Documentation of the Evaluation of CALPUFF and Other Long ...

Documentation of the Evaluation of CALPUFF and Other Long ... Documentation of the Evaluation of CALPUFF and Other Long ...

20.04.2013 Views

Figure C‐ ‐3. Global model m perforrmance statistics for nine HYSPLIT INITD sensitivity tests for CAPTEX RRelease 3. 4

The final panel in Figure C‐3 (bottom right) displays the overall RANK statistic. The RANK statistics orders the model performance of the HYSPLIT INITD configurations as follows: 1. INITD1 (1.25) 2. INITD2 (1.21) 3. INITD104 (1.19) 4. INITD130 (1.19) 5. INITD0 (1.18) 6. INITD103 (1.15) 7. INITD4 (1.14) 8. INITD3 (1.11) 9. INITD140 (1.10) The RANK performance statistics results presented above raise some interesting questions about the RANK metric. The puff based configurations (INITD1 and INITD2) are the highest ranking with scores using the RANK metric with values of 1.25 and 1.21 respectively. However, each of these options had the worst (highest) NMSE and FB scores, while puff‐particle configurations ranking slightly less using the RANK metric (1.1 to 1.19) have NMSE scores that are much better (only one‐third) those for the puff configurations as well as slightly lower FB scores. On the basis of RANK scores, the INITD1 and INITD2 configurations are the best performing, but based upon other model performance statistics that are not included as the four statistical metrics that make up the RANK metric (i.e., PCC, FB, FMS and KSP), the puff‐ particle hybrid configurations are better performing. Thus care must be taken in interpreting model performance based solely on the RANK score and its use in performing model intercomparisons and we recommend examining the whole suite of statistical performance metrics, as well as graphical representation of model performance, to come to conclusions regarding model performance. C.2.3 HYSPLIT SPATIAL STATISTICS FOR CAPTEX RELEASE 5 Figure C‐4 displays the spatial model performance statistics for the HYSPLIT INITD sensitivity tests for CAPTEX Release 5. Overall, the spatial performance for this experiment is very similar to the results obtained from the ETEX INITD sensitivities for HYSPLIT. The puff configurations (INITD1 and INITD2) exhibited the poorest performance across all of the spatial statistics. INITD2 had the poorest FMS score with 5%, followed by INITD1 with 9.6%. INITD3 had the best FMS score of 19.66%, but less than 2% separated all of the remaining particle and puff‐particle INITD configurations. The particle mode (INITD0) exhibited the best TS with 24.4% with less than 1.5% separating INITD103, 130, and 140 from INITD0. Consistently, the puff configurations exhibited the lowest TS among the nine configurations, both with 7.9%. 5

The final panel in Figure C‐3 (bottom right) displays <strong>the</strong> overall RANK statistic. The RANK<br />

statistics orders <strong>the</strong> model performance <strong>of</strong> <strong>the</strong> HYSPLIT INITD configurations as follows:<br />

1. INITD1 (1.25)<br />

2. INITD2 (1.21)<br />

3. INITD104 (1.19)<br />

4. INITD130 (1.19)<br />

5. INITD0 (1.18)<br />

6. INITD103 (1.15)<br />

7. INITD4 (1.14)<br />

8. INITD3 (1.11)<br />

9. INITD140 (1.10)<br />

The RANK performance statistics results presented above raise some interesting questions<br />

about <strong>the</strong> RANK metric. The puff based configurations (INITD1 <strong>and</strong> INITD2) are <strong>the</strong> highest<br />

ranking with scores using <strong>the</strong> RANK metric with values <strong>of</strong> 1.25 <strong>and</strong> 1.21 respectively. However,<br />

each <strong>of</strong> <strong>the</strong>se options had <strong>the</strong> worst (highest) NMSE <strong>and</strong> FB scores, while puff‐particle<br />

configurations ranking slightly less using <strong>the</strong> RANK metric (1.1 to 1.19) have NMSE scores that<br />

are much better (only one‐third) those for <strong>the</strong> puff configurations as well as slightly lower FB<br />

scores. On <strong>the</strong> basis <strong>of</strong> RANK scores, <strong>the</strong> INITD1 <strong>and</strong> INITD2 configurations are <strong>the</strong> best<br />

performing, but based upon o<strong>the</strong>r model performance statistics that are not included as <strong>the</strong><br />

four statistical metrics that make up <strong>the</strong> RANK metric (i.e., PCC, FB, FMS <strong>and</strong> KSP), <strong>the</strong> puff‐<br />

particle hybrid configurations are better performing. Thus care must be taken in interpreting<br />

model performance based solely on <strong>the</strong> RANK score <strong>and</strong> its use in performing model<br />

intercomparisons <strong>and</strong> we recommend examining <strong>the</strong> whole suite <strong>of</strong> statistical performance<br />

metrics, as well as graphical representation <strong>of</strong> model performance, to come to conclusions<br />

regarding model performance.<br />

C.2.3 HYSPLIT SPATIAL STATISTICS FOR CAPTEX RELEASE 5<br />

Figure C‐4 displays <strong>the</strong> spatial model performance statistics for <strong>the</strong> HYSPLIT INITD sensitivity<br />

tests for CAPTEX Release 5. Overall, <strong>the</strong> spatial performance for this experiment is very similar<br />

to <strong>the</strong> results obtained from <strong>the</strong> ETEX INITD sensitivities for HYSPLIT. The puff configurations<br />

(INITD1 <strong>and</strong> INITD2) exhibited <strong>the</strong> poorest performance across all <strong>of</strong> <strong>the</strong> spatial statistics.<br />

INITD2 had <strong>the</strong> poorest FMS score with 5%, followed by INITD1 with 9.6%. INITD3 had <strong>the</strong> best<br />

FMS score <strong>of</strong> 19.66%, but less than 2% separated all <strong>of</strong> <strong>the</strong> remaining particle <strong>and</strong> puff‐particle<br />

INITD configurations. The particle mode (INITD0) exhibited <strong>the</strong> best TS with 24.4% with less<br />

than 1.5% separating INITD103, 130, <strong>and</strong> 140 from INITD0. Consistently, <strong>the</strong> puff configurations<br />

exhibited <strong>the</strong> lowest TS among <strong>the</strong> nine configurations, both with 7.9%.<br />

5

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!