Abstract

This article presents the design and functional validation of deep neural network-based approximators for the control policy of constrained model predictive control applied to a distributed energy resource unit operating in grid-supporting mode within an AC microgrid. The control scheme follows a conventional cascaded architecture, consisting of a zero-level control loop that regulates the DER output voltage, and a primary control loop responsible for balancing power supply and demand. Simulations show that the control policy approximated using neural networks achieves functional performance equivalent to a conventional implicit formulation of MPC. Moreover, the execution time of neural networks is expected to scale better to high-dimensional optimization problems compared to conventional iterative solvers. Experimental validation was carried out on a lab-scale plant prototype with controllers executed on dSPACE MicroLabBox platform. To meet execution-time requirements for a target control interval of $200\mu $ s, the conventional implicit formulation requires simplifying the constraints, which restrict the maximum actuation voltage, and adopting a short prediction horizon. In contrast, the neural-network-based controller enables the use of more complex constraint sets and a longer prediction horizon, thereby maximizing the utilization of the variable inverter voltage.