摘要
量测数据的采集、传输、转换各个环节都有可能发生故障或受到干扰,导致数据出现缺失。传统重建方法仅考虑单一数据分布规律,忽略了电力系统中量测点、采集变量之间的相关性及历史的负荷变化规律,重建精度低。该文提出了基于改进生成式对抗网络(wassersteingenerative adversarial networks,WGAN)的量测缺失值重建方法,并设计了适用于该问题的WGAN网络结构。通过WGAN的无监督训练,神经网络将自动学习到量测之间相关性、负荷波动规律等难以显式建模的复杂时空关系。利用真实性约束及上下文相似性约束优化隐变量,使得训练后的生成器将能够生成高精度的重建数据。文中方法完全依靠数据驱动,不涉及显式建模步骤,在大量量测出现缺失的情况下仍具有较高的重建精度。算例中分析了量测缺失数量与重建误差之间的关系,证明了文中方法性能稳定。对于算例中长期缺失的特定量测,文中方法所重建的数据能够体现量测真实的时序特性。
The collection,transmission and conversion of measured data may be interrupted or disturbed, leading to the loss of data. The traditional reconstruction method only considers the distribution of single data, ignores the correlation between measurement points, measured variables and the historical load change in power system. In this paper, an improved Wasserstein generating adversarial networks(WGAN) method for the reconstruction of missing value was proposed, and a WGAN network structure was designed. Through the unsupervised training of WGAN, the neural network could automatically learn the complex space-time relation which was difficult to model explicitly, such as the correlation between measurement and load fluctuation rule. The hidden variables were optimized by using the real constraint and context similarity constraint, so that the generator after training could generate high precision reconstruction data. The method in this paper is completely data-driven and does not involve explicit modeling steps. It still has high reconstruction accuracy when a large number of measurements are missing. In the calculation example, the relationship between the number of measurements missing and the reconstruction error is analyzed in the example, and it’s proved that the data reconstructed can reflect the real time sequence characteristics of the measurement for the long term missing specific measurement.
作者
王守相
陈海文
潘志新
王建明
WANG Shouxiang;CHEN Haiwen;PAN Zhixin;WANG Jianming(Key Laboratory of Smart Grid of Ministry of Education (Tianjin University),Nankai District,Tianjin 300072,China;State Grid Jiangsu Electric power Research Institute,Nanjing 210024,Jiangsu Province,China)
出处
《中国电机工程学报》
EI
CSCD
北大核心
2019年第1期56-64,共9页
Proceedings of the CSEE
关键词
电力系统量测
生成式对抗网络
缺失数据重建
卷积神经网络
时序特性
power system measarement
generating adversarial networks
missing data reconstruction
convolution neural network
time series characteristics
