Clustering Analysis of Daily Load Curve Based on Convolutional Variational Autoencoder

The users' daily electricity consumption data can reflect their electricity consumption behavior characteristics, and the clustering task can extract users' representative daily load from a large amount of operating data, it can provide such a basis for tasks as power system planning and scheduling. In allusion to such defects in traditional clustering methods as low efficiency and the difficulty in extracting potential representations from daily load data scenarios with huge data size and high data dimensions, a clustering method based on convolutional-variational autoencoder (C-VAE) was proposed to cluster the load curves. Firstly, the potential characteristics of daily load data was extracted by dimensionality reduction of C-VAE. Secondly, cooperated with K-means the load clustering task was conducted. Finally, based on the distance between each load curve with the clustering center, each sort of clustering center was revised by weighted correction to obtain more representative typical day load curve. The actually collected data from Portuguese users in UCI dataset was utilized to conduct the examples validation, and the results show that the Davies-Bouldin Index (DBI) of this method decreased significantly than such traditional clustering methods as K-means, PCA+ K-means and so on, it also showed that the data within each type was more compact, and the distance among each type was further away, so the clustering quality was improved. Afterward, the Gaussian distance weighting was utilized to improve the clustering center, and a more typical daily load curve was extracted to make the analysis on the characteristics of users' electricity consumption behavior more accurate. Thus, the effectiveness of the convolutional variational autoencoder clustering method in daily load curve clustering task was verified.