Differentiable Predictive Control for Power Electronic Systems

This letter presents the first implementation of differentiable predictive control (DPC) in power electronic systems. By embedding the predictive control cost function directly into a differentiable neural network policy, the proposed approach eliminates the need for labeled control data and online optimization solvers. The control policy is trained solely from system trajectories and deployed on a shallow network suitable for an embedded execution. Experimental validation demonstrates that DPC achieves an accurate reference tracking, strong generalization under unseen conditions, and significantly lower computational cost compared to conventional model predictive control, particularly for extended prediction horizons. These results highlight the potential of DPC as a data-efficient and computationally scalable alternative for advanced power converter control.