A regression model to forecast the daily peak 8-h ozone concentration for the Louisville Metropolitan Statistical Area

A hybrid, nonlinear regression model was developed for the Louisville Metropolitan Statistical Area to forecast 8-h ozone concentrations. The hybrid model consisted of a standard model for use on most days, a "HI-LO" model designed for making predictions on days of expected high ozone, and an independent switching criteria for model selection. The standard and HI-LO models were calibrated on ozone concentration data covering 620 days over a five year period (1996--2000). The models included a total of 12 parameters consisting of meteorological data, deterministic properties, and derived meteorological products. The hybrid model had an overall R2 of 0.755, and the MAE for the hybrid model was 7.6 ppb. The model was comparable to a previously developed model for 1-h ozone concentration. The hybrid model demonstrated a good ability to accurately predict high ozone days when using the recommended 75 ppb alarm threshold. It produced a detection rate of 89%, while giving false alarms 29% of the time. Raising the alarm concentration to the National Ambient Air Quality Standard of 85 ppb diminished the detection rate of the model to 58%, while increasing the false alarm rate to 34%. Field test results are included for the month of May during the 2001 ozone season. Preliminary statistics indicate the model performed as expected. The forecast model errors were approximately equal to the prediction error. Several special studies investigating possible parameters for future model use were conducted. The parameter designated as OZ48 shows potential in improving model predictions on days following a regional accumulation of ozone in the atmosphere.