The aim of this master thesis is to investigate enhanced safety measures for reinforcement learning (RL) in grid control, with a focus on addressing the challenges posed by complex solution spaces (e.g., non-convex solution space). In grid systems, complex physical and operational constraints often result in search space, where standard RL approaches might inadvertently select actions that compromise safety. This thesis will explore various approximation techniques (e.g., convex approximation), projection methods, and learning-based approaches to better capture the true boundaries of these constraints.

The proposed methods will be tested with grid scenarios to assess their effects on safety, performance, and overall efficiency. The insights gained from this evaluation are expected to contribute to safer and more effective RL applications in grid management.

This Thesis will be done in cooperation with the OFFIS Insitute. I will forward your applications to the colleague working there who will supervise your thesis. You can also reach him directly by writing an e-mail to sharaf.aldin.alsharif@offis.de

• Identify existing measures for power grid safe operation with focus on complex solution spaces.
• Explore various approximation techniques (e.g., convex approximations), projection methods, and learning-based approaches to accurately capture constraint boundaries.
• Test the proposed methods using grid scenarios and compare their effects on safety, performance, and overall efficiency.

Proficiency in programming

Helpful knowledge, but not essential:

• Familiarity with power systems
• Familiarity with Python
• Basic knowledge of reinforcement learning
• Experience with simulation and modeling of complex systems