Analysis of Electricity Consumption Behaviors Based on 1D-CAE-ResBlock and Deep Embedded Clustering

Abstract

With the rapid development of smart grids, accurate identification of user electricity consumption behavior patterns has become an essential foundation for demand-side management and personalized power services. However, traditional clustering methods struggle to effectively capture the underlying regularities of daily load curves. To this end, this paper constructs a Deep Embedded Clustering model integrating a one-dimensional convolutional autoencoder and residual blocks (1D-CAE-ResBlock-DEC). The model adaptively extracts low-dimensional feature representations from load curves through a two-layer one-dimensional convolutional structure combined with down-sampling and residual connections. Building on this, a clustering embedding layer is designed to unify feature extraction and clustering objectives within an end-to-end training framework. The experimental results show that the proposed 1D-CAE-ResBlock model significantly outperforms traditional methods in terms of electricity consumption pattern recognition and clustering interpretability. The reconstruction loss of CAE-ResBlock is as low as 0.000568, and its reconstruction performance outperforms both AE and CAE models. The clustering results exhibit a SC of 0.8215 and a DBI of 0.2196, both of which are superior to those of other comparison models.