Boosting Communication Efficiency in Federated Learning for Multiagent-Based Multimicrogrid Energy Management.

Area of Science:

• Energy Systems
• Artificial Intelligence
• Cybersecurity

Background:

• User privacy is critical for efficient multi-agent energy management in multi-microgrids (MMGs).
• Federated learning (FL) offers privacy protection but faces communication burdens and convergence inconsistencies.
• Layer convergence speed differences in FL can degrade overall agent performance.

Purpose of the Study:

• To propose a communication-efficient federated learning (CEFL) algorithm for MMG energy management.
• To address communication burdens and layer inconsistency problems in existing FL approaches.
• To enhance the performance and convergence rate of agents in MMG systems.

Main Methods:

• Developed a layer evaluation (LE) mechanism using Shapley value (SV) to identify and upload only high-contribution layers.
• Implemented a communication-efficient parameter aggregation method with an aggregation model (AM) for global model (GM) updates.
• Validated the CEFL algorithm through numerical analysis of MMGs with 3-8 microgrids.

Main Results:

• The proposed CEFL algorithm significantly reduces communication overhead compared to standard FL.
• CEFL demonstrates improved convergence rates and overall agent performance.
• Experiments confirmed performance enhancements against four state-of-the-art algorithms.

Conclusions:

• The CEFL algorithm effectively balances privacy protection with communication efficiency in MMG energy management.
• The layer evaluation and aggregation model mechanisms are key to improving FL performance in this context.
• CEFL offers a promising solution for scalable and efficient decentralized energy management systems.