Relative importance of model uncertainty components is compared. Uncertainty in the dispersion model results generally comes from the following three components: (1) random turbulence in the atmosphere, (2) input data errors, and (3) model errors and uncertainties. Because of the turbulent nature of the atmosphere, the same meteorological conditions do not always lead to the same pollutant concentrations. Input data errors can be due to instrument errors or unrepresentative instrument siting. Model errors and uncertainties can be due to factors such as inadequate physical formulation or uncertainties in the parameters used in these physical formulations. This paper investigates the relative contributions from random turbulence and from input data errors to the uncertainty in concentration predictions.

The Dipole Pride 26 (DP26) field data and the Second-order Closure Integrated Puff (SCIPUFF) model are used to address the above question. DP26 involved instantaneous point releases of sulfuric hexafluoride (SF₆), where 15-min average SF₆ concentrations were measured at three sampling lines, each with 30 whole-air samplers. The three sampling lines were roughly 5, 10, and 20 km away from the source. Meteorological data were measured by a dense network of eight surface, one radiosonde, and two pibal stations. Because of its second-order turbulence closure formulation, SCIPUFF is one of the first operational dispersion models that can predict both the mean and variance of concentration fields.

In this study, the concentration fluctuation predicted by SCIPUFF is used to estimate the uncertainty due to random turbulence. The data withheld technique, instead of a full-blown Monte Carlo uncertainty analysis, is used to estimate the uncertainty due to input data errors, where sensitivity runs are made with meteorological data from one surface station withheld at a time. Uncertainty in the source term is not included because it is well-defined for DP26. The results, based on the maximum dosage (concentration integrated with time) along a sampling line, show that random turbulence is more important than is input data errors in contributing to model uncertainty at the closest sampling line, and less important at the farthest sampling line. Implications of the results are discussed.