Daily air quality forecasts are used extensively throughout the country as a way to protect public health and to encourage reduction of emission-producing activities. As more cities begin forecasting air quality, new methods are needed to improve the reliability and accuracy of the forecasts. Such improvements can be obtained through the development of objective forecasting techniques. To this end, the U.S. Environmental Protection Agency (EPA) sponsored a project to develop forecasting tools to predict ozone concentrations for four U.S. cities: Columbus, Ohio; Minneapolis, Minnesota; and Nashville and Memphis, Tennessee.

To develop these tools, five years of historical meteorological and air quality data were compiled into a developmental data set for each city. Conceptual models were developed for each city using synoptic weather maps and the historical data sets. Predictor variables for the equations were selected using factor analysis to eliminate co-linear variables and statistical software was used to generate several possible linear regression equations. The predictor variables in each equation were checked for accessibility, accuracy, and consistent physical relationships between meteorological variables and ozone. Verification of the equations on small independent subsets of data showed that the equations for Columbus, Memphis, and Nashville predicted Air Quality Index categories as accurately as human forecasters, but human forecasters better predicted ozone concentrations. In addition, the equations tended to underpredict when observed ozone concentrations were actually high. Verification of the final equations for Minneapolis showed poor results, so forecasting guidelines were created as an expansion of the conceptual model for Minneapolis. The guidelines account for two limitations in the data set that statistical methods did not: transport from upwind regions and a low number of high ozone episodes. The forecasting guidelines were not tested against human forecasters; however, they correctly predicted high ozone on 10 of 13 days, and correctly predicted low ozone on 16 of 18 days.