基于卷积自编码器的日负荷深度嵌入聚类方法

负荷聚类是电力大数据分析的重要基础.针对高维日负荷数据时序特征提取困难,以及特征提取与聚类处理分离降低负荷聚类准确性的问题,文章提出了一种基于一维卷积自编码器的日负荷深度嵌入聚类方法(deep embedding clustering method based on one dimensional convolutio...

Deep Embedding Clustering Method for Daily Load Based on Convolutional Auto-Encoder

Clustering of load data is an important foundation for analyzing electrical big data. Aiming at the difficulty of extracting sequential features of high-dimensional daily load data, and the reduction of accuracy of load clustering due to the separation of feature extraction and clustering processing, a deep embedding clustering method based on one dimensional convolutional auto-encoder (DEC-1D-CAE) is proposed for daily load data in this paper. Firstly, a one-dimensional convolutional auto-encoder is used to extract sequential features contained in the load curve. Then, a user-defined clustering layer is used for soft division of the extracted load feature vector. Finally, the Kullback-Leibler divergence (KLD) is used as loss function to jointly optimize convolutional auto-encoder and the clustering layer to obtain the clustering result. A numerical experiment were carried out and the results of the proposed method are better than K-means, 1D-CAE+K-means and DEC-1D-CAE on both Davies-Bouldin index (DBI) and Calinski-Harabasz index (CHI), which indicate that the proposed method can effectively improve the accuracy of daily load clustering.