基于改进深度自编码网络的异常用电行为辨识

为准确检测异常用电行为以降低电力公司的运营成本，提出一种基于改进深度自编码网络的异常用电行为辨识方法。首先将正常用户的用电数据作为训练样本，自编码网络逐层学习数据的有效特征；然后重构输入数据以计算检测阈值，而由于异常用电行为破坏数据的特征规则，再通过对比重构误差与检测阈值的差异即可实现异常用电行为辨识。为了改善自编码网络的特征提取能力与鲁棒性，分别引入了稀疏约束和噪声编码，并利用粒子群算法优化网络的超参数以提高模型的学习效率和泛化能力。选用福建省某地区居民用电和商业用电数据集进行了验证，这一模型的异常行为检测的准确率高于92%。实验表明所提方法具有优异的特征提取能力和异常用电行为辨识能力。

Abnormal Electricity Consumption Behaviors Detection Based on Improved Deep Auto-Encoder

In order to accurately detect the abnormal electricity consumption behaviors for reducing the operating costs of power companies, a detection method of abnormal electricity consumption behaviors is proposed based on the improved deep auto-encoder (DAE). Firstly, the data of normal electricity users are employed as training samples, and the effective features of the data are automatically extracted by AE; and then the data is reconstructed to calculate the detection threshold. Because the effective data characteristics are destroyed by the abnormal behaviors, the abnormal behaviors can be detected through comparing the difference between the reconstruction error and the detection threshold. To improve the feature extraction ability and the robustness of AE network, the sparse restrictions and the noise coding are introduced into the auto-encoder, and the hyper-parameters of AE network are optimized through the particle swarm optimization algorithm to improve the learning efficiency and generalization ability. The proposed model is validated by the electricity consumption dataset of domestic and business users of a region in Fujian province, and the abnormal detection accuracy is higher than 92%, which indicates that the proposed method has a powerful ability in feature extraction and abnormal behavior detection.