Abstract

—Theft detection in the power sector is a significant challenge for power distribution companies worldwide. The power losses are mainly due to dissipation from wires or theft done by manipulating energy meters or tapping cables at the consumer end. With power theft becoming a global issue, automatic detection of robbery is the need of the hour. This paper presents a deep learning-based solution for automated detection of power theft using consumers’ consumption data. In this work, fully connected neural networks are trained on daily consumption data, and customized convolutional neural networks (CNN) and residual networks are trained on weekly consumption data. The models are evaluated using the area under the receiver operating characteristics curve (AUC) metric, which measures the degree of separation between the predicted classes. The results obtained on the real dataset indicate that residual networks provide better results than other methods, and ResNet34 outperforms the existing methods in the literature. The proposed system has a high potential to detect power theft in households, which can help the authorities cut down non-technical losses occurring in the power sector.

摘要

—电力部门的盗窃检测是全球配电公司面临的重大挑战。功率损耗主要是由于电线的耗散或通过在消费者端操纵电能表或分接电缆而造成的。随着电力盗窃成为全球性问题，自动检测抢劫已成为当务之急。本文提出了一种基于深度学习的解决方案，用于使用消费者的消费数据自动检测电力盗窃。在这项工作中，全连接神经网络在日常消费数据上进行训练，定制的卷积神经网络 (CNN) 和残差网络在每周消费数据上进行训练。使用接收者操作特征曲线 (AUC) 指标下的面积评估模型，该指标衡量预测类别之间的分离程度。在真实数据集上获得的结果表明，残差网络提供了比其他方法更好的结果，并且 ResNet34 优于文献中的现有方法。拟议的系统在检测家庭电力盗窃方面具有很高的潜力，这可以帮助当局减少电力部门发生的非技术损失。