考虑样本不平衡的窃电检测模型

Electricity Theft Detection Model Considering Sample Imbalance

摘要: 随着全社会用电量持续增加，窃电问题日益严重，给电力企业造成巨大的经济损失，并影响电网的运行。与此同时，窃电样本不足导致基于大数据的窃电检测方法受到限制，本文针对现实情况中窃电案例收集困难、数量稀少问题，提出了一种考虑样本不平衡的窃电检测模型。首先，通过生成对抗网络(GAN)生成器与判别器的对抗训练，学习窃电数据的时序相关性，制造出与窃电样本近似的样本，使得窃电数据集中正负样本趋于平衡。然后结合卷积神经网络、长短期记忆递归神经网络和注意力机制(CNN-LSTM- Attention)对用户进行窃电检测，将经过样本不平衡处理后的用户用电信息经过CNN进行特征提取，通过LSTM捕捉数据的时序变化信息，使用Attention对LSTM的输出赋予权重，强化有利于窃电检测的特征数据，弱化无关数据。算例分析表明，本文提出的方法能有效避免样本不平衡问题，更好地检测出用户窃电行为。

Abstract: With the continuous increase of electricity consumption in the whole society, the problem of power theft is becoming more and more serious, which causes huge economic losses to electric power enterprises and affects the operation of power grids. At the same time, the lack of power theft samples leads to the limitation of big data-based power theft detection methods. In this paper, we propose a power theft detection model considering sample imbalance to address the problem of difficulty in collecting and scarcity of power theft cases in the real situation. Firstly, the temporal correlation of electricity theft data is learned through the adversarial training of generator and discriminator of Generative Adversarial Network (GAN), which creates samples that are close to the electricity theft samples, so that the positive and negative samples in the electricity theft dataset tend to be balanced. Then combine the convolutional neural network, long and short-term memory recurrent neural network and attention mechanism (CNN-LSTM-Attention) to detect power theft to the user, after the sample imbalance processing of the user’s electricity consumption information through the CNN for feature extraction, through the LSTM to capture the temporal change of the data information, the use of Attention on the output of the LSTM to give weight, and strengthen the features that are favorable to the detection of power theft. Strengthen the feature data conducive to power theft detection and weaken the irrelevant data. Case analysis shows that the method proposed in this paper can effectively avoid the sample imbalance problem and better detect the user’s power theft behavior.