基于动态融合注意力机制的电力负荷缺失数据填充模型  

Power Load Missing Data Imputation Model Based on Dynamic Fusion Attention Mechanism

作　　者：赵冬[1] 李亚瑞 王文相 宋伟[4] ZHAO Dong;LI Yarui;WANG Wenxiang;SONG Wei(School of Software,Zhongyuan University of Technology,Zhengzhou 450007,China;School of Computer,Zhongyuan University of Technology,Zhengzhou 451191,China;Xuchang Xuji Software Technology Co.,Ltd.,Xuchang 461000;School of Computer and Artificial Intelligence,Zhengzhou University,Zhengzhou 450001,China)

机构地区：[1]中原工学院软件学院,河南郑州450007 [2]中原工学院计算机学院,河南郑州451191 [3]许昌许继软件技术有限公司,河南许昌461000 [4]郑州大学计算机与人工智能学院,河南郑州450001

出　　处：《郑州大学学报(工学版)》2025年第2期111-118,共8页Journal of Zhengzhou University(Engineering Science)

基　　金：国家重点研发计划项目(2023YFC2206400);国家高能物理科学数据中心开放课题(HT-HEPS-T7-01050200-21-0008);河南省高等学校重点科研项目(22A520010)。

摘　　要：为了提高电力负荷数据的缺失值填充精度,保障后续数据分析与应用的高效进行,首先,提出一种基于动态融合注意力机制的填充模型(DFAIM),该模型由注意力机制模块和动态加权融合模块构成,通过注意力机制模块的两种不同注意力机制挖掘特征与时间戳之间的深层关联;其次,通过动态加权融合模块将可学习的权重赋予注意力机制模块的两个输出以得到特征表示;最后,利用特征表示来替换缺失位置的值,从而得到准确的填充结果。使用纽约市某地区的气象及负荷数据集及UCI电力负荷数据集对提出的模型进行验证,实验结果表明:相较于统计学、机器学习和深度学习填充模型,DFAIM在评价指标MAE、RMSE和MRE上均具有一定优势。In order to improve the imputation accuracy of power load missing value data and guarantee the efficiency of subsequent data analysis and application,firstly,a imputation model based on dynamic fusion of attention mechanism dynamic fusion of attention mechanism imputation model(DFAIM)was proposed.The model consisted of an attention mechanism module and a dynamic weighted fusion module,where the deep correlation between features and timestamps was mined through the two different attention mechanisms of the attention mechanism module.Secondly,feature representations were obtained by assigning learnable weights to the two outputs of the attention mechanism module through the dynamic weighted fusion module.Finally,replacing the values of the missing locations with the obtained feature representations to obtain the imputed values.The proposed model was validated using the meteorological and load dataset and the UCI electric load dataset for an area in New York City,and the experimental results showed that DFAIM had certain advantages in evaluating metrics such as MAE,RMSE,and MRE compared to statistics,machine learning,and deep learning models imputation models.

关 键 词：缺失值填充 注意力机制 电力负荷 时序特征 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]