Introduction
Models for studying historical climate change and for projecting future
climate have increased in complexity and sophistication in recent years and
the importance of including atmospheric composition as a component of such
models is now well established e.g.. Gas-phase
pollutants, such as tropospheric ozone (O3), exert a positive radiative
forcing on climate , while the radiative
forcings associated with aerosol–radiation and aerosol–cloud interactions
are partly masking the strong positive forcing associated with long-lived
greenhouse gases . A changing climate, in turn, has an impact
on both natural emissions e.g. and chemistry
and aerosol processes themselves e.g.,
influencing atmospheric composition. Atmospheric composition and near-surface
air quality are intricately linked and poor air quality has impacts on human
health e.g.. In addition, surface O3 can adversely
impact crop growth , while aerosols can potentially promote
global plant productivity by increasing the diffuse fraction of
photosynthetically active radiation .
Given the interactions between atmospheric composition, air quality, and
climate, it is essential that the development of climate change mitigation
policies and air quality abatement strategies are developed jointly and
consider the full spectrum of co-benefits and trade-offs
e.g.. As a result, there is a strong need
for models that can simulate both climate and air quality. Likewise, it is
also necessary to develop modelling frameworks which can dynamically
downscale global climate and air quality projections to the regional scale,
on which population centres and crop locations vary significantly.
Downscaling allows a greater level of detail to be made explicit and
analysed. Air pollutant concentrations exhibit a high degree of spatial
inhomogeneity compared to meteorological fields such as temperature and wind,
and more highly resolved regional modelling can improve the representation
and evolution due to more highly resolved emissions and the dependence of
reaction rates on concentrations of reactive species. A further imperative
for higher-resolution modelling concerns the sensitivity of composition
projections to the difference in meteorology. For example,
discuss the sensitivity of O3 under regional climate change to cumulus
cloud parametrizations. In their
review article, cite a number of other examples where
significantly differing model predictions are attributed to differences in
air pollution meteorology between global and higher-resolution regional
models.
Various modelling configurations have been employed in studies of regional
air quality in the context of present-day climate and under future climate
change scenarios. A common approach has been to use a global–regional
climate model nest to provide meteorology and then use the stored fields to
drive an offline chemistry transport model (CTM) e.g.. This approach was used, for example, to investigate the impacts
of emission changes on UK O3 and European air quality by
and , respectively. Another example is ,
which nests the WRF-CMAQ (Weather Research and Forecasting – Community
Multi-scale Air Quality) air quality model over the UK
domain inside a European regional model but takes initial and lateral
boundary conditions (LBCs) for composition and climate from two different
global models. Some examples of future climate and air quality simulations
are those carried out by , , and
. Recognizing the advantages of more closely coupled
meteorology and composition, online models have increasingly been developed.
Initially this was mainly in the context of global general circulation models
(GCMs) for climate modelling, where long timescale simulations potentially
render even small feedback mechanisms between composition and meteorology
important. Results from some of these models have been used in the latest
Intergovernmental Panel on Climate Change (IPCC) Assessment reports
. Online regional chemistry models are
a more recent development, with applications to air quality forecasting
e.g. and impacts from a changing climate
e.g.. , for example,
nest the WRF-CMAQ online regional model inside the atmospheric component of the Community Earth System Model
CESM; and referred to the configuration as CESM-NCSU
CESM – North Carolina State University;. Single online
chemistry models that can be used at all scales, from global through regional
and even to urban-scale resolutions, represent the most advanced modelling
configuration. The first model with this capability was GATOR-GCMM
gas, aerosol, transport, radiation, general circulation and mesoscale
model;, which linked existing global and regional versions of
the GATOR model such that the gas, aerosol, and radiative parts of the two
scales were the same, although the meteorological and transport parts
differed. This capability has also since been implemented more recently in
GU-WRF/Chem , which started from a mesoscale model WRF-Chem
e.g. re-configured for the global scale. These models
are capable of running regional models nested within a consistent global
chemistry model.
In this paper we describe and evaluate a new modelling framework which uses a
more consistent set of models to go from the global scale down to the UK
national scale. We employ the Met Office's Unified Model, MetUM
, to downscale from a global composition-climate model
(GCCM) configuration to the UK national scale, via a regional
composition-climate model (RCCM) configuration. At each scale, model
configurations of MetUM appropriate to the resolution are employed, but the
use of a single framework results in a higher degree of consistency across
the scales. The global climate model used is based on the Global Atmosphere
3.0 (GA3.0) configuration of HadGEM3 and the RCCM is a
limited area version, described by . The inner nest is
the AQUM regional air quality forecast model. This operates at a resolution
of 12 km and is used operationally to provide the UK national air quality
forecast. The forecasts generated by AQUM are evaluated against hourly
pollutant measurements on a daily basis . Whilst we have
sought to maximize consistency between the models, there do remain some
differences and these are noted and described in subsequent sections. The
purpose of the present paper is to describe the new modelling framework and
to evaluate simulations of present-day air quality by comparing against UK
observations. The paper is structured as follows. Section
describes the modelling framework employed in this study. Section
describes the experimental set-up of the present-day simulations.
Section presents results on the performance of the nested
configurations and a discussion with concluding remarks can be found in
Sect. . This modelling framework has also been used to
downscale global climate and air quality projections for the 2050s onto the
UK national scale and is discussed in .
Modelling system description
In this section, we provide a brief overview of each of the scientific
configurations of the MetUM employed in this study (this is presented in
tabular form to allow comparison of the model configurations in
Table in Appendix A). We give a summary description of the model dynamics
and model physics, and details of the two-step, one-way nesting approach
developed. A discussion of the chemistry and aerosol schemes is also
included.
Global Composition-Climate Model (GCCM)
The GCCM is based on the Global Atmosphere 3.0/Global Land 3.0 (GA3.0/GL3.0)
configuration of the Hadley Centre Global Environmental Model version 3
HadGEM3,, of the Met Office's Unified Model
MetUM,. Soil–vegetation–atmosphere interactions are
calculated using the Joint UK Land Environment Simulator
JULES, and a full description of the GCCM can be found
in . The model has a horizontal resolution of
1.875∘ × 1.25∘, which translates to approximately
140 km × 140 km at the mid-latitudes. The model has 63 levels in the
vertical, spanning up to 41 km with the first 50 levels below 18 km. The
model's dynamical time step is 20 min.
The GA3.0 configuration of HadGEM3 incorporates an
interactive aerosol scheme, CLASSIC (Coupled Large-scale Aerosol Simulator
for Studies in Climate; ). CLASSIC is a
mass-based aerosol scheme in which all the aerosol components are treated as
external mixtures. The scheme simulates ammonium sulfate, mineral dust, soot,
fossil-fuel organic carbon (FFOC), biomass burning (BB), and ammonium nitrate
in a prognostic (evolving) manner, and biogenic secondary organic aerosols
are prescribed from a climatology. Sea salt is treated as a diagnosed
quantity over sea points in the model; a limitation of this is that it does
not contribute to particulate matter predictions over land points. The
aerosols can influence the atmospheric radiative and cloud properties through
aerosol–radiation and aerosol–cloud interactions, but for this study, these
interactions have been switched off. The reasons for this were 2-fold:
(1) the primary focus of this study was on the simulation of air quality, and
not on the impact of air quality on model dynamics, and (2) for statistical
significance, much longer simulations are required when radiative and
microphysical feedbacks are active (typically 20–30 model years as opposed
to 5–7 years without these feedbacks).
The gas-phase chemistry in the GCCM is simulated by a tropospheric
configuration of the United Kingdom Chemistry and Aerosol (UKCA) model
. However, for this study, the two
tropospheric chemistry schemes described in were replaced
by an extended tropospheric chemistry scheme, called UKCA-ExtTC. This version
of UKCA applies a more detailed gas-phase chemistry scheme that has a
significantly larger number of chemical species – 89 chemical species in
comparison to the 41 and 55 in the StdTrop and TropIsop chemistry schemes in
, respectively – and chemical reactions – 203 in
UKCA-ExtTC in comparison to the 121 and 164 described in .
The UKCA-ExtTC chemical mechanism has been designed to represent the key
chemical species and reactions in the troposphere in as much detail as is
necessary to simulate atmospheric composition and air quality, while
retaining the capability to conduct decade-long climate simulations. As a
result, it is more suitable for air quality studies and has been applied
successfully in previous studies e.g..
Of the 89 chemical species that UKCA-ExtTC considers, 63 are transported as
“tracers”. For the remaining 26 species, transport is negligible in
comparison to chemical transformation during one model time step, and hence
they are treated as “steady-state” species. UKCA-ExtTC uses the same
backward Euler solver, a chemical time step (5 min), wet and dry deposition,
large-scale and convective transport, and boundary layer treatment of tracers
as described in . A separate, detailed description of this
extended version of UKCA is in preparation .
Although UKCA has two options in relation to photolysis ,
the photolysis reactions in this configuration are handled using offline
rates, calculated in the Cambridge 2-D model using the
two-stream approach of . We used this option in the GCCM and
RCCM configurations mainly for two reasons. First, the extended tropospheric
chemistry version of UKCA, UKCA-ExtTC, has been developed and extensively
evaluated only with the 2-D photolysis model, and there was no time within
the scope of this work for development and evaluation of UKCA-ExtTC coupled
to the Fast-J online photolysis model. Second,
there is a non-negligible, albeit not prohibitively large, extra cost
attached to using the Fast-J online photolysis scheme over the 2-D photolysis
scheme. With the offline photolysis scheme, the photolysis rates are read in
by UKCA-ExtTC on the first time step of the model integration and
interpolated in time and space at each model grid box. The impact of cloud
cover, surface albedo, and aerosols is included in the form of a
climatological cloud cover, prescribed albedo, and aerosol loading,
respectively.
A two-way coupling between the UKCA-ExtTC chemistry scheme and the CLASSIC
aerosol scheme is applied through the provision of simulated oxidant species
(ozone (O3), the hydroxyl (OH) and hydroperoxyl (HO2) radicals, and
hydrogen peroxide (H2O2)) and the provision of nitric acid (HNO3) as
a nitrate aerosol precursor. Oxidation of sulfur dioxide (SO2) and
dimethyl sulfide (DMS) occurs in both the gas phase and the aqueous phase to
form sulfate aerosol and the HNO3 generates ammonium nitrate aerosol with
any remaining ammonium ions after reaction with sulfate. The coupling is
two-way because gas-phase concentrations of both H2O2 and HNO3 are
depleted, following sulfate and
nitrate aerosol formation.
Although UKCA does include an aerosol microphysics scheme, GLOMAP-mode
, the simpler mass-based CLASSIC aerosol scheme
was used across the three MetUM
configurations for the following reasons: (1) the UKCA-ExtTC chemistry scheme
has historically only been coupled to the CLASSIC scheme and there was no
time within the scope of the current study to couple it to GLOMAP-mode,
(2) the operational air quality forecast model, AQUM, also uses CLASSIC as
its aerosol scheme, and one of the aims of this work was to maximize the
consistency in the treatment of both meteorology and composition across the
three model domains, and (3) the computational cost of running both
UKCA-ExtTC and GLOMAP-mode would have been prohibitively expensive.
Regional Composition-Climate Model (RCCM)
The RCCM, referred to as the HadGEM3-A “regional” (HadGEM3-RA)
configuration, is described in detail in , and is also
based on the GA3.0/GL3.0 configuration of HadGEM3 . The
RCCM has a horizontal resolution of 0.44∘ × 0.44∘
(roughly 50 km × 50 km) with a domain covering most of Europe and
northern Africa (Fig. ) and the same 63 vertical levels as the
GCCM. The RCCM closely follows the GCCM configuration
(Sect. ), with the same dynamical solver, radiation,
precipitation, and cloud (PC2) schemes. The same principal components are
included: the UKCA-ExtTC chemistry model, the CLASSIC aerosol model, and the
JULES land-surface model. The model dynamical time step was reduced to
12 min (20 min in GCCM) to account for the increase in resolution and
shorter turnaround of dynamical processes and interactions. The chemical time
step is 5 min. Boundary conditions, used to drive the RCCM from the GCCM,
will be discussed in Sect. .
AQUM
The final, high-resolution nest employed is the AQUM (Air Quality in the
Unified Model) air quality forecast model. AQUM, like both the GCCM and the
RCCM, is based on the MetUM. AQUM has a horizontal resolution of
0.11∘ × 0.11∘ (approximately 12 km × 12 km) on a
“rotated pole” grid, covering the UK and nearby western Europe (see
Fig. ), with 38 vertical levels up to 39 km. The LBCs, provided
by the RCCM, are on 63 levels but interpolated onto the 38 levels of AQUM.
The dynamical and chemistry time steps are both 5 min.
The set-up of this model is described in detail in and uses
the same parametrization schemes as the global and regional CCMs described
above, apart from large-scale cloud, where AQUM uses the diagnostic cloud
scheme as described by . As with the GCCM and RCCM, AQUM uses
the CLASSIC aerosol scheme () and the UKCA
model for its gas-phase chemistry. This helps to improve consistency between
many aspects of the models. For example, large-scale and convective
transport, boundary layer mixing, and wet and dry deposition are similar
between all the nests. However, a different chemistry mechanism, the Regional
Air Quality (RAQ) scheme, is used and the photolysis scheme also differs.
Photolysis rates in AQUM are calculated with the Fast-J online photolysis
scheme , which is coupled to the modelled liquid
water and ice content, and sulfate aerosols at every time step.
Nested modelling domains. The rectangular boundary of the figure is
an extract of the GCCM (resolution 140 km) containing the RCCM domain
(resolution 50 km) plotted in blue and the AQUM domain (resolution 12 km)
in red.
The RAQ chemistry scheme pre-dates the ExtTC scheme and has been used in AQUM
throughout its development and use as a forecast model. The experience
developed with AQUM and the understanding of model performance established
relies on the continuing use of this scheme and therefore we chose to retain
this scheme for the final nest. The scheme has 40 transported species,
18 non-advected species, 116 gas-phase reactions, and 23 photolysis
reactions; 16 of the transported species are emitted: nitrogen oxide (NO),
methane (CH4), carbon monoxide (CO), formaldehyde (HCHO), ethane
(C2H6), acetaldehyde (CH3CHO), propane (C3H8), acetone
(CH3COCH3), isoprene (C5H8), methanol (CH3OH), hydrogen
(H2), ethene (C2H4), propene (C3H6), butane (C4H10),
toluene, and o-xylene. As was the case in the GCCM and the RCCM, there is
two-way coupling of oxidants between CLASSIC and the RAQ chemistry scheme.
Further details of the RAQ scheme can be found in .
A comparison of the MetUM settings for all three configurations described
above can be seen in Table .
Experimental set-up
In this section, a description of the experimental set-up for modelling
present-day air quality using the configurations of MetUM is provided,
covering meteorological lower boundary conditions, emissions, upper boundary
conditions, and lateral boundary conditions.
Model simulations and model calibration
Both the GCCM and the RCCM were initialized using meteorological fields from
a pre-existing 20-year simulation of the standard HadGEM3 configuration. The
simulations for both these model configurations cover a total period of 6
model years representative of the decade centred around the year 2000, for
both meteorology and emissions. The first year is considered as an additional
spin-up and the last 5 years are used in the analysis. The GCCM was used to
produce the offline lateral boundary conditions (LBCs) at 6-hourly intervals
to drive the RCCM, together with the emissions and upper and lower boundary
conditions described below. LBCs include meteorological drivers (3-D winds,
air temperature, air density, Exner pressure, humidity, and cloudiness),
important chemical tracers from UKCA-ExtTC (O3, NO, nitrogen dioxide
(NO2), HNO3, dinitrogen pentoxide (N2O5), H2O2, CH4, CO,
HCHO, C2H6, C3H8, CH3COCH3, and peroxy
acetyl nitrate (PAN)), gas-phase aerosol
precursors (SO2, DMS) and aerosols (dust, sulfate, nitrate, soot, FFOC,
and BB) from CLASSIC. In turn, the RCCM produced meteorological and
composition LBCs required to drive the AQUM national-scale air quality model.
Simulations with AQUM were initialized from the last month of the first year
of the RCCM and were continued for 5 model years by applying the LBCs
supplied by the RCCM offline at 6-hourly intervals. The chemical and aerosol
species provided in the LBCs are dust, SO2, DMS, SO4, soot, OCFF,
nitrate, O3, NO, NO2, N2O5, HONO2, H2O2, CH4, CO,
HCHO, C2H6, PAN, and C3H8.
For lower boundary conditions the GCCM used monthly mean distributions of sea
surface temperature (SST) and sea ice cover (SIC), derived for the present
day (1995–2005) from transient coupled atmosphere–ocean simulations
of the HadGEM2-ES model . It should be
pointed out here that the entire set-up is intended to represent a decadal
climatological mean state of near present-day conditions encompassing the
period from 1995 to 2005 and centred on the year 2000. This particularly
applies to the meteorological drivers (sea surface temperature, SSTs, and sea
ice cover) and the anthropogenic emissions of pollutants. The latter will be
discussed in more detail in Sect. . The vegetation
distribution for each of the simulations was prescribed using the simulated
vegetation averaged for the same decade from this transient climate run, on
which crop area, as given in the 5th Coupled Model Intercomparison Project
(CMIP5) land use maps , was superimposed. The same
present-day SST and SIC climatologies developed for the GCCM were regridded
to the RCCM and the AQUM domains using a simple linear regridding algorithm.
The GCCM was calibrated against O3 measurements from the monitoring
station located at Mace Head Atmospheric Research Station in western Ireland at
53.3∘ N and 9.9∘ W. It is part of the Automatic Urban and
Rural Monitoring Network (AURN) which is run by a number of institutions
coordinated by Defra. The Mace Head monitoring station is representative of
rural background conditions. Model output has been compared to the annual
cycle of monthly mean O3 which is based on a multi-year climatology of
observed near-surface O3 concentrations. The parameterized O3 surface
dry deposition was used to perform the calibration as the model shows very
high sensitivity to deposition. The model has been optimized to reproduce
both the magnitude and seasonal cycle of O3 at the Mace Head site in the
global model domain as closely as possible by varying the O3 surface dry
deposition flux within its uncertainty limits. An increase in the O3 dry
deposition by 20 % yielded the best agreement, with respect to both O3
monthly mean surface concentration and seasonal cycle, with the observed
climatology at the Mace Head station, which is representative of the O3
background concentration in the lower troposphere, in the study area.
As the RCCM uses the same code base as the GCCM, this calibration is
inherited by the former automatically. The model calibration has been applied
to optimize consistency between the individual configurations in the
global-to-national model nesting chain.
Due to the different chemistry scheme used in AQUM, the calibration used by
the GCCM was not incorporated into AQUM as the RAQ scheme has been developed
with performance over the UK as its main focus. This is unlike the GCCM,
where performance usually has to be taken into account over the entire globe,
which may lead to worse performance in some regions such as the UK. The
calibration was performed to ensure that the best possible boundary conditions are applied
to the innermost, national-scale domain. Mace Head station was chosen because
it is representative of the large-scale background tropospheric ozone level
in the study area and includes the impact of transcontinental influx of
pollution from North America.
Emissions
A consistent set of emissions has been used for all three model
configurations through using the same source data, but then regridding to the
required resolution for each model.
The emissions of reactive gases and aerosols from anthropogenic and biomass
burning sources used in this study are based on the dataset used for CMIP5
simulations and described by . The models are all driven
by decadal mean present-day emissions from CMIP5, representative of the
decade centred on 2000. An example of the emissions for the different
domains is given for NO in Fig. , while a full set of emission
totals can be seen in Tables , , and
.
UKCA-ExtTC takes into account emissions for 17 of its chemical species:
nitrogen oxides (NOx = NO + NO2), carbon monoxide (CO),
hydrogen (H2), methanol, formaldehyde, acetaldehyde and higher aldehydes,
acetone (CH3COCH3), methyl ethyl ketone, ethane (C2H6), propane
(C3H8), butanes and higher alkanes, ethene, propene and higher alkenes,
isoprene, (mono)terpenes, and aromatic species. Of these butanes and higher
alkanes, propene and higher alkenes, terpenes, and aromatics are treated as
lumped species. Surface emissions are prescribed in most cases. The only
exception is the emission of biogenic volatile organic compounds (BVOCs)
which are calculated interactively in JULES using the iBVOC emission model
. The emission of biogenic terpenes, methanol, and acetone
follows the model described in . As summarized in
Table , global annual total emissions of biogenic isoprene and
monoterpenes interactively computed with the iBVOC model of, for instance,
480 Tg(C) yr-1 and of 95 Tg(C) yr-1 are in reasonably good
agreement with most other state-of-science interactive biogenic VOC emission
models e.g.
and global bVOC emission inventories e.g..
A detailed evaluation of the model performance is presented in
.
Emissions of NOx from lightning are taken into account in UKCA. Lightning
NOx emissions are calculated interactively at every time step, based on
the distribution and frequency of lightning flashes following
, , and . In this
parametrization the lightning flash frequency is proportional to the height
of the convective cloud top in all the models. For cloud-to-ground (CG)
flashes lightning NOx emissions are added below 500 hPa, distributed from
the surface to the 500 hPa level, while NOx emissions resulting from
intra-cloud (IC) flashes are distributed from the 500 hPa level up to the
convective cloud top. The emission magnitude is related to the discharge
energy where CG flashes are 10 times more energetic than IC flashes
. The scheme implemented in the GCCM produces a total global
emission source of around 7 Tg(N) yr-1, which is in good agreement
with the literature cf. e.g..
Soil-biogenic NOx emissions are taken from the monthly mean distributions
from the Global Emissions Inventory Activity
(http://www.geiacenter.org/inventories/present.html), which are based
on the global empirical model of soil-biogenic NOx emissions of
giving a global annual total of 5.6 Tg(N) yr-1.
For CH4, the UKCA model can be run by prescribing surface emissions or
prescribing either a constant or time-varying global mean surface
concentration. For the simulations being evaluated here, a time-invariant
CH4 concentration of 1760 ppbv was prescribed at the surface.
The sea salt and mineral dust emissions are computed interactively at each
model time step based on instantaneous near-surface wind speeds
. Mineral dust is a fully prognostic, advected
species but, as mentioned in Sect. , sea salt is not advected
and makes no contribution to model aerosol concentrations over land.
Similarly the ocean DMS emissions are computed based on wind speed,
temperature, and climatological ocean DMS concentrations from
, using the sea–air exchange flux scheme from
.
Emissions for AQUM are derived by re-gridding emissions from the regional
model to the required 0.11∘ resolution. The ExtTC and RAQ chemistry
schemes emit different anthropogenic VOC species; consequently, some
conversion is required. Our approach is to sum the anthropogenic VOC emission
from ExtTC and apportion this total according to the values given in
Table . These values were derived using the tabulated VOC
emission fraction data over the UK for 2006 given by . A
consequence of this is that for some species the emission total in the
smaller AQUM domain exceeds that of the larger RCCM domain. However, the
total VOC emitted is conserved between AQUM and the corresponding part of the
RCCM domain. For biogenic isoprene emissions, AQUM uses an offline, monthly
varying climatology which was derived from the online isoprene emission
fluxes generated by the RCCM. A diurnal cycle is applied to account for
daylight hours.
VOC split to convert total emitted VOCs from ExtTC to RAQ emitted VOCs.
These factors sum to 1.0.
Species
Conversion factor
HCHO
0.055
C2H6
0.156
CH3CHO
0.015
C3H8
0.110
CH3COCH3
0.078
CH3OH
0.116
C2H4
0.079
C3H6
0.034
C4H10
0.238
toluene
0.095
o-xylene
0.024
AQUM with higher-resolution emissions
Following an initial evaluation of results, an additional model run was also
carried out using AQUM. This run was identical to the main AQUM run (using
the same RCCM LBCs), with the exception of the anthropogenic emissions used.
A new set of the latter was produced based on the higher-resolution datasets
which AQUM uses for its operational air quality forecast; these are described
further in . Figure shows the impact of these
emissions for NO. The highest-resolution input data to these emissions are at
1 km over the UK, although regridded to the 12 km resolution required by
AQUM. These are based on 2006 emissions, but the total emission has been
rescaled to match the year 2000 decadal mean areal totals given by
(as described in Sect. ). For the
remainder of the paper, this additional run will be referred to as AQUM-h.
Upper boundary conditions
While the chemistry is calculated interactively up to the model top in each
configuration, upper boundary conditions are applied at the top of each model
domain to account for missing stratospheric processes such as those related
to CH4 oxidation and bromine and chlorine chemistry. These boundary
conditions are described in detail in and are only
briefly discussed here. For O3, the field used in the radiation scheme
by MetUM in the absence of interactive chemistry is used to overwrite the
modelled O3 field in all model levels that are 3–4 km above the
diagnosed tropopause . For stratospheric odd nitrogen
species (NOy), a fixed O3 to HNO3 ratio of 1.0 / 1000.0
kg(N) / kg(O3) from is
applied to HNO3 in the same vertical domain. Finally, for CH4, an
additional removal term is applied in the three uppermost levels of the
model. This CH4 loss term was calculated in to be
50 ± 10 TgCH4 yr-1 in a global configuration.
Results
Our aim is to evaluate the air pollutant concentration output from the RCCM
and AQUM simulations using different datasets representative of the true air
quality in the UK. In this way, we also aim to assess the potential for
improving modelled air pollutant concentrations by increasing model spatial
resolution. The datasets we use include (i) in situ observations of hourly
air pollutant concentrations from the UK Automatic Urban and Rural Network
(AURN) and (ii) annual mean surface pollutant concentrations produced by the
Pollution Climate Mapping (PCM) model which also takes into account
observations described by . This model produces gridded fields
at a spatial resolution of 1 km over the whole of the UK.
NO emissions for all models: GCCM (a), RCCM (b), AQUM (c), and
the higher-resolution emissions run (AQUM-h) (d).
Another aspect of the analysis undertaken is to employ two different
approaches to model assessment. The first uses standard verification metrics
such as bias based on site-specific comparisons averaged over the 5-year
modelled period. The second approach uses neighbourhood verification
techniques which consider the area surrounding a particular point and thus
allow for some mismatch in the spatial positioning of elevated pollutant
values, thereby avoiding the well-known “double penalty” problem
.
We begin with a qualitative comparison of the GCCM against
the two limited-area models in order to illustrate the need for improved resolution
over that of the GCCM for air quality applications.
Comparison to GCCM
Figure compares UK monthly mean NO2 concentrations for
June calculated from runs of the GCCM, RCCM, AQUM, and AQUM-h models. In the
GCCM plot the resolution is wholly insufficient to realistically represent
the elevated NO2 levels around the UK urban centres (London, West
Midlands, Greater Manchester, West Yorkshire, Edinburgh) and in the busiest
shipping lanes and ports (English Channel, Bristol Channel, Southampton,
Liverpool). The representation improves qualitatively as we move to the right
in this plot. It can clearly be seen that higher-resolution modelling is
essential for providing realistic pollutant representations at more localized
spatial scales.
Monthly mean NO2 concentrations over the UK for June for the four
different model runs. From left to right: GCCM, RCCM, AQUM, AQUM-h.
Comparison against in situ observations
In this section we compare results from the RCCM, AQUM, and AQUM-h
simulations with suitable averages derived from observations from the UK
Automatic Urban and Rural Network (AURN,
https://uk-air.defra.gov.uk/networks/network-info?view=aurn) for
2001–2005. Note that the years here refer only to the observation time
series and have no intrinsic meaning for the model simulations. As discussed
in Sects. and , the simulations represent
climatological mean states representative of the decade from 1995 to 2005 and
centred on the year 2000. We compare the model to the AURN 2001–2005
observational record because it represents the most complete record for the
selected period available. The individual model years do not correlate with
the corresponding years in the observational record. We performed the
multi-year simulations to obtain a statistical sample to investigate
interannual variability to some degree. The variability, of course, will be
reduced due to the fact that composition and climate have been decoupled, but
there is still variability in the atmospheric chemistry. From
AURN we only
consider “background” sites which include the site classifications of
remote, rural, suburban, and urban background. We are therefore excluding
sites which we expect to be un-representative of a large area, such as
roadside or industrial sites. As the models are driven by climatological
meteorology, we do not expect the model results to match the hourly AURN
observations; hence, we compare values averaged over the 5-year period with
corresponding averages derived from the hourly observations.
Statistics comparing modelled air pollutant concentrations to AURN
observations, for the period of the observational record
1 January 2001–31 December 2005 (for the correlation between model years and
the observational record compare the discussion in the text).
RCCM
AQUM
AQUM-h
NO2
Number of sites
65
65
65
Bias (µg m-3)
-4.76
-5.47
-0.80
% Observations > threshold (= 65.0 µm-3)
6.21
6.21
6.21
% Model > threshold (= 65.0 µg m-3)
1.86
2.07
5.64
O3
Number of sites
65
65
65
Bias (µg m-3)
6.23
13.94
9.96
% Observations > threshold (= 100.0 µg m-3)
2.39
2.39
2.39
% Model > threshold (= 100.0 µg m-3)
3.18
8.54
7.07
PM10
Number of sites
40
40
40
Bias (µg m-3)
-12.45
-13.32
-14.41
% Observations > threshold (= 50.0 µg m-3)
4.18
4.18
4.18
% Model > threshold (= 50.0 µg m-3)
0.99
0.87
0.85
PM2.5
Number of sites
2
2
2
Bias (µg m-3)
0.33
-0.75
-2.46
% Observations > threshold (= 35.0 µg m-3)
1.08
1.08
1.08
% Model > threshold (= 35.0 µg m-3)
3.93
3.11
2.40
SO2
Number of sites
49
49
49
Bias (µg m-3)
2.61
1.44
1.59
% Observations > threshold (= 25.0 µg m-3)
2.89
2.89
2.89
% Model > threshold (= 25.0 µg m-3)
3.98
3.71
5.31
Frequency distribution of the main pollutants: (a) NO2,
(b) O3, (c) PM10, (d) PM2.5, and
(e) SO2. Observations are shown in black, RCCM in red, AQUM in
blue, and AQUM-h in green.
Monthly mean concentrations of (a) NO2 and
(b) O3. Observations are shown in black, RCCM in red, AQUM in
blue, and AQUM-h in green.
NO2
Figure a shows a frequency distribution of hourly observed
concentrations of NO2 with corresponding frequency distributions for
modelled concentrations from the RCCM, AQUM, and AQUM-h configurations. It is
clear that the AQUM-h model distribution more closely matches the observed
distribution than the other model configurations, illustrating the importance
of increased spatial resolution and emissions for this pollutant.
Corresponding statistical measures of model skill are given in
Table . The bias in RCCM and AQUM against AURN observations
is -4.76 and -5.47 µg m-3, respectively, but is reduced
to -0.80 µg m-3 in AQUM-h. In Table a
comparison of the percentage of observations/model values greater than the
65.0 µg m-3 threshold is also included; it illustrates that
AQUM-h simulates observed frequencies of higher NO2 concentrations well,
making it better suited to calculating health burdens due to elevated levels
of NO2 e.g.. However, shown in
Fig. a is a comparison of the seasonal cycle of observed
and modelled NO2 concentrations, averaged over all AURN sites considered.
This shows that none of the models is able to fully capture the seasonal
cycle of NO2, with wintertime modelled concentrations biased low, while
the RCCM and AQUM straddle the observed concentrations during summer. This is
possibly due to the poor representation of the monthly variation of emissions
over the UK in the global model which is then inherited by the higher
resolution models. However, other processes such as boundary layer mixing or
chemistry could equally contribute. Further work would be required to
elucidate this clearly.
O3
Relevant statistics are given in Table , while a frequency
distribution plot, showing the distribution of hourly O3 concentrations
over the entire period for models and observations, is shown in
Fig. b and the seasonal cycle is given in
Fig. b. The latter plot illustrates that the pattern of the
seasonal cycle of O3 is captured reasonably well; however, the modelled
spring–summer maximum persists too long and does not replicate the gradual
decline in monthly mean concentrations as indicated by observations. This has
implications for the use of modelled O3 to quantify health impacts from
long-term exposure to O3 during warmer months, as indicated by studies in
North America . In the frequency distribution
plots in Fig. b, it can be seen that all models are able to
reproduce the shape of the observed distribution quite well but differ in
their most frequent concentration, corresponding to different model biases.
The RCCM exhibits the smallest bias against observations of
+6.23 µg m-3, and AQUM the greatest at
+9.96 µg m-3 (see Table ). However, the
RCCM used an offline photolysis scheme , whilst both
configurations of AQUM used the interactive Fast-J scheme .
Given the different photolysis schemes used, a sensitivity experiment for a
single month of July was carried out, in which AQUM-h was re-run with offline
photolysis. The O3 bias for this month is 7.33 µg m-3 for
the RCCM, 22.48 µg m-3 for AQUM, and
13.95 µg m-3 for AQUM-h. Although the photolysis rates
relevant to O3, j(NO2) → NO, and
j(O3) → O1D are known to be biased low in the
offline photolysis scheme relative to both observations and online photolysis
, the modelled O3 bias in AQUM-h is reduced to
+6.99 µg m-3 with the offline scheme, which is marginally
better than the RCCM. However, the sensitivity of surface O3 to the choice
of photolysis scheme found here differs from two previous studies
. Both of these studies found that O3
decreased in the Northern Hemisphere by less than 5 % when switching from
offline to online photolysis and, indeed, the changes in the tropospheric
O3 budget were consistent between the two studies. In addition,
found no significant change in modelled O3 evident at
NH mid-latitude sites (e.g. Mace Head). However, both and
were global studies rather than the regional scale
considered here. Another conflicting factor is the calibration which has been
applied to the RCCM for the O3 dry deposition, which would have an impact
on the O3 concentrations, although this would have impacted AQUM through
the LBCs. This calibration was not included in the papers described above,
which may help to explain the conflicting results. Consequently, these
factors make it difficult to isolate and quantify the impact of the higher
resolution third nest on model performance.
PM10
Relevant statistics are given in Table , while a frequency
distribution plot, showing the distribution of hourly PM10 values over
the entire period for models and observations, is shown in
Fig. c.
For PM10, none of the models are able to reproduce the shape of the
observed distribution, and there is a significant negative bias across all
the model configurations (between -12.45 and
-14.41 µg m-3), with AQUM-h exhibiting the poorest
performance. Poor modelling performance for PM10 is a common feature of
many global composition and regional air quality models
e.g. and is often attributed to the unreliability
of primary emissions of coarse component aerosol, both from anthropogenic and
biogenic sources. In our simulations the lack of sea salt in modelled values
over land points plays a significant role in this underprediction.
estimate that over north-western Europe sea salt contributes
on average between 7 % (kerbside sites) and 12 % (rural sites) of the
observed annual mean PM10. In periods of strong winds and at sites close
to the coast downwind of the sea, values may be considerably higher. A
related consequence of our lack of inclusion of sea salt is that our aerosol
modelling does not include sodium nitrate, and so this coarse component of
secondary aerosol is also missing from our estimates. These underpredictions
could potentially affect the quantification of health effects due to
short-term and long-term exposure of PM10, as documented by the
.
PM2.5
Relevant statistics are given in Table , while a frequency
distribution plot, showing the distribution of hourly PM2.5 values over
the entire period for models and observations, is shown in
Fig. d.
For the finer PM2.5 component of aerosol, the models perform
significantly better in capturing the shape of the observed distribution than
for PM10; there is a small positive bias for PM2.5 in the RCCM
(+0.33 µg m-3), whereas AQUM becomes slightly negative
(-0.75 µg m-3) and AQUM-h more negative still
(-2.46 µg m-3).
However, the observed frequency distribution is only based on two background
observational sites available for PM2.5 in the UK for the 2001–2005
time period. The introduction of PM2.5 monitoring stations in the UK
increased significantly from 2009 and we explored the possibility of using
observations from 2011 to 2015 to generate a proxy for the 2001–2005
frequency distribution. However, we found that the PM10 distribution
changed significantly over the 10 years and concluded that it was not valid
to use the more recent PM2.5 observations in place of 2001–2005
observations. Consequently, due to the paucity of PM2.5 observations for
the 2001–2005 time period against which to compare, for the remainder of
this paper, we shall no longer consider PM2.5 results.
SO2
Relevant statistics are given in Table , while a frequency
distribution plot, showing the distribution of hourly SO2 values over the
entire period for models and observations, is shown in Fig. e.
For SO2, the model configurations exhibit similar distributions to the
observed distribution, with generally positive biases of between +1.44 and
+2.61 µg m-3, with AQUM-h exhibiting the best performance.
Comparison against PCM
In order to assess the variation in the quality of modelled air pollutant
concentrations between the different model configurations, it is necessary to
consider full spatial fields rather than the site comparison afforded by in
situ observations described in the preceding section. Therefore, it is
essential to compare the models against a realistic spatial field and, for
this purpose, we use fields derived from the PCM model, as described in
. This sophisticated model combines information from a variety
of sources, including emission inventories and
observation datasets, to produce estimated annual mean surface pollutant
concentrations on a 1 km × 1 km grid over the entire UK for NO2,
SO2, PM10, and PM2.5. The data are freely available at
https://uk-air.defra.gov.uk/data/pcm-data. These results are widely
used in the UK to provide the background pollutant concentrations for local
air quality modelling studies and new site impact assessment studies. O3
is also modelled by PCM, but the output available is the number of days
exceeding 120 µg m-3 (as required by the European Union
ambient air quality directives
(http://eur-lex.europa.eu/LexUriServ/LexUriServ.do?uri=OJ:L:2008:152:0001:0044:EN:PDF)
rather than pollutant concentrations, and so cannot be used in our analysis.
In view of the lack of AURN PM2.5 observations (also used in deriving
the PCM maps) during the period 2001–2005 (as described in
Sect. ), we have not considered PM2.5 in the following
analysis.
PCM data for NO2 and PM10 are available for 2001–2005, while SO2
data are only available from 2002 onwards. A comparison (not shown) of the
PCM against the in situ AURN observations as done for the models in
Sect. proved the PCM verifies better than any of the other
models. PCM data from the available years were processed to produce 5-year
means (4 years for SO2) for comparison with the similarly averaged model
fields.
Comparisons between MetUM modelled annual mean concentrations and PCM annual
mean concentrations are shown for NO2, SO2, and PM10 in
Fig. . In these plots nearest neighbour regridding is used to
interpolate the model fields and the PCM fields onto the 12 km AQUM grid.
Spatial correlations have been calculated between the regridded model and PCM
fields (only at valid PCM data points, i.e. UK land points) and are shown at
the top of each figure.
Model and PCM mean fields for
different pollutants, regridded onto a 12 km AQUM grid. From left to right
the models are RCCM, AQUM, AQUM-h, a 12 km version of the PCM, and finally
the 1 km PCM for comparison. Plots also show the correlation between the
fields and the 12 km version of the PCM. Pollutants shown are
(a) NO2 (top row), (b) SO2 (middle), and
(c) PM10 (bottom).
Fractional skill score for the 95th percentile for
(a) NO2 and (b) SO2. The RCCM is shown in red, AQUM
in blue, and AQUM-h in green. The “Random” (dot-dashed) line represents the
FSS for a random forecast with the same fraction of points over the domain
exceeding the percentile threshold as the truth field. The “Uniform”
(dashed) line represents a forecast with the same fraction of points above
the percentile threshold in the neighbourhood surrounding each grid point as
the truth field for every grid point. Above this line the forecast is
considered skilful.
For the primary pollutants of NO2 (Fig. a) and SO2
(Fig. b), there is an improvement in correlation with the PCM
as we move from the RCCM to AQUM and finally AQUM-h: for NO2 the
correlations are 0.822, 0.824, and 0.836, respectively, while for SO2 the
correlations are 0.664, 0.743, and 0.761, respectively. For SO2, the
introduction or removal of strong point sources can influence the comparison
via a calculated spatial correlation. This is apparent in the AQUM-h plots in
Fig. b, where two new strong point sources in south-eastern
England are present in the 2006 data used to generate the AQUM-h emissions.
These modest increases in correlation with PCM (as our proxy for “truth”),
as model resolution increases, illustrate the benefits of increased
resolution modelling, with respect to both the model grid and the underlying
emissions data, in better capturing the strongly inhomogeneous spatial
distribution of these pollutants.
For PM10, however (Fig. c), this improvement in
correlation with higher resolution is not as clear. The correlation values
with the PCM are 0.841 for the RCCM, 0.912 for AQUM, and 0.883 for AQUM-h.
PM10 has a large secondary contribution which contributes a relatively
smoothly varying background to the PCM maps in Fig. c. This is
likely to be the reason for the lack of a clear improvement in PM10
modelling with the high-resolution AQUM-h model.
Beyond the figures shown above, we also investigated the correlation scores
by just considering data above fixed threshold concentration values (plots
not shown). However, these results were very variable, depending on the
threshold values considered, partly due to the biases (as given in
Sect. ).
Analyses based on neighbourhood comparisons: the fractional skill score
In evaluating a comparison of modelled air pollutant concentrations against
some gridded representation of true concentrations (such as the PCM fields
described above), small offsets in the spatial location of elevated values
can give an exaggerated contribution to simple metrics such as bias and root
mean square error evaluated at each grid point. This is commonly referred to
as the “double penalty” problem. The resulting analysis may then give a
misleading indication of the comparison between the two fields. So-called
“neighbourhood” verification techniques have
been developed to avoid these problems. Here, we consider the use of the
fractional skill score (FSS) (explained in detail in ) to
analyse the variation in model skill in representing spatial patterns. This
statistic has mainly been employed in evaluating the improvements offered by
high-resolution precipitation forecasts, where a “double penalty” problem
occurs if rain is forecast in a neighbouring grid box to where it was
actually observed (hence an incorrect forecast in both grid boxes). A lower
resolution forecast might place the forecast and observed shower in the same
grid box, resulting in an apparently improved forecast. Similar issues are
found in pollution modelling due to the high degree of inhomogeneity of air
pollutant concentrations and evaluation of the FSS may offer improved
comparisons.
The FSS is calculated by computing, for each grid box, the fraction of
neighbouring grid boxes which exceed a given threshold value (or percentile).
This is done both for the gridded model fields that are to be evaluated and a
gridded benchmark field representative of the “truth”, which in this case
is the PCM fields, as described in Sect. . This can be
repeated for varying neighbourhood sizes. As the size of the neighbourhood
increases, the fractional skill score should increase towards unity. A
forecast may be considered “skilful” at the grid scale where the model has
the correct fraction of points above the percentile threshold in the
neighbourhood surrounding each grid point as the truth field for every grid
point.
We have calculated the FSS using output from the three model configurations
(RCCM, AQUM, and AQUM-h) and compared it to the PCM for various threshold
values, based on both fixed thresholds and percentile values. An example set
of results is shown in Fig. . In these plots, the variation of
FSS against the spatial scale is shown for the RCCM, AQUM, and AQUM-h, using
a 95th percentile threshold. For NO2, there is little difference between
the three model configurations, and the same is found for PM10 (not
shown). Calculations using other fixed thresholds and different percentile
thresholds also show little difference. However, for SO2, AQUM-h shows the
best performance, crossing the threshold value of 0.5 at the shortest spatial
scale, and reflects the strong point sources of SO2 in contrast to NO2
emissions. The use of neighbourhood verification techniques to compare our
different nests has therefore not offered any obvious increased insight into
the differences between the models and the consequent impacts on improved
predictions across the spatial scales. This may be an indication that the
resolution differences between the models may not be the key factor in
determining performance, particularly for NO2 and PM10.
Summary and conclusions
This study describes the initial development of a more consistent framework
for dynamic downscaling of climate and air quality from a global
composition-climate model to the national scale, via a regional
composition-climate model and thence to a higher-resolution regional air
quality forecast model. In this attempt, some of the difficulties in
presenting a clear-cut, quantitative demonstration of the value of higher
resolution modelling have been made apparent. All three models use a single
modelling framework – the MetUM – but some differences between the models
do remain. The most notable of these are the different chemistry mechanisms,
photolysis schemes, and the calibration factor that have been used in the
GCCM and RCCM compared to AQUM. AQUM has been developed with forecasting air
quality over the UK as its primary aim, and performance has been optimized
for predicting in situ UK observations on an hourly timescale with a focus on
high impact, more extreme events. By contrast, the GCCM and RCCM have been
developed to predict global and regional climatologies, giving a faithful
representation of seasonal and annual means across the entire globe. These
differences have resulted in some of the inconsistencies highlighted in this
paper. This has led to a challenge in determining the benefits of a
three-level nest for downscaling to the regional scale, but has highlighted
important areas for consideration in future work.
The comparison of modelled air pollutant concentrations against in situ UK
observations was conducted initially by a traditional site-specific analysis,
with standard metrics such as bias. In addition, the impacts of model
resolution on pollutant spatial patterns were assessed via comparison to the
gridded PCM annual average pollution maps. In order to guard against the
susceptibility of the traditional verification methods to the double penalty
problem, an analysis was also carried out using a neighbourhood approach,
utilizing the fractional skill score (FSS), although the results from this
were generally inconclusive.
For NO2, significantly improved modelled concentrations can be
quantitatively demonstrated for the higher-resolution models, using higher
resolution emissions (biases of -4.76, -5.47, and
-0.80 µg m-3 for RCCM, AQUM, and AQUM-h, respectively).
This is readily understood, given the dependence of surface concentrations of
this primary pollutant on local emissions. For another primary pollutant,
SO2, a modest benefit of high-resolution modelling is demonstrated by the
small increase in spatial correlation of AQUM-h with the PCM climatology maps
(correlations compared to the PCM of 0.664, 0.743, and 0.761 for RCCM, AQUM,
and AQUM-h). However, the benefit is less pronounced for SO2 than for
NO2. The main reason for this is likely to be that in the UK, SO2
levels have fallen dramatically over the last 25 years and ambient
concentrations are now generally the result of relatively low magnitude
traffic emissions and much stronger emissions from a small number of
industrial point sources. This results in an annually averaged mean
concentration map over the UK which shows relatively little spatial
structure, but with a small number of locations having much higher
concentrations due to strong local emission sources (see the PCM 1 km plot
in Fig. b). This low level background with little overall
spatial structure limits the quantitative increases in spatial correlation
with the PCM climatologies. Another reason may be the impact of the
introduction and removal of strong point emissions sources affecting the
comparison, as noted in Sect. .
Conclusions regarding the benefits of high-resolution modelling for
PM2.5 have been hampered in the present study due to the lack of
observations over the study period. This pollutant has both primary and
secondary contributions and one might expect improvements in the modelling of
the primary component by higher-resolution modelling. However, the magnitude
of the improvement will depend on the relative sizes of primary and secondary
components and it may well be that the contribution of the large secondary
component masks any improvement in the representation of the primary
component. For PM10, model performance remains poor regardless of model
resolution, with all three regional models (RCCM, AQUM, and AQUM-h) failing
to capture the observed frequency distribution and having negative biases in
the range -14.41 to -12.45 µg m-3. Based on the observed
PM values analysed by , it is estimated that the lack of sea
salt lowers the modelled PM10 annual mean values by around 12 %.
Additional important factors in the underprediction of PM10 magnitudes
include the absence of coarse component sodium nitrate aerosol, the poor
representation of other coarse component primary emissions, and poor
modelling of the growth of aerosols to sizes in the coarse range.
For O3, all regional models were able to reproduce the shape of the
observation distribution well, but the offset of the modelled from the
observed central location varied. Tests showed that the differences are
likely to be largely due to differences in the photolysis schemes employed.
However, given the modest benefits of higher-resolution modelling found for
the other secondary pollutants, it seems unlikely that high-resolution
modelling with AQUM would offer significantly improved performance for O3
predictions beyond those demonstrated by the RCCM.
The model simulations described in this paper have been evaluated in their
air quality performance under present-day climate. However, the same
techniques can be applied for projecting future climate and air quality from
the global scale to the UK national scale . The ability to
model air quality at the regional scale will be particularly important for
health impact modelling where high spatial resolution is important to allow
the concentration variations to be matched to population locations. Indeed,
the techniques in this paper have already been applied to 2050s climate and
air quality in for assessing potential changes in UK
hospital admissions.