January 20, 2011 Meeting Summary
Dr. Chanh Kieu from the Lab of Climate and Weather Research at Vietnam National University presented his work on "Representing Model Errors in Hurricane Ensemble Assimilation Systems by Multiple Physical Forcings". Dr. Kieu began his presentation by describing some idealized experiments. He first explained his motivation for these experiments. By comparing four different experiments using different microphysics schemes, four different results were achieved. Dr. Kieu then posed the question of whether there was a way to isolate model errors from errors in the initial conditions. To answer this question, Dr. Kieu explained his "perturbed forcing" method which using the definition of the tangential linear model can be used to find the exact model error. Dr. Kieu then explained how he would apply this forcing method in a practical implementation.
Next, Dr. Kieu described his experiment involving a Lorenz 40 variable model. Using the Lorenz 40 variable model as the base model, he also used the tangential linear model with random Gaussian white noise as observational data and also used perturbed forcing. Compared to the multiplicative inflation (MF) method (dotted line in the plot), the perturbed forcing (PF) method (bold line) performs better with lower RMS values throughout the time steps. Looking at model sensitivities, Dr. Kieu compared the PF method to the MF method using different assimilation windows (in the bar chart on the left). Here, PF outperforms MF for each of the different assimilation windows. In the chart on the right, the assimilation window is as long as it could be for MF. Here, the RMS values for the PF method are very similar whether using 10 ensemble members or 100. This indicates that the PF method works well with a smaller number of ensemble members as well as beyond the divergent limit of the MF method.
Dr. Kieu then explained his experiment involving the WRF model. He first detailed some issues with WRF implementation such as isolating model errors from initial condition errors and a lack of clear method to perturb model forcing. He also mentioned that the degree of freedom number for a full rank Kalman filter was too large for the kind of experiment he wanted to perform, which is why the local ensemble transformed Kalman filter (LETKF) is used here. Dr. Kieu then went over the LETKF algorithm concentrating on the point that the basis of the LETKF transforms model space spanned by a number of grid points around a local patch to the ensemble space, where the analysis is performed at each grid point. Model error can also be taken into account by introducing an inflation factor. In the idealized design of this WRF experiment, GFS data was used from July 2010 with seven microphysics schemes used for perturbation. Looking at the results in bar chart form, Dr. Kieu explained that for the first 18h, multiple physics (MP) peformed better. However, at a later time, the MF experiments were better. Also, the performance of MF and MP did not seem to be dependent on the number of ensemble members. Dr. Kieu then showed ensemble spread results for MP (at the top) and MF (at the bottom) and noted that for MP, the ensemble spread decreased with time, meaning observations would not be used in the model. For the MF method, with time, the area of small spread increased but observation information would still be used in the model.
Next, Dr. Kieu detailed his real-time experiments using the WRF-LETKF. In the design of the WRF-LETKF, observations are first input into quality control. This is combined with the previous 12-hour short range ensemble WRF forecasts and go into ensemble perturbations (LETKF). Then, the updated global forecasts (GFS) go into the updated initial condition and are used to replace the ensemble mean. This goes into the WRF model and its output is used as the next cycle's 12-hour WRF forecast. This creates a very stable system, mostly from the use of GFS data to replace the ensemble mean. Dr. Kieu then described his four experiments with the WRF-LETKF. Super Typhoon Megi was the first experiment. Tracks for ensemble members are shown compared to the best track (in black). Here, the members all have an eastward bias, however they do capture the sharp turn made by the typhoon. Next, heavy flooding in October was used for experiment two. Overall, a few members were able to capture the heavy rainfall amounts of 700 mm but most missed the rainfall center location. Another flood was used for experiment three, and again a few ensemble members captured the accumulated rainfall of 550 mm, however most missed the rainfall center. The fourth experiment used an abnormal cyclogenesis event from December 2010. Of the four different microphysics schemes shown, Lin et al. (1st panel on left) overestimated the TC genesis.
Dr. Kieu concluded by mentioning some of the issues he faces during his work. For instance, there is a lack of real-time observations for the Vietnam area as well as a lack of computer resources. Future work includes objective verification of model results as well as fixing some issues with LETKF. There is also a plan to couple to a wave model in the future.