Artificial recurrent model for parameter identification of dynamic systems

Abstract

AbstractThe process of determining parameter values causing a specific system response usually relies on the existence of a mathematical model of the system, generally derivable from physical principles. This model may then be used to tune parameter values until the corresponding model output coincides with a desired one. To eliminate the necessity of setting up a mathematical model by causal deduction starting from physical principles, machine learning strategies may be applied. Here, the Long Short‐Term Memory (LSTM) network is the basis for a recurrent surrogate which predicts the solution starting from a specific initial condition being influenced by parameter values under investigation. With the trained model it is then possible to perform parameter identification by adjusting values of the model parameters either by use of classical optimization or machine learning (ML)‐based parameter training until the desired model response is achieved.