Comparing an Ensemble Kalman Filter to a 4DVAR Data Assimilation System in Chaotic Dynamics
This is an incremental comparison of two data assimilation methods on a simple chaotic model, providing limited new insights for practitioners.
The paper compares Ensemble Kalman Filter (EnKF) and 4DVAR data assimilation methods on the Lorenz model. With 10% initial error both methods perform well; with 20% error 4DVAR outperforms EnKF; with 40% error both fail. In a sparse observation scenario, 4DVAR shows perfect tracking while EnKF diverges after 80 time steps.
In this paper, the Ensemble Kalman Filter is compared with a 4DVAR Data Assimilation System in chaotic dynamics. The Lorenz model is chosen for its simplicity in structure and its dynamical similarities with primitive equation models, such as modern numerical weather forecasting. It was examined whether the Ensemble Kalman Filter and 4DVAR are effective in tracking the control for 10%, 20%, and 40% of error in the initial conditions. With 10% of noise, the trajectories of both methods are almost perfect. With 20% of noise, the differences between the simulated trajectories and the observations, as well as the true trajectories, are rather small for the Ensemble Kalman Filter but almost perfect for 4DVAR. However, the differences become increasingly significant at the later part of the integration period for the Ensemble Kalman Filter, due to the chaotic behavior of the system. For the case with 40% error in the initial conditions, neither the Ensemble Kalman Filter nor 4DVAR could track the control with only three observations ingested. To evaluate a more realistic assimilation application, an experiment was created in which the Ensemble Kalman Filter ingested a single observation at the 180th time step in the X, Y, and Z Lorenz variables, and only in the X variable. The results show a perfect fit of 4DVAR and the control during a complete integration period, but the Ensemble Kalman Filter shows disagreement after the 80th time step. On the other hand, a considerable disagreement between the Ensemble Kalman Filter trajectories and the control is observed, as well as a total failure of 4DVAR. Better results were obtained for the case in which observations cover all the components of the model vector.