基于卷积自编码器的日负荷深度嵌入聚类方法

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

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 convolutional auto-encoder,DEC-1D-CAE)。首先,采用一维卷积自编码器网络提取负荷曲线蕴含的时序特征。然后,利用自定义聚类层对所提取的负荷特征向量进行软划分。最后,采用KL散度(Kullback-Leibler divergence,KLD)为损失函数,联合优化卷积自编码器与聚类层,得到聚类结果。算例分析表明所提方法在DBI(Davies-Bouldin index)、CHI(Calinski-Harabasz index)指标上均优于K-means、1D-CAE+K-means、基于堆叠式编码器的深度嵌入聚类方法(deep embedding clustering method based on stacked auto-encoder,DEC-SAE),所提方法可以有效提升日负荷聚类的准确性。

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-1 D-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-1 DCAE 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.