The rapid integration of distributed energy resources—solar panels, wind turbines, battery systems, and other local generators—has fundamentally altered how distribution networks operate. Unlike traditional grids with centralized generation, modern networks experience constant fluctuations in voltage and power flow as these resources turn on and off throughout the day. Maintaining voltage within safe operating limits has become increasingly difficult.
Deep reinforcement learning (DRL) offers promise for automating voltage regulation. AI agents learn control policies by trial and error in simulated environments, potentially outperforming manual human operation. However, purely autonomous DRL methods suffer from critical weaknesses in power systems: they explore control actions that violate safety constraints, they learn slowly, and their decisions lack transparency—all problematic in grid operations where equipment damage and service disruption are costly.
This research bridges the gap by pairing AI with human expertise. The system uses a Soft Actor-Critic algorithm as its core AI engine while embedding voltage safety constraints directly into the learning process. Critically, human operators provide guidance based on their understanding of network physics: if the AI proposes a suboptimal control action, human feedback corrects it, and the AI learns from these corrections to improve future decisions. Control actions involve coordinating capacitor banks and battery storage systems to stabilize voltage.
Testing on a simulated 13-node distribution feeder showed the approach reduced voltage violations significantly compared to baseline AI methods. Power losses also decreased, indicating more efficient operation. The framework operates in an open-source simulation environment, making it reproducible and accessible.
This hybrid model reflects a pragmatic approach to grid automation: rather than replacing human operators, AI augments their capability by handling rapid, routine decisions while remaining accountable to human oversight. As distribution networks grow more complex with higher renewable penetration, such collaborative frameworks may prove essential for maintaining reliability while leveraging computational advantages of modern AI.



