基于实值深度置信网络的用户侧窃电行为检测

用户侧窃电行为造成的非技术性损失对电网企业危害重大,不仅会影响电力系统的供电质量,还会增加电网的运营成本。为了辅助电网公司提高用电稽查效率、管理用户规范化用电,提出了基于实值深度置信网络的用户侧窃电行为检测模型。实值深度置信网络具有提取抽象特征的功能,并通过前馈神经网络微调后可实现较高分类精度。为了优化实值深度置信网络因随机初始化产生的局部最优化问题,该模型通过萤火虫算法对网络参数全局寻优。针对用户窃电行为检测,该模型利用因子分析进行数据降维,利用随机欠采样和套索算法应对数据不平衡问题,并利用ROC(receiver operatingcharacteristiccurve)曲线选取该模型的检测阈值。最后仿真实验验证了所提出模型的有效性和精确性。

Electricity Theft Detection for Customers in Power Utility Based on Real-valued Deep Belief Network

Non-technical losses cause great harm to power utility due to electricity theft, affecting power supply quality and increasing operating cost. In order to assist power utility to improve efficiency of electric inspection and normalize power consumption, an electricity theft detection model for customers is proposed in this paper based on real-valued deep belief network(RDBN). RDBN has the function of extracting abstract features, and high classification accuracy after fine-tuning by back propagation network. Facing to the local optimization problem in RDBN caused by random initialization of network parameters, the proposed model optimizes the network parameters with firefly algorithm. Aiming at electricity theft detection, factor analysis is used in the proposed model for data dimension reduction. Random under-sampling and LASSO are employed to deal with data imbalance, and the detection threshold is chosen through ROC curve. Simulink experiment verified effectiveness and accuracy of the proposed model.