Abstract: In the process of guiding demand response in an orderly manner, the historical response data of demonstration project users is less accumulated. To quantitatively evaluate the response potential of residential users under aggregate control, aiming at the problem of small samples caused by the lack of historical response samples, this paper proposes a demand response potential evaluation method that adapts to the characteristics of small samples. Firstly, given the dynamic characteristics of time series data on the time axis, weighted dynamic time warping was used to identify the typical daily load patterns of residents, and the selection criteria for potential users were set according to the distribution characteristics of the load response period; On this basis, the deep clustering algorithm is used to achieve accurate classification of potential users under the action of multi-dimensional influencing factors. Secondly, the potential user response feature generation model is constructed to generate load information with a similar distribution pattern through the load information of the potential user's response time to construct an enhanced dataset. At the same time, considering the small-sample characteristics caused by the absence of real labels, the Co-regression algorithm was used to predict the user response labels of the enhanced dataset to weaken the dependence of the potential evaluation model on the response labels. Finally, the method's effectiveness is verified through a case study, which provides a reference for guiding the demand response of urban residents under the new situation.