A Localised Markov Chain Particle Filter (LMCPF) for the Global Weather Prediction Model ICON
A localised adaptive particle filter (LAPF) has been introduced into DWD’s experimental data assimilation environment for evaluation and comparison with the operational LETKF. The evaluation (compare Potthast et al 2018) has shown that the LAPF can be run stably over months and is functional. However, the analysis of the method by construction also shows some difficulties which it has when the observations are outside of the convex hull of the ensemble members.
Here, we introduce the key features of a localised Markov Chain particle filter (LMCPF), which resolves major issues of the LAPF. The basic idea of the LMCPF is to consider each particle as a representative of some local distribution, where the width of the distribution is related to the forecast uncertainty. A part of this uncertainty is captured by the ensemble spread, but an important further part needs to be included by additional tools. Here, we employ an estimate of the forecast uncertainty by Gaussian uncertainty in ensemble space which is attributed to each individual particle. Then, exact Bayesian steps are carried out in ensemble space on the resulting Gaussian mixture.
Experimental Results for the LMCPF are shown which demonstrate the strong improvements with respect to the LAPF.
Ms. Anne Sophie Walter German Meteorological Service (DWD)