Abstract: As an incentive mechanism, demand response (DR) is of great importance to the promotion of the balance between power supply and demand. However, the uncertainty and diversity of DR also lead to large deviations between actual and expected DR response volumes. To this end, a stochastic optimization strategy for DR is proposed based on an integrated semi-trapezoidal cloud model. First, considering that the actual degree of participation in demand-side response is closely related to the user's willingness, a logit-XGBoost user willingness linear regression model is established to strip the behavioral information in the user's information that has a low correlation with the willingness to respond, and express the user's uncertain willingness to participate in the response explicitly. Subsequently, for all the information of demand-side users that cannot be obtained in the actual scheduling process, we build a comprehensive half-trapezoidal cloud model based on the extracted user willingness to obtain the numerical features of user responses, and jointly characterize the uncertain mapping relationship between the benefits and response behaviors of users of the three types of demand response: curtailable, interruptible, and pannable. Then, considering that the dispersion and diversity of demand-side resources make their dispatchability have great limitations, a generic virtual battery (VB) model of three types of uncertain demand-side resources is established to quantify the demand response resource output, and optimize the peak shaving and valley filling effects of the demand-side scheduling model. Finally, the effectiveness and feasibility of the method are verified by arithmetic simulation.