Physics-Informed Neural Network for Parameter Identification: a Buck Converter Case Study

Abstract

System-level condition monitoring methods estimate the electrical parameters of multiple components in a converter to assess their health status. The estimation accuracy and variation can differ significantly across parameters. For instance, inductance estimations are generally more accurate and stable than inductor resistance in a buck converter. However, these performance differences remain to be analyzed with a more systematic approach otherwise the condition monitoring results can be unreliable. Therefore, this paper analyzes the training loss landscape against multiple parameters of a buck converter to provide a systematic explanation of different performances. If the training loss is high and smooth, the estimated circuit parameter typically is accurate and has low variation. Furthermore, a novel physics-informed neural network (PINN) is proposed, offering faster convergence and lower computation requirements compared to an existing PINN method. The proposed method is validated through simulations, where the loss landscape identifies the unreliable parameter estimations, and the PINN can estimate the remaining parameters.