Abstract: Under the traditional cloud power quality disturbance identification mode, massive disturbance data has caused great storage and calculation pressure on the cloud server, and cloud disturbance identification has delay and poor real-time performance; Edge side disturbance identification can alleviate cloud pressure and reduce delay, but cross domain data sharing cannot be achieved between edge sides. To solve the above problems, this paper proposes a edge computing framework based on federated learning. First, the edge side uses local data to train the model, then upload the model parameters to the cloud for aggregation, update the model and send it to the edge side for deployment, and identify and classify power quality disturbances on the edge side. The experimental results show that, compared with the cloud disturbance recognition mode, the storage requirements of the edge side disturbance recognition based on federated learning on the cloud are reduced by 97.58%, the data communication cost is reduced by 53.68%, and the transmission rate requirements of single disturbance recognition are reduced by 99.994%, meeting the real-time requirements of disturbance recognition; At the same time, compared with the traditional federated learning algorithm, the accuracy of disturbance identification of the optimized federated learning algorithm is improved by 1.72% to 3.64%.