Data assimilation and air quality - Air Pollution Modeling and Its Application - page 250

model predictions. It then generated the third day's bias-adjusted forecasts by
combining with the third day's raw model forecasts with the updated KF parameters.
All the updated KF parameters for each hour and at each site were saved into a file
for use in the next KF run. The KF runs then continued by reading the previous
day's KF parameters and 2 preceding days' observations and raw model predictions
to continuously generate the next day's bias-adjusted forecasts through combining
with the next day's raw model forecasts. The KF simulations run daily when the
preceding day's observations and the raw model forecasts for next day (issued on
current day) were available. In our implementation, if 2 consecutive days' data
were missing at a site, the method would automatically drop this site from future
bias-adjustment forecasts; however, if a new site with 2 consecutive days' data
appeared in the observation data set, the KF would initialize the site with initial
values of KF parameters and generate bias-adjusted forecasts further on. This
implementation is very adaptable to the variable nature of monitoring stations
which report hourly observations to the AIRNOW network and can be easily
combined with AQF system to perform real-time bias-adjusted forecasts. The
bias-adjusted forecasts were initialized on January 4 and April 3 for PM 2.5 and O 3
forecasts, respectively, and the programs were run daily on a Linux system; it took
less than 10 min of computation to create a bias adjusted forecast.
3. Performance Evaluations
Table 1 presents a summary of domain (Dom) and sub-regional mean discrete
statistics for the raw model and the KF forecast daily maximum 8-h O 3 mixing
ratios during the study period. Table 2 presents similar model performance statistics
for daily mean PM2.5 concentrations for warm and cold seasons; in each cell, the
value on the left of slash (/) is for warm season and the value on the right of the
slash is for the cool season; the values in the rows with white background marked
with “-mod” represent Table 1. Regional summary of discrete statistics for raw
model and KF bias-adjusted daily maximum 8-h O 3 forecasts during 2008 summer
season statistics associated with raw model forecasts, while those in the rows with
shaded background and with the extension “-kf” represent the statistics associated
with the KF bias-adjusted forecasts. As seen in Table 1 , for daily maximum 8-h O 3
forecasts, the Root Mean Square Error (RMSE) values associated with raw model
forecasts ranged from 10.4 to 16.0 ppb. The application of the KF bias-adjustment,
reduced the RMSE to 8.5-10.5 ppb; on average, this corresponds to more than
25% reduction. Similar reduction was reflected by Normalized Mean Bias (NMB).
More remarkable forecast improvement by the KF forecasts over raw model is
reflected by the Mean Bias (MB) and Normalized Mean Bias (NMB); the MB values
were reduced from several ppb to less than 1 ppb for all the regions, and NMB
from as high as 17% to less than 2%. The correlation coefficients (r) also increased
systematically from 0.5-0.7 for the raw model to 0.7-0.84 in the KF forecasts.
Similar forecast skill improvement in PM 2.5 forecasts by the KF forecasts over