In view of the flaw of privacy leakage problem in the process of uploading user data by users during traditional power load decomposition, a scheme has thus been proposed to train the power load model through federated learning by using the Conditional Generative Adversarial Network (cGAN) model. By using a small amount of data set from each local user model training, users upload the relevant parameters of the trained model to the server, which subsequently organizes and aggregates the collected parameters, with the model parameters further distributed to the users. It is guaranteed that local user data remains private while still enabling model training, with the effectiveness of this method validated through experiments on the publicly available UK_DALE dataset.