Entropy-Regularized Weighted Federated Meta-Learning Approach for Power Load Forecasting in Non-IID Scenarios

Authors:
Publisher website
Google Scholar
Abstract
Accurate power load forecasting is crucial for smart grid security and efficiency, yet traditional centralized methods fall short in addressing data privacy and communication demands. Federated Learning (FL) offers a promising solution but faces dual challenges: the Non-Independent and Identically Distributed (Non-IID) nature of load data limits model generalization and invites security threats like model poisoning attacks. To address these challenges, this paper proposes a Federated Meta-Learning (FedML) framework integrated with Entropy-Regularized Gradient-Weighted (ER-Grad) aggregation. The framework employs FedML to construct high-precision personalized forecasting models for clients while maintaining training efficiency, thereby significantly enhancing predictive performance. Additionally, the ER-Grad aggregation algorithm is introduced to dynamically calibrate client weights, effectively suppressing abnormal or malicious updates and strengthening the system’s robustness against model poisoning attacks. Experimental results demonstrate that the proposed method substantially improves overall and peak forecasting accuracy, ensures comprehensive robustness against malicious attacks and communication dropouts, and achieves training with high efficiency.
Funding Information
  • National Natural Science Foundation of China (62473204)
  • Natural Science Foundation of Nanjing University of Posts and Telecommunications (NY222144)
  • Huali Program for Excellent Talents in Nanjing University of Posts and Telecommunications

This publication has 33 references indexed in Scilit: