Sylvain Robert, ETH Zürich, seminar for statistics
Local Ensemble Transform Kalman Particle Filter
CoauthorsHans R. Künsch
Abstract: Ensemble Kalman Filters (EnKFs) are state of the art algorithms for ensemble data assimilation in weather forecasting. Particle Filters (PFs), on the other hand, are a more general class of ensemble methods which can handle fully non-linear and non-Gaussian systems, but are not directly applicable to large-scale data assimilation. The Ensemble Kalman Particle Filter (EnKPF) is a hybrid algorithm which combines the EnKF with the PF in a way that maintains scalability and sample diversity and was shown to perform well on the Lorenz 40 model (Frei and Künsch, Biometrika 100(4),2013)
Two key ideas that allow EnKFs to perform well in large-scale applications are localization and square-root schemes, which are best examplified by the Local Ensemble Transform Kalman Filter (LETKF). We propose a novel algorithm, the Local Ensemble Transform Kalman Particle Filter (LETKPF), which is based on a reformulation of the EnKPF in ensemble space and replaces the previous stochastic update with a square-root scheme. Furthermore, the new algorithm can be localized similarly to the LETKF, making it efficient and applicable in realistic settings.
We present results obtained with the COSMO model in a 10 days period of strong cumulus convection over the swiss domain. The project is an ongoing cooperation with Meteoschweiz and Deutscher Wetterdienst.