Different approaches for parameter estimation based on Ensemble Transform Kalman Filter using the Lorenz's model

RUIZ J. J.; M. PULIDO

Foz Iguazu (Brazil)

Conferencia; The Meeting of the Americas, AGU; 2010

American Geophysical Union

In this work a parameter estimation technique based on the Ensemble Transform Kalman Filter (ETKF) is applied for parameter estimation in the 3 variables Lorenz's model. Two main different implementations of this technique were examined: one is the use of a single ensemble in which initial conditions and parameters were perturbed, this will be refered as the simultaneous estimation; the other consist of two ensembles one with perturbations in the model state only and the other with perturbations in the parameters only. The initial conditions ensemble uses the most recent estimation of the parameters for all the members, and the parameter ensemble uses the analysis ensemble mean as the initial condition for the state variables for all the members. This approach will be refered as the parallel estimation. The accuracy and convergence of the different approaches were evaluated using experiments where the true evolution of the system is asumed to be known and consist of a long integration of the Lorenz's model. Observations were generated from the true evolution of the system, at fixed intervals, asuming a normally distributed random observational error. 2000 assimilation cycles where simulated using this observations for the different experiments. Some experiments describe the sensitivity of the estimation accuracy to the ensemble size and parameter ensemble dispersion. It has been found that the ETKF approach can produce an accurate estimation of the parameters involved in the Lorenz's model. From the comparison of the different implementations it is shown that the simultaneous estimation of initial conditions and parameters leads to a more accurate estimation of the state variables and the parameters and also to a faster convergence of the parameters towards their true value. The convergence of the parameters can be accelerated introducing an outer loop within the ETKF which also helps to deal with non-linearities.