The Australian Air Quality Forecasting System: Exploring First Steps Towards Determining The Limits of Predictability For Short-Term Ozone Forecasting

Abstract

Physical parameterisations of turbulent transfer processes in the atmospheric boundary layer, such as the stability parameterisations developed by Joost Businger, and recent advances in computing capabilities, have been important factors leading to the emergence of operational, numerical air quality forecasting systems. The present paper investigates the performance of the Australian Air Quality Forecasting System (AAQFS) in forecasting the peak 1 h ozone for the current or next day. These 24/36 h forecasts are generated for the Sydney and Melbourne regions and issued twice daily. Quantitative evidence is presented of the potential for the AAQFS to provide accurate numerical air quality forecasts. A second goal is to provide an initial benchmark for investigating the limits of predictability for air quality in the Sydney and Melbourne regions by looking at the dependence of the forecasts on the domain spatial scale (while maintaining the same model grid resolution), the starting time and length of the forecast (0000 UTC starts are 36-h forecasts and 1200 UTC starts are 24-h forecasts), and the sophistication of the photochemical mechanism (simple chemistry, Generic Reaction Set (GRS) and complex chemistry, Carbon Bond IV (CBIV)). The probability of detection by the forecast model is much better than persistence, showing considerable skill. The normalised bias, in general, decreases going from regional scale to sub-regional scale and becomes negative at the station scale. In Melbourne the gross error increases as the domain spatial scale decreases, but in Sydney there is a dip in the error at the sub-regional scale due to a sampling artifact. Better results are obtained at the smaller domain scales for 1200 UTC forecasts in Sydney. These are attributed to the shorter forecast period and secondarily to greater model spin-up effects at 0000 UTC. In Melbourne the results are ambiguous. Similar conclusions are derived from scatter plots of forecasts versus observations. Dividing the scatter plots into four sections by plotting vertical and horizontal lines (at 60 ppb) forms contingency tables for categorical forecasting. These plots show the increase in missed forecasts due to underprediction and the decrease in the number of extreme events detected as the spatial scale decreases. A comparison of the highly condensed GRS photochemical mechanism with the comprehensive CBIV mechanism indicates that, in general, GRS performs well for predicting ozone in urban situations provided that the background concentrations are appropriately specified. The potential to improve the forecasts at the smaller spatial scales, particularly for extreme events at high ozone concentrations, may require moving to a more complex mechanism as computer resources become available.