面向配电异构数据的生成对抗式数据增殖技术研究

Generative Adversarial Learning Based Data Generation Technology for Distribution Heterogeneous Data

配电大数据跨媒体、小样本特征愈发凸显,为配电数据分析带来了新的挑战。为了解决样本分布不均匀的海量配电数据资源难以得到有效利用的问题,提出了面向配电异构数据的生成对抗式数据增殖方法。该方法通过引入生成对抗网络对数据完备的配电异构数据空间进行稠密化,而后利用峰值聚类算法实现样本空间的有限开覆盖,实现数据样本的有益增殖。仿真实验验证了所提出的生成对抗式数据增殖方法对于配电设备台账数据与巡检影像数据的有效性与稳定性,为配电数据分析提供了新的视角与思路。

The cross-media and small sample features of power distribution network big data have become more and more prominent,which brings new challenges to data analysis of power distribution network. Due to uneven distribution of the targeted sample groups,the massive data resources of power distribution network are difficult to be effectively utilized. In order to solve the problem,this paper proposes a data generation method for power distribution network heterogeneous data based on Generative Adversarial Network(i.e.GAN).The newly proposed method densifies the data-complete heterogeneous data space of power distribution network by introducing the GAN model,and then employs the peak clustering algorithm to realize the limited open coverage of the sample space,which can acquire beneficial generation of the data samples. The simulation experiments verify the validity and stability of the proposed GAN based data generation(i.e.GANDG)method for the index data and inspection image data of power distribution equipment,which provides a new perspective and ideas for power distribution network data analysis.