Parameter Estimation of Power Electronics by Forward and Backward Solutions

A detailed knowledge of system parameters is crucial in modern industrial power electronics. Applications ranging from condition monitoring and control to data-driven modeling, in short, a reliable power electronics application. In this work, we present the method of Automatic Differentiation Shooting (ADShooting) and give the construction of a lightweight Physics-Informed Neural Network (PINN) allowing non-invasive extraction of power electronics parameters. In particular, we present dedicated training and noise handling strategies, both being major aspects of the given application. Exemplary, we apply both approaches for power electronics parameter estimation of a DCDC buck converter. In particular, we investigate the dependence of sampling rates, as well as, the influence of noise. As the execution time is a limiting factor in any online application, we investigate the algorithm’s runtime in a continuous operation, indicating possible detectable physical effects and application scenarios.