![[logo] US EPA](https://webarchive.library.unt.edu/eot2008/20081107114341im_/http://www.epa.gov/epafiles/images/logo_epaseal.gif){kind=logo}
Final Report: Development and Demonstration of a Methodology for Characterizing and Managing Uncertainties in Emission Inventories
EPA Grant Number: R826766Title: Development and Demonstration of a Methodology for Characterizing and Managing Uncertainties in Emission Inventories
Investigators: Frey, H. Christopher , Houyoux, Marc , Loughlin, Daniel
Institution: North Carolina State University , MCNC / North Carolina Supercomputing Center
EPA Project Officer: Shapiro, Paul
Project Period: October 1, 1998 through September 30, 2001
Project Amount: $553,298
RFA: Air Pollution Chemistry and Physics (1998)
Research Category: Engineering and Environmental Chemistry
Description:
Objective:We hypothesized that quantification of uncertainty in Emission Inventories (EIs) will lead to new insights regarding the quality of EIs, the best resource allocation to improve EIs, and decisionmaking for air quality management. The objectives of this research project were to: (1) develop and refine methods for quantitative analysis of variability and uncertainty in EIs; (2) demonstrate the methods via application to a detailed case study of an EI; and (3) characterize the benefits of the techniques for environmental and research management.
Summary/Accomplishments (Outputs/Outcomes):This project resulted in the following major findings:
· Parametric probability distribution models, such as the lognormal, Weibull, and gamma, often are useful means for quantifying inter-unit variability in emissions within a source category (e.g., Frey and Zheng, 2002b; Frey and Bammi, 2002b; Frey and Bammi, 2003; Frey and Li, 2003).
· In cases in which a parametric distribution is not an adequate representation of inter-unit variability, an empirical distribution can be used instead (e.g., Frey and Zheng, 2002c) or a mixture distribution may provide an improved fit (Zheng and Frey, 2003).
· For mixture distributions, there is a tradeoff in the apparent goodness-of-fit versus uncertainty introduced as a result of over-parameterization of the distribution. In other words, there are diminishing returns to adding additional components to the mixture. Moreover, although the numerical method for estimating parameters of a mixture was shown to be robust in the case of a two-component mixture (Zheng and Frey, 2003), it is less robust as the number of components increases or as the degree of separation between components decreases.
· For data sets that include nondetected values, Maximum Likelihood Estimation is an asymptotically unbiased and robust method for fitting parametric distributions that describe inter-unit variability (Zhao and Frey, 2003).
· Uncertainty in the mean or other statistics of emission factors or activity factors can be quantified using the numerical method of bootstrap simulation, as demonstrated in specific case studies (e.g., Frey and Zheng, 2002b, 2002c; Frey and Bammi, 2002b; Frey and Bammi, 2003; Frey and Li, 2003; Zheng and Frey, 2003; Zhao and Frey, 2003). This numerical method for estimation of uncertainty in a statistic is robust and provides appropriately skewed distributions of uncertainty in situations with small sample size or large sampling error.
· Autocorrelation is statistically significant and important to the accurate estimation of uncertainty, at least for the time-series case studies considered based on hourly utility NOx emissions for coal-fired baseload units (Abdel-Aziz and Frey, 2003e).
· Autocorrelation in data can be identified using statistical tests and can be quantified using autoregressive time-series models. Such models provide an accurate method for accounting for autocorrelation in hourly emissions data required as inputs to air quality models (Abdel-Aziz and Frey, 2003a, 2003b, 2003c).
· Inter-unit dependence, as well as intra-unit autocorrelation, can be accounted for using vector auto-regressive time-series models, as illustrated for the example of hourly utility NOx emissions (Abdel-Aziz and Frey, 2003c).
· Comparison of the results of the independent units approach versus an approach in which inter-unit dependence is accounted for implies that erroneous conclusions could be drawn regarding ambient air quality predictions if the independent approach is used. For some grid cells, the difference in estimated probabilities based on the two approaches was as large as 0.38 (Abdel-Aziz and Frey, 2003d). A previous comparison of the inventories obtained using both approaches showed that the dependent approach more accurately represents the observations (Abdel-Aziz and Frey, 2003c). Therefore, decisionmaking should be based on results obtained from the dependent approach.
· The range of uncertainty in emission factors varies from as little as plus or minus 10 percent in some cases, to as large as a factor of three in other cases, based on detailed quantitative assessments made regarding source categories, such as power plants, natural gas-fueled engines, consumer solvents, architectural coatings, gasoline terminal loading and storage, asphalt paving, lawn and garden equipment, construction equipment, farm equipment, agricultural equipment, and light duty gasoline vehicles, for pollutants including nitrogen oxides and hydrocarbons (e.g., Abdel-Aziz and Frey, 2003a, 2003b, 2003c, 2003e; Bammi and Frey, 2001; Frey and Zheng, 2001a, 2001b, 2002a, 2002b, 2002c; Frey and Bammi, 2002a, 2002b, 2003; Frey and Li, 2001, 2002, 2003; Frey, Bharvirkar, and Zheng, 1999; Li and Frey, 2002).
· Even for emissions sources for which there are substantial data, such as in the case of utility NOx emissions measured using continuous emissions monitoring, there is uncertainty regarding what the emissions will be in the near-term future. The historical record of the range of variability in emissions for a given hour of the day provides a basis for quantification of the lack of knowledge regarding what the actual emissions will be for the same time of day in the near-term future (Frey and Zheng, 2002b; Abdel-Aziz and Frey, 2002a).
· Emissions testing procedures should attempt to represent as closely as possible real-world operating conditions, so that average emission factor estimates are accurate, as compared to average real-world emissions (e.g., Frey and Zheng, 2002c).
· Comparison of uncertainty estimates for individual driving cycles for light duty gasoline vehicles reveals that in some cases, two or more driving cycles produce statistically similar results and, therefore, are redundant with each other. The identification of redundant cycles, such as the LSP1 and LSP2 cycles, or the ART-AB and FWY-D, -E, -F, and -G cycles, presents opportunities to reduce the cost of data collection by focusing on a smaller set of nonredundant cycles (Frey and Zheng, 2002c).
· Probabilistic models, such as vector-autoregressive time-series models, can be used to gain insight into the proportion of variability in emissions that can be explained based on available activity data, such as capacity factor or heat rate in the case of utility NOx emissions (Abdel-Aziz and Frey, 2002b, 2002c).
· Software tools for quantifying variability and uncertainty have been extensively tested as part of this work, especially with regard to mixture distributions and fitting of distributions to censored data (e.g., Zheng and Frey, 2003; Zhao and Frey, 2003).
· Software tools for developing air quality-ready probabilistic emission inventories have been developed (Houyhoux, et al., 2003) and applied to specific case studies (Abdel-Aziz and Frey, 2003d).
· Not all emission modeling uncertainties are included in this effort, such as uncertainties associated with temporal allocation, chemical speciation, emission controls, and computation of layer fractions. Once the use of the uncertainty-enabled sparse matrix operator kernel emissions (SMOKE) system becomes more widespread and the benefits of including uncertainties in air quality modeling is more widely realized, it may be useful to make additional modifications to the system to include these other uncertainties. In addition, the existing system is based on MOBILE5b, which has since been replaced by MOBILE6 in newer versions of SMOKE, and will later be replaced by the MOVES model now in development. The two versions should be integrated to make uncertainty analysis using the latest vehicle emissions model possible. Also, the uncertainties in biogenic emissions, which play a large role in ozone formation in the Eastern United States, should be included to allow air quality modelers to establish a more realistic picture of the range of impact biogenic emissions have on air quality models (Houyoux, et al., 2003).
· The probabilistic air quality modeling case studies based on quantification of uncertainty in the emission inventory demonstrate the following: (1) any given statistic of predicted ambient ozone concentrations can be predicted for a given grid cell, such as the range of uncertainty in maximum values realized over iterative Monte Carlo simulation; (2) the probability with which each grid cell exceeds a specific ambient air quality standard can be quantified; and (3) key sources of uncertainty in the inventory can be identified, thereby enabling targeting of data collection to reduce uncertainty or control strategy development to improve the confidence with which an ambient standard will be attained (Abdel-Aziz and Frey, 2003d).
· A potential use of probabilistic air quality modeling results is in regards to the siting of monitoring stations. Monitoring stations could be spatially allocated where the estimated probability of exceeding the NAAQS is estimated to be substantial. However, the comparison of 1-hour and 8-hour standards implies that the identification of an MSA as nonattainment is less sensitive to location in the latter case, because many more grid cells have a substantial probability of exceeding the standard. In the latter case, it would be possible to design a monitoring network that is sufficient to detect nonattainment with a relatively small number of stations. Of course, the location of hotspots could differ for different meteorological episodes, and this could be addressed by conducting analyses for several such episodes to identify common locations of potential exceedences (Abdel-Aziz and Frey, 2003d).
· Uncertainty analysis is facilitated if sufficient information is reported, upon which statistical analysis can be conducted. For example, knowledge of the mean, standard deviation, and sample size for variability in emissions can be applied to estimate uncertainty in mean emissions based on random sampling error. In addition, knowledge of measurement errors or detection limits further aids in the quantification of uncertainty. A key recommendation of this work is that such information should be reported routinely (e.g., Frey and Zheng, 2002b, 2002c; Frey and Bammi, 2002b, 2003; Frey and Li, 2003).
· Although the ranges of uncertainty in the emission factors can be large, this does not imply that the emission factors are meaningless. The significance of the range of uncertainty is context-dependent. Aside from using uncertainty analysis as a tool for prioritizing additional data collection, uncertainty analysis can be applied to emission inventories. A quantitative assessment can be made of the likelihood with which an emission budget will be met. A decisionmaker can use this information to make tradeoffs between emissions management strategies and the confidence with which the budget will be met. Furthermore, probabilistic emission inventories can be used as input to air quality models to determine the likelihood that ambient air quality management goals will be achieved, and to develop strategies that produce an acceptable confidence level of air quality benefits (Frey and Zheng, 2002c).
· The large range of quantified uncertainty in emission factors, and in the example case studies for emission inventories, suggests that it is important to quantify uncertainty and to account for this when developing emission inventories. As the National Research Council noted in its recent report on modeling mobile source emissions, it is not possible to quantify all sources of uncertainty. Nonetheless, the quantifiable portion of uncertainty should be taken into account when reporting and using emission factors. Nonquantifiable contributions to uncertainty should be acknowledged qualitatively. Decisionmakers should be aware of both the strengths and limitations of emission factors and emission inventories, so that decisions regarding air quality management can be made that are robust to uncertainty. Furthermore, by understanding the key sources of uncertainty in an emission inventory, resources can be prioritized to reduce uncertainty, such as by collecting better and more data. Thus, the probabilistic methodology presented here is part of an overall approach to developing policy, program management, and research planning.
Relevance to the U.S. Environmental Protection Agency (EPA). The work conducted under this grant has been relevant to various parts of the EPA, including the: Office of Research and Development, Office of Air Quality Planning and Standards (OAQPS), and Office of Transportation and Air Quality. Related work regarding quantification of uncertainty, conducted under separate funding, has benefited from the methods and case studies developed in this project. For example, we have conducted separate projects regarding development of software tools for quantification of variability and uncertainty in exposure and risk assessment (Frey, Zheng, Zhao, Li, and Zhu, 2002; Zheng and Frey, 2002a, 2002b, 2003b); quantification of variability and uncertainty in emission factors and emission inventories (Frey, Bharvirkar, and Zheng, 1999; Frey and Zheng, 2002, 2001a, 2001b); quantification of uncertainty in vehicle emissions estimates (Frey, 2003; Frey, Unal, and Chen, 2002, 2003; Frey, Unal, Chen, and Li, 2003; Frey, Unal, Chen, Li, and Xuan, 2002); and recommendation of requirements for new software tools to deal with sensitivity and uncertainty in multimedia models (Loughlin, Frey, Hansiak, and Eyth, 2003). These publications are listed separately under the heading "Publications Prepared Under Separately Funded Projects for EPA Pertaining to Quantification of Uncertainty." In addition, we currently are performing work for the EPA regarding uncertainty analysis of the 1996 National Air Toxics Assessment and are preparing state-of-the-science and guidance documents regarding a hierarchy of methods for uncertainty analysis for OAQPS.
Journal Articles on this Report: 11 Displayed | Download in RIS Format
| Other project views: | All 44 publications | 17 publications in selected types | All 12 journal articles |
| Type | Citation | ||
|---|---|---|---|
|
|
Abdel-Aziz A, Frey HC. Development of hourly probabilistic utility NOx emission inventories using time series techniques: Part 1-univariate approach. Atmospheric Environment 2003;37(38):5379-5389. |
R826766 (Final) |
not available |
|
|
Abdel-Aziz A, Frey HC. Development of hourly probabilistic utility NOx emission inventories using time series techniques: Part II-multivariate approach. Atmospheric Environment 2003;37(38):5391-5540. |
R826766 (Final) |
not available |
|
|
Abdel-Aziz A, Frey HC. Quantification of hourly variability in NOx emissions for baseload coal-fired power plants. Journal of the Air & Waste Management Association 2003;53(11):1401-1411. |
R826766 (Final) |
not available |
|
|
Abdel-Aziz A, Frey HC. Propagation of uncertainty in hourly utility NOx emissions through a photochemical grid air quality model: A case study for the Charlotte, NC, modeling domain. Environmental Science & Technology 2004;38(7):2153-2160. |
R826766 (Final) |
|
|
|
Frey HC, Zheng J. Probabilistic analysis of driving cycle-based highway vehicle emission factors. Environmental Science & Technology 2002;36(23):5184-5191. |
R826766 (2001) R826766 (Final) R826790 (Final) |
|
|
|
Frey HC, Bammi S. Quantification of variability and uncertainty in lawn and garden equipment NOx and total hydrocarbon emission factors. Journal of the Air & Waste Management Association 2002;52(4):435-448. |
R826766 (2001) R826766 (Final) R826790 (2002) R826790 (Final) |
|
|
|
Frey HC, Zheng JY. Quantification of variability and uncertainty in utility NOx emission inventories: Method and case study for utility NOx emissions. Journal of the Air & Waste Management Association 2002;52(9):1083-1095. |
R826766 (2001) R826766 (Final) |
not available |
|
|
Frey HC, Bammi S. Probabilistic nonroad mobile source emission factors. Journal of Environmental Engineering 2003;129(2):162-168. |
R826766 (2001) R826766 (Final) R826790 (2000) R826790 (2001) R826790 (2002) R826790 (Final) |
|
|
|
Frey HC, Li S. Quantification of variability and uncertainty in AP-42 emission factors: case studies for natural gas-fueled engines. Journal of the Air and Waste Management Association. |
R826766 (Final) |
not available |
|
|
Zhao YC, Frey HC. Quantification of variability and uncertainty for censored data sets and application to air toxic emission factors. Risk Analysis 2004;24(4):1019-1034. |
R826766 (Final) |
not available |
|
|
Zheng J, Frey HC. Quantification of variability and uncertainty using mixture distributions: evaluation of sample size, mixing weights, and separation between components. Risk Analysis 2004;24(3):553-571. |
R826766 (Final) R826790 (Final) |
|
air, mobile sources, nitrogen oxides, volatile organic compounds, VOC, public policy, engineering, modeling, agriculture, business, transportation, industry. , Air, Scientific Discipline, Engineering, Chemistry, & Physics, Mathematics, Physics, Chemistry, innovative emissions estimation models, decision making, quatitative analysis, predict uncertainty, environmental monitoring, variability, characterizing uncetainties, propagation of uncetainty, emissions inventory, air quality standards
Relevant Websites:
http://www4.ncsu.edu/~frey/ 
http://www.mcnc.org
Progress and Final Reports:
2000 Progress Report
2001 Progress Report
Original Abstract