An improved prediction model for residential user response potential based on DTW-K-medoids double-layer clustering and Bi-LSTM

Abstract

As demand response (DR) participants gradually loosen up, the number of new users participating in DR will continue to increase. Due to the small electricity consumption and wide distribution of residential users, they cannot participate in DR individually and need to be represented by load aggregators. However, for users who haven’t participated in DR before, the lack of historical response data makes it more difficult for load aggregators to predict their response potential. In this context, a response potential prediction model was constructed for such users. Firstly, based on the user's electricity consumption load and energy-saving awareness as the clustering criteria for the upper and lower layers, the DTW-K-medoids model is used to perform double-layer clustering on users, grouping users with similar electricity consumption behaviors and concepts into one group. Then, based on the clustering results, the electricity consumption data of residential users who have participated in DR among various types of users are used as the training set, and the Bi-LSTM algorithm is used to sequentially construct a response potential prediction model for each type of user. Finally, case analysis shows that the proposed model can effectively solve the problem of low prediction accuracy of response potential for users who haven’t participated in DR, and further reduce the assessment cost of load aggregators.