生成对抗网络与胶囊网络协同优化的窃电检测方法
Collaborative Optimization of Generative Adversarial Networks and Capsule Networks for Electricity Theft Detection

在现代智能电网系统中，窃电是导致经济损失的主要因素。然而，现有窃电检测方法面临数据量不足的问题，尤其在窃电样本稀缺的情况下，检测器往往会偏向大多数类，从而降低窃电样本的检测精度。为了解决这一问题，本文提出一种在小样本条件下基于深度学习的窃电检测方法(CapsGAN)。该方法融合胶囊网络与卷积神经网络。卷积神经网络用于提取用电数据中的局部浅层特征，识别基本的用电模式，而胶囊网络通过动态路由机制对这些特征进行整合，进一步构建更深层次的高级特征，最终实现对窃电行为的精准识别。同时为了进一步解决数据不平衡问题，本文还引入了一种基于生成对抗网络的三方结构模型，该模型结合了鉴别器、生成器和分类器，并引入了决策边界约束。生成器与分类器通过协同训练，共同生成窃电检测中的异常样本，逐步扩展少数类的决策边界，从而有效缓解窃电样本的类别不平衡问题。通过在中国国家电网公司提供的真实数据集上的实验，结果充分验证了该方法的有效性和准确性。
In modern smart grid systems, electricity theft is a major cause of economic losses. However, existing electricity theft detection methods face the problem of insufficient data, especially when theft samples are scarce. In such cases, detectors tend to be biased toward the majority class, leading to a reduction in the detection accuracy of theft samples. To address this issue, this paper proposes a deep learning-based electricity theft detection method under small sample conditions (CapsGAN), which integrates Capsule Networks (CapsNet) and Convolutional Neural Networks (CNNs). In this approach, CNNs are used to extract local shallow features from electricity consumption data, recognizing basic consumption patterns, while the CapsNet integrates these features through dynamic routing, further constructing deeper-level high-level features to accurately identify theft behaviors. Additionally, to further address the data imbalance problem, this paper also Introduces a tripartite structure model based on Generative Adversarial Networks (GANs), which combines a discriminator, generator, and classifier, and introduces decision boundary constraints. The generator and classifier are collaboratively trained to jointly generate abnormal samples for theft detection, gradually expanding the decision boundary of the minority class, thus effectively alleviating the class imbalance issue in theft samples. Experiments on a real dataset provided by the State Grid Corporation of China validate the effectiveness and accuracy of this method.