PNNL

Abstract

A data-driven method is developed to approximate a nonlinear time-varying system (NTVS) by a linear time-varying system (LTVS) based on Koopman Operator and deep neural networks. The approximation error in system states of the proposed method is analyzed and quantified. It is further shown by simulation on a simple yet representative NTVS that the resulted LTVS approximates the original system with reasonable accuracy. Optimal linear controller can be readily developed based on the proposed deep Koopman learning. Simulations on a cartpole showed that the controller worked well on the original NTVS.