Parameter estimation in the presence of model error

JUAN J. RUIZ; MANUEL PULIDO

Albany

Workshop; 5th EnKF Workshop; 2012

Parameter estimation using data assimilation is a promising extension of data assimilation tecniques which takes into account model error and partially correct it. The evaluation of parameter estimation using data assimilation techniques is usually performed under the perfect model assumption (i.e. the only source of model error is associated with the unknown parameters). In this case it is shown, that the techniques are able to estimate the optimal value of model parameters if the model sensitivity to changes in the parameters is aproximately linear. In this work, the impact of other model error sources on the parameter estimation and on analysis quality is explored using a local ensemble transform Kalman filter applied to a simple general circulation model. The forecast skill of the system for different forms of accounting model error in a data assimilation cycle is compared. The spread-skill relationship is also examined under the perfec and the imperfect model assumptions. It is shown that under the perfect model assumption the estimated parameter converge to the true parameter values (i.e. the values of the parameters used for the generation of the true system evolution). In this case, the analysis error is reduced and the forecast skill as well as the error spread relationships are improved. Under the imperfect model assumption the parameters do not converge to the true parameters, however the analysis error is reduced and the forecast skill is improved. It is also shown that parameter estimation can be easly combined with other techniques designed to account for model error and that the inclusion of the on-line estimation of some model parameters can provide an additional improvement of the analysis and the forecast.