Research on Federated Learning and Its Security Issues for Load Forecasting

Abstract

The safe and effective management and utilization of the load data of electricity has become one of the important issues for power supply and distribution departments as electricity is an important part of industry 4.0. Accurate forecasting of power load is of great significance for the safety and stability of power grid dispatching and economical operation. However, many of the current power data sets have serious problems of data island; furthermore, the centralized storage of large amounts of data may cause privacy leakage of the original data owners and faces regulations of security supervision. Therefore, federated learning is introduced to address these issues. Nevertheless, this approach is not sufficient to provide adequate data privacy protection. The present research proposes a federated learning model based on improved differential privacy algorithm. The model uses multi-scale Laplacian algorithm to analyze data distribution and generate noises in accordance with data patterns. Moreover, the parameters of the model are protected by attribute-based access control (ABAC). The simulation results show that the model proposed by the present research makes accurate forecasting and the improved differential privacy algorithm has less influence on the model's accuracy; the model also shows a good resistance to attacks, which ensures the security of data while having a high precision.