A Privacy-Preserving Decentralized Partition-Autonomous Planning Method for Interconnected Multi-Microgrid Distribution Systems Driven by Federated Proximal Policy Optimization

INTRODUCTION: The rapid integration of distributed renewable energy and ICT infrastructure has made interconnected multi-microgrid distribution systems essential for low-carbon power grids, yet centralized coordination exposes sensitive operational data to privacy risks, including eavesdropping and membership inference attacks under 5G and edge computing environments. OBJECTIVES: This study resolves the tension between data privacy protection and decentralized collaborative planning by developing a unified federated reinforcement learning framework with formal privacy guarantees. METHODS: A privacy-preserving decentralized partition-autonomous planning method driven by federated proximal policy optimization is proposed. The network is divided into autonomous units, each performing local proximal policy optimization on local observations. A proximal term constrains update steps against Non-IID drift. A time-period adaptive differential privacy budget mechanism dynamically adjusts noise according to gradient sensitivity, complemented by a Shamir secret sharing-based secure aggregation protocol forming dual-layer protection. RESULTS: Experiments on the Baran-Wu 33-node test system show daily operational cost within 3.0% of the centralized benchmark, per round communication reduced by approximately 80% to 9.6 MB, voltage compliance at 98.4%, and membership inference attack AUC suppressed to 0.521 DISCUSSION: The adaptive privacy budget reduces utility loss by 3.2 percentage points under strict privacy constraints versus fixed schemes, while secure aggregation maintains reliability under one PAU dropout with only 0.38% deviation.  CONCLUSION: The proposed framework provides a feasible technical path for secure collaborative planning of large-scale interconnected multi-microgrid systems under practical ICT constraints.