A Program for the Statistical Evaluation of Point Source Dispersion Models Using ASTM D 6589 ------- EPA-454/R-03-006 A Program for the Statistical Evaluation of Point Source Dispersion Models Using ASTM D 6589 U.S. ENVIRONMENTAL PROTECTION AGENCY Office of Air Quality Planning and Standards Emissions, Monitoring, and Analysis Division Research Triangle Park, North Carolina 27711 August 2003 EPA - RIP LIBRARY ------- DISCLAIMER The information in this document is being reviewed in its entirety by the U.S. Environmental Protection Agency (EPA), for approval for publication as an EPA document. Mention of trade names, products, or services does not convey, and should not be interpreted as conveying official EPA approval, endorsement, or recommendation. 11 ------- PREFACE The American Society for Testing and Materials (ASTM) has published a Standard Guide D 6589 for Statistical Evaluation of Atmospheric Dispersion Model Performance. Within the Annex to this ASTM Guide, a procedure is outlined for assessing the performance of atmospheric transport and diffusion models to predict the average maximum "centerline" concentration values of a chemical species that has been released from a point source. The EPA has developed a FORTRAN program called "Design 2" that implements the procedure documented within the Annex to ASTM Standard Guide D 6589, for the statistical evaluation of the performance of transport and diffusion models to simulate the average "centerline" concentration values. This report describes the input files and formats of these input files for running Design 2. The report describes the information provided by the various output files created by Design 2. The report does not describe the actual statistical procedures of Standard Guide D 6589, as these are documented in detail within this Guide's Annex. in ------- ACKNOWLEDGMENTS This research described in this report was funded wholly or in part by the United States Environmental Protection Agency (EPA) through EPA Contract NO. 68-D-98-006, Work Assignment NO. 4-02 with EC/R Inc., who subcontracted with the Pacific Environmental Services, Inc, a MACTEC Company (PES). The EPA Work Assignment Officer was John S. Irwin. The EC/R Project Leader was Steve Fudge. The PES Project Leader was James Paumier. IV ------- TABLE OF CONTENTS Page Preface iii Acknowledgments iv 1. INTRODUCTION 1 2. OBSERVED CONCENTRATION DATA 4 3. MODELED CONCENTRATION DATA 8 4. REGIMES 10 5. CONTROL FILE 13 6. RESULTS 17 7. REFERENCES 23 APPENDIX - DIRECTORY STRUCTURE OF DISTRIBUTION CD 24 ------- SECTION 1 INTRODUCTION Air quality transport and diffusion models have been in use for several decades to estimate the impacts of pollutants released from point sources. Some of the simpler models are steady-state, Gaussian plume models that include little or no chemical transformation. During the model development phase and in subsequent upgrades, model performance is constantly evaluated. In the past, the emphasis of the statistical evaluation comparisons has been on the "intended use." For instance, one of the uses for modeling results is to estimate the highest concentration values to be expected over a 5-year period, resulting from the operation of a proposed new power plant (Cox and Tikvart, 1990). Other statistical measures have also been employed to compare the concentration (or dose) values of "intended use," such as number of values within a factor of two, linear least-square fits to scatter plots of observed and predicted values, and normalized mean- squared errors of observed and predicted values (Fox, 1981). Implicit in such statistical comparisons is an assumption that the predicted and observed distributions of concentration values are from the same population, which may not be a well-founded assumption (Oreskes et al, 1994). Work is underway to develop a new generation of evaluation metrics that takes into account the statistical differences (in error distributions) between model predictions and observations. As a result, a shift in philosophy is occurring as to how models of environmental processes can be acceptably evaluated. Most models provide estimates of the first moment of conditions to be expected for each ensemble (e.g., average time-space variation of the meteorological conditions, average time-space variation of the surface-level concentration values). The key to the next-generation evaluation metrics is that they will no longer assume that the modeled and observed values come from the same statistical population of values. They will assume that they "share" certain fundamental properties, but are inherently different. To fill a part of this void, the U.S. Environmental Protection Agency (EPA) has participated within the American Society for Testing and Materials (ASTM) to develop a consensus on a philosophical basis that could be used in developing new statistical methods for model evaluation. In doing this, a statistical procedure was drafted and documented within the Annex to Standard Guide D 6589 for Statistical Evaluation of Atmospheric Dispersion Model Performance (ASTM, 2000). This procedure implements the idea that the distributions of model predictions and observations "share" certain fundamental properties, but are inherently different. The procedure assesses the performance of transport and diffusion models to simulate the average "centerline" concentration values from a point source release. The EPA has developed a FORTRAN program called "Design 2" that implements the -1- ------- procedure described in the Annex to D 6589. This report describes the input files and formats of these input files for running Design 2. The report describes the information provided by the various output files created by Design 2. The report does not describe the actual statistical procedures of Standard Guide D 6589, as these are documented in detail within this Guide's Annex. As noted in the ASTM standard, many practical problems must be resolved before the software can be run (see Section XI .4), including how to group the data into regimes. If we consider each observation period as an experiment and groups of monitors/receptors (nearly) equidistant downwind as an arc, then we can define an experiment-arc as a single experiment and single arc combination. An experiment-arc pair can then be included in a regime and the software can connect observed data to modeled data. Some regime definition strategies are mentioned in Section 4. The purpose of this document is to provide guidance on the procedures to prepare the data and run the evaluation software. Figure 1.1 shows a flow chart of the process. The first input file required is the observed concentration data. The structure of this file is described in Section 2. The second file is the modeled concentration data. Since this software determines a best performing model, results from at least two models should appear in this file. The structure of this file is described in Section 3. The third file defines ranges of conditions, or regimes, based on one or more parameters that subdivide the results. In defining regimes, the behavior of the dispersion should be similar within each subdivision. A more complete discussion of regimes and the structure of this file is described in Section 4. For the software to function, a control file is required that has the names of the input and output file, file formats, and processing options. This file is short (less than 20 lines) and the structure of this file is described in Section 5. Once the software is run, several output files are produced: 1) the primary file with the information on the best performing model (the ASTM analysis), 2) a file of the bootstrap results, and 3) a file of results by regime. An optional output file is the fitted observed data. The analysis and structure of these files is discussed in Section 5. In Section 6, the results from the evaluation software are presented. The software was coded in FORTRAN. Although a FORTRAN 95 compiler was used to create the executable, attempts were made to keep the code backward compatible with FORTRAN 77 standards. In several places, the user is required to provide a format for the data to be read by the program. These formats must follow FORTRAN conventions and MUST be enclosed in parentheses; for example, (IX, 15,15, 3F12.0). The characters in the FORTRAN statement can be upper or lower case; the spaces after the commas are not necessary and are shown here only to provide better legibility. The example used throughout this documentation is from the 1985 Indianapolis field experiment (Murray and Bowne, 1988). The field experiment was performed in a flat, urban to suburban environment. A tracer, SF6, was injected into the buoyant gases exiting from an 86- meter point source. Data were collected for approximately 170 hours over a five-week period. Monitors were located in arcs from about 250 meters to 12 kilometers downwind. -2- ------- Meteorological data were collected from a 94-meter and 10-meter tower in the urban environment, a 10-meter tower in a suburban location, and 10-meter tower in a rural location. National Weather Service meteorological data were from Indianapolis, IN (hourly weather observations) and Dayton, OH (upper air data). For the dispersion modeling, receptors were placed along a single line due east of the release point. To obtain centerline concentrations the wind directions were all modified to be blowing from the west (270°). INPUT Observed Concentration Data (2) Modeled Concentration Data (3) OUTPUT Fitted Observed Data (optional) Regime Definition (4) DESIGN 2 ASTM Analysis (6) Control File (5) Bootstrap Results (6) Regime Results (6) Figure 1.1 File I/O for the ASTM analysis software, DESIGN2. The numbers in parentheses indicate the section in which the input/output is described. -3- ------- SECTION 2 OBSERVED CONCENTRATION DATA The observed concentration data are required to be in a specific format for the analysis software to correctly read and process the data. The general structure is shown in Table 2.1. The first record in the file is Header 1 and contains a title that can be up to 40 characters long and only appears once. Note that this title must be enclosed in single quotation marks. Headers 2 and 3 appear before each arc of observations. These arc header records are followed by the x- and y-coordinates and observed concentration, for the number of receptors on the arc (the first field on Header 3). The coordinates and concentration must be specified as x-coordinate, y- coordinate and then concentration value. The FORTRAN format for this information must remain fixed for all values in the file. This FORTRAN format is defined by the user as part of the control file information (see Section 5). Currently, the number of observed concentration records per experiment-arc is limited to 250 records. An example of the observed concentration file is shown in Figure 2.1. Header records 2 and 3 are read 'free' format, i.e., without a format statement. This means that a comma or at least one space (or a combination of the two, as seen in Figure 2.1) must separate each entry on each record. Since the date and time are the only character variables in these two records, they must be enclosed in single quotation marks. The header 2 records provide general information on the field experiment period and arc combination. The experiment and arc (or traverse) pair identify an observation period and arc of data. The pair must be unique. Usually, the experiment numbers are unique (possibly a simple sequential numbering), but the arc numbers likely repeat from field experiment period to field experiment period. The important thing to remember is that these two parameters are used to connect the observed data with the modeled concentration data and the grouping into regimes, so the pair must be unique. The date, start and stop time, and nominal distance to the arc are not used in the program, but provide additional identifying information within the archive. The header 3 records contain information on how to process this experiment-arc combination. The first three values define the total number of receptors on the arc and which receptors to use in the analysis. Receptors can be included or excluded from the analysis by varying the starting and ending point for the analysis. The location of the receptors must be entered in the same coordinate system used in defining the receptor locations, namely Cartesian (x,y) or polar coordinates (r,0). The type of coordinate system in use is defined by the value of IXYARC, which is the fourth parameter of -4- ------- Header 3. If the receptor locations are entered as polar coordinates, then the program converts them to Cartesian coordinates internally. The program works with coordinates expressed in meters, so if the units are other than meters, a conversion factor from the user's units to meters must be applied. This factor is provided as the fifth parameter on header 3 records. The sixth and seventh parameters identify the location of the source. If the x- and y- coordinates are entered as polar coordinates, then the source is likely to be located at an origin, i.e., (0,0). If the x- and y-coordinates are entered as Cartesian coordinates (e.g., UTM coordinates), then the source is likely to be located at something other than (0,0). The software translates the source and receptors to a (0,0)-based system internally to perform its analysis. The tracer release rate is specified next and is used to normalize the fitted observed concentrations. The tracer height and the height of the receptors are not used in this version of the software, but may provide useful information possibly interpreting the results, and serve as a check on the source and receptor heights specified in the dispersion modeling. The last parameter on this record converts the tracer concentration from user's units to micrograms per cubic meter (ng/m3). Since this value is applied to every tracer concentration without regard to meteorology, careful consideration should be used as to the appropriate value to enter. For example, converting from parts per trillion to ng/m3 requires atmospheric pressure and ambient temperature. The header records are followed by the location and observed concentration for the number of receptors on the arc (defined by the first value on header type 3). At a minimum there are three values on these records. However, since the format is specified in the control file (see Section 5), any amount of information can be included and skipped with the format statement the user provides. For example, if the coordinate location is in polar coordinates, Cartesian coordinates can be included on the record but skipped with the appropriate format statement (as is done in Figure 2.1). During the fitting process, the concentrations are divided by the emission release rate (eighth parameter in Header 3). -5- ------- TABLE 2.1 STRUCTURE OF THE OBSERVED CONCENTRATION DATA FILE. Record Type Header 1 Header 2 Header 3 Observa- tion Data Title (up to 40 characters) Experiment number Traverse number Date Time (Start & Stop) Nominal distance to arc Number of receptors on the arc Starting data point for integration Ending data point for integration How x- and y-coordinates are entered: polar or Cartesian (IXYARC) Conversion from user units to meters X-coordinate of source Y-coordinate of source Tracer release rate Tracer release height Altitude of airborne traverse Constant multiplier to convert tracer concentration from user's units to |ig/m3 X-coordinate Y-coordinate Observed Concentration Data Type C I I C C R I I I I R R R R R R R R R R Description Appears only once, as the first record in the file, in single quotation marks Field experiment period Identifier for the arc 10 characters, in single quotes ' MM-DD-YY ' 10 characters, in single quotes ' HHMM-HHMM' (not used) Total number of receptors that will be read by the program Receptors prior to this point are omitted from analysis Receptors after this point are omitted from analysis <= 0, cartesian > 0, polar Converts user units to meters, e.g., for kilometers to meters, enter 1000. Coordinate in base system Coordinate in base system Release rate in grams/second Height of release in meters (not used) Height of receptors in meters (not used) Enter 1.0 if no conversion is needed, otherwise enter the multiplier that will be applied to all observed concentrations If IXYARC <= 0, x is in meters If IXYARC > 0, x is in degrees If IXYARC <= 0, y is in meters If IXYARC > 0, y is radial distance to arc Concentration in user units Data Types: C=character, I=integer, R=real ------- 'IND SF61 1, 1, ' 1, 1, 24.71 1, 2, ' 8, 1, 5.10 21.13 34.63 57.90 73.93 309.61 325.45 347.88 1, 3, ' 4, 1, 38.25 332.79 338.05 356.90 1, 4,' 6, 1, 250.85 290.40 298.74 323.61 335.41 344.65 1, 5, ' 11, 1, 6.07 15.99 257.66 263.77 269.59 280.49 287.25 293.21 302.81 341.93 358.25 1, 6,' 17, 1, 10.43 27.36 44.34 51.53 68.32 250.99 257.43 268.08 274.44 282.63 290.17 297.51 305.40 313.61 323.38 332.62 340.45 09-16-85 ',' 1100-1200', 1, 1 , 1000.000, 0.0, .31 .13 .28 09-16-85 ' , ' 1100-1200' , 8, 1 , 1000.000, 0.0, .47 .04 .47 .47 .17 .44 .40 .23 .33 .45 .38 .24 .43 .42 .12 .50 -.39 .32 .55 -.31 .45 .48 -.10 .47 09-16-85 ',' 1100-1200', 4, 1 , 1000.000, 0.0, .66 .41 .52 .71 -.32 .63 .68 -.25 .63 .72 -.04 .72 09-16-85 ',' 1100-1200', 6, 1 , 1000.000, 0.0, 1.43 -1.35 -.47 1.03 -.97 .36 .94 -.82 .45 1.01 -.60 .81 1.01 -.42 .92 1.02 -.27 .98 09-16-85 ',' 1100-1200', 11, 1 , 1000.000, 0.0, 1.41 .15 1.40 1.48 .41 1.42 . 1.50 -1.46 -.32 1.47 -1.46 -.16 1.38 -1.38 -.01 1.81 -1.78 .33 1.52 -1.45 .45 1.55 -1.42 .61 1.77 -1.49 .96 1.47 -.46 1.40 1.47 -.05 1.47 09-16-85 ',' 1100-1200', 17, 1 , 1000.000, 0.0, 1.97 .36 1.94 1.93 .88 1.71 1.92 1.34 1.37 1.95 1.52 1.21 1.92 1.79 .71 2.06 -1.94 -.67 1.97 -1.93 -.43 1.79 -1.79 -.06 2.20 -2.19 .17 1.97 -1.92 .43 2.03 -1.91 .70 2.01 -1.79 .93 2.07 -1.69 1.20 2.04 -1.48 1.41 2.13 -1.27 1.71 1.95 -.90 1.73 1.96 -.66 1.85 0.2 0.0, 4.94000, 83.80, 1.50, 7.0 0.5 0.0, 4.94000, 83.80, 1.50, 155.0 9.0 .0 .0 .0 .0 18.0 143.0 0.7 0.0, 4.94000, 83.80, 1.50, .0 38.0 25.0 84.0 1.0 0.0, 4.94000, 83.80, 1.50, .0 .0 .0 6.0 116.0 206.0 1.5 0.0, 4.94000, 83.80, 1.50, 225.0 59.0 .0 .0 9.0 .0 .0 .0 .0 79.0 444.0 2.0 0.0, 4.94000, 83.80, 1.50, 55.0 .0 .0 .0 .0 .0 .0 .0 .0 .0 .0 .0 .0 .0 6.0 15.0 40.0 1.000 1.000 1.000 1.000 1.000 1.000 Figure 2.1. Sample of the observed concentration file. -7- ------- SECTION 3 MODELED CONCENTRATION DATA The second component required to perform the analysis is a file of modeled concentration data. This file contains the concentration estimates from two or more dispersion models. Results from as many model runs can be included (within DOS or compiler limitations) in this file since the user must specify in the control file the number of model runs to include in the analysis and the format to read the data. However, only results from up to 15 models can be included in a single analysis. The structure of the file is shown in Table 3.1 and an example is shown in Figure 3.1. The format of these data is contained in the control file (described in Section 5). Specifying the format in this manner allows the user to identify which model estimates to include in the analysis. It should be mentioned that the units of the modeled concentration estimates MUST match those in the fitted observed data, otherwise the results will be completely unreliable. It may be necessary to develop an external program that converts the modeled concentrations to the units in the observed concentration file, as was done for the example used in this documentation. TABLE 3.1 STRUCTURE OF MODELED CONCENTRATION DATA FILE. Record Description Header record; can be any descriptive language; the information in this record is not used in the program Header record; can be any descriptive language; the information in this record is not used in the program 3...#of experiment- arc periods Modeled concentration data; multiple columns of data; repeated for as many arc-experiment periods to be included in the analysis Required data on each record for N model runs (N is specified in the control file): Experiment period, Arc number, Concentration 1,..., Concentration N ------- (3x,2i5,fll.3,2x,i6.6,i4. EXP TRAV 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 3 3 3 3 4 4 4 4 4 4 4 4 4 4 4 4 1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12 X 0.200 0.500 0.700 1.000 1.500 2.000 3.000 4.000 6.000 8.000 10.000 12.000 0.200 0.500 0.700 1.000 1.500 2.000 3.000 4.000 6.000 8.000 10.000 12.000 0.200 0.500 0.700 1.000 1.500 2.000 3.000 4.000 6.000 8.000 10.000 12.000 0.200 0.500 0.700 1.000 1.500 2.000 3.000 4.000 6.000 8.000 10.000 12.000 4,5fl2.5) DATETIME 8509161100 8509161100 8509161100 8509161100 8509161100 8509161100 8509161100 8509161100 8509161100 8509161100 8509161100 8509161100 8509161200 8509161200 8509161200 8509161200 8509161200 8509161200 8509161200 8509161200 8509161200 8509161200 8509161200 8509161200 8509161300 8509161300 8509161300 8509161300 8509161300 8509161300 8509161300 8509161300 8509161300 8509161300 8509161300 8509161300 8509161400 8509161400 8509161400 8509161400 8509161400 8509161400 8509161400 8509161400 8509161400 8509161400 8509161400 8509161400 13 124 129 106 72 52 31 21 13 9 7 5 9 116 122 101 68 48 28 19 11 8 6 5 6 110 120 99 66 46 26 18 10 7 5 4 6 109. 122. 102. 67. 47. 26. 17. 9. 6. 5. 4. ADMS .52654 .19214 .02158 .83546 .80358 18434 09959 46810 23311 56996 39425 92507 67782 65285 54763 43091 40762 55746 52670 32401 43619 15481 31754 09892 93544 24163 36591 47187 36179 57782 95085 02947 29495 16038 51101 45420 02123 91783 56912 79445 80977 12925 90252 84781 99287 80607 18173 17275 AER02161 15.88650 62.27761 62.43041 51.13265 35.57238 27.28770 18.67307 14.77599 11.21865 9.23732 7.90674 6.94058 17.32178 67.36076 67.21519 53.96894 35.32141 24.98616 15.65525 11.41748 8.05074 6.53249 5.56385 4.87275 16.03329 69.93465 70.52279 56.41741 36.37708 24.99461 14.90788 10.54413 7.00644 5.59702 4.74401 4.14563 0.01478 50.51303 89.26309 84.29244 57.40505 39.04949 20.83178 12.90603 6.44580 3.97508 2.83168 2.22595 20 120 115 88 55 40 27 21 16 13 11 10 16 117 112 85 52 34 20 15 11 9 7 6 11 102 103 81 51 34 19 13 9 7 6. 5. 3. 76. 92. 82. 56. 39. 21. 14. 8. 6. 5. 5. HPDM 58027 69160 09613 37265 99589 92309 57795 56590 25815 34078 56248 34749 85009 75447 84499 90490 63741 97790 92225 39928 10693 14303 83048 94412 07608 03912 26080 77794 53038 29897 38106 86609 48170 80348 67414 88584 08309 51933 48806 91390 95367 30663 43528 23599 91191 91433 94636 25265 ISCST3 0 11 51 76 68 57 42 34 26 22 19 17 0 10 48 70 54 41 30 25 19 16 14 12 0 11 48 68 50 37 26 21 16 13 12 10 0 21 64 75 50 34 23 18 14 12 10 9 .00000 .44278 .72357 .59892 .48630 .06482 .60660 .89255 .69948 29126 47426 48786 00000 78590 39931 07381 24477 89233 92515 32321 37537 17562 13122 68949 00000 46303 85902 63754 91323 08468 71908 87267 73271 96850 20238 95750 00000 24328 44592 67673 77602 06233 24096 99567 52074 11810 58417 50299 AMODNoObsZi 22.98427 67.62427 65.56213 51.24015 32.48832 22.16314 13.06510 9.14368 5.99775 4.77061 4.03602 3.52151 19.70494 67.53592 66.77747 52.49763 33.22327 22.42547 12.81057 8.82312 5.54923 4.33742 3.65312 3.18173 17.17133 68.71230 69.20486 54.65300 34.47306 23.10219 12.85674 8.71999 5.30995 4.06644 3.40525 2.95946 0.01169 48.07773 86.99314 83.30546 56.88217 38.64830 20.55332 12.70320 6.32857 3.90839 2.79318 2.20017 Figure 3.1. Example of modeled concentration file The first two records can contain any information the user wants to include since neither record is used by the analysis software. In this example, the FORTRAN format of the data appears on the first record. The second record identifies the fields. In this example, there are two model runs with the AERMOD dispersion model: the first with the column header AER02161 and a second with the column header AMODNoObsZi. The former run utilized observed mixing heights, and the latter allowed AERMET to calculate the mixing heights. -9- ------- SECTION 4 REGIMES The third file required to perform the ASTM analysis is a file of 'regime' definitions. This file groups the experiment-arc combinations into categories that the user feels are appropriate to answer the question(s) under investigation. Comparing separately averaged observed and modeled concentrations within grouped data "provides an empirical estimate of the combined deterministic error associated with input uncertainty and formulation errors" (ASTM, 2000). An extended discussion of establishing the regimes is provided in the ASTM standard guide (ASTM, 2000). Regimes can be defined based on almost any property, such as atmospheric stability, type of air mass, or wind speed. The definitions of the regimes appearing in the example in this document were based on z/L, where zi is the mixing height and L is the Monin-Obukhov length. Two additional restrictions were placed on the data to insure there were sufficient data to perform the analysis. The first was that a minimum of five nonzero observed concentrations must appear on an arc to be included in a regime and that there must be a minimum of five experiment-arcs per regime. Regimes, including the one presented here, are not hard and fast definitions, and in fact, may require modification if the initial results are ambiguous or the meteorology changes. Additionally, which model's meteorology (or other model-specific data) to use to define the regimes must be considered when developing the regimes. The structure of the regime definition file is shown in Table 4.1. There are three record types, with the first type defining the number of regimes (and appearing only once, at the beginning of the file), and the second type defining the number of experiment-arc combinations in a regime. The third record type consists of the experiment-arc pairs within the regime. All required data are read as free format integers. An example of the regime definition file is shown in Figure 4.1. Note that in the example, additional information follows the required data on each record. Only the first field is read on record types 1 and 2, so any descriptive language can follow these values. For example, the criteria for the data grouping can appear after the first field. In record type three, only the first two fields are required and are read free format. Additional information can appear after these two fields that may be useful in reviewing and understanding the results. In this particular example, z /L is used as the criterion to group the data. The first regime is defined for z/L < -50.0. The values that follow the experiment and arc number on each of the type 3 records is the number of nonzero observed concentrations. -10- ------- TABLE 4.1 STRUCTURE OF THE REGIME DEFINITION FILE Record Type 1 2 3 Description Number of regimes to use; appears only once as the first record Number of experiments-arc combinations in the regime; appears before grouping of experiment-arcs each Experiment period and arc to include in this regime -11- ------- 2 3 19 20 21 28 32 33 34 113 114 116 12 2 3 19 20 21 28 32 33 34 113 114 116 29 Number of Regimes 12 Number of Exps in this regime 6 Cases where Zi/L is less than: -50.0 7 6 6 9 8 13 10 5 Number of Exps in this regime 6 6 7 5 9 7 13 15 9 than: 22 23 24 25 29 30 31 35 55 14 22 Number of -25.0 Exps in this regime Cases where Zi/L is greater than: -50.0 and less 9 9 7 8 7 8 10 5 5 Number of 4 8 Exps in this regime Figure 4.1. Example of the regime definition file. The required fields and information are shown in bold and the information that is not required is shown in a lighter font. -12- ------- SECTION 5 CONTROL FILE Once the data files are in place, the remaining file required to run the ASTM evaluation software is the control file. This file contains file names, data formats, and processing options. Table 5.1 shows the structure of the control file and Figure 5.1 shows an example. The output from the analysis software will be presented in Section 6. The first six records in the control file pertain to the fitting of the observed concentration data. The fourth record is needed only if the results of the fit are to be retained in a file. Otherwise, the data are retained in memory for the remainder of the analysis and lost when the program terminates. The fifth record (IPHIY) indicates how oy is to be retained, either in user units or in degrees. The sixth record defines the minimum number of nonzero concentration values required to attempt a Gaussian fit of the observed data. The remaining controls are relevant to the analysis of the modeled data and the actual evaluation analysis. Records 7-9 apply to the sampling scheme (NPAIR) and method (NWIDE), and the number of bootstrap samples (NBOOT). NPAIR defines whether to sample the data as pairs or individually; NWIDE defines whether to sample modeled and observed concentrations simultaneously or independently. Record 10 defines the number of models or models runs (NMODEL) that are to be analyzed with this run. The analysis can be performed with up to 15 models in any single run. Note too that different runs of the same model or versions of a model can be included in an analysis. This record is followed by the names of the NMODEL models. The names are only used to identify the models in the output files. It is recommended to keep the identifying names to 10 characters or less. If the name is longer than 10 characters, only the first 10 characters will be used. The order of the names must match the order the modeling results are read in the modeled concentration file (Section 3). The record following the model names is the random number seed and should be at least five digits. This value is used during the bootstrapping process. The seed value is followed by the filtering option. It is an integer value greater than or equal to zero and controls the maximum number of near-centerline concentrations to be considered that satisfy -0.67 < y/0y < 0.67. Entering a value of 0 selects all near-centerline concentrations that satisfy this criterion. The last six records (shown in Table 5.1 as records 13+NMODEL through 18+NMODEL) in the control file are input and output file names and the format of the modeled concentration file. The names should follow the standard personal computer (PC) naming -13- ------- conventions. The FORTRAN code and executable version accompanying this documentation accommodates filenames as long as 70 characters in length. The format specified on the fifth record from the end can be as long as 30 charcters in length, must follow FORTRAN format statement rules for the data to be read correctly, and must include the initial and closing parentheses. -14- ------- TABLE 5.1 STRUCTURE OF THE INPUT CONTROL FILE FOR THE ASTM EVALUATION SOFTWARE. Record 1 2 3 4 5 6 7 8 9 10 11 10+NMODEL 11+NMODEL 12+NMODEL 13+NMODEL 14+NMODEL 15+NMODEL 16+NMODEL 17+NMODEL 18+NMODEL Control* filename FRMXYC ISPLUS output filename IPHIY MINNOK NPAIR NWIDE NBOOT NMODEL model name model name ISEED NFILTER input filename data format input filename output filename output filename output filename Description File where arcs of observed data are stored Format to read the location (x- and y-coordinates) and observed concentration (can be up to 30 characters in length, including parenthetses). Option to save results from the PLTFIT routines in a file: = 0, do not save results to file # 0, save results to file File with fitted observed data - an optional record based on the value given on the 3rd record oy output units: = 0, list output ay in user units, # 0, list oy in degrees Minimum number of values required to attempt Gaussian fit Sampling scheme: = 2, sample by pairs * 2, do not sample by pairs Sampling method = 2, sample simultaneously observation & modeled values * 2, do not sample simultaneously Number of bootstrap samples Number of models for which values will be provided First model name Last model name Random number seed value (make sure it is 5 or more digits) Filtering option (if zero, then no limit; otherwise this is the max number of values to be selected) File of modeled centerline concentration values for each model Format of the data of modeled centerline concentration values File defining which experiment numbers and arcs are in each Regime Name of output file for listing results - the primary listing for evaluating model results with the ASTM standard File with bootstrap results File with regime results * Controls in capital letters show the variable name used in the source code, otherwise a generic 'name' is used. -15- ------- INDARCS.DAT (lx,F8.0,F8.0,16x,F8.0) 1 INDSPLUS.DAT 2 3 2 2 500 5 ADMS AMOD02161 HPDM ISCST3 AMODNoObZi 12345 1 INDMODEL.DAT (3x,2i5,13x,10x,5fl2.0) INDREGM.DAT INDASTM.OUT INDBOOT.PLT INDREGIM.OUT Figure 5.1 Sample control file. -16- ------- SECTION 6 RESULTS In the earlier sections, portions of the various input files were shown. In this section, the results from the analysis software are presented based on the example files presented in the previous sections. Due to the size of the files, not all sections of the output files will be shown. A brief description of the general layout of the files will be provided. The output filenames are specified by the user, as shown in Table 5.1 and Figure 5.1. The last two files in the control file are results from the bootstrap method and information by regime. The third file from the end (identified as 16+NMODEL in Table 5.1, INDASTM.OUT in Figure 6.1) contains the listing of results used to apply the ASTM standard. The first part of this file has the following structure: 1) Run summary of the information provided in the control file; 2) Results for each regime, including the number of nonzero values for each experimental period and the number of centerline values that satisfy the following -0.67 < y/ay < 0.67, where y is the receptor distance from centerline and sy is the standard deviation of the distances for all nonzero concentration estimates; 3) Average and standard deviation of Ci, and the geometric average and standard deviation of Ci/Cavg, where the near-centerline concentrations and averages use only the values resulting from the filtering process (i.e., limit on number of near-centerline concentrations). Items 2) and 3) are repeated for all regimes. Figure 6.1 shows an example of these initial records in the output file. -17- ------- ASTM D22.ll Boot. for Draft Z6849Z VERSION: SETUP INFORMATION: Iseed (random number seed value) : Npair (0=individual, 2=pair sampling): Nwide (0=individual, 2=concurrent) : 063002 12345 2 2 Nboots (number of bootstrap samples) : 500 Nmodel (number of models) : Model 1 ADMS Model 2 AMOD02161 Model 3 HPDM Model 4 ISCST3 Model 5 AMODNoObZi Iphiy (0=m, l=degrees) : Nfilter (Limit on number selected) : Model file: Output file: Regime Definition File: Results for Group: 1 5 1 1 indmodel indASTM. .dat out indregm.dat N Exp Arc NumValues NumCenterline 123 13 233 14 3 19 3 17 4 20 3 19 5 21 3 18 6 28 3 15 7 32 3 18 8 33 3 18 9 34 3 18 10 113 3 10 11 114 3 10 12 116 3 10 Total number of values: 180 Computed Group Sy: 19.10 (deg) Avg Std Num Ci Ci — OBS 12 5.86E+01 4.49E+01 ADMS 12 1.28E+02 4.63E+01 AMOD02161 12 7.41E+01 1.77E+01 HPDM 12 1.20E+02 3.64E+01 ISCST3 12 7.17E+01 3.94E+01 AMODNoObZi 12 6.09E+01 3.73E+01 1 1 1 1 1 1 1 1 1 1 1 1 Num >0 11 12 12 12 10 12 NumModels 5 5 5 5 5 5 5 5 5 5 5 5 GeoAvg Ci/Cavg 6.30E-01 7.41E-01 9.63E-01 9.26E-01 1.14E+00 4.17E-01 GeoStd Ci/Cavg 3.31E+00 3.37E+00 1.34E+00 1.58E+00 1.37E+00 9.58E+00 Figure 6.1. The initial records in the output file used to determine the best performing model (associated with record NMODEL+16 in the control file). -18- ------- These statistics by regime are followed by the statistical measures over all regimes for each model and maximum observed value. Currently, there are 12 statistical measures calculated in the software: fractional bias, absolute fractional bias, normalized mean squared error (NMSE), mean squared error (MSB), slope, intercept, r2, (unsystematic MSE)/MSE, (systematic MSE)/MSE, Willmott-d, average observed value, and average modeled value. For each statistical measure, an abreviated description is provided of the distribution of values generated during the bootstrap: average, standard deviation, minimum, low hinge (25-th percentile), median, high hinge (75-th percentile), and maximum value. It is these statistics that are used to determine the "best performing model." An example of the summary information provided for each statistical measure is shown in Figure 6.2. For each statistical measure, the "best performing model" is indicated by an asterisk or an excalamation point after the model name (and before the column labeled Avg). An asterisk is used for those measures for which the smallest (absolute) average value identifies the best performing model. The exclamation point is used for those measures for which a value closest to 1.0 is appropriate, such as the correlation coefficient and the slope. Note that the first three and the last three statistical measures are shown in the figure; there are another six measures not shown. -19- ------- Summary Over All Groups (No Inverse Variance Weighting Employed) Fractional Bias Results Model Avg MAX 0.11905 ADMS * 0.00008 AMOD02161 -0.29222 HPDM -0.03237 ISCST3 0.03411 AMODNoObZi -0.22308 t-value compared to the base 2 AMOD02161 5.7244 3 HPDM 0.8565 4 ISCST3 0.8028 5 AMODNoObZi 5.5705 Degrees of freedom (DF) : 28 Std 0.02346 0.06137 0.06333 0.06574 0.05951 0.06185 Nboots 500 500 500 500 500 500 Min 0.06918 -0.18474 -0.52664 -0.22245 -0.17623 -0.40081 HL 0.10212 -0.04431 -0.33212 -0.08129 -0.00250 -0.26444 Med 0 -0 -0 -0 0 11892 00080 29712 03377 03322 -0.22674 0 0 -0 0 0 -0 HH .13289 .04321 .25111 .01350 .07034 .17864 Max 0 0 -0 0 0 -0 20091 22366 * 09162 17498 21014 02809 model : ADMS (at the 90% confidence level n significant difference in the metrics between the base model and the model being tested) Absolute Fractional Bias Results Model Avg MAX 0.11905 ADMS 0.55608 AMOD02161 0.60549 HPDM * 0.46838 ISCST3 0.48538 AMODNoObZi 0.54344 t-value compared to the base 1 ADMS 2.0452 2 AMOD02161 3.0677 4 ISCST3 0.3725 5 AMODNoObZi 1.6021 Degrees of freedom (DF) : 28 (at the 90% confidence level Std 0.02346 0.04422 0.04837 0.04687 0.05049 0.04826 Nboots 500 500 500 500 500 500 Min 0.06918 0.43668 0.46769 0.30258 0.32799 0.40345 HL 0.10212 0.52596 0.57155 0.43679 0.45442 0.51097 Med 0 0 0 0 0 0 11892 55828 60523 47103 48618 54504 0 0 0 0 0 0 HH .13289 .58757 .63789 .50124 .51749 .57555 Max 0 0 0 0 0 0 20091 68484 74218 62682 * 66683 67938 model: HPDM a significant difference in the Normalized Mean Squared Error Model Avg MAX 0.02203 ADMS 0.51363 AMOD02161 0.45289 HPDM * 0.34060 ISCST3 0.42333 AMODNoObZi 0.42698 t-value compared to the base 1 ADMS 2.7531 2 AMOD02161 1.4852 4 ISCST3 0.7391 5 AMODNoObZi 1.0565 Degrees of freedom (DF) : 28 (at the 90% confidence level value > metrics 1.701 indie. between the ites that there is a base model and the model being tested) Sesults Std 0.00956 0.06581 0.09301 0.07401 0.11733 0.09067 Nboots 500 500 500 500 500 500 Min 0.00478 0.33369 0.24901 0.10480 0.18697 0.22267 HL 0.01491 0.46615 0.38521 0.29104 0.34874 0.36160 Me 0 0 0 0 0 0 •d 02013 51186 44552 33899 40906 42129 0 0 0 0 0 0 HH .02714 .55525 .51077 .38767 .48065 .48415 Me 0 0 0 0 1 0 X 06065 72839 84992 63089 * 04803 79317 model: HPDM a significant difference in the Willmott - d Model Avg MAX 0.97037 ADMS 0.59000 AMOD02161 0.61483 HPDM ! 0.64662 ISCST3 0.59026 AMODNoObZi 0.60632 t-value compared to the base 1 ADMS 1.1011 2 AMOD02161 0.5605 4 ISCST3 0.7755 5 AMODNoObZi 0.6283 Degrees of freedom (DF) : 28 (at the 90% confidence level value > metrics 1.701 indie between the ates that there is a base model and the model being tested) Results Std 0.01428 0.05730 0.06993 0.07320 0.08255 0.07067 Nboots 500 500 500 500 500 500 Min 0.91508 0.41442 0.42199 0.46643 0.33910 0.42730 HL 0.96186 0.55072 0.56497 0.59056 0.53556 0.55701 Med 0 0 97311 59026 0.61813 0 0 0 64853 59158 60583 0 0 0 0 0 0 HH .98161 .62879 .66311 .69900 .65071 .65867 Max 0 0 99403 76611 0.80535 0.89654 ! 0 78874 0.82291 model: HPDM a significant difference in the Average of Observed Values Model Avg MAX 49.65137 ADMS 49.65137 AMOD02161 49.65137 HPDM 49.65137 ISCST3 49.65137 AMODNoObZi 49.65137 value > metrics 1.701 indie between the ites that there is a base model and the model being tested) Results Std 2.83214 2.83214 2.83214 2.83214 2.83214 2.83214 Nboots 500 500 500 500 500 500 Min 42.61565 42.61565 42.61565 42.61565 42.61565 42.61565 HL 47.59573 47.59573 47.59573 47.59573 47.59573 47.59573 Med 49 49 49 49 49 49 72298 72298 72298 72298 72298 72298 51 51 51 51 51 51 HH .62337 .62337 .62337 .62337 .62337 .62337 Max 59 59 59 59 59 59 85370 85370 85370 85370 85370 85370 Figure 6.2. A portion of the output file with the statistical measures for the ASTM analysis. The "best performing model" is indicated by an asterisk or exclamation point for each statistical measure. -20- ------- To assess whether or not there is a significant difference in the statistical measure between the "best performing model" and the other models, the bootstrap results for the statistical measure are used to compute the t-value(s). The null hypothesis that the average of the model bootstrap differences is greater than 0 can be tested for a given confidence level (ASTM, 2000). The t-values are shown immediately below the statistical measures in Figure 6.2. Examining the Normalized Mean Square Error in Figure 6.2, HPDM is the "best performing model," but there is no statistical difference at the 90% confidence level between HPDM and both versions of AERMOD and ISCST3. However, there is a significant difference between HPDM and ADMS. The other two output files are a file of bootstrap results and results by regime. Each contains different information that may be useful in interpreting the results. The data used to compute the t-value(s) come from the bootstrap file. The file of bootstrap results is divided into five parts: 1) Run summary of the information provided in the control file; 2) Averages and standard deviations of observed and modeled concentrations by regime; 3) Averages and standard deviations of the fractional bias and absolute fractional bias for each regime. The definitions of fractional bias can be found in the discussion of the ASTM standard (ASTM, 2000); 4) Averages computed for each boot and regime; 5) Comparison statistics for each boot; for each statistical measure. In this example, there are nearly 21,000 records in the output file. Due to the length of the initial records and the number of records in the file, a sample is not shown here. Figure 6.3 shows a sample of the information by regime. The standard deviation of the concentration values along the arc is listed first. Then for each experiment-arc combination an analysis is provided of the observed concentration values satisfying the criterion of being within -0.67 < y/0y < 0.67. Ci is a concentration value for this arc that satisfies the requirements of NFILTER and Cavg is the average of all N values with zero values excluded for this regime. The asterisk next to a Ci value indicates this value is the maximum observed concerntation found anywhere along the arc. There can be multiple occurrences for an experiment-arc pair depending on the value of NFILTER and the number values that are within ±0.67y/0y. Some additional statistics appear below this grouping and are based on the nonzero values. These statistics are followed by similar results where NFILTER was applied in the modeling. In this example, NFILTER was set to one, so only one near-centerline concentration is selected for analysis. -21- ------- Process Regime : Sy(deg) : 19.10201 i of N Exp Arc 1 27 2 3 2 27 2 3 3 27 2 3 4 27 3 3 5 27 19 3 6 27 19 3 7 27 19 3 8 27 20 3 9 27 20 3 10 27 20 3 11 27 21 3 12 27 21 3 13 27 28 3 14 27 28 3 15 27 28 3 16 27 32 3 17 27 32 3 18 27 33 3 19 27 33 3 20 27 33 3 21 27 34 3 22 27 34 3 23 27 113 3 24 27 113 3 25 27 114 3 26 27 116 3 27 27 116 3 Avg Std Ci Ci 54.0696 40.2394 1 Ci 48.5830 92.7125* 35.4251 78.5425* 34.5494* 26.8240 30.2575 53 .8627 108.7983* 72.9614 148.1799* 134.4754 0.0000 0.0000 30.5376* 60.6452* 56.7742 72.4731 109.6774* 68.3871 21.7204 13.7634 5.5794 5.7940 3.2189 81.3305* 64.8069 GeoAvg Ci/Cavg 0.7105 Cavg 58.3952 58.3952 58.3952 58.3952 58.3952 58.3952 58.3952 58.3952 58.3952 58.3952 58.3952 58.3952 58.3952 58.3952 58.3952 58.3952 58.3952 58.3952 58.3952 58.3952 58.3952 58.3952 58.3952 58.3952 58.3952 58.3952 58.3952 GeoAvg Ci/Cavg 2.6809 Ci/Cavg 0.8320 1.5877 0.6066 1.3450 0.5916 0.4594 0.5182 0.9224 1.8631 1.2494 2.5375 2.3029 0.0000 0.0000 0.5229 1.0385 0.9722 1.2411 1.8782 1.1711 0.3720 0.2357 0.0955 0.0992 0.0551 1.3928 1.1098 Results from Subroutine Filter i of N Exp Arc 1 12 2 3 2 12 3 3 3 12 19 3 4 12 20 3 5 12 21 3 6 12 28 3 7 12 32 3 8 12 33 3 9 12 34 3 10 12 113 3 11 12 114 3 12 12 -116 3 Avg Std Ci Ci 58.5942 44.8805 Process Regime : Sy(deg) : 19.13367 i of N Exp Arc 1 30 2 4 2 30 2 4 Ci 92.7125* 78.5425* 26.8240 108.7983* 134.4754 0.0000 56.7742 109.6774* 21.7204 5.5794 3 .2189 64.8069 GeoAvg Ci/Cavg 0.6298 2 Ci 66.1943* 41.9028 Cavg 63.9209 63 .9209 63 .9209 63.9209 63 .9209 63 .9209 63 .9209 63 .9209 63 .9209 63.9209 63 .9209 63.9209 GeoAvg Ci/Cavg 3.3111 Cavg 54.5058 54.5058 Ci/Cavg 1.4504 1.2287 0.4196 1.7021 2.1038 0.0000 0.8882 1.7158 0.3398 0.0873 0.0504 1.0139 Ci/Cavg 1.2144 0.7688 Figure 6.3. Sample of the output by regime. -22- ------- SECTION 7 REFERENCES ASTM, 2000: Standard Guide for Statistical Evaluation of Atmospheric Dispersion Model Performance. D 6589-00. American Society for Testing and Materials, West Conshohocken, PA. Cox, W. M. and J. A. Tikvart, 1990: A Statistical Procedure for Determining the Best Performing Air Quality Model, Atmospheric Environment, 24A, 2387-2395. Fox, D. G., 1981: Judging Air Quality Model Performance, Bulletin of the American Meteorological Society, 62, 599-609. Murray, D. R. and N. E. Bowne, 1988: Urban Power Plant Plume Studies. EPPJ Report No. EA- 5468, Research Project 2736-1, Electric Power Research Institute, Palo Alto, CA. Oreskes, N., K. Shrader-Frechette and K. Belitz (1994): Verification, validation, and confirmation of numerical models in the earth sciences. Science, 263: 641-646. -23- ------- APPENDIX DIRECTORY STRUCTURE OF DISTRIBUTION CD ASTM Root Directory Indianapolis ASTM Evaluation Dispersion Modeling ADMS AERMOD AERMOD_NoObZi HPDM ISCST3 Meteorology ADMS AERMOD AERMOD_NoObZi HPDM ISCST3 Kincaid ASTM Evaluation Dispersion Modeling ADMS AERMOD AERMODJTurb HPDM ISCST3 Meteorology ADMS AERMOD AERMODJTurb HPDM ISCST3 MixHts -24- ------- PrairieGrass ASTM Evaluation Dispersion Modeling ADMS AERMOD HPDM ISCST3 Meteorology ADMS AERMOD HPDM ISCST3 MixHts Documentation Software -25- ------- TECHNICAL REPORT DATA (Please read Instructions on reverse before completing) 1. REPORT NO. EPA-454/R-03-006 3. RECIPIENTS ACCESSION NO. 4. TITLE AND SUBTITLE A Program for the Statistical Evaluation Of Point Source Dispersion Models Using ASTM D 6589 5. REPORT DATE August 2003 6. PERFORMING ORGANIZATION CODE 7. AUTHOR(S) 8. PERFORMING ORGANIZATION REPORT NO. 9. PERFORMING ORGANIZATION NAME AND ADDRESS Pacific Environmental Services, Inc, a MACTEC Company Research Triangle Park, NC 10. PROGRAM ELEMENT NO. 11. CONTRACT/GRANT NO. 12. SPONSORING AGENCY NAME AND ADDRESS Office of Air Quality Planning and Standards Emissions, Monitoring, and Analysis Division U.S. Environmental Protection Agency Research Triangle Park, NC 27711 13. TYPE OF REPORT AND PERIOD COVERED Technical Report 14. SPONSORING AGENCY CODE 15. SUPPLEMENTARY NOTES 16. ABSTRACT The ASTM published a Standard Guide D 6589 for Statistical Evaluation of Atmospheric Dispersion Model Performance, which provides a procedure for assessing performance of atmospheric transport and diffusion models to predict the "average centerline"concentration values of a chemical released from a point source. The EPA has developed a FORTRAN program entitled "Design 2" that implements this procedure. This report describes the input files and formats of these input files for running Design 2. The report describes the information provided by the various output files created by Design 2. The report does not describe the actual statistical procedures of Standard Guide D 6589, as these are documented in detail within this Guide's Annex. 17. KEY WORDS AND DOCUMENT ANALYSIS DESCRIPTORS b. IDENTIFIERS/OPEN ENDED TERMS c. COSATI Field/Group Meteorological Data Air Dispersion Models Statistical Model Evaluation 18. DISTRIBUTION STATEMENT Release Unlimited 19. SECURITY CLASS (Report) Unclassified 21. NO. OF PAGES 34 20. SECURITY CLASS (Page) Unclassified 22. PRICE EPA Form 2220-1 (Rev. 4-77) PREVIOUS EDITION IS OBSOLETE ------- United States Office of Air Quality Planning and Standards Publication No. EPA-454/R-03-006 Environmental Protection Planning and Standards August 2003 Agency Research Triangle Park, NC 27711 ------- |