Coupling of Recurrent and Static Neural Network Approaches for Improved Multi-step Ahead Time Series Prediction

Abstract

A novel nonlinear system identification approach is presented based on the coupling of a neuro-fuzzy model (NFM) with a multilayer perceptron (MLP) neural network. Therefore, the recurrent NFM is employed for multi-step ahead predictions, whereas the MLP is subsequently used to perform a nonlinear quasi-static correction of the obtained time-series output. In the present work, the proposed method is applied as a reduced-order modeling (ROM) technique to lower the effort of unsteady motion-induced computational fluid dynamics (CFD) simulations, although it could be utilized generally for any nonlinear system identification task. For demonstration purposes, the NLR 7301 airfoil is investigated at transonic flow conditions, while the pitch and plunge degrees of freedom are simultaneously excited. In addition, the sequential model training process as well as the model application is presented. It is shown that the essential aerodynamic characteristics are accurately reproduced by the novel ROM in comparison to the full-order CFD reference solution. Moreover, by examining the results of the NFM without MLP correction it is indicated that the new approach leads to an increased fidelity regarding nonlinear ROM-based simulations.