A crucial consideration when implementing a limited-area ensemble Kalman filter is how to generate an ensemble of lateral boundary conditions. We propose two classes of methods to generate this boundary condition ensemble: one class that has the correct short-term error characteristics such as from a global EnKF and another class that perturbs around a deterministic estimate of the state using assumed spatial and temporal covariances.

This second class of methods are both flexible and easy to implement when a global ensemble of boundary conditions does not exist. We perform experiments for both an idealized low-dimensional model and limited-area simulations over the United States from the Weather Research and Forecasting (WRF) model using simulated observations under the perfect model assumption. These experiments show that errors are greater than a corresponding global EnKF ensemble near the lateral boundaries, but away from the boundaries the errors are indistinguishable for all methods tested.

Error differences arise from a lack of observations near the lateral boundaries, linear tendencies and boundary spatial covariances. Methods that use assumed spatial covariances on the boundary have larger analysis errors than methods that employ state-dependent estimates because observations are assimilated in a state-independent manner, especially for unobserved fields and in sparse observation networks. These experiments suggest that the lateral boundaries for a limited-area EnKF can be perturbed around a statistical estimate without significant penalty in the domain interior if the lateral boundaries are placed far enough away from the area of interest.

Back to: Gregory Hakim's Home Page.