A dispersion/receptor model coupled using a genetic algorithm (GA) is validated using synthetic data and comparison with a least squares method. Defining sources of atmospheric pollution can be approached using either a backward-looking or a forward-looking model. A receptor model begins with monitored data and looks backward to apportion pollutant to its sources, given differential source emission characteristics. In contrast, a dispersion model begins with emission data and computes downwind transport and dispersion to predict the concentration at a receptor. This project combines the strengths of these two types of models through optimizing the calibration factor that ties them together. Meteorological data from various time periods are used to compute the expected dispersion for each averaging period. The calibration factors are then optimized to best match given receptor data. This coupled model is different from a standard receptor model since it includes known information on meteorological data and computes transport and dispersion of the pollutant to the receptor. When the calibration factor is multiplied by the computed dispersed emissions, it denotes the amount of pollutant from a particular source. The computed calibration factor can also be interpreted as the combined error of the input data plus the two coupled models, producing a total uncertainty. In addition, multiple calibration factors can be computed that each contain a portion of that uncertainty through contingencies, thus separating out the sources of uncertainty. Optimizing a single calibration factor can be accomplished by various techniques. However, once contingency becomes part of the problem, AI methods are required to obtain a solution. Those cases will no longer be solvable with standard minimization approaches since the cost function will embed a contingency for higher order matrices of calibration factors.

The coupled model optimization approach is validated using synthetic data for source emissions and meteorological conditions. The optimization is compared for two different methods: 1. a genetic algorithm, and 2. a standard least squares technique. Synthetic data is constructed for both a circular source configuration and for an actual source/receptor configuration using data obtained from Cache Valley, Utah. The GA performs as well as the least squares method but takes more CPU time. Thus, we have confidence that when this technique is applied to more difficult cases that it will produce a correct solution.

The GA coupled model is then applied to a synthetic case where the source characteristics change with time. In this case, a contingency applies to the calibration vector. We demonstrate that the GA can separate the uncertainty between multiple calibration vectors.