Use of Monte Carlo uncertainty analyses to evaluate differences in observed and predicted ozone concentrations

Abstract

It is now possible to carry out Monte Carlo uncertainty studies with large photochemical grid models such as UAM-IV. The current study uses information on the effects of uncertainties in 109 input variables on UAM-IV model predictions of maximum daily one-hour averaged ozone concentrations at 29 monitoring sites in the New York City UAM domain for the 8 July 1988 ozone episode. Fifty Monte Carlo runs were made with the UAM-IV model making use of simple random sampling from assumed distributions for each input variable. The resulting sets of 50 predicted values of ozone concentration allowed Cumulative Distribution Functions (CDFs) to be determined for each site. These CDFs were typically log-normal with a standard deviation of about -10% or -20%. Note that these uncertainties are due only to uncertainties in input variables and do not include contributions due to errors in model physics or stochastic fluctuations. The Lewellen-Sykes-Parker (LSP) model evaluation methodology has been applied to the predictions of maximum daily one-hour averaged ozone concentrations at the 29 monitoring sites using the CDFs from the Monte Carlo study. The assumption is tested that the model prediction is correct, in the sense that the observations are not significantly different from the predictions. The hypothesis is tested that the Monte Carlo CDFs of predicted concentrations represent a set of 50 realisations from the population of observed concentrations. Therefore the Monte Carlo results were used to generate many possible realisations of CDFs of model residuals, i.e. observed minus predicted concentration, which were then compared with the actual CDF of model residuals at the 29 sites. It was found, as expected, that the single actual CDF line for the model residuals covered a slightly broader range (a standard deviation of about 20% compared to about 10%) than the 95% confidence intervals for the CDFs from the 50 Monte Carlo runs.