基于改进生成式对抗网络的电气数据升频重建方法  

Frequency-increased Reconstruction Method for Electrical Data Based on Improved Generative Adversarial Network

作　　者：李富盛 林丹 余涛[1] 王克英[1] 吴毓峰 杨家俊 LI Fusheng;LIN Dan;YU Tao;WANG Keying;WU Yufeng;YANG Jiajun(School of Electric Power,South China University of Technology,Guangzhou 510640,China)

机构地区：[1]华南理工大学电力学院,广东省广州市510640

出　　处：《电力系统自动化》2022年第3期105-112,共8页Automation of Electric Power Systems

基　　金：国家自然科学基金资助项目(U2066212);广东省普通高校基础研究与应用基础研究重点项目(2018KZDXM001)。

摘　　要：高频电气数据是提高电网态势感知准确度、监测水平和辅助服务质量等的数据基础之一,但是,传统重建算法难以实现高精度的数据重建。因此,文中利用改进生成式对抗网络将低频电气数据重建为高频。通过将时序数据转化为电气图像,实现神经网络方法对电气图像特征的高效提取。利用基于深层残差网络的生成器和改进的残差块结构,提高生成器的特征学习能力。此外,生成器损失函数考虑真实样本与生成样本在低维或高维特征的差别。以公开数据集为例进行算法验证,验证结果表明,相比于传统重建方法,所提方法具有更高的峰值信噪比、结构相似性和更低的平均绝对误差、平均绝对误差百分数,以及更高的高频细节还原度、重建精度,能够对不同数据集实现泛化。High-frequency electrical data is one of the data bases to improve the accuracy of situation awareness,monitoring level and auxiliary service quality of power grids.However,the traditional reconstruction algorithm is difficult to achieve the highprecision data reconstruction.Therefore,this paper uses the improved generative adversarial network to reconstruct the lowfrequency electrical data into the high-frequency ones.By transforming the time series data into electrical images,the efficient extraction of electrical image features by neural network method is realized.The generator based on deep residual network and improved residual block structure are used to improve the feature learning ability of the generator.In addition,the generator loss function considers the difference in low-or high-dimensional features between real samples and generated samples.The public data set is taken as an example to verify the algorithm.The verification results show that compared with the traditional reconstruction methods,the proposed method has higher peak signal-to-noise ratio,structural similarity,lower mean absolute error and mean absolute percentage error,as well as the higher high-frequency detail reproduction and reconstruction accuracy,and can be generalized for different data sets.

关 键 词：升频重建 低频电气数据 生成式对抗网络 深度残差网络 图像特征 数据驱动 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TM73[自动化与计算机技术—计算机科学与技术]