Transformation Method of Nonlinear Mathematical Models of the DC Series Drive into the Form of Modified Recurrent Neural Network

Abstract

The method of transformation of a nonlinear mathematical model of an electromechanical object to the form of a modified artificial recurrent neural network has been further developed. The method makes it possible to use knowledge about the object for the synthesis of a recurrent neural network (RNN) structure and the computation of their coefficients. Nonlinearities in the proposed RNN were realized by expanding the input signal space of the network, using the normalized signals of polynomial terms. Mathematical transformations were performed for a model of thyristor-based electric drive with a dc motor of series excitation. In the electric drive model, different nonlinearities were set, namely, the magnetic flux and inductance of the motor winding dependence on the motor current and its derivative, the thyristor converter gain from the reference voltage, and the dependence of the moment of inertia on the speed. An accuracy estimation for the models in the form of an RNN was made.