A Hierarchical Personalized Attention Mechanism for Federated Learning in Smart Grid Load Forecasting

In recent years, smart grids have increasingly be-come the lifeline of national economies. Faced with intermittent power generation and dynamic demand patterns, accurate load forecasting plays a critical role in power systems. Federated learning, with its benefits in privacy protection, has been grad-ually adopted in the field of load forecasting. However, the heterogeneity of data from different power users severely impacts the performance of the model. To achieve high accuracy in load forecasting and ensure data privacy protection, this paper proposes a personalized federated learning model, FedDuM, based on the attention mechanism and a two-layer structure. Firstly, the attention mechanism is used to capture the feature correlation among high-dimensional power load data. Secondly, a two-layer local model is designed on the client side, preserving a unique personalized attention layer for each client and using the other layers for global model aggregation, realizing the design of personalized modules. Finally, through simulation experiments, it is demonstrated that the proposed FedDuM model outperforms other federated learning models in the performance of smart grid load prediction, significantly improving the convergence speed while maintaining model accuracy.