A system for forecasting future air quality cannot, by itself, solve the problems described above. Forecasts, if they are reliable and sufficiently accurate, can however play an important role as part of an air quality management system working in concert with more traditional emissions-based approaches. The applications of air quality forecasts fall into the following broad areas:

• Health Alerts – Many cities currently provide warnings to the public when air pollution levels exceed specified levels. The more reliable the forecast is the more effective it is. These warnings are directed at specific populations that are particularly sensitive to air pollution (e.g., asthmatics). Interest in finding innovative ways to protect these individuals has heightened in recognition of lack of a discernable health threshold for exposure to ozone or fine particles, which implies that no level of emissions reduction will protect all individuals.
• Supplementing Existing Emission Control Programs – In many parts of the country, the air quality standards are exceeded only infrequently, a few days out of the year. The availability of reliable air pollution forecasts affords local environmental regulators the option of “on demand” or intermittent emission reductions on those days, thus avoiding the high cost of continuous emission control. This approach is currently being successfully employed in several areas of the country and could be expanded were reliable forecasts available. Many cities also offer free access to public transportation on “ozone alert” days to reduce automobile emissions. The accuracy of these forecasts is critical due to the high cost associated with these programs.
• Operational planning – Regional haze can impair and even endanger activities such as private and commercial aviation. Aerial photography and visits to many National Parks are significantly impacted by the presence of haze. A reliable visibility forecast could improve safety and efficiency by permitting the scheduling of these activities during the most favorable periods. The U.S. Forest Service (U.S.F.S.) is planning a 10-fold increase in prescribed burns. Since these activities are regulated under the Clean Air Act, the U.S.F.S. will have to demonstrate to local regulators that they can schedule these burns so that no National Ambient Air Quality Standards will be violated, requiring some form of air quality forecast.
• Emergency response – Wildfires consumed more than 4 million acres of forest in the United States during 2000. The vast amount of smoke generated by these burning forests affects the visibility in the area that can cause accidents, increase traffic congestion and even jeopardize aviation safety. The availability of reliable smoke forecasts offers rerouting options for automobiles and air traffic to reduce the possibility of accidents. The information provided by these forecasts can also provide an initial assessment of the impact of these wildfires.

A wide variety of techniques, ranging from the simple to the complex, have been used to produce air quality forecasts. To date, most of these efforts have focused on producing 1-to 3-day ozone forecasts. The techniques that have been used to produce these forecasts are described in a recent report [U.S. EPA, 1999]. The techniques used to forecast ozone concentrations are representative of those that can, or could, be used for other pollutants. They fall into three broad categories:

Climatology – The use of climatology to predict air quality is based on the assumption that the past is a good predictor of the future. This approach relies on the association of elevated pollution levels with specific meteorological conditions. The application can be as simple as assuming persistence (i.e., if pollution levels are high today they will also be high tomorrow) or can involve the development of complex weather typing schemes (i.e., identifying recurring weather patterns that are accompanied by high pollution levels) to forecast air quality. These approaches are usually used to predict exceedances of specific thresholds not ambient concentrations. These approaches do, however, have the advantage of being reasonably simple and inexpensive to implement and operate.

Statistical Methods – The association between specific meteorological parameters and air quality can be quantified using a variety of statistical techniques. In fact, these are probably the most common techniques in use for ozone forecasting. In their survey, EPA, 1999 has identified three statistical approaches that are in use:

• Classification and Regression Tree (CART) – This technique uses specialized software to identify those variables (meteorological or air quality) that are most strongly correlated with ambient pollution levels. These variables are then used to predict future pollution levels based on current air quality and forecasted meteorology.
• Regression analysis – The association between pollutant levels and meteorological and aerometric variables can be quantified by analyzing historical data sets using standard statistical analysis packages. The resultant multi-variant linear regression equation can be used to forecast future pollution levels.
• Artificial Neural Networks – Another way of analyzing historical data is to identify atmospheric parameters that influence air quality and quantify that association through the application of adaptive learning and pattern recognition techniques, such as neural networks. Neural networks are intended to mimic the way the human brain recognizes recurring patterns. Networks have been developed that identify weather patterns that are associated with elevated ozone levels (see U. S. EPA, 1999 and references therein). Presumably, the same technique could be applied to other pollutants.

These approaches, while more complex than the ones discussed in the previous group, are reasonably simple to develop and use, requiring only modest computing resources and specialized knowledge.

Three Dimensional (3-D) Models – Although the techniques described above have many strong points, they have a common weakness. They assume a certain amount of stability in terms of the processes that affect air quality. Any change in emissions or climate (short and long-term) will serve to diminish the skill of these techniques. One way around this problem is to employ a more deterministic approach to the prediction of air quality. Deterministic 3-D air quality models seek to mathematically represent all of the important processes that affect ambient pollution levels. These models are actually comprised of several submodels that work together to simulate the emission, transport, and transformation of air pollution. Examples of submodels include:

• Emissions models – These models simulate the time-dependent, spatially distributed emissions of the pollutant in question, and/or (in the case of secondary pollutants such as O3) its precursors, from both anthropogenic and natural sources.
• Meteorological models – These models forecast meteorological conditions that determine transport and mixing and influence chemistry (solar intensity, temperature, humidity, etc.), emissions (e.g. temperature), and deposition. Trajectory models use the 3-D meteorology from these models in consort with emissions data to forecast ambient levels of reasonably unreactive pollutants like dust and smoke.
• Chemical models – These models use fundamental chemical kinetic rate parameters, spectroscopic properties, and thermodynamic relationships to simulate the transformation of primary (emitted) pollution into secondary pollution, including the composition and morphology (size distribution and optical properties) of aerosols.

Three-dimensional air quality models are classified as being either Lagrangian or Eulerian depending on the method used to simulate the time-dependent distribution of pollution concentrations. Lagrangian models follow individual air parcels over time using the meteorological field to advect and disperse the pollutants. This approach results in a computationally efficient system. However, it is difficult to properly characterize the interaction of a large number of individual sources when nonlinear chemistry is involved. Eulerian models use fixed grids (vertically and horizontally) and solve the appropriate chemical equations simultaneously in all cells, including exchange of pollutants between cells. Typically the computational requirements are reduced through the use of nested grids, with a coarse grid used over rural areas (where concentrations tend to be reasonably homogenous) and a finer grid used over urban areas (where concentration gradients tend to be more pronounced). These models can also accommodate a plume-in-grid treatment by performing a semi-Lagrangian calculation for large point sources (e.g., power plants) during the early stages of plume dilution. These models can produce three-dimensional concentration fields for several pollutants but require significant computational power and expertise.

Virtually all of the techniques described above start with a meteorological forecast. Therefore, the reliability of the air quality forecast is dependent on the reliability of the weather forecast. Weather forecasters use a number of tools to predict tomorrow's weather. Local forecasters will typically use the output from several different models in combination with local knowledge and experience to produce an accurate forecast. The same must be true for an air pollution forecast. A skilled forecaster will combine several of the techniques described above to ensure that the prediction is as accurate as possible.

As with a weather forecasting system, an air quality forecasting system must contain a compatible combination of two components. These components are a suite of predictive models/techniques tailored to specific needs of the customer community and an observation network capable of providing real-time measurements of atmospheric composition needed to initialize the models and evaluate the quality of the forecast.

Fig 1.1 Schematic showing the interrelation among the main elements of an air quality forecasting system. A well designed forecast system includes both a process for producing the forecast and an observing system to evaluate the quality of the forecast and identify areas where improvement is needed.

A comprehensive observing system must be an integral part of any forecast system that is developed to predict future air quality. It is only by comparing actual and forecast pollution levels that we can assign confidence to future forecasts and identify areas where improvements are needed.

Traditionally the evaluation of air quality models (whatever their form) has fallen into two broad categories:

Operational evaluation – An operational evaluation involves a direct comparison between the forecast pollutant fields and the observed pollutant distribution. For example, in the case of ozone the model predictions would be compared against the concentrations measured in the regulatory network, and some skill score calculated based on a point-by-point comparison at the monitoring sites. The existing regulatory network for ozone is well suited to such an evaluation. Unfortunately the situation for other pollutants (e.g., fine particles) is not as good since those monitoring networks are less dense and the instruments not as sensitive, or in some cases as selective, as is desired.

Diagnostic evaluation – An operational evaluation will tell you how close the model came to the correct answer; a diagnostic evaluation will tell you if you got the right answer for the right reason. As the name implies, a diagnostic evaluation requires the measurement of parameters (both meteorological and chemical) that control pollutant formation and distribution, not just the concentration of the pollutant that is being forecasted. To perform a diagnostic evaluation, concentrations of pollutant precursors and key intermediates need to be tracked to evaluate the performance of the emissions model and the chemical processor (if one is used), while meteorological parameters such as mixing height and winds aloft will aid in the evaluation of the meteorological processor.

As techniques for forecasting air quality improve and their use expands, we should not overlook the opportunity to use this process to improve our understanding of the processes that control the formation and distribution of air pollution. The meteorological research community has benefited enormously from the experience gained through an operational forecasting enterprise. The ongoing evaluation of the daily weather forecasts is used to identify areas of insufficient understanding and guide research.

The same opportunity exists for advancing the understanding of atmospheric processes that control ambient pollution levels. By evaluating the success of forecasts produced by different techniques we have the opportunity not only to evaluate the relative merits of these techniques but also to test our knowledge of key processes and identify areas where more information is needed.

Adams, R.M., J.D. Glyer, S.L. Johnson, and B.A. McCarl, A reassessment of the economic effects of ozone on U.S. agriculture, J. Air Pollution Control Assoc. 39, 960-968, 1989.

Banta, R.M., C.J. Senff, A.B. White, M. Trainer, R.T. McNider, R.J. Valente, S.D.Mayor, R.J. Alvarez II, R.M. Hardesty, D. Parrish, and F.C. Fehsenfeld, Daytime buildup and nighttime transport of urban ozone in the boundary layer during a stagnation episode, J. Geophys. Res., 103, 22519-22544, 1998.

Cannon, J.S., The Health Costs of Air Pollution: A Survey of Studies Published 1984-1989, American Lung Association, 1990.

Grell, G.A., S. Emeis, W.R. Stockwell, T. Schoenemeyer, R. Forkel, J. Michalakes, R. Knoche, and W. Seidl, Application of a multiscale, coupled MM5/chemistry model to the complex terrain of the VOTALP valley campaign, Atmos. Environ., 34, 1435-1453, 2000.