A comprehensive review of reinforcement learning-based voltage control in smart grids

Voltage control is a critical challenge in modern smart grids, especially with the increasing integration of distributed energy resources, electric vehicles, and variable renewables. Reinforcement learning (RL), as a data-driven and adaptive control paradigm, has gained significant traction for real-time voltage regulation. This review presents a comprehensive and systematic survey of RL-based voltage control strategies, covering value-based, policy-based, model-assisted, meta-learning, multi-agent, and hierarchical approaches, with a focus on studies published between 2017 and 2025. The novelty of this work lies in its unified evaluation framework, which benchmarks RL methods not only by their control accuracy and training efficiency, but also by their scalability, generalization, safety guarantees, and suitability for deployment in realistic grid environments. Unlike prior reviews that focus narrowly on control performance or algorithm categories, this work integrates recent advances such as physics-informed learning, risk-aware RL, and federated learning into a broader discussion of real-world applicability. The analysis shows that modern RL controllers can regulate voltage within 1–2 % of nominal values, scale to networks with over 100 buses, and adapt to disturbances using relatively few training episodes. Model-based and off-policy strategies improve sample efficiency, while multi-agent and hierarchical designs enable decentralized scalability. We also address practical considerations, including communication overhead, cyber-resilience, and control architecture integration, to bridge the gap between academic research and operational deployment.