Since their introduction, Physics-guided Neural Networks (PGNN), a novel class of hybrid models, have already been successfully implemented in several domains of application. As a result, both synergetic effects as well as physically sound models were obtained. Within the context of this thesis, for the first time, the potential of PGNNs for the identification of dynamic systems is investigated from a control engineering point of view. Various approaches are identified and first investigations are performed. By combining a recurrent neural network with a physical dynamics model, a Physics-guided Recurrent Neural Network (PGRNN) is constructed. It is demonstrated that the PGRNN generally outperforms a purely data-driven approach by a substantial margin. Further investigations indicate that the quality of the introduced dynamics model is merely of minor importance to the resulting performance benefits. Consecutively, the PGRNN is augmented by a physics-based constraint, inciting energy conserving solutions as well as a function library of nonlinear terms. The latter resulted in the full compensation of prediction error deficiencies due to inaccurate dynamic models.

Full Text (in German)

To be uploaded