Until recently, grid-based photochemical modeling systems have been used predominantly to simulate historic high-ozone events (typically lasting a few days), i.e. for hindcasting. Recent developments in computing technology have now made it possible to apply these modeling systems using mesoscale meteorological forecasts to provide real-time air quality forecasts. However, real-time numerical air quality forecasting is still in its infancy, and operational air quality predictions by federal, state and local agencies are based on a combination of weather predictions, statistical analyses, and expert judgment. Numerical models provide the potential benefit of higher spatial and temporal resolution than the methods currently used, but before photochemical models can be more widely used for real-time air quality predictions, it is necessary to evaluate the quality of these predictions and estimate the modeling systems' uncertainty.

In this paper, ozone predictions from a hindcast simulation for the summer of 1995 generated by the RAMS/UAM-V modeling system are first analyzed to establish a "best case" scenario for model performance since in this case meteorological observations were assimilated for this hindcast simulation. The results indicate that the modeling system's forecasting skill is lower than that of air quality forecasts generated by current established methods (e.g., statistical analysis, weather forecasts, expert judgment). Because of the presence of inherent uncertainty associated with the outputs of grid-based models, a method to transform the deterministic model predictions into a probabilistic form that takes into account known sources of model uncertainty is presented in this paper. Specifically, this method was developed based on past model evaluation studies using spectral decomposition techniques that found poor agreement between observations and model simulations (even in hindcast mode) on short time scales. Thus, the magnitude of the fluctuations on these time scales is estimated from observations and is used as a measure of uncertainty associated with real-time air quality forecasting using numerical models. The method is applied to the summer 1995 hindcast simulation for illustration.

A companion paper describes the application to a real-time forecasting project for the summer of 2001 (Cai et al., "An Experimental Air Quality Forecast Modeling System (AQFMS) for the Northeast United States: A Demonstration Study", abstract submitted for presentation at the Symposium on Observations, Data Assimilation, and Probabilistic Predictions). The results are discussed in terms of strengths and weaknesses associated with numerical air quality predictions. The need to direct further research towards quantifying model uncertainty in an effort to increase both public acceptance and the chances of long-term success of numerical air quality predictions is also discussed.