A Load Forecasting Method of New Power System Based on Personalized Federated Learning

Abstract The emerging distributed generation technology in new power systems encounters the challenges of unstable efficiency and high accuracy of prediction models. The generalization ability of prediction models is hindered by variations in users’ behavioral characteristics. Furthermore, inability to share power data across regions poses substantial impediments to generation arrangement and power dispatch. This paper proposes a load forecasting technology based on federated learning (FL), which can avoid uploading or sharing the users’ data to protect data privacy. A multi-task module was added to traditional FL to improve user accuracy (UA) rather than global model accuracy, where the client trains a separate personalized model by keeping the local Layer-Normalization (LN) private. Moreover, in order to fast model convergence, the local LSTM prediction algorithm was added with the Grey Wolf optimization (GWO) algorithm and the attention mechanism. The experimental results show that the overall model training time of the improved LSTM algorithm is shortened by 26%. The Mean absolute percentage error (MAPE) of the proposed multi-task FL is 9.79% lower than traditional FL, and the MAPE of clients with small data volume and large feature deviation is reduced by 18.07% at most.