Transferability of a discrepancy model for the dynamics of electromagnetic oscillating circuits

Modelling of dynamic systems plays an important role in many engineering disciplines. Two different approaches are physical modelling and data‐driven modelling, both of which have their respective advantages and disadvantages. By combining these two approaches, hybrid models can be created in which the respective disadvantages are mitigated, with discrepancy models being a particular subclass. Here, the basic system behaviour is described physically, that is, in the form of differential equations. Inaccuracies resulting from insufficient modelling or numerics lead to a discrepancy between the measurements and the model, which can be compensated by a data‐driven error correction term. Since discrepancy methods still require a large amount of measurement data, this paper investigates the extent to which a single discrepancy model can be trained for a physical model with additional parameter dependencies without the need for retraining. As an example, a damped electromagnetic oscillating circuit is used. The physical model is realised by a differential equation describing the electric current, considering only inductance and capacitance; dissipation due to resistance is neglected. This creates a discrepancy between measurement and model, which is corrected by a data‐driven model. In the experiments, the inductance and the capacity are varied. It is found that the same data‐driven model can only be used if additional parametric dependencies in the data‐driven term are considered as well.