42 Can Assimilation of Ground Particulate Matter Observations …
269
where x is a vector of analysis, x b is the background (forecast) vector, y is an observation vector, B is the background error covariance matrix, H is an observation operator,
and R is the observation error covariance matrix. The background error covariance
matrix B is separated into vertical and horizontal components, and is represented as a
product of error variances and spatial correlation matrices. The correlation matrices
simulate Gaussian shapes in space and in the GSI are modelled with recursive filters.
The observational data of PM2.5 concentrations, available at 1 h resolution, are provided by the Chief Inspectorate of Environmental Protection (http://powietrze.gios.
gov.pl). There were 42 stations available for the study period.
42.2.2 The Simulation Period and Verification
The study period covers episode of a very low air quality and high concentrations
of particulate matter in Poland on 11–25 February 2017. The episode was not reproduced by a standard forecasting system. More details are given in [2]. The most
recent forecasts are available here: powietrze.uni.wroc.pl.
In this study we run and compared two simulations. The first simulation was
initialised each day at 00 UTC, using the chemistry cycling from a previous forecast
(GOCART). The second run the simulation was also initialised at 00 but the initial
conditions were modified by incorporation of observational data by the GSI system
(GOCART_GSI). Both forecasts were run for the 72 h (as a standard procedure in
our system), however only first 24 h were analysed in this study. To verify the results
we used factor of two (FAC2), mean bias (MB), mean gross error (MGE), normalised
mean bias (NMB), and correlation coefficient (R).
42.3 Results
Mean observed concentrations of PM10 and PM2.5 were respectively 57.2 and
46.4 µg m
−3 during the period of 11–25 February 2017. Very high concentrations
were observed in the first part of the analysed episode. PM2.5 concentrations reached
or exceeded 150 µg m
−3 at many Polish stations between 13th and 18th of February
(Fig. 42.1).
In general the WRF-Chem model underestimated observed PM2.5 concentrations.
However, all statistics were improved after assimilation of ground based observations of PM2.5. MB decreased from −11.4 to −4.8 µg m
−3 , FAC2 increased from
0.68 to 0.74 and R from 0.63 to 0.73 (Table 42.1). Time series in Fig. 42.1 shows that
GOCART_GSI better fits in the observations during the peak of concentrations, however for the warmer period it overestimates concentrations at the Warsaw station. An
improvement between GOCART_GSI and GOCART is also observed for PM10 but
the changes are smaller than for PM2.5. FAC2 and R increased from 0.63 to 0.65 and
from 0.56 to 0.63, respectively. The GOCART simulation underestimated observed
269
where x is a vector of analysis, x b is the background (forecast) vector, y is an observation vector, B is the background error covariance matrix, H is an observation operator,
and R is the observation error covariance matrix. The background error covariance
matrix B is separated into vertical and horizontal components, and is represented as a
product of error variances and spatial correlation matrices. The correlation matrices
simulate Gaussian shapes in space and in the GSI are modelled with recursive filters.
The observational data of PM2.5 concentrations, available at 1 h resolution, are provided by the Chief Inspectorate of Environmental Protection (http://powietrze.gios.
gov.pl). There were 42 stations available for the study period.
42.2.2 The Simulation Period and Verification
The study period covers episode of a very low air quality and high concentrations
of particulate matter in Poland on 11–25 February 2017. The episode was not reproduced by a standard forecasting system. More details are given in [2]. The most
recent forecasts are available here: powietrze.uni.wroc.pl.
In this study we run and compared two simulations. The first simulation was
initialised each day at 00 UTC, using the chemistry cycling from a previous forecast
(GOCART). The second run the simulation was also initialised at 00 but the initial
conditions were modified by incorporation of observational data by the GSI system
(GOCART_GSI). Both forecasts were run for the 72 h (as a standard procedure in
our system), however only first 24 h were analysed in this study. To verify the results
we used factor of two (FAC2), mean bias (MB), mean gross error (MGE), normalised
mean bias (NMB), and correlation coefficient (R).
42.3 Results
Mean observed concentrations of PM10 and PM2.5 were respectively 57.2 and
46.4 µg m
−3 during the period of 11–25 February 2017. Very high concentrations
were observed in the first part of the analysed episode. PM2.5 concentrations reached
or exceeded 150 µg m
−3 at many Polish stations between 13th and 18th of February
(Fig. 42.1).
In general the WRF-Chem model underestimated observed PM2.5 concentrations.
However, all statistics were improved after assimilation of ground based observations of PM2.5. MB decreased from −11.4 to −4.8 µg m
−3 , FAC2 increased from
0.68 to 0.74 and R from 0.63 to 0.73 (Table 42.1). Time series in Fig. 42.1 shows that
GOCART_GSI better fits in the observations during the peak of concentrations, however for the warmer period it overestimates concentrations at the Warsaw station. An
improvement between GOCART_GSI and GOCART is also observed for PM10 but
the changes are smaller than for PM2.5. FAC2 and R increased from 0.63 to 0.65 and
from 0.56 to 0.63, respectively. The GOCART simulation underestimated observed