Machine learning methods have been widely used in the field of short-term load forecasting of power systems. However，it is difficult for load operators to obtain high-performance forecasting models due to insufficient data samples，poor model generalization ability, and high data privacy protection requirements in the application pro? cess. In this paper, meteorological, date, and historical load are used as input features to construct a short-term load forecasting model based on Long Short-Term Memory(LSTM). A federated learning(FL) based coordinated training method of a short-term load forecasting model is proposed. The proposed method mainly iteratively updates model pa? rameters through decentralized training and aggregation of centers, so as to realize cooperative construction of the pre? diction model by all load operators under the condition of data privacy. The simulation results based on the GEF?Com2012 dataset show that the proposed method not only ensures the data privacy of operators but also effectively im? proves the forecasting accuracy of the load forecasting model, and the trained model has satisfied generalization abil? ity in multiple scenarios.