Reproducible Reinforcement Learning on a Real-World Power Grid Control Problem

Type: Project

Status: running

Date: March 10, 2026 - June 10, 2026

Supervisors: Julian Oelhaf, Andreas Maier

Reinforcement Learning for Control and Protection in Power Distribution Systems

This project studies reinforcement learning (RL) methods for control and coordination tasks in power distribution systems, with a focus on reproducible, paper-driven implementation. A representative RL approach from the literature is re-implemented, including the design of state/action representations, reward formulation, and constraint handling.

The implementation is systematically validated on a controlled benchmark to ensure correctness, reproducibility, and consistent evaluation. Subsequently, the approach is assessed on simulation data from a realistic medium-voltage grid model to analyze performance under practical conditions.

Depending on progress, extensions towards more advanced RL formulations and protection-related applications are considered. The work emphasizes structured evaluation and reproducibility in safety-critical learning-based control.