Prediction of Linear Time-varying Stochastic Systems based on the Subspace Identification Method and its Recursive Algorithm

Abstract

In this paper, a prediction method is newly proposed for time-varying stochastic linear systems in the subspace identification framework. The key to this subspace-based prediction is to regard the change of the extended observability matrix yielded by the time-varying parameters of system as the rotation of the principal vectors that span the basis of the signal subspace. The rotation rate is evaluated from the angle between the past and current signal subspaces, and the future signal subspace is predicted by rotating the current subspace. A recursive algorithm is derived and its efficacy is tested by simulation experiments.