University of California, Berkeley
Thursday, December 6, 2007
Mesa Laboratory, Chapman Room
Lecture 3:30pm (Refreshments at 3:15pm)
Particle Filters and Their Potential Use in Numerical Weather Forecasting
Numerical weather forecasting can be viewed as a statistical prediction problem in a high-dimensional state-space model. Variants of the Kalman filter have been used with some success (Ott et al 2004; Anderson 2007), despite the failure of Gaussianity and linearity in these models.
In the present work we try to adapt particle filters, a non-parametric data assimilation method. Although, as shown in (Bengtsson et al, 2007; Snyder et al.), particle filters tend to degenerate in high-dimensional situations, we intend to use these methods locally and then combine the local state variables according to their spatial relationship, in an appropriate fashion. We will show some preliminary experimental results on a 40-dimensional system