Order model reduction for two-time-scale systems based on neural network estimation

A new order model reduction technique for two-time-scale systems is presented in this paper. This reduction technique provides the advantage of forcing the dominant poles of the original system to be the dominant poles of the reduced order model. The reduction technique is performed based on the two-time-scale system reduction technique, while the dominant eigenvalue preservation is achieved by the implementation of a neural network and the use of the matrix reducibility concept. The eigenvalues of the reduced order model are selected as a subset of the full order model eigenvalues. Simulation and comparison with other techniques for a third order system along with its results are presented as part of this paper.