Evaluation of CALPUFF, HPAC, and VLSTRACK with the Dipole Pride 26 Field Data

The CALPUFF (version 5.0), HPAC (version 3.1), and VLSTRACK (version 3.0) dispersion models are evaluated with the Dipole Pride 26 (DP26) field data. DP26 is relatively unique in terms of its scale and data coverage. It involves instantaneous releases of SF6, where concentrations are measured along three sampling lines that are roughly 5, 10, and 20 km from the source. Each line extends about 10 km and has 30 samplers. Meteorological data are measured at eight surface, one radiosonde, and two pibal stations.

Because the three models are quite different in various aspects, such as data requirements and wind field generation, the framework for comparison is carefully designed so that an objective, meaningful evaluation can be performed. This mainly involves the use of the same observed meteorology and modeling domain. Whenever applicable, default model options are also chosen. However, it is recognized that, despite of the efforts, there are still fundamental inter-model differences that cannot be reconciled. Two additional HPAC configurations, in addition to the base run, are considered to study the effects on model predictions due to uncertainty in input data.

The maximum SF6 dose anywhere along a sampling line and the summation of SF6 doses over all samplers along a sampling line are chosen for model evaluation. A systematic model evaluation methodology is used to measure model performance. Overall, the performance for CALPUFF and HPAC is comparable. VLSTRACK tends to overpredict and has a larger scatter. For the maximum dose, the three models yield values of the geometric mean bias (MG) corresponding to 50% overprediction to 40% underprediction, and values of the geometric variance (VG) corresponding to a factor of 4 to 10 random scatter. The fraction of predictions within a factor of two of observations (FAC2) is about 40 to 50% for CALPUFF and HPAC, and 20% for VLSTRACK. Model performance deteriorates slightly for the cross-line summed dose, with values of MG corresponding to 15% underprediction to a factor of two overprediction, and values of VG corresponding to a factor of 4 to 19 random scatter. This is because all models tend to predict a larger lateral dimension for puffs, thus a larger predicted value for the summed dose. FAC2, however, remains relatively unchanged for the three models for the summed dose. The results for various HPAC runs, although comparable, are often significantly different at the 95% confidence limits.

Some suggestions for future work are also given, such as the use of a common gridded wind field to isolate the influence of meteorology, and the study of the effects of stochastic uncertainty versus the uncertainty due to the choice of input data.