Aiming at the problems of complex signal types and heterogeneous data of power quality disturbance in novel power system, a power quality disturbance classification method using Federated learning and prototype learning is proposed. This method includes two types of work nodes: server and client. The server collects the local prototype output from the local model of clients. The local prototype cannot be reverse reconstructed to get the original data. Instead, the server aggregates the local prototype to get the global prototype and sends it back to the client to regularize the local model training. Compared with the convolutional neural network model, this method does not require a lot of training data, and the model is not vulnerable to slight heterogeneous data disturbance, and has strong robustness to unknown disturbance signals. The simulation experimental results show that, compared with existing methods, the proposed method is suitable for small-scale power quality disturbance samples, with a classification accuracy of 0.998 3, which has high application value in the new distributed power grid environment.