![[logo] US EPA](https://webarchive.library.unt.edu/eot2008/20081013073324im_/http://www.epa.gov/epafiles/images/logo_epaseal.gif){kind=logo}
Advanced Space-Time Techniques for Model Evaluation
Research Programs
One approach that treats space and time simultaneously is called Principal Component Analysis (PCA). The main objective of this approach is to identify, through a reduction in data (which will become increasingly important as model simulation periods continue to lengthen), the characteristic, recurring and independent modes of variation across a myriad of spatial and temporal scales. Such analysis is useful in that it facilitates understanding of the probable mechanisms (related to either meteorological, chemistry, or emissions) responsible for the unique behavior of the model simulation. An examination of the time series of these modes of variation could then be performed using Spectral Density Analysis (SDA) to decompose the modes of variation's time series into a sum of cosine and sine waves or varying amplitudes and wavelengths. Such temporal decomposition can yield a measure of the distribution of the variance over a continuous spectrum of all possible wavelengths. Application of PCA and SDA to both CMAQ model output and observational data from various network will facilitate evaluation of the models ability to accurately capture the major modes of variability (both in space and time). Sulfate aerosols will first be used in this analysis, and then other aerosol species will be tested once the approach has been demonstrated.
Model Evaluation
Diagnostic Evaluation:
Indicator Metrics & Instrumented CMAQ Investiguations for Inorganic Fine Particle System
Another approach for spatial analyses that is being tested is Bayesian-Kriging, which compares observed data with the CMAQ output over space. This technique is similar to classical kriging, which has become quite common in environmental statistics applications. Both kriging approaches interpolate data based on a statistical model of the spatial correlation that exists, allowing the user to obtain statistically defensible estimates at points for which no data are available. The major difference that is introduced in the Bayesian kriging approach is that the parameters of the statistical model are treated as random variables, and the uncertainty associated with their values is included in the error estimates for the kriging predictions. Both techniques meet the basic objective of assessing uncertainty in the CMAQ predictions and comparing the CMAQ results with the observations from the monitors. The results from this approach is a spatial field based on the observations, which is produced by Bayesian kriging, and a spatial field predicted by CMAQ. With results using this method, we will consider whether the model predictions have bias in certain regions of the country or along other patterns (geographic features, urban settings, etc.). Specific case studies for sulfate have successfully demonstrated the method. Further testing and refinement of the method is needed to determine what chemical species are appropriate for application of this method. For example, nitrate aerosols may be too noisy spatially for this method given the sparseness of the data. As this method continues to be developed, additional methods development will demonstrate how the model predictions can be used in the Bayesian kriging of the observations to develop a merged dataset based on the observations and spatial variability from the model where observations are not available.