With more renewables being integrated into distribution grids, the problem of voltage fluctuation has become prominent. Effective Volt/VAR regulation is a commonly used method to ensure the safe operation of distribution networks. Model-based approaches tend to work well only if detailed network parameters are available, while data-driven approaches can suffer from overfitting and may not generalize well. We created the PHY-GAT-SAC framework to address these issues. Physics-regularized reinforcement learning uses bidirectional graph attention, which combines a physics-informed model with a safety projection method that relies on sensitivity matrices. This makes it so that the voltage regulation is practical, interpretable, and secure. The framework works with two combined branches. One branch takes care of the nonlinear mapping from power injections to voltage states using a forward graph encoder and a reverse consistency constraint. At the same time, another branch extracts features directly from the voltages to improve the perception of system violation risk. The framework has a sensitivity-based safety layer as well. This layer projects every control action into a feasible area formed by linearized voltage restrictions, thus securing operation safety. Experiments on an IEEE 33-node system show that the framework works well. A safety layer guarantees a safe operating range without exact impedance values. And PHY-GAT-SAC greatly lowers voltage violations compared to multi-agent deep reinforcement learning. By successfully combining physics with learning, this study gives a unified framework for merging graph neural networks and reinforcement learning within intricate grid management.
Qin et al. (Mon,) studied this question.