多目标优化时域卷积神经网络的窃电行为高准确检测算法

High-accuracy Detection Algorithm of Electricity Theft Behavior Based on Multi-objective Optimization Time-domain Convolutional Neural Network

随着当前用户用电行为的多样性及窃电行为隐蔽性的增加,窃电检测检出率高的同时误检率也增加,致使现有算法难以满足高准确窃电稽查的工程实际需求,对此该文提出多目标优化时域卷积神经网络的窃电行为高准确检测算法。首先基于时域卷积神经网络构建深度模型,引入侧输出融合结构提取用户用电信息的高低维特征,并通过注意力机制进行特征融合,提高模型窃电检测准确率;然后通过两阶段训练模型参数,第一阶段基于传统梯度下降算法优化模型权重与阈值,第二阶段基于窃电检测的混淆矩阵建立准确率、检出率与误检率3个维度的目标函数,采用第三代非支配排序遗传算法(non-dominated sorting genetic algorithm,NSGA-Ⅲ)进行模型优化训练,提高窃电检出率并降低误检率,据此建立多目标优化时域卷积神经网络的窃电行为高准确检测算法,并开发窃电检测实验平台实际验证所提算法;最后通过电网实测数据与实验平台验证,结果表明该文所提算法在降低误检率的同时提高10%检出率,比现有算法检出率提高同时误检率降低,更符合当前工程应用高准确窃电的检测实际需求。

With the diversity of current users'electricity consumption behaviors and the increasing concealing of electric theft behaviors,the detection rate of electric theft detection is high,and the false detection rate is also increasing,which makes it difficult for existing algorithms to meet the actual engineering requirements of high-accuracy electric theft inspection.In this paper,a high-accuracy electric theft detection algorithm based on multi-objective optimization of time-domain convolutional neural network is proposed.Firstly,the depth model is constructed based on the time-domain convolutional neural network,and the side output fusion structure is introduced to extract the high-low dimensional features of the user's electricity consumption information,and the feature fusion is carried out by the attention mechanism to improve the model's detection accuracy.Then model parameters are trained in two stages.In the first stage,the weight and threshold of the model are optimized based on the traditional gradient descent algorithm;in the second stage,objective functions of accuracy,detection rate and false detection rate are established based on the confusion matrix of power theft detection.The third-generation Non-dominated Sorting Genetic Algorithm(NSGA-Ⅲ)was used to optimize the model training,improve the detection rate of electric theft and reduce the false detection rate,and based on this,a high-accuracy detection algorithm for electric theft behavior of multi-objective optimization time-domain convolutional neural network was established.The experimental platform of electric theft detection is developed to verify the proposed algorithm.Finally,through the actual data of the power grid and the experimental platform verification,the results show that the proposed algorithm can reduce the false detection rate and increase the detection rate by 10%,which is more in line with the actual demand of highly accurate detection of electricity theft in current engineering applications.