PINN-based Dynamical Modeling and State Estimation in Power Inverters

Thermal modeling in the field of control and state estimation of fluid cooled power inverters is a challenging task since pure physical models come with a lot of computational effort and simple lumped-parameter thermal models often lack sufficient precision. Also analytical modeling of nonlinearities, as faced in the case of forced convection in heat exchangers of inverters, is prone to errors. Neural networks, on the other hand, provide a nonlinear modeling approach for regression. This article shows the application of physics-informed neural networks for the task of modeling the thermal dynamics of a 3-phase inverter. After obtaining robustness by training on a large artificial training set in combination with a physically motivated differential equation, transfer learning is used for subsequent training on a limited measurement set. The accuracy and robustness of the physics-informed neural network for control related tasks is shown by validation in state estimation using a second order divided difference filter. Results on measurement data from a test bench indicate a significant improvement in the estimation accuracy in comparison to state of the art modeling methods.