Robustness of Reinforcement Learning-Based Congestion Management in Low-Voltage Grids

This paper presents a decoupled framework for curative congestion management in low-voltage distribution grids. A random-forest pre-classifier detects likely constraint violations, while an actor-critic controller selects curtailment actions. The evaluation uses a real low-voltage grid with synthetic future operating scenarios under sparse observability and limited controllability. With accurate grid parameters, total violation magnitude is reduced by 98.9%, and the result remains nearly unchanged across the tested measurement-noise settings. Grid-parameter mismatch is more difficult, although most violations are still mitigated under the tested assumptions.