Ensemble Air Quality Forecasting Over the Pacific Northwest

Ensemble weather forecasts have been extensively evaluated over the past decade, and found to provide better accuracy than any single numerical model run. Different Numerical Weather Prediction (NWP) models usually perform better for different situations, and often their behavior cannot be anticipated. Hence, their combination into a multi-model ensemble is usually fruitful. NWP ensembles have been created with different ICs and/or BCs, different parameterizations within a single model (physics packages, parameter values), different numerics within a single model (finite difference approximations and solvers, grid resolutions, compiler optimizations), and different models, trying to take into account different sources of uncertainties.

The ensemble technique can potentially yield similar benefits to air-quality (AQ) modeling, because there are similar code complexity and constraints. Different AQ models can be better for different air-pollution episodes, also in ways that cannot always be anticipated. For AQ, the ensemble-mean can be created similarly with different inputs (background concentrations, emissions inventories, meteorology), different parameterizations within a single model (chemistry mechanisms, rate constants, advection and dispersion packages), different numerics within a single model (finite difference approximations and solvers, grid resolutions, compiler optimizations), and different models. Given the nonlinear nature of photochemical reactions, the ensemble spread may rapidly account for the uncertainties associated with each component of the modeling process.

Results of an AQ ensemble forecast system will be presented. The system includes the Community Multiscale Air Quality Model (CMAQ), driven by the Fifth-Generation NCAR / Penn State Mesoscale Model (MM5), and the Mesoscale Compressible Community Model (MC2). CMAQ is run with a resolution of 12 and 4 km. Furthermore, for each of the four mesoscale model/resolution combinations CMAQ is run four times with different settings, leading to sixteen different ensemble members. The spatial domain considered in the simulation includes the Lower Fraser Valley (LFV) of British Columbia. In this region, the Emergency Weather Net (EmWxNet) meteorological data and the Quality-Controlled AQ Data Set (from Environment Canada and the Greater Vancouver Regional District) are provided each day for several locations, and include hourly time series of meteorology, ozone, and particulate matter (PM). This data set allows extensive testing, in a wide range of meteorological scenarios and air-pollution episodes.

Ideally the ensemble should be composed of state-of-the-art photochemical models that are run starting from the best possible emission scenario, as well as with the best possible meteorological fields. The meteorological fields can be indeed different for different photochemical models, since each of them is obtained differently (from different mesoscale models, and then different starting analyses, map projections, domain grids, etc.). Moreover, the different model formulations, i.e., the different advection and turbulence transport schemes and the different chemical mechanisms implemented in each model, should assure a good ensemble spread, which is desirable to define the likely bounds of possible pollutant-concentration fields. The uncertainty in each of those components should average out by the ensemble approach.

The ensemble tested in this study has some of those desirable features. For example, there are differences in the emission data of each ensemble member, partly because the hourly emission values (i.e., biogenic and mobile sources) depend on the meteorology that differs from one mesoscale model to another. These differences can take into account the uncertainty in the emissions estimate, which is often a factor of three or more, and which is the dominant limitation in the photochemical model performance. For the same reason, the different meteorological input fields from MM5, and MC2 allow the ensemble to filter out some of the unpredictable components of the weather. Furthermore, the fact that different ensemble members run at different resolutions lead to different parcel trajectories, and this allows the ensemble to take into account the uncertainties related to the different but plausible choices of the grid location and resolution. Finally, the four runs done for each of the four possible driving meteorology/resolution combinations, may account for part of the uncertainties in the chemistry module and parameters.