Decentralized Voltage Regulation in Low-Voltage Grids Using Projection-Constrained Multi-Agent Deep Reinforcement Learning : A Case Study in E.ON’s Swedish Distribution Grid

This thesis presents a decentralized inverter-based voltage control strategy for low-voltage (LV) power grids with high photovoltaic (PV) penetration, using Multi-Agent Deep Reinforcement Learning (MADRL). Inspired by recent advancements in safe reinforcement learning, the pro-posed method integrates a projection layer into a MADRL framework to ensure that smart inverter actions respect physical and regulatory constraints. A real-world LV grid operated by E.ON in southern Sweden serves as the case study. The model is implemented in Python us-ing open-source libraries and trained on real consumption and simulated PV data. Simulation results show that the method decreases voltage deviations from the reference-voltage, but it increases the power losses in the grid, and outperforms simpler voltage control methods, evenunder high R/X-ratio conditions and varied PV-penetration levels. The study also discusses the technical, legal, and economic feasibility of deploying such AI-driven control in practice.