Economic autonomous operation of microgrids on islands with renewable energy interactions: A satellite-aided over-the-air federated reinforcement learning approach

For microgrids (MGs) in pelagic islands, the economic autonomous operations considering renewable energy interactions across different islands can be achieved through the available communication and computation support of the low Earth orbit satellites (LEOSats). However, the limited onboard resources of LEOSats and unstable satellite-terrestrial channel conditions require resource and parameter optimization during communication to ensure reliable data exchange and the effective implementation of MG scheduling methods. Therefore, this paper proposes a LEOSat-aided over-the-air computation (OAC) federated reinforcement learning (FRL) approach with two key advantages. First, the FRL approach leverages LEOSats as the central server for coordinated global updates and distributed training across edge nodes to enhance the economy of autonomous MG operations considering inter-island renewable energy interactions, while using LEOSats in a lightweight and temporary manner that avoids long-term resource occupation. Second, by integrating dynamic adaptive resource allocation, parameter optimization, and OAC into the communication process, the proposed FRL approach achieves convergence with satisfactory performance subject to the limited communication and computation capacities of LEOSats. The case study demonstrates that optimizing LEOSat communication parameters can effectively ensure FRL training performance. The proposed FRL-based approach is more suitable for achieving economic autonomous operation of MGs, particularly in pelagic island settings by enabling secure data sharing through LEOSats compared with alternative methods.