Scholarlink

Short-term load forecasting is essential for the construction of smart grids and optimizing power dispatch. This paper innovatively proposes a federated learning framework that fuses dynamic Fisher personalization with adaptive constraints to address the three major challenges faced by traditional methods: the contradiction between privacy and utility, the rigidity of personalization strategy, and the difficulty of convergence due to differential privacy. This scheme intelligently screens parameters through layer-by-layer Fisher information and combines it with adaptive constraints to enhance noise resistance. It significantly improves prediction performance while guaranteeing privacy. Experiments show that, compared to the FedAvg+DP benchmark method, this scheme reduces RMSE, MAE, MSE, and MAPE by 38.72%, 39.77%, 63.96%, and 41.14%, respectively, while striking an optimal balance between privacy preservation and prediction precision.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DP-DPFL: Short Term Load Forecasting Based on Differential Privacy and Dynamic Personalized Federated Learning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Shuhui Zhang,
  • Abiao Yuan,
  • Lianhai Wang,
  • Shujiang Xu,
  • Wei Shao,
  • Qizheng Wang

摘要

Short-term load forecasting is essential for the construction of smart grids and optimizing power dispatch. This paper innovatively proposes a federated learning framework that fuses dynamic Fisher personalization with adaptive constraints to address the three major challenges faced by traditional methods: the contradiction between privacy and utility, the rigidity of personalization strategy, and the difficulty of convergence due to differential privacy. This scheme intelligently screens parameters through layer-by-layer Fisher information and combines it with adaptive constraints to enhance noise resistance. It significantly improves prediction performance while guaranteeing privacy. Experiments show that, compared to the FedAvg+DP benchmark method, this scheme reduces RMSE, MAE, MSE, and MAPE by 38.72%, 39.77%, 63.96%, and 41.14%, respectively, while striking an optimal balance between privacy preservation and prediction precision.