Improving The Accuracy Of Dispersion Models
Ever since dispersion models were first developed over 30 years ago, their accuracy has been subject to scrutiny. If automobile speedometers are accurate within five percent and orifice meters are accurate within about one percent, why can’t dispersion models be just as accurate? The early documents on dispersion modeling referred to accuracy as a “factor of 3.”(1) By the 1970’s, dispersion modeling started to be used as a tool for assessing compliance with ambient air quality standards. The definition of accuracy shifted from the comparison of predicted to measured concentrations for the period of an hour or so to the ability to predict the magnitude of the highest expected hourly or daily concentrations during the period of a year. The U. S. EPA commissioned some studies, and one of these included this cumulative frequency distribution, comparing measured and computed hourly concentrations.(2) The plot, Figure 1, is merely the hourly values arranged in ascending order. The vertical ordinate is the log of concentrations, while the horizontal is the log of the frequency. Since there were 8,079 observations, the 8th highest would be at the 99.9 percentile, the 81st highest the 99.0 percentile, and the 162nd highest the 98.0 percentile. No connection exists between the time of observation and computation. For example, the 50th highest value may have been observed on March 21, while the 50th highest computation may have occurred on November 5. Note that the prediction and the observed values agree quite well above 98 percentile, but agreement is quite poor below this percentile.