12.1 Meteorological Grid Ensemble

The uncertainty of the concentration calculation can be addressed in a similar way in which we earlier addressed trajectory uncertainty, by slightly offsetting the meteorological data to test the sensitivity of the advection calculation to the gradients in the meteorological data fields.

1. retrieve captex_control.txt and captex_setup.txt from the files directory.

2. Recall from the trajectory discussion that the internal grid offset is in the horizontal and vertical. Therefore, to take full advantage of the ensemble, the release height should be changed to 200 m from the Setup Run / Starting Locations menu. Also to reduce the number of output frames, open the grid menu and change the sampling start time to 83 09 26 06 00 to match observations at sampler 510, rename the output file to ensemble, then save to exit. To further speed up the calculation, we will only compute results through 18 (or 12) UTC 27 September. Therefore, change the run duration from 68 to 49 (or 43) hours. For future reference, save the CONTROL file as ensemble_control.txt.


3. Open the Advanced / Configuration Setup / Concentration / Menu #4. Because the GUI script will cycle through 27 simulations, we can speed up the calculation by reducing the particle release number from 50000 to 5000. Also the particle position output files are not required. For future reference, save the namelist file as ensemble_setup.txt and then save to close all menus.


4. To start the model simulation, press the Special Runs / Ensemble / Meteorology menu tab and accept the prompts to continue and when finished, close the ensemble complete menu. The calculation results are saved to the \working directory and the base output file name, ensemble is appended with a simulation sequence number from .001 to .027 representing different meteorology grid offsets.


5. To analyze the simulation results, a special pre-processing program needs to be run to convert the individual member concentration output files to files of concentration probability at each grid point. The Display / Ensemble / Create Files tab opens a menu to select the input file base name, species, and level. Only one pollutant or level can be processed per pass. If you set the time aggregation period to some value greater than 1, then the ensemble results will represent multiple time periods. The completion of this step leaves a variety of different files (prob??, cmax??, cmean) in the working directory.


6. Press the Display / Ensemble / View map menu tab. This opens up the ensemble display menu. The Output Selection Options can be used to change the maps between different metrics such as the number of members, mean, variance, or the probability of exceeding a certain concentration level. Selecting the Concentration at Percentile of 50% means that the output maps will show the median concentration of all the ensemble members. Essentially this menu is used to select which of the probability files from the last step will be displayed. There is one graphic for each time period (starting 26 06) but only the 4th time period is shown here.


7. These files can also be displayed as box plots. Start from the Display / Ensemble / box plot tab. This opens a simple menu where you can set the location of the data extraction point. Use the same location, sampler 510 (Little Valley, NY), that we examined in the concentration utility tutorial by entering 42.25 -78.80, then Execute Display.


8. The resulting box plot shows the ensemble concentration predictions at Little Valley for each 3-hr averaging period during the simulation. The box bottom and top shows the 25th and 75th percentile concentrations, the single lines the 10th and 90th percentiles, and the circles the 5th and 95th percentile concentrations. The plus symbol and middle line show the mean and median concentrations. Note there are no unit conversion options for any of these programs so the results are all shown as g m^-3. An additional graphic with the same axis labels is also created that shows the individual member concentrations indicated by member number.


9. To relate the box plot with the original graphic of median concentrations, the median box plot concentration on the 26th between 15 and 18 is perhaps 1x10^-9. Box plots are labeled according to the end time of the sample. Examining the graphic at 42.25 -78.80 shows the location to be very close to either the blue (10^-10) or yellow (10^-9) contours. Your results may differ from those shown here depending upon how many particles were released in your simulation.


10. To view how the measured data fall within the ensemble estimates, we can overlay the data points on the boxplot. In the concentration utility tutorial we created a file with just the measurements from station 510 for display purposes. Unfortunately, the box plot program has no overlay capability, but we can manually (e.g. Paint) add the measurements to the box plot graphic after converting the numbers from pg back to g. Model predictions drop off much more quickly than the measurements which is not accounted for by the meteorological grid ensemble. Note that 10^-11 is approximately the detection threshold for the measured concentrations. Values lower than this are effectively zero. The ensemble also provides an estimate of the potential maximum concentration.


11. For later reference, examine the box plot again at station 510 and note the concentration range (5%-95%) for the initial time period of the highest predicted concentrations (26 18). It is approximately 5x10^-10 to 1.5x10^-9. The range is much larger for many of the other time periods because a significant number of members predicted zero concentrations.

There are many causes of concentration uncertainty, such as how well the meteorological data represent the true flow field, uncertanties and errors in the parameterizations of physical processes in the model, and uncertainty arising from estimating concentrations from the distribution of computational particles. In this case we looked at the issue of how well discrete data points represented the flow field. In the subsequent sections we will explore some of the other factors.