Can reinforcement learning revolutionize control technology? The Cognitive Train/Test System for Reinforcement Learning using Labs (CTRL) project sets up an infrastructure for independent control in the distribution grid.
The project offers an infrastructure for training control concepts for distributed plants. Agents are prepared for their areas of application from simulated pretraining and realistic hardware-in-the-loop methods to field tests. As a result, an agent independently maintains voltage stability in a local grid transformer.
The project is interesting to:
Power plant manufacturers
- Designing a tool chain for the development, testing, and establishment of a reinforcement learning (RL) agent for control concepts in energy system technology
- Developing a validation plan for offline training, RCP optimization, and testing, hardware-in-the-loop optimization and validation as well as field testing
- Integrating adaptive black-box approaches and flexible communication links between the reinforcement learning agent, the controlled test system, and the energy system
- Testing the tool chain by developing, testing, and establishing an RL agent for the intelligent control of local grid transformers
+49 (0) 561 7294-103
Control applications are usually based on a heterogenous method, i.e. a combination of complex systems that are partly independent of each other. The methods used in the project are particularly suitable for working with this complex system.
The following steps were taken to identify a suitable RL agent:
- Offline training: Developing an RL agent (type, size, parameterization)
- RCP training: Mapping real interfaces and real-time performance
- HIL training/testing: Testing and optimization in realistic application
- Field validation: Testing in the realistic area of application
Project participants, Fraunhofer IEE
Dr. Ron Brandl, Juan Montoya