Research Grants/Fellowships/SBIR

Statistical Post-Processing of the National Air Quality Forecast Capability to Optimize the Value and Layout of Ozone Monitors in the Washington-Baltimore Region

EPA Grant Number: FP917299
Title: Statistical Post-Processing of the National Air Quality Forecast Capability to Optimize the Value and Layout of Ozone Monitors in the Washington-Baltimore Region
Investigators: Garner, Gregory G
Institution: Pennsylvania State University - Erie Behrend College
EPA Project Officer: Zambrana, Jose
Project Period: August 1, 2011 through July 31, 2014
Project Amount: $126,000
RFA: STAR Graduate Fellowships (2011) RFA Text |  Recipients Lists
Research Category: Academic Fellowships , Fellowship - Emerging Environmental Approaches and Challenges: Social Sciences



Air shed managers rely on accurate air quality forecasts when making decisions regarding actions and policy on pollution emissions and exposure in their region. The National Air Quality Forecast Capability (NAQFC) produces these forecasts; however, there are known problems with this model. This project will apply statistical post-processing algorithms known as Model Output Statistics (MOS) to the NAQFC, similar to what is done with meteorological models, to improve the forecast skill and economic value. These algorithms then will be used to optimize pollutant monitor layout in the Washington- Baltimore area.


Observed and forecasted 8-hour surface ozone mixing ratios, along with various meteorological and chronological variables, will be assembled from the Washington-Baltimore region during the ozone seasons (April-October) of 2005 through 2010. These data will be used to train Classification and Regression Tree (CART) models for each monitor in the region. The CART models are trained by recursively splitting the observed ozone into homogeneous groups depending on the state of the included variables. A multivariate regression model will be fit to each group. This process will be repeated numerous times using a bootstrap process. This process produces a distribution of possible outcomes with the mode of the distribution yielding the forecast conditioned on known local statistical relationships. The resulting CART and regression models are the MOS for the given monitor. Cross-validation will be used to assess the error of the MOS forecasts. These MOS can then be applied to methods in determining optimal monitor layout using spatial correlation and value metrics.

Expected Results:

The product resulting from this research will be a robust tool for forecasting ozone events in the Washington-Baltimore region. The increased skill of these forecasts will provide sound information to decision makers regarding actions and policy that reduce emission of and exposure to pollutants. The additional value produced by these forecasts will provide quantifiable evidence of how these decisions affect the local economy. By applying these MOS to methods of optimizing pollutant monitor layout, maximum data quality can be attained with minimum cost.

Potential to Further Environmental / Human Health Protection

By increasing forecast skill, decision makers can have higher confidence in their choices to mitigate emissions and exposure. This will minimize human health problems and environmental impacts related to exposure to poor air quality events while maximizing resource efficiency.

Supplemental Keywords:

air quality, forecasting, modeling, decisions, decisions under uncertainty, model output statistics, MOS, regression tree, CART, bootstrap, cross-validation, value