基于双层深度残差收缩网络的台区窃电用户识别方法

Transformer Area Electricity Theft User Identification Method Based on Double Layer Deep Residual Shrinkage Network

针对现实情况下台区数据海量且冗余的导致训练的深度神经网络梯度弥散与网络退化问题,本文引入深度残差网络模型,构造天然的恒等映射,通过将残差项加入网络进行训练,降低了深度神经网络训练的难度。为了在学习的过程中增强有用的特征抑制无用的特征,引入Squeeze-and-Excitation Networks(SENet)模型,学习每个通道的重要程度,自适应实现不同通道赋不同的权重。提出基于双层深度残差收缩网络的台区窃电用户识别方法,按照周期性设立父类与子类双层阈值,解决台区用电数据月份变化和周内变化导致的周期性问题,利用嵌入的非线性软化阈值,剔除不重要的特征参数,提升了在海量冗余数据下算法寻找特定异常解的能力。经过数据训练并结合台区现场运维结果,所提出的窃电用户识别方法能有效地提升台区内窃电用户辨识率。

Aiming at the problem of gradient dispersion and network degradation caused by massive and redundant data in the transformer area in reality,this paper introduces the deep residual network model,constructs a natural identity map,and reduces the difficulty of deep neural network training by adding the residual term to the network for training.In order to enhance useful features and suppress useless features in the process of learning,the sequence-and-exception networks(SENet)model is introduced to learn the importance of each channel and adaptively assign different weights to different channels.A method of transformer area electricity theft identification based on double-layer deep residual shrinkage network is proposed.The double-layer thresholds of parent and child classes are set periodically to solve the periodic problem caused by the monthly and weekly changes of transformer area power consumption data.The embedded nonlinear softening threshold is used to eliminate unimportant characteristic parameters,which improves the ability of the algorithm to find specific abnormal solutions under massive redundant data.After data training and combined with the on-site operation and maintenance results of the transformer area,the proposed method can effectively improve the identification rate of electricity theft in the transformer area.