Geophysical Research Letters

Online versus offline air quality modeling on cloud-resolving scales

Authors

  • Georg A. Grell,

    1. Cooperative Institute for Research in Environmental Sciences (CIRES), University of Colorado/NOAA Research–Forecast Systems Laboratory, Boulder, Colorado, USA
    Search for more papers by this author
  • Richard Knoche,

    1. Institute for Meteorology and Climate Research, Atmospheric Environmental Research (IMK-IFU), Forschungszentrum Karlsruhe, Garmisch-Partenkirchen, Germany
    Search for more papers by this author
  • Steven E. Peckham,

    1. Cooperative Institute for Research in Environmental Sciences (CIRES), University of Colorado/NOAA Research–Forecast Systems Laboratory, Boulder, Colorado, USA
    Search for more papers by this author
  • Stuart A. McKeen

    1. Cooperative Institute for Research in Environmental Sciences (CIRES), University of Colorado/NOAA Research–Aeronomy Laboratory, Boulder, Colorado, USA
    Search for more papers by this author

Abstract

[1] Computational advances now allow air quality forecast models to fully couple the meteorology with chemical constituents within a unified modeling system – online – that allows two-way interactions. The more common approach is the offline system, which only allows one-way coupling from the meteorology – sampled at fixed time intervals – to the chemistry. To study the information loss between online and offline numerical forecasts, a next-generation nonhydrostatic air quality modeling system has been developed that can be used both offline or online. First, a control online air quality simulation is conducted and the meteorology and chemical data are saved at a 10 s time interval. Subsequently, three offline simulations are conducted with meteorological data updates at 10, 30, and 60 min time intervals. Analysis of the wind velocity power spectrum and chemical profiles indicate that the offline simulations are susceptible to large errors in the vertical mass distribution.

1. Introduction

[2] The prediction and simulation of the coupled evolution of atmospheric transport and chemistry will remain one of the most challenging tasks in environmental modeling over the next decades. The dramatic increase in computer power enables us to use higher resolution to explicitly resolve fronts, convective systems, local wind systems, and even clouds. The enhanced computational power also allows us to increase the complexity of the numerical models. However, traditional air quality modeling procedures may not be valid anymore, or at least will require reevaluation. One fundamental assumption in air quality modeling is that it is possible to make accurate air quality forecasts (and simulations) even while ignoring much of the coupling between meteorological and chemical processes. This commonly used approach, termed offline, requires initially running a meteorological model independently of the chemical transport model. The output from the meteorological model, typically available once or twice an hour, is subsequently used to drive the transport in the chemistry simulation. Although this methodology has computational advantages (discussed later), the separation of meteorology and chemistry can also lead to a loss of potentially important information about atmospheric processes that often have a time scale much smaller than the meteorological model output frequency (e.g., wind speed and directional changes, cloud formation, and rainfall). This may be especially important in future air quality prediction systems, since horizontal grid-sizes on the order of 1 km may be required to match the operational models. At these cloud-resolving scales (horizontal grid spacing less than 5 km), it is important to note that almost all (with the exception of small-scale turbulence) vertical transport comes from resolved, explicit vertical motion fields and not from convective parameterizations. These explicit vertical motion fields usually exhibit very large variability on short time and space scales. Furthermore, in an offline approach the feedback from the chemistry to the meteorology cannot be considered. These deficiencies may make the alternative online approach more attractive. In online modeling systems, the chemistry is integrated simultaneously with the meteorology, allowing feedback at each model time step both from meteorology to chemistry and from chemistry to meteorology. This technique more accurately reflects the strong coupling of meteorological and chemical processes in the atmosphere. Online models can also be used to study the influence of air quality on regional climate and weather.

[3] Even though online modeling may permit better characterization of the time-resolved dispersion of air pollutants and allow for the feedback of the chemistry to the meteorology, there are also disadvantages in online approaches. One such disadvantage is seen in operational weather forecasting, where a longer computational time is required to produce the meteorological forecast with an online air quality prediction. In addition, retrospective offline chemical transport simulations only require a single meteorological data set to produce many chemical transport simulations to examine a scientific research question. These research issues may only be related to air chemistry, thus it may not be necessary to run the modeling system online.

[4] In this study we try to determine how much information - in particular for the vertical mass transport – is lost in cloud-resolving simulations during transition from an online to an offline approach. This is accomplished by using a single modeling system capable of both online as well as offline simulations. In conjunction with this process we also test the sensitivity toward the coupling interval. The model configuration is identical for online and offline simulations with various coupling intervals. Since no horizontal or vertical interpolation is required in this type of offline approach, any differences in the model simulations should be attributed to the different coupling intervals. In the next section we briefly describe the numerical model used in this study. Section 3 describes the methodology. Section 4 presents results, and conclusions are given in section 5.

2. Model Description

[5] Here we only summarize the modeling system used during the experiment, but for a more detailed description the reader is referred to Grell et al. [2000]. The fifth-generation Penn State/NCAR Mesoscale Model (MM5) [Grell et al., 1994] is the cornerstone of this modeling system, but a three-dimensional positive definite advection scheme developed by Smolarkiewicz and Grabowski [1990] was implemented for the advection of tracers. The RADM2 gas-phase chemical mechanism [Stockwell et al., 1990], originally developed for the Regional Acid Deposition Model version 2 (RADM2) [Chang et al., 1989], is used to treat the interaction of the chemical species with each other. The Madronich [1987] photolysis scheme provides the photolysis rates. Deposition is calculated as in the “flux-resistance” analogy [Wesley, 1989], and is directly tied to the turbulence and soil/vegetation/snow parameterization. Sulfate is present in the form of aerosol particles, and its deposition is described according to Erisman et al. [1994]. Biogenic emissions are also calculated online, following Simpson et al. [1995] and Guenther et al. [1993, 1994].

[6] Anthropogenic emissions data are derived from the U.S. Environmental Protection Agency (EPA) county-wide emissions for a base year of 1996 (EPA NET-96, version 3.12, ozone season day) [EPA, 2000]. For both point and area sources, the hourly allocation factors and hydrocarbon speciation of the Volatile Organic Compounds (VOCs) from the EPA [EPA, 1989] are used.

3. Experimental Setup

[7] During the summer of 2002, this modeling system was run in real-time at the Forecast Systems Laboratory (FSL) over the central and eastern United States, using three domains. Figure 1 shows the highest resolution domain. This domain is used for this study. The horizontal resolution of this domain is 3 km with 30 vertically stretched levels. The center of the lowest layer is at about 8 m. The thickness of the layers are slowly stretched to approximately 250 m at about 2 km Above Ground Level (AGL), with even coarser resolution above. One day of particular interest is July 23, 2002 when southwesterly flow ahead of a cold front accompanied ozone values in excess of 120 ppbv [Angevine et al., 2004]. This case has been selected for our model runs. All simulations used in this study are performed without the use of a convective parameterization. During the online simulation, meteorological wind fields are saved at every time step (10 s) for frequency analysis. In addition, three offline simulations are performed with 10, 30, and 60 minute coupling intervals. For the offline simulations, the meteorological fields at the output intervals are interpolated in time to provide the meteorological fields necessary for the chemical transport calculations. Because of the immense amount of data generated by these model runs, the forecasts are restricted to a 12-h period.

Figure 1.

Domain over which cloud-resolving simulations are performed. Also shown is box A and box B, which were used for analysis of data (see text for explanation). Horizontal grid resolution is 3 km (100 × 100 grid points), with 30 vertically stretched levels.

[8] Data analysis is performed in three different ways. First power spectra are calculated for every horizontal grid point and for the vertical velocity (w) at one level (level 8 corresponding to about 450 m AGL, usually just above the oceanic PBL). This analysis is limited to a 4 hour period centered around the time that the front passed through the domain (1700 UTC TO 2100 UTC). To get an idea of the average “behaviour” of the power spectra they are averaged over 30 × 30 grid points (box A in Figure 1).

[9] In a second analysis we first average w to different resolutions, then calculate the power spectra of the averaged fields. Because of the limited domain, we only compare a single grid point (center of box A).

[10] Finally, in order to estimate the integrated effects, we compare composite mixing ratios of Carbon Monoxide (CO) and Ozone (O3) at a time after the front has passed (2100 UTC). Box B in Figure 1 shows the area over which the averaged profiles are calculated. The domain is limited to the southeastern part, since at upper levels the north-west flow is strong enough to flush the pollutants out of the small domain quickly.

4. Results

[11] A power spectrum can be used as a measure for what scales of motion contribute to the analyzed field. Figure 2a shows a plot of the average power spectrum of the vertical velocity at level 8 for box A. From Figure 2a it is clear that the vertical velocity exhibits large variability on short time scales, even when averaged over a large number of grid points. Scales of motion with time frequencies smaller than 10 minutes significantly contribute to the total variability. For offline simulations this would mean that a meteorological output interval of less than 10 min would be required to capture most of the energy. It must be noted that these results are based on the MM5 model which uses a significant amount of filtering and diffusion. It is likely that for models with less filtering, there will be even more energy in the small scales. This should also be expected with any further increase in resolution.

Figure 2a.

Power spectrum of the vertical velocity (at model level 8, in the lowest 1 km of the atmosphere) from the online simulation. The spectrum is calculated over the 4 hr period from 1700 to 2100 UTC, using 10 s output from the model simulation. Displayed is the average power spectrum for 30 by 30 grid points (box A). The corresponding (period) meteorological output intervals are shown on the x-axis.

[12] A different way to look not only at the dependence on the meteorological output interval but also the horizontal resolution is seen in Figure 2b, which displays the percentage of the variability of the vertical velocity that is captured as a function of horizontal resolution and output frequency. Here 100% is defined as the sum of the spectral power for all frequencies. For an output interval of 1 h, less than 40% of the total variability is captured by both the original 3-km resolution run as well as the 9-km results. Again, for the cloud-resolving simulation, it appears that the meteorological output interval must be less than 10 min to capture more than 85% of the variability of the vertical velocity. On the other hand, for the vertical velocities averaged to 27 km horizontal resolution, results look somewhat better, especially at 30 min output interval, where more than 90% of the variability is captured.

Figure 2b.

Percentage of variability of the vertical velocity (same level as in Figure 2a) that can be captured as a function of the meteorological sampling interval and the horizontal resolution. Variability was derived with power spectra calculated using vertical velocities that were representative (through averaging) of 3 km, 9-km (dotted), and 27-km (dashed) horizontal resolution at the center of box A. Percentage is a fraction of 1, and output frequency is in minutes.

[13] Another concern is whether it might be enough to capture 50% of the variability. This goal, for this particular level and model run, could be reached with an output interval of 30 min. To answer this question, we looked at the integrated effect of the differences in vertical mass transport and computed average mixing ratios of carbon monoxide as well as ozone over a larger region of the domain and at 2100 UTC. The resulting profiles are shown in Figures 3a and 3b as a function of the output interval and height. Clearly, an output interval of 30 min is not sufficient. Differences for carbon monoxide – an almost inert tracer – are dramatic, in particular for the upper levels. It becomes clear that in order for cloud resolving simulations to capture the vertical transport of mass properly, very high frequency coupling intervals are required. Note that even for a 10-min output interval at this resolution, differences are still as high as 20 ppb (more than 10%), especially in the lowest few layers and in the upper troposphere. The online simulation is much more effective in cleaning out the air in the Planetary Boundary Layer (through transporting polluted air upward to the free troposphere but also transporting cleaner air downward into the Planetary Boundary Layer). For ozone (Figure 3b) the results are similarly dramatic, but less obvious. This is caused by the more variable ozone distribution in the vertical. While carbon monoxide sources are generally restricted to the boundary layer, ozone values are highest in the stratosphere, resulting in a strong source of ozone for the high levels in case of sinking motion.

Figure 3a.

Domain averaged carbon monoxide mixing ratio in ppb. Mixing ratios are averaged over box B. Shown are averages for the online simulation (solid line), and offline simulations with 1-hr (long dashed), 30-min (short dashed), and 10-min (gray) coupling intervals.

Figure 3b.

As in Figure 3a, except for ozone.

5. Conclusions

[14] A nonhydrostatic cloud-resolving coupled weather/air quality modeling system is used to test the sensitivity of air quality forecasts to online versus offline modeling methodologies. The focus is the vertical mass transport, since this is likely the most affected parameter in cloud resolving simulations. Although this study presents what might be considered a severe test case – a frontal passage, results should also be applicable in other situations with a large variability in the vertical velocity fields (complex terrain or any situation where convection occurs). Convective systems may have a particularly severe effect on the vertical redistribution of mass, since cloud-resolving simulations can not use any convective parameterizations; hence, all mass transport has to come from the model predicted vertical velocity fields. In this case it is not only important to capture most of the variability, but also most of the peaks (positive and negative) of the vertical velocity fields, especially in low levels. For a constituent that has its sources and highest concentration in the PBL, both upward motion out of the PBL and downward motion into the PBL will decrease its concentration in the boundary layer. For this reason it may not be helpful to use time-averaged meteorological fields for the interpolation to the chemical transport model.

[15] If in the future we can trust the capabilities of high resolution cloud-resolving NWP models to predict the complex flows in the atmosphere with reasonable accuracy, online simulations may be required for most applications. Although the accuracy of the simulated vertical winds cannot be guaranteed, capturing of most of the variability in the vertical velocity field is still essential in order to properly simulate the vertical redistribution of mass. Clearly, this will have the most noticeable influence on simulations that use large domains to prevent air from quickly flushing beyond their perimeters, or that feed on themselves (e.g., air quality forecasts that are reinitialized with forecasts). It may also be critical to run with very high time resolution meteorological output when performing dispersion simulations in the case of the release of toxic substances.

[16] As indicated above, it is not clear how air quality forecast will depend on online versus offline modeling. Although the observed variability, and the observed magnitudes of the vertical velocity will most likely be larger than what is simulated by the model used in this study (MM5), because of the uncertainty in the prediction of the vertical velocity fields there is no guarantee that the value of air quality forecasts could actually be improved. In addition, in some situations the air quality forecasts may not be changed much at all. In our case, prediction of 1 hr peak ozone values would change little for either simulation, while the 8 hour average values would be changed significantly. These issues will be looked at in a follow-up study.

Acknowledgments

[17] The authors gratefully acknowledge funding support from the NOAA funded AIRMAP (Atmospheric Investigation, Regional Modeling, Analysis and Prediction) and air quality programs. We would also like to thank John Brown as well as two anonymous reviewers for a careful review of the manuscript, and Nita Fullerton for editorial assistance.

Ancillary