In order to solve the problem of lack of analysis and prediction of user's response behavior under different environments and incentive signals in the current power demand response analysis, a data mining method based on power user's power consumption behavior is proposed. Construct and analyze the incentive-based demand response architecture, establish a user response flexibility model based on the existing abstract formula of user response cost, and propose a double-layer longshort-term memory network to identify user response behavior model; The proposed model is compared with random forest (RF), support vector machines (SVM), recurrent neural network (RNN) and long short-term memory (LSTM). The results show that the proposed model has excellent performance, the accuracy rate is 94.83%, the F1 score is 95.45%, and the quality factor is 39.42%, which can provide a reference for the development of safe operation and management of electric power.