Adaptive intelligence in smart microgrids: a systematic review and comparative analysis of RL-based control, communication, and simulation frameworks

Abstract

Abstract This article is framed as a systematic review with original illustrative benchmarking of reinforcement learning (RL) for smart microgrid (SMG) control, communication, and simulation-to-real deployment. Although prior surveys have reviewed RL for microgrid energy management, they have not jointly analyzed control-communication co-design, adversarial robustness, federated and multi-agent coordination, and Sim2Real validation under one reproducible review protocol. To address this gap, we synthesize more than 160 peer-reviewed studies published between 2015 and 2025 and organize the literature through an explicit taxonomy covering RL family, control objective, communication dependency, validation level, cyber-resilience mechanism, and deployment maturity. The review compares algorithms, protocols, simulation platforms, and validation practices across scalability, latency sensitivity, robustness, privacy preservation, convergence behavior, and real-world readiness. In addition to the systematic synthesis, we report an illustrative benchmark using SUMO-RL-derived demand traces coupled to pymgrid to compare representative RL controllers under normalized control and communication metrics. The analysis shows that policy-gradient and multi-agent variants often provide stronger adaptability under dynamic operating conditions, but no single algorithm dominates across cost, latency, resilience, and deployment complexity. The paper concludes by identifying reproducibility, communication-aware validation, safety-constrained RL, privacy-preserving aggregation, and high-fidelity cyber–physical digital twins as the main barriers to trustworthy RL deployment in SMGs.