Identification of Dynamic Systems Assisted by an Autoregressive Recurrent Model

The identification of parameters in dynamic systems usually first requires modeling the system as nonlinear differential equations based on physical principles, which then can be evaluated to search for a set of optimal parameters that enable the mathematical model to reproduce some desired behavior of the real system. To overcome the necessity of complex modeling and to possibly reduce the number of extensive experimental evaluations of the real system or numerical evaluations of the differential equations, strategies from machine learning may be applied. The proposed method uses recurrent architectures such as Long Short‐Term Memory (LSTM) networks as core to predict the system development from initial conditions and parameter values. The trainable weights of the model are optimized based on a set of training data containing parameter values and corresponding solution trajectories generated by evaluation of the system to be investigated. The trained model may then be used to identify unknown system parameter values related to a specific solution trajectory by solving an optimization problem for the inputs of the machine learning model.