Another obvious ensemble is to run the dispersion simulation with different meteorological data files.
For instance all the members of the NOAA Global Ensemble Forecast System (GEFS) or High Resolution Ensemble Forecast (HREF) could be used.
Here we will simply use five different meteorological datasets that cover the same time period.
There is no one-click option through option in this in the GUI. Each simulation must be configured manually, but the GUI can be used to analyze the products.
For this example, we will run the full 68 h simulation rather than the 49 h used in the previous ensemble examples.
- Press Reset
- retrieve the previously saved captex_control.txt
and captex_setup.txt settings into the GUI menu.
We will run the simulation for the full CAPTEX period rather than a subset as in the previous ensemble examples.
- Before running the first simulation, open the Setup Run / Grid Menu and change the output file name from hysplit2.bin to hysplit2.001. The suffix will be manually incremented with each new simulation, following the same procedure used in the automated ensemble scripts.
- open the Advanced / Configuration Setup / Concentration / Menu #4 and change the particle release per cycle to 10000. Or you can use less particles if you want the simulations to run faster. Save the changes and run the model.
- Once the simulation has completed, open the Concentration / Utilities / Convert to / DATEM menu and as before select the measured data file captex2_meas.txt, convert the model output to the DATEM format, enter a unique character string such as wrf27uw to rename the output files, and then Compute Statistics. The statistical results file will then be named statA_wrf27uw.txt.
- Now repeat the same process with each of the other four meteorological data files, giving each output file and resulting statistical file a unique name:
Remember the Convert to DATEM step must be performed with each new simulation, otherwise the previous file will be used for the statistical analysis.
- When all the simulations have been completed, open the Setup Run / Grid Menu and rename the output file from hysplit2.00? to just hysplit2. The base name will then be passed through the GUI to the ensemble scripts, where the programs automatically search for the 3-digit suffix. Save the changes, run the Display / Ensemble / View Map to generate the probability files, then create the boxplot at Little Valley, NY (42.2N 78.8W). With only five members, once the concentrations are measurable, the variation between members becomes smaller, and the ensemble also captures the time persistence of the concentrations. A view of the member plot shows that simulation 002 (the ERA40) gave the highest concentrations. The highest measured concentration was around 1.3x10-9 g/m3.
- Although there are a variety of possible metrics, the complete statistics for each of the simulations compared with the measured data can be extracted from the statA.txt file. The correlation coefficient and the rank (0-4) are shown below. The higher the rank, the better the overall performance. The rank consists of the sum of the normalized correlation, fractional bias, figure-of-merit, and Kolmogorov-Smirnov parameter.
- 0.81 (3.03) WRF27UW
- 0.65 (2.66) ERA40
- 0.68 (2.87) NARR
- 0.81 (3.19) WRF27
- 0.82 (3.12) WRF09
The results suggest that the best performance is when using one of the WRF meteorological data fields. In most modeling situations, we usually don't know the correct answer. One approach is to use the ensemble mean concentration for the simulation. This field has already been computed in the View Map step.
- Open the Concentration / Setup Run / Grid menu and change the base name from hysplit2 to cmean. Then open the Utility / Convert to / DATEM menu and run the statistical analysis for the ensemble mean, which shows a correlation of 0.83 and a rank of 3.19, a similar performance to the better ensemble members. A similar, but more qualitative comparison, can be made for the scatter plots, with the ensemble mean plot also showing excellent results.
The results shown here suggest that a practical application for ensemble products may be simply to compute the ensemble mean and use that as one would a single deterministic simulation. The use of an ensemble does not require the choice of any one member. Do not delete any of these simulations as they will be used in the next section.
4 m 17 s