Abstract:Machine learning methods are widely used in the field of short-term load forecasting of power systems, but the existing forecasting methods don't consider the following problems: one is the insufficient number and characteristics of each load operator data in the distribution network; the other is the ability to transform of the extensive load forecasting model is poor; third, because load data have privacy protection requirements, participants can't share data to train models together.Therefore, there's an urgent need for a method to achieve collaborative construction of the load forecasting model under the premise of protecting the privacy of load data. To solve this problem, we proposed the cooperative training method of a short-term load prediction model based on federated learning(FL). This method considers the LSTM load forecasting model with the characteristics of short-term load forecasting scenarios so that all participants can meet the premise of ensuring data privacy, and train the short-term load forecasting model together. The simulation results on multiple regional load data sets of the GEFCom2012 competition show that the proposed framework effectively improves the prediction performance, and has excellent generalization ability in multiple scenarios.