD）中的重要性日益得到认可。然而，现有方法获取深度特征的能力有限，在可靠有效地检测功耗数据中的异常方面一直存在挑战。因此，本研究提出了一种混合模型，将卷积神经网络（CNN）和Transformer网络结合起来，以解决这一问题。具有双尺度双分支（DSDB）结构的CNN模型包含周期内和周期间卷积块，对不同维度的序列进行浅层特征提取。这使得模型能够以从局部到全局的方式捕获多尺度特征。具有高斯加权（GWT）的Transformer模块有效地捕获了功耗数据中存在的整体时间依赖性，能够在深度上提取序列特征。许多研究表明，所提出的方法在特征提取方面具有更高的效率，产生高F1分数和AUC值，同时还具有显著的鲁棒性。
