Spatio-Temporal Graph Transformer MARL for Resilient Point of Common Coupling Control in Networked Grid-Tied Micro-grids

The increasing penetration of renewable energy sources introduces significant variability and uncertainty into grid-connected microgrids (MGs), affecting voltage stability, frequency regulation, power quality, fault ride-through capability, and operational resilience at the Point of Common Coupling (PCC). These challenges become more pronounced in networked MGs with changing electrical and communication topologies, heterogeneous operating conditions, communication constraints, and variable loads. This study proposes an adaptive, distributed control framework that integrates a Spatio-Temporal Graph Transformer (ST-GT) for learning evolving network relationships, cooperative Multi-Agent Reinforcement Learning (MARL) for decentralized control with physics-informed constraints, and differential-privacy-enhanced federated selective sharing for secure collaboration among MGs. The framework is evaluated on modified IEEE 33-bus and IEEE 69-bus networked MG systems using high-fidelity digital twins implemented in pandapower and pymgrid. Experiments consider renewable penetration of 50% to 80%, short-circuit ratios below 3, communication failures, forecast uncertainty, islanding, and fault ride-through disturbances. Compared with benchmark control approaches, the proposed framework reduces PCC voltage deviations and total harmonic distortion by 26% to 33%, achieves fault ride-through success rates above 96%, improves economic performance by 17% to 23%, and maintains robust operation under forecast errors exceeding 40%. These findings demonstrate that the framework provides a scalable, resilient, and privacy-preserving approach for coordinated control of networked grid-connected MGs.