Safe Reinforcement Learning-Based Control in Power Electronic Systems

Data-driven approaches such as reinforcement learning (RL) allow a controller design without a priori system knowledge with minimal human effort as well as seamless self-adaptation to varying system characteristics. However, RL does not inherently consider input and state constraints, i.e., satisfying safety-relevant system limits during training and test. This is challenging in power electronic systems where it is necessary to avoid overcurrents and overvoltages. To overcome this issue, a standard RL algorithm is extended by a combination of constrained optimal control and online model identification to ensure safety during and after the learning process. In an exemplary three-level voltage source inverter for islanded electrical power grid application, it is shown that the approach does not only significantly improves safety but also improves the overall learning-based control performance.