Based on the monitoring data of household smart meters, a sample database of residents’demand response is constructed. Based on the sample database, feature engineering is carried out to fully mine the characteristics of responsive users, such as family attributes, response behavior and power consumption behavior. On this basis, the self-learning optimization model of resident power demand response neural network is constructed. According to the data of different family labels and historical response results, the participation of resident demand response is predicted. With the continuous development of demand response in typical scenarios, the model is iterated and optimized. Finally,according to the regulation objectives, the demand response regulation strategy is intelligently formulated. Results show that the proposed demand response strategy can accurately identify the residential demand response participation and reduce the demand response incentive cost.