In recent studies, Non-Intrusive Load Monitoring (NILM) methods based on deep learning have received widespread attentions and achieved promising results. Most existing NILM models are typically trained on data from an individual user, and thus may have poor performance when applied for load recognition of other users. Achieving load recognition across different households requires joint training of models with data collected from different users. However, due to concerns regarding user data privacy, it is challenging to directly access and utilize electricity consumption data from various users in practice. Motivated by such a problem, this study designs a federated learning method for NILM based on Fed-Prox and Bi-GRU. By aggregating and optimizing models trained by various local households at the central server multiple times, the resulting global model achieves load recognition across different households. Experiments are conducted using the UK-DALE and REFIT datasets to validate the proposed framework’s effectiveness in load recognition across multiple users. The experimental results demonstrate that compared to NILM methods tailored for an individual user, the proposed approach exhibits better generalization performance for load recognition across multiple diverse users.