摘要
为改善现有窃电检测时由电力数据特征复杂、数据样本不均等导致的检测效率、精度低等问题,提出了一种基于卷积神经网络(CNN)的电力数据分析模型。首先,考虑到窃电数据样本不均衡、数量有限,提出条件变分自动编码器的窃电曲线数据增强方法。其次,在分析了电力系统攻击模型的基础上,设计了基于CNN的窃电检测模型,从而提高算法执行效率及模型鲁棒性。试验阶段,以中国某电网公司采集的数据为例,对所提模型进行了分析和验证。与现有的数据增强方法相比,所提模型可以根据窃电功率曲线的实际形状和分布特征扩展训练集,并且可对基础CNN模型的性能进行改善。与原始数据集相比,模型准确率、F1分数和G均值分别提高了7.00%、6.65%和6.01%。试验结果验证了所提模型的有效性及实用性。所提模型为电力数据分析及安全故障隐患的发现提供了借鉴。
To improve the problems of low detection efficiency and accuracy caused by the complex features of power data and uneven data samples in the existing power theft detection,a convolutional neural network(CNN)-based power data analysis model is proposed.Firstly,considering the uneven and limited number of power theft data samples,a power theft curve data enhancement method with conditional variational autoencoder is proposed.Secondly,based on analyzing the power system attack model,a CNN-based power theft detection model is designed to improve the algorithm execution efficiency and model robustness.In the experimental phase,the proposed model is analyzed and validated using data collected from a power grid company in China as an example.Compared with existing data enhancement methods,the proposed model can extend the training set according to the actual shape and distribution characteristics of the power theft power curve and improve the performance of the base CNN model.Compared with the original dataset,the model accuracy,F1 score,and G-mean are improved by 7.00%,6.65%and 6.01%,respectively.The experimental results validate the effectiveness and practicality of the proposed model.The proposed model provides reference for power data analysis and the discovery of hidden safety faults.
作者
黄朝凯
吴丹妍
郑惠哲
黄小奇
HUANG Chaokai;WU Danyan;ZHENG Huizhe;HUANG Xiaoqi(Shantou Power Supply Bureau,Guangdong Power Gird Co.,Ltd.,Shantou 515041,China)
出处
《自动化仪表》
CAS
2023年第10期65-69,74,共6页
Process Automation Instrumentation
基金
南方电网公司科技基金资助项目(030500KK52200003)。
关键词
电网
电力系统
数据分析
深度学习
窃电
数据增强
卷积神经网络模型:G均值
Power grid
Power system
Data analysis
Deep learning
Power theft
Data enhancement
Convolutional neural network(CNN)model
G-mean