基于改进Transformer的电力系统不良数据辨识  

Identification of Bad Data in Power Systems Based on Improved Transformer

作　　者：程慧琳 张晶 胡建一 卢志刚[3] 甄晓晨 CHENG Huilin;ZHANG Jing;HU Jianyi;LU Zhigang;ZHEN Xiaochen(State Grid Shijiazhuang Power Supply Company,Shijiazhuang 050000,China;State Grid Jinzhou County Power Supply Company,Jinzhou 052200,China;Key Laboratory of Power Electronics Energy Conservation and Motor Drive of Hebei Provine,Yanshan University,Qinhuangdao 066004,China)

机构地区：[1]国网河北省电力有限公司石家庄供电分公司,河北石家庄050000 [2]国网河北省电力有限公司晋州市供电分公司,河北晋州052200 [3]河北省电力电子节能与传动控制重点实验室(燕山大学),河北秦皇岛066004

出　　处：《河北电力技术》2024年第6期36-43,共8页Hebei Electric Power

基　　金：国网河北省电力有限公司科技项目(kjcb2021-019)。

摘　　要：针对目前电网状态估计时存在的不良数据辨识率低的问题,提出了一种基于改进Transformer的电力系统不良数据辨识方法。首先,改进传统Transformer编码器结构,在自注意力机制的基础上引入高斯核函数,以提高模型对不良数据邻近点数据的检测能力;然后,提出了一种基于JS散度极大极小值训练策略的损失函数,通过两阶段的互相优化,使高斯分布权重和注意力权重达到动态平衡;最后,采用无监督学习方法,以正常量测数据训练模型,对输入数据进行重构,并计算重构误差和重构得分,实现对不良数据的有效辨识。仿真结果表明:该方法在不良数据检测精确率、召回率、F1分数和总体准确率方面具有较好的性能。To improve the identification rate of bad data in power grid state estimation,this paper proposes a novel method based on an improved Transformer model.First,the traditional Transformer encoder structure is enhanced by incorporating a Gaussian kernel function into the self-attention mechanism,enhancing the model's capability to detect neighboring points of bad data.Second,a loss function is introduced based on a JS divergence maximization-minimization training strategy.This approach achieves a dynamic balance between Gaussian distribution weights and attention weights through mutual optimization in two stages.Utilizing an unsupervised learning framework,the model is trained with normal measurement data to reconstruct the input,allowing the calculation of reconstruction errors and scores for effective bad data identification.Finally,simulation results demonstrate that the proposed method outperforms existing approaches in terms of precision,recall,F1score,and overall accuracy,thereby validating its effectiveness in bad data detection for power systems.

关 键 词：不良数据辨识 Transformer网络 无监督学习 高斯核函数 重构分数 

分 类 号：TM71[电气工程—电力系统及自动化]