基于LSTM-SAE与支持向量机的窃电识别方法研究

Studies on Electricity Theft Detection Approach Based on LSTM-SAE and Support Vector Machine

用电行为的有效特征提取作为提升异常用电检测准确率的关键要素,近年在窃电检测的研究中常被忽略。基于此,文章提出基于长短期记忆堆叠自编码器的特征提取方法。基于长短期记忆神经网络对高维非线性时间序列的特征提取能力与深度自编码器的数据信息展示能力实现对负荷序列的深度特征挖掘,利用人工蜂群算法优化的支持向量机将提出的特征量映射至是否窃电的标签。借助实际数据,以真正率、假正率为评价指标验证了所提模型的有效性。

Effective feature extraction of power consumption behavior is frequently disregarded in the research of power theft detection in recent years,despite its importance in enhancing the accuracy of anomalous power consumption detection.This paper suggests a feature extraction method based on long and short term memory stacking auto encoder(LSTM-SAE) in light of this.The deep feature mining of load series is realized,and the artificial bee colony algorithm-optimized support vector machine (SVM) is used to map the proposed feature quantity to the tag of whether to steal electricity.These capabilities are based on the feature extraction capability of long and short term memory (LSTM) neural network for high-dimensional nonlinear time series and the data information display capability of deep auto encoder.The correctness of the suggested model is confirmed with the aid of actual data by using the true positive rate (TPR) and the false positive rate (FPR) as evaluation markers.