Developing an Updated Statistical Ozone Model for Operational Air Quality Forecasting in the Philadelphia Metropolitan Area

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Sunday, 4 January 2015
Alexandria J. Herdt, Pennsylvania State University, University Park, PA; and W. F. Ryan and A. K. Huff

An updated statistical linear regression ozone model was developed to aid air quality forecasters in generating accurate maximum 8-hour average ozone concentration forecasts for the Philadelphia metropolitan area. A statistical ozone model was previously utilized in the Philadelphia area, but it was discontinued in 2011 since a steady decline in emissions of ozone precursors made the model, trained on data prior to 2003, progressively less skillful. Recent stabilization of ozone precursors provides an opportunity to explore an update the statistical ozone model for Philadelphia. The statistical ozone model calculates a maximum 8-hour average mixing ratio in parts per billion (ppb) to be used as forecast guidance for “tomorrow's” operational ozone forecast. The foundation of the model was a historical dataset of over sixty key parameters for the period 2004–2012, which were selected based on local climatology and atmospheric chemistry. A series of rejections and transformations were made to the historical dataset using the SYSTAT general purpose statistical software program to determine the 25 variables that were most highly correlated with observed maximum 8-hour average ozone concentrations. Two versions of the statistical model were developed: a stepwise regression and a handmade regression. The stepwise regression model was generated by SYSTAT and included the 7 parameters from the subset of 25 that gave the most accurate and consistent predictions of maximum 8-hour average ozone concentrations. The handmade statistical model was created based on expert knowledge of the 6 parameters that have historically given the most accurate and consistent predictions of maximum 8-hour average ozone concentrations. The 13 parameters used among the two models consisted of both predicted and observed data. Both models were tested during a 3-week pilot study period during the height of the ozone season in Philadelphia to determine their forecast skill in an operational setting. The two versions of the statistical models were run separately using predicted meteorological values from the NAM and the GFS, as well as from expert synthesis of the models; these 5 parallel model runs were considered ensembles of the statistical models. By all measures, the computationally derived stepwise model outperformed the handmade model. The stepwise model had lower bias and higher accuracy during the entire pilot period. The stepwise model also had higher skill on days when 8-hour average ozone concentrations exceeded the National Ambient Air Quality Standard of 75 ppb. In particular, the stepwise model predicted fewer false alarms, which was when an ozone exceedance was forecasted but not observed. These initial results suggest that the stepwise model has the potential to be a useful tool for operational ozone forecasting in Philadelphia. Both versions of the updated model will be tested during the entire 2015 ozone season in order to assess their skill during the full range of summer meteorological and emissions conditions that impact local ozone concentrations and determine if the results of the pilot study prove consistent for an entire season. The superior performance of the stepwise model is most likely due to recent changes in regional emissions that have shifted the predictor variables of interest to produce high ozone events, which suggests that using SYSTAT to select the most highly correlated predictor variables will yield a more accurate model instead of relying on expert judgment. This study implies that other forecast regions where statistical ozone models have also been discontinued may benefit from using a statistical software program to update them.