مشخصات پژوهش

The importance of accurate load forecasting in modern smart grids has grown significantly for planning and operational purposes. Furthermore, the unpredictability of residential energy consumption adds to the complexity of the forecasting task. Smart meters and deep learning models enable more precise predictions than conventional methods. However, privacy concerns arise due to the need to transmit data. As an alternative, a new learning approach known as federated learning has emerged, enabling the training of a global model without sharing raw meter data. Nevertheless, federated learning methods face challenges such as dealing with heterogeneous datasets and potential poisoning attacks. Therefore, this paper proposes a personalized federated learning load forecasting method incorporating a distance defense mechanism to tackle these issues. This paper presents extensive experiments with real-world data and compares the proposed method with centralized and federated averaging (FedAvg) approaches. The results demonstrate that the proposed framework achieves a 16 % improvement with only a 16 % higher cost, outperforming FedAvg and requiring half the computation time of Ditto, while maintaining strong accuracy and robustness.