Mining typical electricity consumption patterns of industries and users is an important guarantee for power system planning and operation based on the massive historical electricity data collected and stored by smart meters and electricity information acquisition systems, as well as refined management of distribution networks. However, there are significant differences in users’ demand response potential across different time periods, and correlations exist between demand response characteristics of time periods with varying lengths. Additionally, users’ willingness to respond to demand is influenced not only by electricity prices but also by factors such as their historical response records and declared response electricity volume. To address these issues, this paper proposes a method for evaluating the typical demand response potential of industrial users based on IEMD and convolutional self-attention Transformer. Firstly, IEMD is used to decompose user load sequences, overcoming the mode mixing problem to extract objective demand response feature sequences. Secondly, considering the local contextual information of features in each time period, a convolutional self-attention Transformer is employed to establish a mapping model between demand response features and users’ demand response potential. Finally, case studies are conducted to verify the effectiveness and accuracy of the proposed method for evaluating the typical demand response potential of industrial users.