Summary
This paper describes a method to construct reduced‐order models for high‐dimensional nonlinear systems. It is assumed that the nonlinear system has a collection of equilibrium operating points parameterized by a scheduling variable. First, a reduced‐order linear system is constructed at each equilibrium point using state, input, and output data. This step combines techniques from proper orthogonal decomposition, dynamic mode decomposition, and direct subspace identification. This yields discrete‐time models that are linear from input to output but whose state matrices are functions of the scheduling parameter. Second, a parameter‐varying linearization is used to connect these linear models across the various operating points. The key technical issue in this second step is to ensure the reduced‐order linear parameter‐varying system approximates the nonlinear system even when the operating point changes in time. Copyright © 2016 John Wiley & Sons, Ltd.