Reinforcement Learning improves the energy efficiency of battery storage systems

The KI-M-Bat project has developed a reinforcement learning-based control strategy and successfully deployed it on a real-world battery storage system.

As part of the KI-M-Bat project, the Smart Energy Systems research group at Kempten University of Applied Sciences has successfully deployed a reinforcement learning (RL) –based optimal power split control on the FENECON multi-string Industry-M battery system.

The RL agent was trained using a fully data-driven battery digital twin developed by the Chair for Electrical Energy Storage, Technische Universität München, enabling robust and realistic learning before field deployment. The RL agent interacts with the industry BESS via OpenEMS OpenEMS Association e.V., ensuring seamless integration in a real-world environment.

Our RL approach, aimed at optimizing round trip efficiency during real-time control, delivers up to 3% improvement in efficiency compared to the existing industry "Near2Equal" droop control strategy — a promising step toward smarter and more efficient multi-string battery energy storage operation.

We thank our industry partners at FENECON and STABL Energy, Professor Dr. Holger Hesse and fellow researchers Vivek Teja Tanjavooru, Martin Cornejo and Dr. Prashant Pant for closely coordinating this deployment. We gratefully acknowledge Bayerische Forschungsstiftung (BFS) for making this research possible through the funded project KI-M-Bat.