Non-intrusive load disaggregation is an important means of sensing the power demand of terminal users. Traditional methods have low accuracy of appliance identification and power disaggregation. A load disaggregation model based on GAN (generative adversarial network) is proposed in this paper. The generative network deconstructs the total power signals by constructing a convolutional self-encoder to generate the power sequence of the appliance, and the discrimination network is used to identify the authenticity of the generated sequence. The two oppose each other to get a more realistic sequence. In view of the shortage of centralized model training method, this paper adopts deep separable convolution instead of traditional convolution, so that it can be applied to smart meters and other terminal equipments, and proposes a network model implementation solution based on federated learning. Training the model in a cloud-side collaborative manner reduces communication transmission pressure and protects user privacy and data security. The validity of the proposed method is verified based on public data sets.