Abstract: Source and load prediction is an important basis for virtual power plant （VPP） to make future dispatching plans.A collaborative optimization scheduling method of VPP generation side and user side based on multi-frequency combination short-term source load prediction is proposed.First of all,ensemble empirical mode decomposition （EEMD） is performed on the load data of the time series and reconstructed into two kinds of frequency,which is then predicted by the graph convolution network and long short-term memory （GCN-LSTM） fusion algorithm.The prediction results obtained from the multi-frequency model are aggregated into an uncertain fuzzy set.Considering the demand response,the VPP day-ahead two-layer optimal scheduling model is established.The upper layer takes the user benefit maximization as the goal,comprehensively utilizes the scheduling function of demand response,and optimizes multiple types of controllable loads based on the established time-of-use price.The lower layer aims to minimize the output cost of distributed power supply and take into account the interests of both sides of supply and demand,so as to optimize the internal resources of VPP.The above model is decomposed into main and sub-problems for solving by using the improved column reduction generation algorithm.The economy,robustness,and effectiveness of the proposed model are verified by a case analysis.