[9-2]

The impact of different model error covariance matrices on the performance of a particle filter applied to a coupled ocean-atmosphere climate model

P Browne (Reading) , PJ van Leeuwen (Reading)

It is a widely held assumption that particle filters are not applicable in high-dimensional systems due to filter degeneracy, commonly called the "curse of dimensionality". However the equivalent weights particle filter (EWPF) has been shown to perform particularly well on systems of dimension up to 2^16 = 6.5 x 10^4 without suffering filter degeneracy. In this talk we will present the use of the EWPF in twin experiments with the global climate model HadCM3 (with state dimension 2.3 x 10^6), and show that it continues to perform well in such a high dimensional system.

The EWPF makes use of the model error covariance matrix Q, which for this work we have chosen a priori based on physical assumptions of its properties. We shall show that the specific choice of Q impacts on the performance of the particle filter, and that the more physically realistic Q that we choose, the better the performance. This motivates the need for methods to systematically approximate Q for use in operational situations. We shall discuss the impact of various free tuning parameters on the performance of the EWPF.