Nonlinear system identification with applications to space weather prediction.

Abstract

<italic>System identification</italic> is the process of constructing empirical mathematical models of dynamcal systems using measured data. Since data represents a key link between mathematical principles and physical processes, system identification is an important research area that can benefit all disciplines. In this dissertation, we develop identification methods for Hammerstein-Wiener models, which are model structures based on the interconnection of linear dynamics and static nonlinearities. These identification methods identify models in state-space form and use known basis functions to represent the unknown nonlinear maps. Next, we use these methods to identify periodically-switching Hammerstein-Wiener models for predicting magnetic-field fluctuations on the surface of the Earth, 30 to 90 minutes into the future. These magnetic-field fluctuations caused by the solar wind (ejections of charged plasma from the surface of the Sun) can damage critical systems aboard satellites and drive currents in power grids that can overwhelm and damage transformers. By predicting magnetic-field fluctuations on the Earth, we obtain advance warning of future disturbances. Furthermore, to predict solar wind conditions 27 days in advance, we use solar wind measurements and image measurements to construct nonlinear time-series models. We propose a class of radial basis functions to represent the nonlinear maps, which have fewer parameters that need to be tuned by the user. Additionally, we develop an identification algorithm that simultaneously identifies the state space matrices of an unknown model and reconstructs the unknown input, using output measurements and known inputs. For this purpose, we formulate the concept of input and state observability, that is, conditions under which both the unknown input and initial state of a known model can be determined from output measurements. We provide necessary and sufficient conditions for input and state observability in discrete-time systems.