Structured Recurrent Neural Network Model Order Reduction for SISO and SIMO LTI Systems

Abstract

Obtaining accurate and less computational demanding reduced models is a continuous challenge with complex systems. We propose a RNN network structure that can model LTI SISO systems of any order. Using this structured RNN model, a complex system of 598 states is reduced to a 10th order system at 9.04e-6 mean-square-error. SISO 4th order outperformed reported results of other MOR techniques. The RNN network structure is extended to model SIMO LTI of any number of output and any system order. Using this RNN SIMO network, RLC interconnect of 108 states was reduced to a 5th system at 9.1e-4 mean-square-error.