Differentially Private Clustered Federated Load Prediction Based on the Louvain Algorithm

Load forecasting plays a fundamental role in the new type of power system. To address the data heterogeneity and security issues encountered in load forecasting for smart grids, this paper proposes a load-forecasting framework suitable for residential energy users, which allows users to train personalized forecasting models without sharing load data. First, the similarity of user load patterns is calculated under privacy protection. Second, a complex network is constructed, and a federated user clustering method is developed based on the Louvain algorithm, which divides users into multiple clusters based on load pattern similarity. Finally, a personalized and adaptive differentially private federated learning Long Short-Term Memory (LSTM) model for load forecasting is developed. A case study analysis shows that the proposed method can effectively protect user privacy and improve model prediction accuracy when dealing with heterogeneous data. The framework can train load-forecasting models with a fast convergence rate and better prediction performance than current mainstream federated learning algorithms.