Methodology to improve the quality of neural network modeling of dynamic objects

Annotation: The problem of neural network modeling of nonlinear dynamic objects using recurrent neural networks is considered. An ap-proach to improve the accuracy of modeling using a static neu-ral network of the «multilayer perceptron» type, that processes correlation dependencies of a dynamic process and approxi-mates the modeling error, is proposed. A technique for synthe-sis and application of the correlation neural network model CCF-MLP improving the quality of modeling of a conventional recurrent neural network, is formulated. Simulation experiments are carried out with a neural network recurrent network of the GRU type, that models the behavior of a nonlinear dynamic object, as well as GRU with the proposed CCF-MLP model. The improvement in the quality of modeling (RMSE, MAPE) is confirmed in the case of using CCF-MLP both in the presence and absence of noise in the observed data. The practical ap-plicability of the proposed method was tested on a real liquid level control system.

Keywords: nonlinear dynamic ob-ject, dynamic object modeling, recurrent neu-ral network, multilayer perceptron, cross-correlation function