9.2
Analog-Kalman filter based post-processing of surface PM2.5 predictions from the Community Multiscale Air Quality (CMAQ) model

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Wednesday, 5 February 2014: 10:30 AM
Room C206 (The Georgia World Congress Center )
Irina V. Djalalova, NOAA/ESRL/PSD and CIRES/Univ. of Colorado, Boulder, CO; and L. Delle Monache and J. Wilczak

An automated method for post-processing surface PM2.5 predictions from the NOAA Community Multiscale Air Quality (CMAQ) forecasting system is developed. It includes three important components: A real-time quality control procedure for surface PM2.5 observations; Model post-processing at each observational site using historical forecast analogs; Spreading the forecast correction over the entire gridded domain. To reach these goals, the CMAQ annual data set of hourly PM 2.5 from December 01, 2009 through November 30, 2010 is used. The model domain covers the contiguous USA, and model data are verified against EPA AIRNow PM 2.5 observations measured at 716 stations over the CMAQ domain. The model bias (model observation) is found to have a strong seasonal dependency, with a large positive bias in winter and a very small bias in the summer months, and also to have a strong diurnal cycle.. Five different post-processing techniques are compared, including a seven-day running mean subtraction, Kalman-filter forecast analogs, and combinations of analogs and Kalman filtering. The most accurate PM2.5 forecasts have been found to be produced when using historical analogs of the hourly Kalman-filtered forecasts, referred to as CMAQ_KFAN. The choice of meteorological variables used in the hourly analog search is also found to have a significant effect. A monthly error analysis is computed, in each case using the remaining 11 months of the data set for the analog searches. The improvement of CMAQ_KFAN errors over the raw CMAQ model errors ranges from 50-75% for MAE and 40-60% for the correlation coefficient. Since the post-processing analysis is only done at the locations where observations are available, the spreading of post-processing correction information over nearby model grid points is necessary to make forecast contour maps. This spreading of information is accomplished with a seven-pass Barnes-type iterative objective analysis scheme. The final corrected CMAQ forecast over the entire domain is composed of a simple summation of the original hourly CMAQ forecasts and the KFAN bias information interpolated over the entire domain.