The rapid evolution in decentralized grids increases the demand for new control methods with higher performance regarding response time, stability and the ability to operate under the presence of safety-relevant constraints. In this contribution, two recent optimal control methods with strongly different operating principles are investigated based on a head-to-head comparison with focus on a general proof of concept and realtime capability. On the one hand, part of the comparison is a continuous control set model predictive control with constraints solved by a quadratic program solver, that relies on a very precise model description of the inverter. On the other hand, there is a reinforcement learning controller with safeguard and continuous control set model, which is, in contrast, purely data-driven and has no a priori knowledge about the underlying control plant. The comparison is implemented for a three-phase, three-level inverter with an LC-filter and a common neutral point. Both control methods operate in dq0-coordinates in order to take advantage of the rotating reference frame and to be able to control DC-offsets. Although both controllers are conceptually contrary, they showed almost identical performance within a real-time hardware-in-the-loop validation delivering rise times of less than 1 ms and a mean absolute control error of less than 0.8 V considering representative application scenarios.

Optimal control approaches, such as model predictive (MPC) and reinforcement learning (RL) control, are promising tools in decentralized grid applications, but face various challenges, e.g., in the presence of model inaccuracy or with concern to safety-related limitations. While contemporary research articles mainly cover simulation-based investigations, this contribution presents the application and head-to-head comparison of both control algorithms on commercial inverter hardware, with the topology being a three-phase, three-level inverter with an LC-filter and a common neutral point. Herein, the safety-shielded RL controller is trained from scratch utilizing an edge learning toolchain while simultaneously interacting with the hardware. Both control approaches benefit from system identification, allowing to decouple achievable controller and safeguard performance from potentially unreliable or incomplete a priori information. During transients, both methods exhibit almost identically fast settling characteristics when operated on the same safety constraints. In steady state, MPC achieves a normalized mean absolute error (w.r.t. the reference voltage amplitude) of 0.33 %, with the RL controller even reaching 0.18 %. Lastly, it is empirically demonstrated that the RL controller manages to compensate inverter non-linearities, which contrary reduce the MPC's tracking performance noticeably in the partial load range if left unchecked.