Title:Probabilistic Reduced-Dimensional Vector Autoregressive Modeling for Dynamics Prediction and Reconstruction with Oblique Projections

Abstract:In this paper, we propose a probabilistic reduced-dimensional vector autoregressive (PredVAR) model with oblique projections. This model partitions the measurement space into a dynamic subspace and a static subspace that do not need to be orthogonal. The partition allows us to apply an oblique projection to extract dynamic latent variables (DLVs) from high-dimensional data with maximized predictability. We develop an alternating iterative PredVAR algorithm that exploits the interaction between updating the latent VAR dynamics and estimating the oblique projection, using expectation maximization (EM) and a statistical constraint. In addition, the noise covariance matrices are estimated as a natural outcome of the EM method. A simulation case study of the nonlinear Lorenz oscillation system illustrates the advantages of the proposed approach over two alternatives.