Communication-robust and Privacy-safe Distributed Estimation for Heterogeneous Community-level Behind-the-meter Solar Power Generation

TL;DR

This paper introduces a novel federated learning approach for estimating behind-the-meter solar power generation that is robust to communication failures, heterogeneity, and malicious privacy attacks, ensuring accurate and privacy-safe community-level predictions.

Contribution

It proposes a multi-task federated learning framework with an updated parameters estimation method and dynamic differential privacy, addressing key challenges in distributed solar power estimation.

Findings

Outperforms traditional FL and localized methods in accuracy and convergence.

Enhances privacy protection against malicious attacks.

Effectively mitigates communication failures and heterogeneity issues.

Abstract

The rapid growth of behind-the-meter (BTM) solar power generation systems presents challenges for distribution system planning and scheduling due to invisible solar power generation. To address the data leakage problem of centralized machine-learning methods in BTM solar power generation estimation, the federated learning (FL) method has been investigated for its distributed learning capability. However, the conventional FL method has encountered various challenges, including heterogeneity, communication failures, and malicious privacy attacks. To overcome these challenges, this study proposes a communication-robust and privacy-safe distributed estimation method for heterogeneous community-level BTM solar power generation. Specifically, this study adopts multi-task FL as the main structure and learns the common and unique features of all communities. Simultaneously, it embeds an updated parameters estimation method into the multi-task FL, automatically identifies similarities between any two clients, and estimates the updated parameters for unavailable clients to mitigate the negative effects of communication failures. Finally, this study adopts a differential privacy mechanism under the dynamic privacy budget allocation strategy to combat malicious privacy attacks and improve model training efficiency. Case studies show that in the presence of heterogeneity and communication failures, the proposed method exhibits better estimation accuracy and convergence performance as compared with traditional FL and localized learning methods, while providing stronger privacy protection.