| 引用本文: | 陈浈斐,包淼,葛磊蛟,马程,焦奇昱.标签优化非对称卷积编码器的电力负荷深度嵌入聚类方法[J].电力自动化设备,2024,44(12):69-75,99. |
| CHEN Zhenfei,BAO Miao,GE Leijiao,MA Cheng,JIAO Qiyu.Deep embedding clustering method of power load based on asymmetric convolutional autoencoder with label optimization[J].Electric Power Automation Equipment,2024,44(12):69-75,99. |
|
| 摘要: |
| 在分析电力系统负荷特性和用电行为的过程中，传统聚类方法面对高维数据的复杂样本往往难以学习负荷样本的深层特征。提出一种加入标签优化的非对称卷积自编码器的负荷深度嵌入聚类方法，取消传统卷积中的池化层和上采样层，以全面保留负荷原始信息，并在编码侧使用残差模块强化特征提取能力；在聚类层的伪标签设置过程中，使用自组织映射-K均值方法提供更精确的初始软聚类标签，从而提高聚类效率。基于实际电力负荷数据集进行实验，结果表明：所提方法能够有效降低数据维度以及增强对负荷特征的学习，提高了聚类的准确性和稳定性；相较于改进的深度嵌入聚类-卷积自编码器方法，所提方法的方差比准则和轮廓系数分别提升了10.881 6和0.06729，戴维森堡丁指数下降了0.133 14。 |
| 关键词: 负荷聚类 卷积自编码器 特征提取 残差模块 深度嵌入聚类 负荷特性 |
| DOI：10.16081/j.epae.202410006 |
| 分类号:TM73 |
| 基金项目:国家自然科学基金资助项目(51907052)；新一代人工智能国家重大专项(2022ZD0116900) |
|
| Deep embedding clustering method of power load based on asymmetric convolutional autoencoder with label optimization |
|
CHEN Zhenfei1, BAO Miao1, GE Leijiao2, MA Cheng1, JIAO Qiyu1
|
|
1.School of Electrical and Power Engineering, Hohai University, Nanjing 211100, China;2.School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China
|
| Abstract: |
| In the process of analyzing power system load characteristic and electricity consumption behavior, the traditional clustering algorithms are often difficult to learn the deep features of load samples when facing complex samples with high-dimensional data. A deep load embedding clustering method based on a label-optimized asymmetric convolutional autoencoder is proposed. The pooling and up-sampling layers in the traditional convolution are removed to fully preserve the original load information, and the residual modules are used on the encoding side to enhance the feature extraction capability. In the pseudo-label setting process of clustering layer, the self-organizing map-K-means method is employed to provide more accurate initial soft clustering labels, thereby improving the clustering efficiency. The experiment is carried out based on the real power load datasets, and the results show that the proposed method can effectively reduce the data dimensionality and enhance the learning of load features, improving the accuracy and stability of clustering. Compared with the improved deep embedding clustering-convolutional autoencoder algorithm, the Calinski-Harabasz index and silhouette coefficient indices are respectively increased by 10.881 6 and 0.06729, and the Davies-Bouldin index is decreased by 0.133 14. |
| Key words: load clustering convolutional autoencoder feature extraction residual module deep embedding clustering load characteristic |
