2. DEVELOPMENT OF AN IMPROVED PROTOCOL FOR EVALUATING THE
2.4 PROMPT Implementation
2.4.2 PROMPT Implementation Phase 2 (PI-P2)
The goal of PI-P2 is to clarify the possible issues in model’s ability to estimate ozone response to precursor control. The emphasis is on the correctness of the source-receptor relationship and on how ozone forms, i.e. the precursor signals in the model predictions. In other words, we are interested in whether a model has emissions that are consistent with observations, and whether the modeling system delivers emitted precursors to where real world winds do. Time series plots and scatter plots should be examined briefly at this stage and more in-depth in the next phase of PROMPT. In generating these plots, special attention should be paid to attributes of the plots such as scales of magnitude (e.g. concentration). One important mistake that often makes these plots ‘dequantifying’ is using improper scales, labels, colors, and more (Tufte, 1997). One good example of dequantifications is over use of contour plots using too many colors. In these plots, users often interpolate spatial data to very fine scales that do not correspond to model cells. Even though these plots seem to be esthetic, they distort the information necessary for high precision analysis. In other words, it may help to get a general or overall impression of model behavior, but it does not help show how spatial gradients are modeled using different grid configurations.
In general, it is good to be aware of ‘overly smoothing’ model outputs to appear to have more spatial and temporal resolution than they actually have. By choosing proper scales and other plot attributes, these graphical measures can provide more information even in
quantitative ways. Unfortunately, there are no commonly accepted methods for setting these plot attributes. Instead, we recommend that these basic settings should be settled upon in the early phase of MPE (or preferably before SIP modeling) among all related groups involved in the SIP modeling processes. In this way, we can at least have an internally consistent
The minimal set of variables for these plots is a collection of plots for surface winds, ozone, nitrogen oxides (plus CO if available), and VOCs (continual measurement data such as auto-GC preferred). When analyzing time series plots, we recommend evaluators divide time series for a day in three temporal sections such as midnight-sunrise, sunrise-noon (or peak ozone hour), noon (or peak ozone hour)-sunset, sunset-midnight. At PI-P2, evaluators need to acquire continual VOCs measurements such as those at Photochemical Assessment Monitoring Stations (PAMS) sites and to convert them properly to compare them with model predictions (i.e. real species converted to model species). If there no PAMS or any
continuousVOCs measurement is available, we recommend evaluators acquire at least canister VOCmeasurements. One can proceed without continual VOCdata because (1) wind errors can indicate whether the chemical signal comparison can be meaningful or not, (2) assessment of NOx bias will narrow the possible issues of VOC bias, and (3) canister
VOCs data are useful for comparison of emission inventory with ambient composition in terms of modeled VOCs species. Depending on the results of analyses suggested below, evaluators can roughly classify the reliability of model in developing control strategies into four classes: “None”, “NOx only”, “VOCs only”, or “Both NOx and VOCs”.
In the case of large biases of wind speed and/or direction near important sources or persistent error during retention time (i.e. the time that average winds take to cross a
modeling domain), ‘none’ will be the likely answer. However, answers are not solely based on scientific assessment because it is not possible. There are no rigid criteria that we can set for our acceptance level of a model since each situation requires different tolerance. For example, a wind direction error of some degree can be acceptable if the error occurs near large area sources of low intensity while the same magnitude of error may not be acceptable
when the same wind direction error happens at very high intensity point sources. Moreover, the acceptance depends on whether that error can influence policy choices. Therefore, we recommend evaluators consult policy analysts or whoever is in charge of developing policy questions for a SIP to get proper information for evaluators’ acceptance claim.
The quality of wind inputs is especially important at this stage because wind biases may lead to the conclusion of a model’s inability to distinguish precursors for ozone control depending on the apparent source-receptor relationship. Biases of chemical concentrations with good wind fields indicate possible biases in emission inputs, but biases of winds likely make emission biases ambiguous. Due to wind biases and error, evaluators may need to request reanalysis of meteorological model results or model episode selection. For testing wind inputs, we recommend inspection of hodograms for observed and modeled winds and hodograms for wind differences and wind speed scatter plots as shown in Figure 2.4. The wind errors shown in the example figure for daytime is over 60 degrees, which may prevent one from performing the proper comparison of model outputs and observations unless the surrounding emissions were very homogeneous. The important attributes of these plots are what the sizes of biases are during each hour window. Large differences in wind speeds during morning hours when ozone is being formed or around peak ozone hours will be very important sources of error especially when the predicted ozone is due to near by precursor sources. If we are dealing with transported ozone problems, wind errors across the modeling domain are important; persistent but tolerable local wind error may end up being
unacceptable because the modeled airmass may undergo a different chemical environment in the model than in real world. Thus, wind errors can influence cause-and-effect and change the control response.
For chemical signals, we propose evaluators focus on the overall features such as when the ozone rise occurs during the day, what the shape of the ozone time series looks like at each monitor on each day, and the associated NO and NO2 time series, e.g. does the model over-predictNO and NO2orVOCs. Other important aspects of time series are whether a model shows NOx inhibition, NOx changes during traffic hours if a monitor is near major line
sources, and so on. For spatial scale, the past MPE often includes 8 cell values surrounding a monitor to show possible small misalignment of ozone plume. This is based on older
practice when there were few monitors in the domain. One problem is that now the size of a model cell varies from 36 km to 1 km or even smaller in a single episode. Thus, we
recommend evaluators use the value of hourly wind speed for an hour as a radius for including cells that might be near by monitors.
In PI-P2, the most difficult class among the four categories of model’s reliability can be “VOCs only” because it indicates there are some issues in NOx predictions. Depending on
the significance of wind errors and the reactivity of VOCs with hydroxyl radicals, however, we can utilize relative compositions of some VOC species in ambient measurements to compare with compositions of those VOC species in the model outputs and in emission inputs. Since we know that the model’s NOx prediction has issues in this case, model
evaluators may want to see some VOC species that have low reactivity with OH· and are not sensitive to NOx in the model. If there are no VOC measurements available, the model’s
performance for VOCs predictions will significantly depend on other factors such as the quality of VOCs emission inventory, and will remain an open question at some level.
Figure 2.4. Illustrative examples of wind scatter plots (top), wind time series plot (bottom left), and wind error time series plot (bottom right) at a monitor site. In these plots, times of a day are encoded with different markers and colors. Model prediction and observation are in different colors. This specific case shows gross (> 60 degree) wind direction differences between modeled winds and observed winds from 1300 to 1700. A series of questions should be asked and answered to investigate if these discrepancies will affect control strategy developments.