MODELS OF DYNAMIC OBJECTS WITH SIGNIFICANT NONLINEARITY BASED ON TIME-DELAY NEURAL NETWORKS

The paper is devoted to modeling nonlinear objects using dynamic neural networks. The aim of the work is to improve the accuracy of modeling dynamic objects with significant nonlinearities using neural network models and to determine the scope of effective application of these models. This aim is achieved using time-delay neural networks. To assess the applicability of the proposed models, the study considers simulation objects with two types of nonlinearities: smooth and piecewise linear (saturation). The investigation of suggested models accuracy in nonlinear dynamic object modeling involves experiments: assessing the models' scalability with different input signals; evaluating their extrapolation capabilities. The findings from these experiments are compared with the modeling results using the compensatory method of deterministic identification based on functional series. The outcomes of the simulation experiments reveal that the suggested models lack invariance concerning the input signal. Nevertheless, when trained on a comprehensive dataset of representative input signals, neural network models can effectively capture the characteristics of nonlinear dynamic objects. The extrapolation abilities of the proposed models tend to degrade as the input signal amplitudes exceed the range covered by the used training set. The scientific novelty is the establishment of a clear relationship between the types of input signals, their amplitudes, and the accuracy of the proposed models. The practical significance of investigation delineates the areas in which time-delay neural networks can be used to address the real-world challenges associated with significantly non-linear objects; demonstrates increasing the accuracy of identifying nonlinear objects compared to functional series models.