Sie sind auf Seite 1von 26

Clim Dyn (2012) 39:24972522

DOI 10.1007/s00382-012-1315-2

WRF high resolution dynamical downscaling of ERA-Interim


for Portugal
Pedro M. M. Soares Rita M. Cardoso
Pedro M. A. Miranda Joana de Medeiros
Margarida Belo-Pereira Fatima Espirito-Santo

Received: 23 August 2011 / Accepted: 9 February 2012 / Published online: 1 March 2012
Springer-Verlag 2012

Abstract This study proposes a dynamically downscaled


climatology of Portugal, produced by a high resolution
(9 km) WRF simulation, forced by 20 years of ERAInterim reanalysis (19892008), nested in an intermediate
domain with 27 km of resolution. The Portuguese mainland is characterized by large precipitation gradients, with
observed mean annual precipitation ranging from about
400 to over 2,200 mm, with a very wet northwest and
rather dry southeast, largely explained by orographic processes. Model results are compared with all available stations with continuous records, comprising daily
information in 32 stations for temperature and 308 for
precipitation, through the computation of mean climatologies, standard statistical errors on daily to seasonally
timescales, and distributions of extreme events. Results
show that WRF at 9 km outperforms ERA-Interim in all
analyzed variables, with good results in the representation
of the annual cycles in each region. The biases of minimum
and maximum temperature are reduced, with improvement
of the description of temperature variability at the extreme
range of its distribution. The largest gain of the high resolution simulations is visible in the rainiest regions of
Portugal, where orographic enhancement is crucial. These
improvements are striking in the high ranking percentiles

P. M. M. Soares  R. M. Cardoso  P. M. A. Miranda 


J. de Medeiros
Instituto Dom Luiz, University of Lisbon, Lisbon, Portugal
P. M. M. Soares (&)
Faculdade de Ciencias da Universidade de Lisboa, Campo
Grande, Ed. C8 (3.26), 1749-016 Lisbon, Portugal
e-mail: pmsoares@fc.ul.pt
M. Belo-Pereira  F. Espirito-Santo
Instituto de Meteorologia, Lisbon, Portugal

in all seasons, describing extreme precipitation events.


WRF results at 9 km compare favorably with published
results supporting its use as a high-resolution regional
climate model. This higher resolution allows a better representation of extreme events that are of major importance
to develop mitigation/adaptation strategies by policy
makers and downstream users of regional climate models
in applications such as flash floods or heat waves.
Keywords Regional climate modeling  WRF model 
High resolution  Climatology  ERA-Interim  Portugal

1 Introduction
In recent years, global numerical weather prediction
models led to the development of an increased number of
global climatological datasets like the reanalyzes from the
European Centre for Medium Range Forecasts (ECMWF)
ERA-40 (Uppala et al. 2005) and ERA-Interim (Berrisford
et al. 2009), from the National Centers for Environmental
Prediction (NCEP/NCAR; Kalnay et al. 1996), and the
Twentieth Century Reanalysis Project (Compo et al. 2011),
and others. Simultaneously, a large number of Global
Climate Models (GCMs) have been used to build climate
change scenarios, e.g. to participate in the Fourth Assessment Report of the Intergovernmental Panel on Climate
Change (Solomon et al. 2007). This effort has been successful in generating global climatic information for the
twentieth century in easy to use regular or quasi-regular
grids.
Reanalysis and climate scenario datasets have coarse
horizontal resolutions, typically between 1 and 4 (in both
latitude and longitude), good enough to reproduce many
aspects of large-scale climate (Meehl et al. 2007), but

123

2498

unable to represent many processes and systems that drive


regional and local climate variability, where the consequences of climate change will be mostly felt. These
limitations are greatly amplified in areas of difficult geomorphology, like complex orography, irregular coastlines,
and regions with heterogeneous land cover, where regional
and local thermal and mechanical circulations are forced
by surface heterogeneity. Other, more fundamental, shortcomings of current model technology come from poor
representation of subgrid scale processes (Randall et al.
2007), which may lead to qualitative inaccuracies in some
of the main processes of the climate system, namely in
midlatitude storm track dynamics and in tropical to subtropical convective processes (Teixeira et al. 2011).
To overcome those problems, different downscaling
approaches have been developed. Statistical methodologies
(e.g. Fowler et al. 2007) use observed relationships
between variables at different scales to estimate finer scale
properties, but have a major drawback: the observed relationships may not persist on a changing climate. Regional
climate models (RCMs) constitute an increasingly popular
alternative (Dickinson et al. 1989; Giorgi and Bates 1989;
McGregor 1997; Christensen et al. 2007). RCMs, forced by
GCMs or by reanalysis data, are able to capture physically
consistent regional and local circulations (Giorgi and
Mearns 1991, 1999; Laprise 2008; Leung et al. 2003;
Wang et al. 2004) at the required horizontal and temporal
scales, allowing for the development of high-resolution
climatologies in any terrain conditions. In complex terrains, available point observations must be used to validate
the RCM results, but these are not limited by the availability or quality of those observations.
Results from different RCMs forced by ERA-40
reanalysis, like those of PRUDENCE (Christensen et al.
2002, 2007) and ENSEMBLES (van der Linden and
Mitchell 2009) European projects are a valuable source of
atmospheric data for Europe, and in particular for Portugal.
However, the horizontal resolutions of those RCMs,
between 25 and 50 km, are still insufficient to represent
fundamental and persistent atmospheric processes associated to convective boundary layer, irregular coastline and
orography in regions like Norway (Barstad et al. 2009;
Heikkila et al. 2010), California (Kanamaru and Kanamitsu
2007; Kanamitsu and Kanamaru 2007; Caldwell et al.
2009) and Portugal. Temperature focused analysis are
almost unanimously stating that higher resolution allows to
improve the description of maximum and minimum temperatures by RCMs, when compared to the driving large
scale forcings. However, higher resolution doesnt guarantee improved RCM results (e.g. Rauscher et al. 2010;
Jacob et al. 2007). Specially in the case of precipitation, the
consensus if far from reached: many studies conclude that
higher resolution improves RCMs predictions of

123

P. M. M. Soares et al.: ERA-Interim for Portugal

precipitation (e.g. Brankovic and Gregory 2001; Barstad


et al. 2009); others have found that RCMs present mix
benefits, failing to add value or even degrading GCMs
precipitation fields in different regions of the globe (e.g.
Castro et al. 2005; Sylla et al. 2009; Jacob et al. 2007). The
degradation may be due to problems in the lateral forcing
of the RCMs (Warner et al. 1997), domain size (Leduc and
Laprise 2009; Vannitsem and Chome 2005), parameterization inadequacies (Derbyshire et al. 2004; Jiao and Caya
2006), or obviously a mix of several of those factors
(e.g. Castro et al. 2005).
The evaluation of high resolution RCMs is seriously
constrained by the lack of representative high-resolution
observational datasets (Rauscher et al. 2010). It is also
important to keep in mind that, while horizontal resolution
is readily increased down to about 5 km, real vertical
resolution in the surface layer is more a function of the
boundary layer scheme than of numerics. State of the art
mesoscale models offer the best tool to describe complex
terrain circulations and local to regional scale processes.
The Weather Research and Forecast model (WRF,
Skamarock et al. 2008) has been developed as a research
and operational numerical weather prediction model, but is
increasingly used as an RCM (Bukovsky and Karoly 2009;
Caldwell et al. 2009; Flaounas et al. 2011; Leung and Qian
2009; Liang et al. 2005; Lo et al. 2008; Zhang et al. 2009).
All these studies showed the ability of WRF to correctly
describe climate statistics in different regions, justifying its
use as an RCM. Recently, Caldwell et al. (2009) used WRF
at resolutions of the order of 10 km to dynamically
downscale NCEP-NCAR reanalysis in California, Zhang
et al. (2009) did a similar study for US North Pacific
region, and Heikkila et al. (2010) looked at the Norway
climate. The latter study emphasizes significant improvement from the use of 10 km resolution, even in the representation of mean climate properties, but more importantly
in the case of extremes.
Portugal is a good example of the need of high resolution climatology, motivated both by geomorphological
complexity and large climate gradients. Portugal (mainland) is located in the southwestern limit of Europe and
presents an elongated coast with about 550 km per 200 km
wide (Fig. 1a). Mainland Portugal, between latitudes 37
and 42N (Fig. 1a), is located in the transitional region
between the sub-tropical anticyclone and the subpolar
depression zones. The most conditioning climate factors in
mainland Portugal are, in addition to latitude, its orography
and the effect of the Atlantic Ocean. Regarding altitude,
the highest values are between 1,000 and 1,500 m, with the
exception of the Serra da Estrela range, which peaks just
below 2,000 m. In spite of the fact that the variation in
climate factors is rather small, it is still sufficient to justify
significant variations in air temperature and, most of all, in

P. M. M. Soares et al.: ERA-Interim for Portugal

(a)

2499

(b)

(c)
Sta_Obs Te mp

Sta_Obs Precipit

NW
NE
CW
CE

SW

(d)

HR Orography

WR _9km

SE

WRF_27km

ERA.Interim

m
3000
2500
2000
1500
1000
800
700
600
500
400
300
200
100
50
20
0

Fig. 1 a WRF model outermost (full, dx = 27 km), and inner (solid


black line, dx = 9 km) domains; b the Portuguese Met Office
weather stations available for temperature; c Portuguese National
Water Institute network of rain gauges; d GTOPO-30 and models

topographies. Lines in (b) delimit the 6 regions considered for


analysis. The number of stations (temperature, precipitation) in each
region are: NW (8, 68), NE (3, 38), CW (6, 87), CE (7, 47), SW
(3, 31), SE (5, 34)

precipitation. While the northwest region of Portugal is one


of the wettest spots in Europe, with recorded mean annual
accumulated precipitation in excess of 3,000 mm (Miranda
et al. 2002), average rain amounts in the SE are of the order
of 400 mm. Like in other Mediterranean regions, climate is
characterized by large interannual variability, with recurrence of drought and vulnerability to desertification.
In this study, WRF was used to perform a high resolution simulation of 20 years, at 27 and 9 km of horizontal
resolution, downscaling ERA-Interim reanalysis to the
Portuguese continental territory. The purpose of this work
is: first, to examine the WRF model performance in
reproducing the available observations; second, to characterize the improvement given by WRF in comparison with
ERA-Interim; third, to understand the added value of
expensive 9 km resolution, when compared to the more
usual 27 km; and fourth, to propose a regional climate
dataset for the Portuguese mainland, with a horizontal grid
of 9 km resolution, which can be used for climate variability and climate change assessment studies.

In Sect. 2 of this paper, the details of the methods,


datasets and model setup design are presented. Results are
shown in Sect. 3, where annual, seasonal and daily analyses
are discussed for the near surface maximum and minimum
temperatures and precipitation. The main conclusions are
presented in Sect. 4.

2 Methods and data


2.1 WRF-RCM design
The present study used the WRF model in its version 3.1.1.
The WRF model (Skamarock et al. 2008) is a non-hydrostatic model, suitable for simulating a wide range of scales,
from thousands of kilometers to a few meters, with a large
number of available options in what concerns the model
core and most physical parameterizations, making it
appropriate for numerical prediction, climate and LES
simulations. In this work, WRF was setup with two nested

123

2500

grids, one at 27 km (WRF27km) and a second at 9 km


(WRF9km) horizontal grid spacing, using one-way nesting.
Both grids are centered in the Iberian Peninsula (IP) and
have, respectively, 162 9 35 and 144 9 111 grid points,
covering the regions shown in Fig. 1a. The outermost
domain was designed to cover a relatively large ocean area,
reducing spurious boundary effects in the inner region.
Some authors pointed out the difficulties RCMs may have
in the representation of the large scale features as they are
forced at their boundaries (e.g. Jones et al. 1995). To mitigate this problem, grid nudging (Staufer and Seaman 1990)
was applied to the full (27 km) outermost grid. Nudging is
performed every 6 h at all levels above the planetary
boundary layer, unless the latter is below grid level 10
(*425 m). The nudging technique was firstly introduced by
Waldron et al. (1996) and von Storch et al. (2000), then
developed and employed in many studies, but its use is not
consensual (Alexandru et al. 2008; Miguez-Macho et al.
2004), due to disagreements regarding the advantage of
removing freedom from RCMs large scales and other
potentially negative side effects. Some of the errors found in
the shortest time scales may be attributed to problems in
intermodel boundary coupling. The innermost, 9 km grid,
implying an increase in computational cost by a factor of 30,
represents an attempt to resolve a number of processes that
are known to be crucial in Iberian climate, namely, coastal
and complex topographic circulations. This change in horizontal resolution is expected to have little impact in the
representation of convection and boundary layer processes,
also recognized to be highly relevant in heterogeneous
environments, but which depend, at both resolutions, mostly
on parametrization schemes (Castro et al. 2005). Vertical
resolution in both WRF grids was chosen to position roughly
20 vertical levels in the boundary layer, with the lowest
model sigma level at approximately 10 m of height and
model top at 50 hPa. This setup uses a total of 49 vertical
levels, well above usual values (typically 30).
The physical parameterizations chosen include the
microphysics WSM 6 class single-moment (mp6) scheme
by Hong and Lim (2006), the planetary boundary layer
scheme of Mellor-Yamada-Janjic (Janjic 2001) and the
Betts-Miller-Janjic cumulus scheme (Betts 1986; Betts and
Miller 1986; Janjic 1990, 1994, 2000). These options have
been supported by other studies for climate applications
(Argueso et al. 2011, personal communication Cardoso
et al. 2010), considering long-term model reliability and
computational cost. The simulations use the Noah LSM
4-layer soil temperature, soil and canopy moistures model
(Chen and Dudhia 2001). The radiation scheme used relies
on the shortwave and longwave schemes (Collins et al.
2004) of the NCAR Community Atmospheric model.
The WRF run started at 0000UTC 1 January 1989 and
ended at 0000 UTC 1 March 2009. The first month was

123

P. M. M. Soares et al.: ERA-Interim for Portugal

regarded as model spin-up. Initial and lateral conditions for


the outer domain were derived from the ERA-Interim
pressure-level reanalysis. The lateral boundary conditions
and sea surface temperature were both updated every 6 h,
from ERA-Interim. In both domains 11 grid points are used
as lateral relaxation areas.
2.2 Observational datasets
This study uses surface observations of daily maximum and
minimum temperatures, recorded at 2 m, and daily precipitation. The observations belong to the Portuguese
Institute of Meteorology and to the Portuguese Water
Institute (SNIRH 2010). In the simulation period, the
available observations were checked for consistency and
continuity, retaining 32 temperature and 308 precipitation
stations for this assessment. Figure 1b shows the distribution of temperature weather stations, along with its altitude,
and Fig. 1c presents the distribution of precipitation stations. Wind is not considered since less than 10 weather
stations had enough quality records to qualify for this
study, a number too small to guarantee representativeness.
2.3 Analysis and data handling
In this paper, the two WRF high resolution (9 and 27 km)
simulation results and ERA-Interim reanalysis are compared
with the local surface observations, using the nearest gridpoint of the model to the observations. ERA-Interim was
interpolated to a regular grid with 0.78 spatial resolution. In
the case of WRF, the surface (2-meters) daily maximum
and minimum temperatures are found from the hourly
temperature model output. For ERA-Interim, minimum and
maximum temperatures at 2 meters are given by previous
post-processing, from the forecasts, starting at 00 and 12
UTC. Model daily precipitation is computed from 9 to
9 UTC, following the standard of Portuguese rain gauges.
The majority of RCMs evaluation exercises recur to
gridded data or reanalysis. However, there is extra information in real observed data, which may be overlooked in
gridded datasets, and may be assessed by high resolution
simulations. Indeed, the 9 km resolution used here is finer
than the resolution of all available gridded observational
datasets with daily records. ENSEMBLES daily observations (Haylock et al. 2008; Klein Tank et al. 2002; Klok
and Klein Tank 2009) are gridded to 0.25, the monthly
Climate Research Unit (CRU) v.TS3 is gridded to 0.5
(Mitchell and Jones 2005), and Belo-Pereira et al. (2011)
offers observed Portuguese precipitation in a 0.2 regular
grid. On the other hand, those datasets were constructed in
certain areas with a relatively small number of stations,
implying that their apparent high resolution results largely
from interpolation, leading to oversmoothed precipitation

P. M. M. Soares et al.: ERA-Interim for Portugal

and temperature fields, with the possibility of dramatic


effects in extreme values (Hofstra et al. 2009a, b). In the
ENSEMBLES dataset, for example, the Portuguese area
relies on interpolation from only 26 stations. For those
reasons, the present study focus on model evaluation
against real point observations, avoiding all interpolation
artifacts, namely their impact on extreme statistics.
Co-location of model and observation data requires the
correction of temperature biases associated with altitude
differences, due to different horizontal positions and, more
importantly, to the smooth model topography (Fig. 1d).
Here, correction is made through a constant lapse rate of
6C km-1 applied to both maximum and minimum temperatures. Zhang et al. (2009) verified that using local lapse
rates to correct temperatures had little additional value, but
this issue is strongly dependent of the region in study. No
terrain corrections were applied to precipitation, since it is
intrinsically much more difficult, due to its dependency on
topography complexity, humidity, buoyancy, and other local
variables (Smith and Barstad 2004). Point to point comparisons of maximum temperature, minimum temperature and
precipitation are done through the computation standard
point error statistics for different averaging time intervals,
from daily to yearly. For a regional assessment, Portugal was
divided in 6 regions (Fig. 1b), characterizing the large north
south and westeast climate gradients, used for the computation of spatial mean values. An additional regional analysis
is done on statistics from pools of stations belonging to each
region. The representation of extreme weather events is
discussed through an explicit analysis of histograms of the
three variables, including high ranking percentiles, and by an
assessment of a set of climate indices: hot days, summer
days, frost days and tropical nights.

3 Results
In this section, WRF model simulation results are presented
and compared with weather station observations and with
ERA-Interim forcing data. The comparisons and analysis
are focused on daily precipitation, and maximum and
minimum temperatures (Tmax and Tmin). These three
variables are the most relevant for the definition of climate
indices, and for impact assessment in many sectors, such as
energy and agriculture. In particular, improvements of
Tmin and Tmax description have a direct impact on the
prediction of the duration of the vegetative cycle, and their
extremes may impact on the carbon fixation mechanism.
3.1 General model evaluation
The 20-years mean daily Tmax (Fig. 2a) is characterized
by a significant northsouth gradient, superimposed with

2501

coastal effects and localized orographic features. The


northsouth gradient and some coastal effects are identifiable even in the ERA-Interim data, whereas topography
only marginally modifies the 27 km simulation and gains
prominence at 9 km. At this resolution, the Tmax field is
rather heterogeneous, with clear signature of the larger
mountain chains, in the north and center, but also of warmer regions at low altitudes, namely within river valleys,
from north to south. These features are still visible at
27 km resolution, but almost completely absent from ERAInterim. In the correspondent weather station values, the
aforementioned spatial heterogeneity is detectable, in spite
of the small number of point observations.
The 20-years average Tmin presents a spatial pattern
somewhat less heterogeneous, still with a milder climate
near the coast, but with less clear orographic effects and a
sharp eastwest gradient in the north, due to the continentality of the NE region. The distribution of Tmin is
very similar at WRF9km and WRF27km simulations,
whereas ERA-Interim is comparable in the northern half
of the country, but much warmer in the south. A careful
inspection of the station observational values supports the
quality of the WRF9km simulation. The spatial distributions of Tmax and Tmin bias for the full 20-years simulation period, computed from the daily values between
models and point observations, are illustrated in Fig. 3a.
For Tmax, WRF9km presents rather small biases, positive
and negative, mostly below 1C, with the majority of
these being within a few tenths of a degree. Only 3 stations exceed a 1C warm bias, remaining in the interval
[1, 2]C. Yet, one of these locations refers to a cape (Cape
Carvoeiro) in the west coast. At 27 km resolution WRF
still reveals a good skill to describe Tmax, the majority of
sites having less than 1C bias, but now with 10 stations
with slightly higher cold bias, and still with warm bias at
Cape Carvoeiro. WRF results, at both resolutions, are
much better than ERA-Interims. Indeed, ERA-Interim
shows a widespread cold bias, having many stations with
bias above 2C, with only two capes showing more significant warm biases. The geographical pattern of Tmax
bias reveals ERA-Interim deficiencies in the description of
Tmax in Portugal, and the good performance of WRF,
especially at 9 km resolution, where the driving model
cold bias is greatly mitigated.
Figure 3b shows the distribution of Tmin bias. WRF
results are rather similar in the two resolutions, although
there is an improvement at 9 km. The majority of sites
have bias less than 1C, but results are generally slightly
worse than those of Tmax. Both positive and negative
biases are present above 1C. ERA-Interim presents
stronger biases, mostly characterized by warm bias in the
south and west regions. In WRF9km, Cape Carvoeiro,
which had a warm Tmax bias (?1.2C) shows the largest

123

2502

P. M. M. Soares et al.: ERA-Interim for Portugal

(a)
WRF_9km

WRF_27km

ERA.Interim

Sta_Obs

Sta_Obs

24
23
22
21
20
19
18
17
16
15
14
13

(b)
WRF_9km

WRF_27km

ERA.Interim

C
17
16
15
14
13
12
11
10
9
8
7
6
5

Fig. 2 The 20-years mean daily a maximum and b minimum near surface temperatures, from WRF9km, WRF27km, ERA-Interim and weather
station observations

Tmin cold bias (-2.1C). This station, located right at the


Cape platform is mostly influenced by the milder ocean
boundary layer air, except under unusual conditions of
easterly wind, a condition that is difficult to simulate even
at resolutions higher than 9 km.
As previously mentioned, it should be emphasized that
the altitude lapse rate correction used to interpolate model
data to each station is the simplest possible, and doesnt
take into account seasonal, diurnal and spatial variability of
the lapse rate. These issues may contribute to some of the
differences detected between the models and observations
(Prommel et al. 2010). An assessment of the impact of the
lapse rate correction was done by analyzing the dependency of temperature biases on altitude differences. Only in
the case of WRF9km the differences (not shown) were
found to be statistically significant, with opposite sign
correlations for Tmax and Tmin, but with absolute correlations below 0.5. However, the definition of a robust
correction method working differently for Tmax and Tmin
is beyond the scope of this paper.
Figure 4 shows the geographical distribution of annual
mean precipitation in the WRF simulation, at 9 and 27 km,

123

ERA-Interim and observations. The observed spatial pattern of precipitation is characterized by a strong southnorth gradient, with a clear topographical dependency.
Maximum values of annual precipitation are found in the
NW, related to high frequency of fronts passage with
orographic enhancement. The observed annual pattern is
well captured by the WRF simulation at 9 km. In the north,
the west-east contrast, characterized by precipitation
varying from [2,200 to [500 mm in less than 100 km, is
completely absent from ERA-Interim and well represented
by WRF9km. At 27 km, WRF captures both the NW wet
spot and the NE dry area, but in a smoother way. The ERAInterim coarse resolution also limits the orographic
enhancement due to its very smooth topography, greatly
reducing the spatial variability of the precipitation field.
However, ERA-Interim represents quite well the north
south precipitation gradient, being slightly better than WRF
at 27 km in the southern region. The southern coastal
mountains, which are generally of smaller height than in
the north, have a clear signature in both observed and 9 km
model results, with realistic relative maximums of annual
precipitation between 600 and 800 mm.

P. M. M. Soares et al.: ERA-Interim for Portugal

(a)

(b)

2503

BIAS WRF9km-Obs

BIAS WRF27km-Obs

BIAS ERA.I-Obs

BIAS WRF9km-Obs

BIAS WRF27km-Obs

BIAS ERA.I-Obs

Fig. 3 Spatial distribution of the 20-years biases of daily a maximum and b minimum temperature
WRF_9km

WRF_27km

ERA.Interim

mm

Sta_Obs

mm

Fig. 4 Annual precipitation climatology (19892008) from WRF9km, WRF27km, ERA-Interim and weather station observations

The spatial distribution of normalized precipitation bias,


defined as the bias divided by the observed mean precipitation at the same point, from WRF and ERA-Interim can
be seen in Fig. 5. In spite of its coarser resolution, ERAInterim performs better than WRF27km in the south and
center west. However, WRF9km presents in general a
smaller bias than both WRF27 and ERA-Interim, with a
tendency to underestimate precipitation by 10% to 30% in
the south and center, while in the north some points show

significant overestimation. The NW region is difficult to


simulate and to observe, due to the very localized (in both
time and space) and intense nature of some precipitation
events. In fact, some of the overestimation in the northwest
may be due to the lack of representativity of some
mountain weather stations, not capturing some known
climatological values above 3,000 mm in this region
(Miranda et al. 2002). It is also important to keep in mind
that in the present study there was no attempt to include

123

2504

P. M. M. Soares et al.: ERA-Interim for Portugal

BIAS% WRF9km-Obs

BIAS% WRF27km-Obs

BIAS% ERA.I-Obs

Fig. 5 Spatial distribution of the 20-years daily normalized precipitation bias

orographic correction in the model to station comparison


procedure.
To quantify the WRF model ability to represent Tmax and
Tmin, the following statistical variables were computed:
correlation coefficient, bias, root mean square error (RMSE),
mean absolute error (MAE), and standard deviation. The
error measures were computed pooling together all the
weather stations, and not from the spatially averaged mean
for two reasons: avoiding error cancelation associated with
spatial averaging, and guaranteeing the ability to trace
localized extreme values. Each station observation is directly
compared with the nearest model point. Table 1 summarizes
the errors found for daily values, 5-days and monthly averages, for WRFs two grids and ERA-Interim. Globally, it is
clear that WRF model performs better than ERA-Interim,
with the large majority of the error measures, for both variables, showing improvements from 27 to 9 km resolution
domain (Table 1 shows in bold the best values for each
indicator from the three models, likewise in all Tables).
The correlation coefficients between the observed and
simulated Tmax are 0.93 for WRF9km, 0.92 to WRF27km
and ERA-Interim. The correlations, when averaging for

increasing number of days, steadily rise. The global bias


for Tmax is negligible (0.1C) at 9 km, whereas
WRF27km reveals an overall cold bias (-0.8C) and ERAInterim a larger one (-1.1C). For WRF9km, as expected,
the errors diminish with the increasing time averaging,
for daily Tmax the RMSE and MAE exceed 2C, and for
monthly averaging reduces to around 1C. Roughly, for
ERA-Interim these errors are half a degree higher.
Similarly to Zhang et al. (2009), Tmin correlations are
always smaller than those of Tmax. These correlations in
both WRF simulations compared with ERA-Interim demonstrates an improvement, consequence of a better representation of local processes. WRF model results indicate a
small cold bias, -0.4 and -0.2C, for 9 and 27 km,
respectively, and ERA-Interim a small warm bias of 0.5C.
The other error measures are always smaller in the 9 km
WRF grid. The comparison of WRF27km and ERAInterim shows similar values, although for daily values
ERA-Interim has slightly smaller errors, but WRF27km is
better for 5-days and monthly averaging. The standard
deviation unveils that WRF is better than ERA-Interim in
the representation of Tmin variability, being slightly better

Table 1 Daily to monthly maximum and minimum temperature errors, anomaly correlations in ()
Modelobs

WRF9km

WRF27km

ERA-Interim

123

Time

CORR

BIAS (C)

Tmax

Tmin

Tmax

Tmin

RMSE (C)

MAE (C)

Tmax

Tmax

Tmin

Norm SD
Tmin

Tmax

Tmin

Daily

0.93 (0.75)

0.90 (0.75)

0.1

-0.4

2.8

2.4

2.1

1.9

1.03

0.98

5-days

0.97 (0.88)

0.94 (0.83)

0.1

-0.4

1.7

1.8

1.3

1.4

1.03

1.00

Monthly

0.98 (0.89)

0.95 (0.85)

0.1

-0.4

1.3

1.5

1.0

1.2

1.03

1.00

Daily

0.92 (0.72)

0.87 (0.73)

-0.8

-0.2

3.1

2.8

2.3

2.1

1.01

0.97

5-days

0.96 (0.86)

0.90 (0.82)

-0.8

-0.2

2.1

2.2

1.6

1.7

1.02

0.98

Monthly

0.97 (0.87)

0.92 (0.83)

-0.8

-0.2

1.8

1.9

1.3

1.5

1.03

1.00

Daily
5-days

0.92 (0.76)
0.95 (0.88)

0.89 (0.74)
0.92 (0.82)

-1.1
-1.1

0.5
0.5

3.1
2.4

2.7
2.2

2.3
1.8

2.0
1.6

0.95
0.96

1.07
1.10

Monthly

0.96 (0.88)

0.94 (0.84)

-1.1

0.5

2.1

1.9

1.6

1.5

0.96

1.13

P. M. M. Soares et al.: ERA-Interim for Portugal

2505

in the case of Tmax. In what concerns 5-days variability,


relevant for synoptic scale analysis, WRF9km is on spot for
Tmin, and overestimates Tmax variability by 3%, whereas
ERA-Interim overpredicts Tmin variability by 10% and
underpredicts Tmax variability by 4%. WRF27km is very
close to WRF9km. Table 1 also shows anomaly correlations, computed from the same data after the removal of the
annual cycle, with smaller but still highly significant
values.
For precipitation the same error measures are used,
added by normalized bias, and following the same methodology (Table 2). At the three time scales, WRF9km is
comparable with ERA-Interim, although it shows less than
half of the bias and experiences improvement with
increasing time accumulation period. WRF27km is slightly
worse than ERA-Interim for all averaging periods. WRF
has been developed mostly for higher resolution mesoscale
studies, and its parameterizations package is better suited
for 9 km than for 27 km. The increased dry bias found at
27 km indicates that the precipitation parameterization is
less active at this scale, penalizing the error statistics. In the
WRF simulation, nudging was applied to the 27 km
domain, suggesting that enhanced errors at this scale come
essentially from problems with the parameterizations and
not from phase errors in the boundary conditions. In the
case of precipitation the anomaly correlations are only
slightly smaller, indicating that the mean annual cycle is
not so relevant in the explanation of observed variance.
One should keep in mind that error statistics are not
entirely fair in the evaluation of model qualities, when
comparing very different resolution models. The smooth
patterns produced by ERA-Interim tend to lead to good
correlations, although its fields are qualitatively different
from the very heterogeneous (observed) reality. WRF
results, especially at 9 km, look very much like the
observed field, but small displacements of precipitation
maxima and minima are highly penalized by error statistics.
The fact that, in spite of that, results from WRF9 km are
able to improve on ERA-Interim statistics is encouraging.
Table 2 Daily to monthly
precipitation statistics, anomaly
correlations in ()

In what concerns the mean fields, results from WRF


dynamical downscaling, at both resolutions, are a good
description of the temperature of Portugal. The increase of
the resolution from 27 to 9 km leads to a clear improvement of the spatial and temporal representation of Tmax
and Tmin, reducing the errors found in the driving model.
WRF9km also gives a good description of precipitation in
Portugal, with a slight improvement of point error statistics
but a remarkable reproduction of observed spatial heterogeneity. However a more detailed analysis of Tmax, Tmin
and precipitation fields, which takes into account seasonal
cycles and extreme weather events, is required to show the
usefulness of WRF simulated data as a climate database.
Those results will be quantified at a regional level.
3.2 Maximum temperature
3.2.1 Mean annual cycle of maximum temperature
Figure 6 shows the maps of seasonal mean Tmax. In winter
Tmax varies from 6C, in the higher mountain (Serra da
Estrela), to 16C in the Sado Valley and other areas also in
the south. The strongest gradients are found in summer,
with places with more than 10C differences in less than
50 km. In both summer and spring, when sea surface
temperature is significantly colder than inland, a narrow
coastal strip is also significantly cooler than the interior.
The opposite effect is found in the northern half of the
territory in fall and winter, with a somewhat wider coastal
warm area, resulting from the inversion of sea-land thermal
contrast. In summer, Tmax varies from 21C in the west
coast cooler strip, and not in the higher mountains, to more
than 33C in interior areas of the SE. These detailed features are realistic, but not well represented by the scarce
observational network. WRF27km gives a smoother picture, as expected, as does, even more, ERA-Interim (not
shown).
Figure 7 displays a selection of seasonal errors of daily
Tmax, pooling together (not averaging) all the points.

Modelobs

Time

CORR

WRF9km

Daily

0.72 (0.71)

5.3

-0.2

-9.8

1.7

5-days

0.84 (0.82)

13.6

-1.1

-9.5

6.0

49.9

Monthly

0.89 (0.87)

43.4

-6.5

-8.9

24.4

33.5

Daily

0.67 (0.65)

5.8

-0.6

-25.3

1.9

78.1

5-days

0.77 (0.75)

15.8

-3.0

-25.1

7.1

59.4

Monthly

0.84 (0.83)

53.4

-17.9

-24.6

31.6

43.4

Daily

0.72 (0.71)

5.1

-0.5

-19.6

1.7

70.7

5-days

0.83 (0.81)

14.5

-2.3

-19.3

6.2

51.1

Monthly

0.87 (0.86)

51.3

-13.6

-18.6

27.0

37.1

WRF27km

ERA-interim

RMSE
(mm)

BIAS
(mm)

BIAS (%)

MAE
(mm)

MAPE
(%)
71.8

123

2506

P. M. M. Soares et al.: ERA-Interim for Portugal

(a) DJF

(b) MAM
27

17

26

16

25
15

24

14

23

13

22
21

12
20
11

19

10

18

17
16

15
7

14

13

(d) SON

(c) JJA
34

24

33

23

31
22
30
21

29
28

20

27

19

26

18

25

17

24
16
23
15

22
21

14

20

13

Fig. 6 Spatial distribution of seasonal mean daily-maximum temperature, results from WRF9km resolution, a DJF, b MOM, c JJA
and, d SON

Correlation coefficients for WRF9km are highest in all


seasons, except winter, but models are comparable. Only in
summer WRF at both resolutions strongly outperforms
ERA-Interim. In what concerns the seasonal bias of daily
Tmax, ERA-Interim exhibits the largest values, except in
winter when it is slightly smaller than the one of
WRF27km. The improvement in bias, and in a smaller
measure MAE, by WRF in comparison with ERA-Interim,

0.9

ERA-Interim
WRF27km
WRF9km

0.6
2.7

0.3
0.0

0.84
0.81

-0.3
-0.6
-0.9

0.78
0.75

2.4

BIAS ( C)

0.87

Correlation

3.0

MAE ( C)

0.90

towards a smaller and warm or less cold bias of Tmax may


be due to a better representation of thermal circulations.
Higher resolution leads to higher correlation coefficients
and smaller absolute errors. However, the smallest bias is
found at intermediate resolution in spring and summer and
at higher resolution in winter and spring. Zhang et al.
(2009), simulating the US Pacific Northwest, argued that
RCM Tmax errors were rooted in the driving model. That
may be the case of WRF27km results, which share the
same sign bias at all seasons with the forcing model,
although significantly mitigated in spring and summer.
Local processes at the 9 km scale seem to be strong enough
to invert the signs of biases in the same seasons, while
reducing its size in the remaining seasons. A simple
bootstrapping technique (Wilks 2006, pag 166ff) using
10,000 samples was used to estimate the 95% confidence
interval of the different error statistics, also shown in
Fig. 7. These results indicate that error statistics from the 3
models are statistical significantly different.
The approach followed here is to examine the statistics
of Tmax data pooled together over each of the 6 regions of
Portugal, rather than using averaged data. For that reason,
error statistics are expected to be worse than those reported
for example by Heikkila et al. (2010) based on averaged
data, but they include relevant information of small scale
features of the simulated fields. In line with Caldwell et al.
(2009), Barstad et al. (2009), and others, the motivation of
the aggregation in regions is to do a fairer comparison
against ERA-Interim, which was never intended to simulate local properties.
A summary of regional error analysis is shown in
Table 3. Bias and RMSE values indicate a good representation of Tmax by the WRF9km. In most of the regions
and for most of the seasons, bias is improved by WRF9km
by more than 1C, when compared with ERA-Interim;
WRF27km is also better than ERA-Interim in most of the
regions. All models present a regional cold bias in winter,
that in the case of WRF9km is rather small, in the range of

2.1
1.8

-1.2
DJF

MAM

JJA

SON

-1.5

DJF

MAM

JJA

SON

1.5

DJF

MAM

JJA

SON

Fig. 7 Global Portuguese seasonal errors for daily-maximum temperature. Horizontal lines indicate the limits of the 95% confidence interval of
the corresponding variable, computed by 10,000 bootstrapping samples

123

P. M. M. Soares et al.: ERA-Interim for Portugal

2507

Table 3 Seasonal errors of daily maximum temperatures, for the 6 regions


Tmax

ERA_Interim

WRF_27km

WRF_9km

DJF

MAM

JJA

SON

DJF

MAM

JJA

SON

DJF

MAM

JJA

SON

NW

-1.2

-1.7

-1.9

-1.7

-1.4

-1.0

-0.7

-1.7

-0.6

0.3

0.4

-0.7

NE

-0.4

-2.2

-3.3

-2.3

-1.1

-1.1

-0.9

-2.2

-0.2

0.5

0.1

-1.0

CW

-0.7

-0.9

-0.3

-0.8

-0.7

0.4

1.1

-0.3

-0.2

1.1

1.7

0.3

CE

-0.9

-0.6

-1.1

-1.4

-1.2

0.5

0.4

-1.4

-0.4

1.4

0.7

-0.7

SW
SE

-0.6
-0.6

0.9
-1.4

3.7
-2.3

0.4
-1.8

-1.1
-0.7

-1.6
-0.3

-1.8
-1.0

-1.7
-1.7

-0.2
0.0

0.1
0.6

0.2
-0.4

-0.3
-1.0
2.7

Region
BIAS (C)

RMSE (C)
NW

2.7

3.1

3.6

3.0

3.0

3.3

4.0

3.3

2.6

2.9

3.2

NE

3.3

3.6

4.6

3.9

3.4

3.2

3.4

3.9

3.4

3.0

3.2

3.2

CW

1.9

2.9

4.0

2.7

2.1

2.6

3.3

2.5

2.0

2.7

3.3

2.4

CE

2.4

2.6

2.7

2.8

2.8

2.9

2.9

3.1

2.4

3.0

2.8

2.5

SW

1.8

3.0

5.4

2.9

2.1

2.9

3.5

2.8

1.9

2.5

2.9

2.2

SE

2.0

2.8

3.5

2.9

2.3

3.0

3.3

3.1

2.2

3.0

3.0

2.5

33

36

NW

33

30

Tmax ( C)

27
24

24

21

21

18

18

Mar

Jun

Sep

21
18

15

Sta_Obs
ERA-Interim
WRF27km
WRF9km

12

12

15

Dec

36

33

CE

Mar

Jun

Sep

Dec

12

36

SW

30

33

27

27

30

24

24

30

CW

27

24

15

Tmax ( C)

30

30

27

33

NE

Mar

Jun

Sep

Dec

Jun

Sep

Dec

SE

27
24

21

21

21

18
15

18

18

12

15

15

Mar

Jun

Sep

Dec

12

12

Mar

Jun

Sep

Dec

Mar

Fig. 8 Seasonal cycle of monthly average of Tmax for each region with error bars, results from WRF, ERA-Interim and weather station
observations

-0.6 to 0C, roughly half of the one of ERA-Interim. The


regions presenting the worst bias for WRF9km are CE and
CW, especially in spring and summer, where the 27 km
presents a better match to observations. ERA-Interim presents a systematic cold bias, in the majority of seasons and
regions, with the exception of the SW region, which from

spring to autumn shows a warm bias. WRF27km also has a


cold bias that is largely improved in the 9 km grid, where a
mix of cold and warm bias can be found; roughly, all
regions have a warm bias in spring and summer. Figure 8
depicts the annual cycle of monthly averaged Tmax for
each of the 6 regions. The inspection of Fig. 8 also shows

123

2508

the regional model and observational spread. It is clear that


WRF9km does a good job representing the monthly means
in most of the regions, CW and CE being the worst cases in
agreement with the errors presented in Tables 1 and 3. The
sparse observational grid clearly underestimates the
regional spatial variability.
In Table 1, ERA-Interim presented a very high monthly
correlation coefficient, not too different from WRF9km or
WRF27km. However, Fig. 8 indicates an important difference between observations and ERA-Interim: in NW the
cold bias of ERA-Interim is present every month,
increasing in the summer months. This is also true in NE,
where ERA-Interim also underestimates the annual Tmax
range in a notorious way. WRF27km partially solves this
monthly bias in NW and NE, yet showing a slight cold
bias. WRF9km essentially cancels the cold bias, revealing
a remarkable quality in describing the monthly means in
those regions. In NW a small cold bias in the winter and
autumn months can be detected, and a slight warm bias in
spring and summer months. In NE, observations and
WRF9km are almost on top of each other, with insignificant cold bias in the first half of the year, and warm bias in
the other half. This same pattern may be seen the CE, but
with more important differences in some months. The CW
is by far the region worst described by WRF at both resolutions, especially from Mars to September. The months
from April to September in this region are the only consecutive ones when ERA-Interim better represents monthly
Tmax.
In the SW region, the agreement between WRF9km and
monthly mean observations is notable, particularly considering the poor performance of (ERA-Interim) boundary
conditions. In fact, ERA-Interim reveals a significant warm
bias, reaching approximately 4C in July. This large bias
may be in part also due to the orographic correction that
doesnt reflect the seasonal cycle of temperature lapse
rates. Finally, in the SE models have the same overall
behavior as in the north regions.
3.2.2 Histograms and extremes of maximum temperature
In the previous section it was possible to see that seasonal
and regional Tmax are well represented by WRF. In this
section the WRF ability to simulate de frequency distribution of daily Tmax and their extremes is assessed. In
Fig. 9a daily Tmax frequency distributions of models and
observations are depicted for the four seasons. The added
value of WRF model, for all seasons and from both
WRF9km and WRF27km, can be easily detected on the
histograms, since these are closer to observations. The
persistent cold bias is substantially reduced by WRF9km
for all seasons, especially in the higher range of Tmax. In
winter, the Tmax frequency between 0 and 15C is

123

P. M. M. Soares et al.: ERA-Interim for Portugal

overestimated by ERA-Interim leading to large underestimation between 15 and 25C; this misrepresentation is
reduced by WRF at both resolutions. The spring histogram
reveals an excellent description by WRF9km, and even
WRF27km, when compared to observations, while ERAInterim shows a large shift towards colder values in almost
all the temperature range. In both summer and autumn
WRF is able to improve significantly the match with
observations and solves the two main ERA-Interim deficiencies related to the cold bias in the higher temperature
bins.
In order to add focus and detail to extremes description
by the models, the observed and modeled quantiles can be
found in Fig. 9b, showing quantiles from 0 to 1 in steps of
0.05, first and the last corresponding to the absolute minimum and maximum of Tmax, respectively. Commonly to
histograms the improvement gained with WRF to simulate
the Tmax quantiles is evident. For the four seasons, with
almost no exceptions, the large number of quantiles presented is much better described by both WRF grids, with
most of the WRF9km quantiles being on spot with observations. Remarkable is the ability of WRF to reproduce the
extreme quantiles; moreover, noting that the first and the
last ones correspond to the absolute minimum and maximum of all the simulation period WRF for the four seasons
reproduces this extremes very well.
Both histograms and quantiles emphasize the good
quality of WRF extreme Tmax, in what concerns their
frequency and absolute values, indicating that finescale
processes or parameterizations resolved in the RCM are
crucial, namely, in reducing the large cold bias in the
forcing model. Furthermore, the higher resolution adds
significant value when compared to the WRF coarser
resolution.
The regional histograms and quantiles (not shown)
illustrate in an even stronger way the ERA-Interim deficiencies in predicting the extremes of Tmax. A regional
illustration of climate indices may be seen in Table 4,
where observed summer days (Tmax [ 25C) and hot days
(Tmax [ 35C) are presented, together with relative model
errors, for the 6 regions and for the whole Portugal. The
observed regional climate indices reveal a strong eastwest
gradient in response to coastal effects, with larger frequency of summer and hot days in the eastern sector. On
average for the country there are 88 summer days, ERAInterim underpredicts this number by 9%, similar to
WRF27km, while WRF9km overpredicts by 6%. There is a
lot of spatial structure in those errors, especially in ERAInterim which benefits from large error cancellation. In
ERA-Interim regional estimates, there is underprediction in
5 regions up to -27%, while in the SW ERA-Interim largely overestimates the number of summer days by ?77%.
WRF27km is better, although it also fails in the SW, with

P. M. M. Soares et al.: ERA-Interim for Portugal

(a)
0.15

0.15

DJF

MAM

0.10

Sta_Obs
ERA-Interim
WRF27km
WRF9km

0.05

0.00

Frequency

Frequency

Fig. 9 a Frequency distribution


of Tmax, and, b quantiles of
daily Tmax (from 0. to 1. in
steps of 0.05, 0. corresponds to
the absolute minimum Tmax,
and 1. to the absolute maximum
of Tmax), observed, ERAInterim and WRF9km and
WRF27km

2509

10

20

30

0.10

0.05

0.00

40

10

20

30

Tmax ( C)

Tmax ( C)
0.15

0.15

SON

0.10

Frequency

Frequency

JJA

0.05

0.00

40

10

20

30

0.10

0.05

0.00

40

10

20

40

25

ERAInterim
WRF27km
WRF9km

15

30

Models ( C)

20

Models ( C)

40

Tmax ( C)

(b)

10
5
0
-5
-5

30
o

Tmax ( C)

DJF
0

10

15

20

20
10

MAM

25

10

20

30

40

StaObs ( C)

StaObs ( C)

50
40
30

Models ( C)

Models ( C)

40
30
20
10

JJA
10

20

30

40
o

StaObs ( C)

-38% underprediction. WRF9km is very good in all


regions, with errors in the range [-1, ? 7]%, except in
CW where it overpredicts the frequency of summer days by
27%. Hot days are not very frequent, being often associated
with easterly flow in summer, a process that seems difficult
to capture. ERA-Interim underestimates their frequency by
more than half, WRF27km underestimates it by only 10%,
while WRF9km overestimates by 17%, in the country

20
10

SON

0
50

10

20

30

40

StaObs ( C)

average. Regional relative errors can be high in all models,


especially in CW and SW, but it is important to note that
absolute errors are small in the two regions, only a couple
of days. Due to the poor sampling of weather stations in
what concerns temperature, which also impacts on model
values because they are computed at the same small
number of points, the representation of extremes may be
somewhat mitigated in this analysis.

123

2510

P. M. M. Soares et al.: ERA-Interim for Portugal

Table 4 Number of Summer days (Tmax [ 25C) and Hot days (Tmax [ 35C), and corresponding relative errors of WRF and ERA-Interim
versus observations, for Portugal and the six regions
Region

Obs (days)

Summer days % error


ERA-Interim

WRF27km

Obs (days)
WRF9km

Hot days % error


ERA-Interim

WRF27km

WRF9km

NW

79

-27

-14

-87

27

NE

111

-27

-13

19

-78

-24

CW
CE

57
106

-16
-10

12
0

27
7

5
17

227
-39

28
6

85
24

SW

59

77

-38

74

242

82

SE

117

-18

-12

21

22

-57

-13

21

88

-9

-10

12

-52

210

17

Portugal

3.3 Minimum temperature


3.3.1 Mean annual cycle of minimum temperature
The 20-years mean seasonal Tmin is presented in Fig. 10.
The most visible gradients are the west-east in the north,
related to the continentality of the NE climate, and those
related with coastal effects. In winter the mean Tmin spans
from 9 (SW) to 0C (NE). Spring presents roughly the
same spatial pattern, but with higher mean temperatures.

(a) DJF

(b) MAM
11

14

10

13

12

11

10

(c) JJA

(d) SON
22

14

21

13

20

12

19

11

18

10

17

16

15

14

13

12

11

Fig. 10 As Fig. 6, but for daily-minimum temperature

123

26

The mean Tmin in summer is quite homogeneous


throughout the Portuguese territory, the vast majority is
characterized by mean temperatures above 14C, only
some of the higher mountainous and the parts of NE
experience lower mean Tmins. Autumn presents a Tmin
spatial distribution similar to spring, but with higher
values.
The statistical errors (correlation, bias and mean absolute error) of daily Tmin at the seasonal scale, pooling
together all the point stations and over the 20-years period,
is presented in Fig. 11. At all seasons WRF9km shows the
highest correlation coefficients, followed by ERA-Interim
and WRF27km. The biggest correlation difference between
WRF9km and the driving model is, like for Tmax, in
summer. ERA-Interim shows the major bias, between all
seasons and all models, of around 0.9C, in summer. In the
other seasons models present small bias, ERA-Interim in
the range of -0.2 and -0.5C, while WRF27km shows
even smaller bias. In winter, a warm bias of 0.1C and in
MAM, JJA and SON a cold bias in the range of 0.10.5C.
WRF9km bias is quite independent of the season, showing
an almost constant cold bias around 0.4. MAE points out as
well to an important improvement added by WRF9km, and
some deterioration for the WRF27km. These results are
different from those reported by Zhang et al. (2009).
Although, it is clear that an RCM doesnt necessarily
perform better than the driving model, WRF is able to
modify the sign of the bias, turning a warm bias in
reanalysis into a cold bias. On the other hand, higher resolution does improve significantly the model performance,
in the current case, unlike what was found by Zhang et al.
(2009). That improvement is not clear in bias, which is
better in JJA and SON, but worse in DJF and MAM, but
stands out in the correlation and MAE values in all seasons.
Nudging, present on WRF27km, is more active in DJF and
much less active in JJA, as traveling weather system vary
in frequency, explaining why WRF27km bias is closer to
ERA-Interim in DJF, similar to WRF9km in JJA, and
somewhere in between in the transitional seasons.

P. M. M. Soares et al.: ERA-Interim for Portugal

2511

0.87
0.84

0.78

2.4
2.3

MAE ( C)

BIAS ( C)

0.3
0.0

0.75

2.2

Correlation

0.6
0.81

2.5

ERA-Interim
WRF27km
WRF9km

0.9

2.1
2.0
1.9

-0.3

0.72
0.69

1.8

MAM

DJF

JJA

-0.6

SON

MAM

DJF

JJA

1.7

SON

DJF

MAM

JJA

SON

Fig. 11 As Fig. 7, but for daily-minimum temperature

Figure 12 shows the seasonal cycle of monthly mean


Tmin for the 6 regions. At the monthly scale, the level of
agreement of WRF9km with observations is noteworthy.
The NW region shows a good description of monthly Tmin
by WRF9km, signal of the good ability of the model to
represent the oceanic and orographic local effects on minimum temperature, and on diurnal temperature range, if
analyzed together with Fig. 8. In the NE, the models dont
cool enough in the winter months, which may be related to
persistent clear sky strong inversions in the NE plateau,
difficult to simulate by models (Molders and Kramm 2010).
In the CW ERA-Interim results largely overestimates the
monthly Tmin in the warmer and colder seasons, more
intensely in the summer months. On the contrary, WRF
behaves quite well in summer, but shows a little cold bias, at
both resolutions, in winter. In the CE region WRF presents a
small cold bias, more relevant on winter and summer. The

A similar behavior was already observed by Tmax bias in


Fig. 7b, although the reversion of Tmax bias by WRF9km
only occurred in MAM and JJA.
Concerning regional differences, the bias and RMSE of
daily Tmin at a seasonal scale are shown in Table 5. As
referred, these errors are computed pooling together all the
point stations and on a daily basis. The good reproducing
skill of Tmin by WRF9km is easy to find in all regions and
seasons, with the bias only attaining 1C, but mostly well
below that level. WRF9km presents somewhat higher
biases in winter and in the inland regions.
The only region where ERA-Interim presents a smaller
bias than WRF is the CE. Observing the RMSE to better
judge synchronized model errors it is obvious that WRF9km
slightly outperforms ERA-Interim and WRF27km in a
regional and seasonal consistent way. Moreover, the inland
regions NE and CE present larger RMSE.

Table 5 Seasonal errors of daily minimum temperatures, for the 6 regions


Tmin

ERA_Interim

WRF_27km

DJF

MAM

JJA

SON

-0.8

-0.6

0.3

-0.2

NE

1.1

0.4

0.2

0.9

CW

1.4

1.0

1.6

1.6

CE
SW

-0.8
0.8

-0.2
1.0

0.4
2.3

-0.2
1.4

1.0

0.9

1.4

0.9

DJF

WRF_9km

MAM

JJA

SON

0.0

-0.3

1.6

0.6

-0.4

0.7

-0.8

-0.8

-0.6

-0.5

-1.2
2.4

-0.9
1.8

-1.1
1.2

-1.2
2.0

-0.1

-0.3

-0.6

DJF

MAM

JJA

SON

-0.3

-0.2

-0.1

0.9

0.1

-0.6

0.2

-1.0

-0.7

-0.3

-0.5

-1.1
0.3

-0.7
0.4

-0.7
0.4

-0.9
0.5

-0.5

-1.0

-0.8

-0.7

-0.9

Region
Bias (C)
NW

SE

0.6

0.2

0.0

RMSE (C)
NW

2.9

2.2

2.1

2.3

2.8

2.2

2.1

2.4

2.5

2.2

2.2

2.3

NE

2.9

2.5

2.4

2.6

3.1

2.6

2.5

2.6

2.8

2.6

2.6

2.5

CW

2.9

2.2

2.7

2.7

2.9

2.4

2.5

2.5

2.6

2.1

2.2

2.2

CE

3.4

2.9

3.2

3.1

3.3

2.9

3.0

3.1

2.9

2.5

2.6

2.7

SW

3.0

2.8

3.3

2.9

4.0

3.3

3.0

3.6

2.6

2.3

2.4

2.5

SE

2.9

2.2

2.2

2.3

3.1

2.3

2.0

2.4

2.8

2.2

1.9

2.2

123

2512

P. M. M. Soares et al.: ERA-Interim for Portugal


18

NW

18

Tmin ( C)

15

NE

18

15
15

12

12

12

Sta_Obs
ERA-Interim
WRF27km
WRF9km

3
0
Mar

Jun

Sep

3
6
0
Dec

Mar

Jun

Sep

Dec

SW

CE

Mar

Jun

Sep

Dec

Jun

Sep

Dec

21

21
18

CW

SE
18

18

Tmin ( C)

15
15

15

12
9

12

12
9
6
3

0
Mar

Jun

Sep

Dec

Mar

Jun

Sep

Dec

Mar

Fig. 12 As in Fig. 8, but for minimum temperature

southern regions have in common a warm bias by ERAInterim, quite strong in in the SW summer and decreasing in
the other months. WRF27km shows warm bias in SW, more
intense in winter, and negligible cold bias in SE. WRF9km
presents an almost constant small warm bias in the SW, and
slightly larger cold bias in the SE.
3.3.2 Histograms and extremes of minimum temperature
The frequency distributions of Tmin, showed in Fig. 13a,
illustrates in a different way the seasonal model behavior,
not focused on synchronized events but in the assessment
of daily variability and extremes. The winter histogram
reveals that WRF27km presents the closest agreement with
observations, in line with its smaller bias (Fig. 11). The
most frequent temperatures, near the mode of the observational histogram, is not well represented by any of the
models, WRF9km being the closest. ERA-Interim shows
some overrepresentation of lower temperatures in DJF, that
appear mostly corrected in WRF. WRF9km presents a left
shift in the histogram towards colder temperatures, but with
a good match of the distribution tails, consistent with the
cold bias of Fig. 11. Spring histograms reveal a much
better agreement for all models, although sharing a small
cold shift. Summer histograms present the worse agreement between ERA-Interim and observations, ERAInterim has an important shift versus warmer Tmins,

123

suggesting a sub-representation of the sea-breeze effect;


WRF represents better these circulations, but it shows a
rather small shift towards colder temperatures. In some
way, autumn histogram reveals the same kind of features,
but in a reduced manner.
The observation of the Tmin quantiles of observations
and models in Fig. 13b allows to reinforce some of the
previous conclusions, namely, the summer and autumn
significant improvement of WRF compared to ERAInterim. On the other hand, a closer look shows that almost
all the percentiles are better represented by both WRF
grids, especially by the WRF9km. Moreover, this
improvement is striking on the description of the absolute
extremes.
Table 6 presents the relative error of frost days and
tropical nights between models and observations. Overall it
can be seen that models have difficulties to skillfully
describe these climate indices, since on average they occur
rather infrequently. Nevertheless, WRF9km is better in
both indices for the whole country, whereas WRF27km
outperforms ERA-Interim in the prediction of tropical
nights and is comparable to it in the case of frost days.
WRF9km frost days relative error is half of ERA-Interims,
both with an overestimation. For tropical nights WRF9km
reduces the relative error to less than one-third of ERAInterims but with opposite sign. Common to all models is
the underrepresentation of frost days and tropical nights in

P. M. M. Soares et al.: ERA-Interim for Portugal


Fig. 13 As Fig. 9, but for
daily-minimum temperature

2513

(a)
0.15

0.15

MAM

Sta_Obs
ERA-Interim
WRF27km
WRF9km

0.10

Frequency

Frequency

DJF

0.05

0.10

0.05

0.00

0.00

-10

10

20

-10

30

10

20

Tmin ( C)
0.15

0.15

JJA

SON

0.10

Frequency

Frequency

30

Tmin ( C)

0.05

0.10

0.05

0.00

0.00

-10

10

20

30

-10

10

(b)

20
o

Models ( C)

Models ( C)

25

ERAInterim
WRF27km
WRF9km

10
5
0
-5
-10

15
10
5
0
-5

DJF

-10 -5

-10
-10 -5

10 15 20
o

5 10 15 20 25

StaObs ( C)
30

30

25

25

20

Models ( C)

20

15

MAM

StaObs ( C)

Models ( C)

30

Tmin ( C)

20
15

20
o

Tmin ( C)

15
10

10
5
0

-5

JJA
0

10 15

20 25 30
o

StaObs ( C)

the NE, although with small absolute errors, and the


overrepresentation of tropical nights in SW, but with much
higher errors in ERA-Interim. WRF9km overrepresents
frost days everywhere but the NE, move severely in the SE.
In the eastern sector, WRF tends to underestimate the
frequency of tropical nights. Considering the rareness of
tropical nights, it is wise to look more to the (small) values
of absolute errors than to the (very large) values of relative
errors.

SON

-10
-10 -5 0

5 10 15 20 25 30
o

StaObs ( C)

3.4 Precipitation
3.4.1 Mean annual cycle of precipitation
The seasonal precipitation, for winter and summer,
(Fig. 14) shows the extremely heterogeneous spatial pattern of precipitation in Portugal, strongly associated with
orography. A main south-north gradient and a secondary
eastwest gradient in the north are present in all seasons.

123

2514

P. M. M. Soares et al.: ERA-Interim for Portugal

Table 6 Number of frost days (Tmin \ 0C) and Tropical nights (Tmin [ 20C), and corresponding relative errors of WRF and ERA-Interim
versus observations, for Portugal and the six regions
Region

Frost days % error


Obs (days)

Tropical nights % error

ERA-Interim

WRF27km

WRF9km

Obs (days)

ERA-Interim

WRF27km

WRF9km

NW

22

116

15

18

14

213

60

NE

23

231

-66

-40

227

-53

-29

CW
CE

1
23

-45
65

43
78

32
44

4
7

146
21

-45
-66

25
-38

SW

-94

-49

35

664

142

73

216

224

244

98

221

-41

12

-9

65

-35

219

SE
Portugal

(a) DJF
WRF_9km

WRF_27km

ERA.Interim

mm

Sta_Obs

mm

WRF_27km

ERA.Interim

mm

StaObs

mm

(b) JJA
WRF_9km

Fig. 14 Spatial distribution of seasonal accumulated precipitation, results from WRF9km, WRF27km, ERA-Interim and point observations,
a DJF, and b JJA

Nevertheless, the accumulated precipitation has a strong


inter-seasonal variability, from around 1,000 mm in winter
to 200 mm in summer in the NW, and above 200 mm in
DJF and less than 10 mm in JJA in the SE. On average, it

123

rains more in SON than in MAM, by up to a factor of two


in some regions.
Figure 15a shows the seasonal correlation coefficients,
MAPE and RMSE for seasonal accumulated precipitation

P. M. M. Soares et al.: ERA-Interim for Portugal

(a)

2515

0.90

60

0.85

50

160

0.80

0.75

RMSE (mm)

MAPE (%)

Correlation

140

40

120
100
80
60

30

40
DJF

MAM

JJA

(b)

DJF

MAM

JJA

20

SON

110

0.75

0.65

DJF

MAM

JJA

SON

0
-5

100

0.70

MAPE (%)

Correlation

20

SON

ERA-Interim
WRF27km
WRF9km

90

BIAS (%)

0.70

80

0.60

-10
-15
-20
-25

70

-30

0.55
DJF

MAM

JJA

SON

DJF

MAM

JJA

SON

DJF

MAM

JJA

SON

Fig. 15 Global Portuguese seasonal errors for a seasonal accumulated precipitation, and b daily precipitation. Horizontal lines indicate the limits
of the 95% confidence interval of the corresponding variable, computed by 10,000 bootstrapping samples

of ERA-Interim, WRF27km and WRF9km. All correlations are quite high, comparable to similar studies (Zhang
et al. 2009). WRF9km (27 km) correlation are in the range
of 0.89 (0.85) to 0.78 (0.71), respectively, in winter and
summer, and ERA-Interim somewhere in between.
WRF27km is marginally worse than ERA-Interim, whereas
WRF9km is consistently and significantly better. WRF9km
seasonal biases (which when normalized are equal to the
daily ones, Fig. 15b) are in the range of (-12,-1%), significantly smaller than the 50% reported in the Third
Assessment Report of the Intergovernmental Panel on
Climate Change by Giorgi et al. (2001) for RCMs, and
similar to the best performing RCMs shown in recent
studies (Heikkila et al. 2010; Rauscher et al. 2010; Zhang
et al. 2009). Even WRF27km presents biases in the range
of (-32, -19%), not small but within the range of referred
studies. Both WRF domains show a consistent dry bias.
MAPE and RMSE point out to a significant improvement
by WRF9km on the seasonal precipitation for all seasons,
while WRF27km shows a small deterioration from ERAInterim, more relevant in SON. WRF9km shows a significant improvement of the representation of the seasonal
variability, but not of the daily one. The interpretation of
the (squared) correlation coefficient as a measure of
explained variance is only applicable for normal distributions, and so it is not relevant for daily precipitation.
However the correlation coefficient is still a measure of the
quality of the forecasts, since it is highly sensitive to phase
errors. At the daily scale (Fig. 15b), the bias obviously

continues to be smaller, but MAPE is comparable to ERAInterims. Also, WRF9km and ERA-Interim have the same
seasonal correlation of daily precipitation (from 0.76 in
winter to 0.69 in summer), larger than the mean correlation
values, in Norway, found by Heikkila et al. (2010) for
WRF at 10 km resolution, 0.63, and for the ENSEMBLES
mean 0.44. The interpretation of the (squared) correlation
coefficient as a measure of explained variance is only
applicable to normal distributions, and so it is not relevant
for daily precipitation. However the correlation coefficient
is still a measure of the quality of the forecasts, since it is
highly sensitive to phase errors. In general, WRF27km
shows the worst performance, in what concerns phase and
errors. It should be noticed that ERA-Interim results are
good in this global seasonal analysis. These features seem
to indicate some loss of performance in the way WRF27km
receives the ERA-Interim forcing, that is compensated, at
least, locally by the higher resolution of the inner grid. This
improvement is more noticeable in the seasons with less
rain, and where the mesoscale and convective processes are
more suitable to contribute to the precipitation regime;
although it seems also that WRF9km is not able to tackle
very well the infrequent summer precipitation. As previously mentioned, Portugal has a large spatial precipitation
heterogeneity, justifying the need of a regional analysis of
the models results. Table 7 shows, at the regional and
seasonal scales, the normalized bias and MAPE of the daily
precipitation. Comparisons are still made station point by
station point, but with a much better spatial sampling (see

123

2516

P. M. M. Soares et al.: ERA-Interim for Portugal

Table 7 Seasonal errors of daily precipitation, for the 6 regions


Precip.

ERA-Interim

WRF-27km

WRF-9km

DJF

MAM

JJA

SON

DJF

MAM

JJA

SON

DJF

MAM

JJA

SON

NW

-33.2

-28.9

-17.1

-29.1

-12.6

-20.3

-43.6

-29.7

3.0

-2.7

-9.6

-5.3

NE

-0.8

15.2

40.0

7.0

14.1

10.9

-10.1

-2.3

-7.2

7.9

16.5

1.5

BIAS (%)

CW

-21.2

-11.3

-5.9

-18.6

-30.3

-30.0

-32.8

-41.7

-20.1

-17.5

-13.6

-23.8

CE

-21.6

-7.1

18.1

-20.1

-21.8

-21.4

-13.0

-37.0

-11.5

-1.6

26.9

-15.1

SW

-27.4

-21.8

2.7

-27.8

-36.6

-38.8

-11.8

-45.2

-21.8

-19.9

4.4

-23.0

SE
-25.6
MAPE (%)

-16.2

5.3

-17.9

-34.3

-33.0

-29.1

-41.9

-20.0

-13.3

-3.1

-17.9

NW

60.8

68.1

94.6

66.7

67.7

74.8

92.0

72.6

61.8

69.0

93.4

67.7

NE

68.4

83.8

133.7

76.2

91.1

93.0

115.6

86.5

65.3

82.0

126.3

77.9

CW

61.8

73.8

110.5

68.4

73.8

79.9

107.9

77.8

64.5

74.1

114.4

71.6

CE

64.2

77.9

126.3

69.2

77.1

82.7

121.6

76.9

62.6

78.1

137.7

70.4

SW

66.1

75.9

119.7

72.2

74.3

83.0

128.8

75.4

67.7

76.6

136.3

75.7

SE

70.9

79.8

123.0

79.3

76.9

84.5

114.3

77.8

69.8

82.8

128.5

80.0

Fig. 1). WRF9km has the smallest bias the majority of the
regions and seasons. For WRF9km, the regional normalized bias is in the range of [-23.8,26.9]%, the extreme
error values appearing during summer and autumn in the
center of the territory, whereas ERA-Interim fails by
[-33.2, ?40]%, indicating an improvement in regional
precipitation, moreover keeping in mind that those errors
refer to pools, not averages, of stations.
With no exception the rainiest region, NW, presents the
smaller MAPE at all the seasons and models, with values
just over 90% in the dry summer, and around 65% in the
other seasons. As expected summer show for all models the
highest MAPE, in agreement with the majority of studies
(e.g. Caldwell et al. 2010; Rauscher et al. 2010), emphasizing the common problem to RCMs in the simulation of
summer precipitation, where convective processes become
relevant. In fact, WRF27km shows slightly smaller summer
MAPEs, in 5 out of the 6 regions, contradicting the common sense of higher resolution leading to better summer
precipitation. However, the suitability of the parameterization used may be a key issue, but one outside the
objectives of this study. In winter, in spite of the similarity
of MAPEs, it is interesting to see that all models perform in
the western sector, ERA-Interim performs slightly better
than the other models in the west regions and WRF9km in
the eastern ones, at this daily scale.
If the regional daily data is filtered to a regional monthly
data the performance of WRF is more impressive. In Fig. 16
the monthly mean daily precipitation for the 6 regions can
be evaluated. In the rainy NW, the agreement of the
observed and WRF9km is almost perfect throughout the
year, even in summer, where in spite of the smaller daily

123

MAPEs WRF27km performs worst. This indicates that


WRF9km has some problems in the timing of rainfall, but
recovers most of the observed precipitation at the monthly
scale. ERA-Interim largely underestimates winter and
spring precipitation in the NW. From November to April,
WRf27km is better than ERA-Interim but not as good as
WRF9km. The secondary peak in April, an important feature in recent Iberian climatology (Paredes et al. 2006), is
well reproduced by WRF9km. In the inland regions, NE and
CE, WRF9km gives again the best fit, better than ERAInterim and much better than WRF27km, although with
some underestimation of October to December precipitation
in all models, less so in WRF9km. In the other regions all
models underestimate the monthly precipitation from
October to April, the best being WRF9km in the vast
majority of the months, and the worst WRF27km.
Wet days are defined differently in some studies,
examples being those when daily precipitation is above
0.1 mm (e.g. Barstad et al. 2009) and those where it is
above 1 mm (e.g. Herrera et al. 2010). The ways models
reproduce wet days contribute to understanding the origin
and consistency of precipitation climate prediction by
models. As in other studies with different models (Barstad
et al. 2009; Heikilla et al. 2010; Zhang et al. 2009), WRF
generates too many days with very light precipitation. This
is clear in Table 8, comparing the models to observations
ratio of the number of days with precipitation above
0.1 mm and 1 mm. ERA-Interim largely overestimates the
number of days with precipitation, both above 0.1 and
1 mm (172, 132%). This deficiency appears diminished in
WRF, the 9 and 27 km domains overestimate the above
0.1 mm days by roughly 37 and 20%, respectively, without

P. M. M. Soares et al.: ERA-Interim for Portugal


5

7
NW
Precipitation (mm/day)

Fig. 16 Seasonal cycle of


monthly mean daily
precipitation for each region,
results from WRF9km,
WRF27km, ERA-Interim and
weather station observations

2517

Sta_Obs
ERA Interim
WRF27km
WRF9km

6
5

CW

2
2

Mar

Jun

Sep

Dec

Mar

Jun

Sep

Dec

Mar

Jun

Sep

Dec

Jun

Sep

Dec

4
SW

CE
Precipitation (mm/day)

1
0

SE

3
2
1
0

Mar

Jun

Sep

Region

Dec

a significant regional spread. This decrease is still more


important in the number of days with precipitation above
1 mm, in which case WRF9km results are good, with
relative errors in the range [-3, 20]% among the different
analysis regions, and 8% for Portugal. WRF27km underestimates these values by 15%, on average, and ERAInterim overestimates it by more than 30%. Table 8
confirms the bigger difficulties models have to represent
the precipitation regime in inland regions (E) compared
with the littoral ones (W), the relative errors being in the
former case consistently higher. Additionally, the number
of wet days allows to state that, in spite of ERA-Interim
having shown good skill for precipitation, its at the cost of
precipitating too often.
This analysis can be extended and deepen looking to the
precipitation quantiles for each region in a logarithm scale
(not shown). In general, in the regions where WRF9km
reproduces better the daily precipitation all the quantiles
are closer to observations, in the other the medium
Table 8 Ratio of the number of
wet days in WRF and ERAInterim to observations, for
precipitation [0.1 mm and
precipitation [1 mm, for the six
regions and Portugal

4
NE

Mar

Jun

Sep

Dec

Mar

quantiles (10 and 50%) of ERA-Interim have a better


agreement, although in the extreme quantiles WRF always
outperforms ERA-Interim (see below).
3.4.2 Histograms and extremes of precipitation
A crucial aspect of climate change assessment is the
understanding of the trend of frequency and intensity of
extreme precipitation events, namely those leading to
droughts and floods. There are many important studies,
either observational (e.g. Giorgi 2002) or model based (e.g.
Zwiers and Kharin 1998; Solomon et al. 2007) indicating
an increased frequency of precipitation extremes, globally
and regionally, in the recent past and in global warming
scenarios. Additionally, models still show difficulties in
representing daily precipitation, and even more precipitation extremes, due to both lack of resolution or/and
parameterization deficiencies. These issues emphasize the
importance of evaluating how RCMs represent extreme

Wet days

Model/Sta_Obs (%)

[0.1 mm

[1 mm

ERA-Interim

WRF27km

WRF9km

ERA-Interim

WRF27km

WRF9km

NW

168.1

109.7

127.7

128.7

87.0

108.5

NE

195.8

127.1

136.6

153.1

106.0

121.9

CW

166.9

127.2

138.7

123.9

74.1

96.9
118.4

CE

172.3

128.3

143.9

135.4

93.7

SW

160.7

125.1

139.7

125.1

73.7

99.3

SE

177.1

123.7

149.8

140.1

84.9

112.6

Portugal

171.8

122.2

137.3

131.9

85.3

107.8

123

2518

P. M. M. Soares et al.: ERA-Interim for Portugal

shows a distribution which is largely left shifted, with


absence of the higher precipitation bins, compensated by
exaggerated frequency of the lower ones. It is clear that this
model rains too often, with very small quantities of rainfall.
This is true for all seasons. WRF27km is much better than
ERA-Interim, but the best results are clearly given by
WRF9km. Remarkably, WRF9km is able to capture the
occurrence of extreme, but rare, precipitation events with
daily values above 200 mm.

precipitation events, namely, how higher resolution


improves or not their description, and if it adds value to the
forcing model results.
The seasonal distributions of daily precipitation,
observed and simulated, can be observed in Fig. 17a, with
bins every 1 mm. In general, WRF, both at 9 and 27 km
resolutions, follows rather well the observed frequency
distribution of precipitation, greatly improving its
description when compared to ERA-Interim. The latter
1

1
DJF

MAM

0.1

0.1

StatObs
ERAInterim
WRF27km
WRF9km

0.01
1E-3

frequency

(a)

frequency

Fig. 17 a Frequency
distribution of daily
precipitation, b quantiles of
daily precipitation
(2.5,10,20,25,30,
40,50,60,70,75,80,90,95,97.5,99
and 99.9), observed, ERAInterim and WRF9 and
WRF27km. Histogram scales
are different to add legibility to
the plots

1E-4
1E-5
1E-6

0.01
1E-3
1E-4
1E-5

50

100

150

200

250

1E-6

300

50

100

250

SON

0.01

frequency

frequency

JJA

1E-4

50

100

150

0.01

1E-4

1E-6

200

50

(b)

ERAInterim
WRF27km
WRF9km

80

150

100

Models (mm)

Models (mm)

100

100

60
40
20

80
60
40
20

DJF
20

40

60

MAM

80 100

20

Sta_Obs (mm)

60

80

100

100

Models (mm)

Models (mm)

40

Sta_Obs (mm)

100
80
60
40

80
60
40
20

20
JJA
20

40

60

80 100

Sta_Obs (mm)

200

250

precipitation (mm/day)

precipitation (mm/day)

123

200

1E-6
0

150

precipitation (mm/day)

precipitation (mm/day)

SON
20

40

60

80 100

Sta_Obs (mm)

300

P. M. M. Soares et al.: ERA-Interim for Portugal

To further investigate the WRF skill to describe the


precipitation extremes, Fig. 17b shows the quantiles distribution, where it is immediately possible to confirm the
outstanding ability of WRF, especially at 9 km resolution,
but even at 27 km, to describe the higher ranking quantiles.
Consistently with Fig. 17a, ERA-Interim underrepresents
the higher quantiles. In winter, ERA-Interim underpredicts
most quantiles, except the lower ranking, while WRF9km
and WRF27km do an excellent fit to observations. Quantile
90 and above are underpredicted by ERA-Interim by
almost 50%, while WRF only shows a slight overprediction
of the strongest events (quantile 99.9). WRF also behaves
well in spring and autumn, although with a slight underprediction of most high ranking quantiles, except 99.9. In
Summer, all models underpredict the higher ranking
quantiles, and the improvement from ERA-Interim to WRf
is minimal, except in the 99 and 99.9 quantiles. The representation of summer precipitation is a clearly difficult
task for any model.

4 Discussion and conclusions


The present study contributes to the understanding of
regional climate features produced by a state of the art
numerical weather prediction model (WRF) forced by the
most recent reanalysis (ERA-Interim), a pre-requisite to the
preparation of high-resolution climate scenarios. The use of
a high horizontal and vertical resolution non-hydrostatic
model, such as WRF, as a downscaling regional climate
model has already been tried in recent studies (e.g. Caldwell et al. 2009; Heikkila et al. 2010; Zhang et al. 2009)
with mixed but overall positive impact in the quality of the
results. Considering their parameterization packages, WRF
and similar models become competitive at resolutions of
10 km or higher, where non-hydrostatic effects gain relevance, and convection is partially resolved, which is in the
limit of what is currently feasible for multidecadal simulations, a good reason to look into these results, as they will
be much easier to produce in the near future, even at higher
resolutions.
Portugal is a relatively small country, characterized by
large spatial heterogeneity of the precipitation fields, due to
rugged topography and coastal effects, and by strong
interannual variability, typical of Mediterranean climate.
Temperature fields are less heterogeneous, but with significant topographic signatures and the impact of strong
thermal circulations, especially in summer. WRF is for the
first time used over Portugal at resolution up to 9 km to
downscale ERA-Interim, in the period 19892008, and
using 49 vertical levels with increased resolution in the
boundary layer. Results concerning maximum and minimum temperature were evaluated against a network of 32

2519

stations with continuous observation records, while the


precipitation fields were compared with observations in a
network of 308 rain gauges.
In what concerns Tmax, WRF9km reduces bias from
-1.1C (ERA-Interim) to 0.1C. Variability of Tmax is
well represented by WRF9km at all timescales from daily
to monthly. WRF9km maximum temperature bias is higher
is spring and summer, where it changes sign from ERAInterim. Correlations with observations are smaller in
winter and higher in autumn, but generally higher than
those of ERA-Interim. Annual cycles of Tmax are better
represented by WRF9km in all regions, with a good fit
to observed cycles, including most regions where ERAInterim largely underpredicts the annual range and the only
region (SW) where ERA-Interim strongly overpredicts that
range. Finally, there is a consistent improvement of the
already good distribution of Tmax extremes of ERAInterim. The good performance of WRF9km is also visible
in the frequency of Tmax related climate indices.
WRF27km is generally better than ERA-Interim but not as
good as WRF9km.
In the case of Tmin, ERA-Interim presents a warm bias
of 0.5C, while WRF9km and WRF27km show cold biases
of -0.4 and -0.2C, respectively. Daily correlations are
only slightly lower than for Tmax, but WRF9km is always
the best performing model. Variability of Tmin is overestimated by ERA-Interim by 10% (on the 5-days scale), and
well captured by WRF at all timescales (1-day, 5-days, and
monthly). Daily seasonal correlations of Tmin are always
larger than ERA-Interim, unlike those of WRF27km. The
seasonal bias of WRF9km is small (-0.45C) and almost
constant, of opposite sign from ERA-Interim, but slightly
better results are found for WRF27km. In the CW, SW and
SE regions ERA-Interim shows significant misrepresentation of the annual cycle, with persistent warm biases, a
feature which is largely diminished by WRF9km. Frequency distributions and quantiles of Tmin also reveal
some improvement by WRF9km, especially concerning the
reduction of warm bias. Considering its coarse resolution,
mean annual precipitation simulated by ERA-Interim is
quite reasonable, with a good but rather smooth representation of the main spatial gradients. However, ERA-Interim
has a dry bias around -20%, a value that is reduced by a
factor of 2 in WRF9km. Correlations are almost equal in
both models, with only a marginal improvement by WRF
at the 5-days and monthly scale. However, WRF9km
significantly outperforms ERA-Interim in all seasonal
correlations, with gains up to 7% in spring and summer,
WRF9km precipitation bias is also reduced in all seasons. At the regional level, WRF9km is generally better.
Annual cycles of precipitation are better described by
WRF9km in all regions, with excellent match in the wettest
NW. WRF27km is mostly worse than ERA-Interim.

123

2520

ERA-Interim reveals a major overestimation of wet days, a


drawback that is partially mitigated by WRF9km and by
WRF27km. In the other end of the spectrum, one finds
what seems to be the highest added-value of WRF at both
resolutions: a major improvement in the representation of
the frequency and intensity of extreme events. ERAInterim completely lacks events with strong precipitation
(above 70 mm/day), and largely underpredicts high precipitation quantiles. These are improved in both WRF
simulations, in the rainier seasons (winter, autumn and
spring), by about a factor of 2, leading to a good description of precipitation extremes. Even in summer WRF
improves the distribution of extremes. Both the match of
observed frequency of wet days and of the higher ranking
quantiles is crucial for impact studies, as they affect biological systems and the hydrological sector.
The present results can be qualified in the context of
other regional climate studies, although a direct comparison is not possible, considering the different regional
contexts and differences in the driving forcings. Jacob et al.
(2007) used the set of 50 km resolution models from
PRUDENCE against Climate Research Unit CRU monthly
mean fields, with mean temperature biases for Iberia in the
range [-1.68, 0.82]C in winter and [-1.28, 1.97]C in
summer; Heikkila et al. (2010), using WRF at 10 km for
Norway reported a mean bias of -0.8C and a mean correlation of 0.95, with ERA40 forcing having -0.7C of
bias and mean correlation of 0.97. Jacob et al. (2007)
reported an Iberian precipitation bias in [-95, 0.9] mm in
winter, and [-45, 50] mm in summer. Rauscher et al.
(2010), using the ENSEMBLES 25 km resolution RCMs
and the same monthly gridded observations, presented an
Iberian precipitation percentage bias in the range [-7.2,
30.2]% in winter and [-17.3, 147.4]% for summer, with
correlations within [0.63, 0.87] in winter and [0.73, 0.93] in
summer. Heikkila et al. (2010) reports a mean percentage
bias of WRF10km precipitation in Norway of 33.4% and a
mean correlation of 0.63, forced by ERA40 with a 42.9%
bias and 0.44 mean correlation.
Present results are better than those of Heikilla et al.
(2010), but this is not a fair comparison, considering the
different location. Indeed, they also show important relative
improvements from the driving reanalysis. Considering the
PRUDENCE and ENSEMBLES model evaluations,
WRF9km appears well placed in their error ranges, suggesting that WRF is a good candidate for regional climate
modeling. Improved description of temperature spatial and
temporal variability by WRF is due to a better representation of regional processes, related with orographic and
coastal forcing. Some problems seem to remain in processes
which are generally misrepresented by numerical models,
such as those occurring in stable boundary layers (Molders
and Kramm 2010) and explain some of the minimum

123

P. M. M. Soares et al.: ERA-Interim for Portugal

temperature errors. In what concerns the precipitation


fields, improvements gained by WRF9km are also as good
as or better than ENSEMBLES models, suggesting a benefit
of the higher resolution in the representation of the orographic enhancement of precipitation, a process that is
crucial in the rainier regions of Portugal. More importantly,
WRF is able to describe well the precipitation extremes, a
feature which has been singled out as a fragility of
ENSEMBLES RCMs in Spain (Herrera et al. 2010).
The overall comparison of ERA-Interim, WRF27km and
WRF9km results support the suggestion by Castro et al.
(2005) pointing to surface boundary forcing as a key factor
in producing atmospheric variability for small-scale, which
has a growing importance as the influence of lateral
boundary forcing decreases. Also, Fischer et al. (2011)
showed that uncertainties in land surface parameters may
lead to large changes in the tails of the distributions in a
climate model. Both the representation of surface features
and of the processes that drive low level climate variability,
namely thermal and orographic circulations, may benefit
from high resolution. Numerical weather prediction models
at resolutions of 10 km or higher show encouraging results,
suggesting the way forward in regional climate modeling.
This attempt to use WRF in Iberia as an RCM, forced by
ERA-Interim, strongly supports its further use. The results
obtained may be used as reference dataset for climate
change studies.
Acknowledgments The authors thank the two anonymous reviewers of this manuscript for their comments and suggestions. This work
was funded by the Portuguese Science Foundation (FCT) under
project REWRITE- PTDC/CLI/73814/2006, and PEST-OE/CTE/
LA0019/2011.

References
Alexandru A, de Elia R, Laprise R, Separovic L, Biner S (2008)
Sensitivity study of regional climate model simulations to large
scale nudging parameters. Mon Weather Rev 137:16661686.
doi:10.1175/2008MWR2620.1
Argueso D, Hidalgo-Munoz JM, Gamiz-Fortis SR, Esteban-Parra MJ,
Dudhia J, Castro-Diez Y (2011) Evaluation of WRF parameterizations for climate studies over Southern Spain using a
multistep regionalization. J Clim 24:56335651. doi:10.1175/
JCLI-D-11-00073.1
Barstad I, Sorteberg A, Flaty F, Deque M (2009) Precipitation,
temperature and wind in Norway: dynamical downscaling of
ERA40. Clim Dyn 33:769776. doi:10.1007/s00382-008-0476-5
Belo-Pereira M, Dutra E, Viterbo P (2011) Evaluation of global
precipitation datasets over the Iberian Peninsula. J Geophys Res.
doi:10.1029/2010JD015481 (in press)
Berrisford P, Dee D, Fielding K, Fuentes M, Kallberg P, Kobayashi S,
Uppala S (2009) The ERA-Interim Archive. ERA report series.
1. Technical report. European Centre for Medium-Range
Weather Forecasts, Shinfield Park, Reading
Betts AK (1986) A new convective adjustment scheme. Part I:
Observational and theoretical basis. Quart J Roy Meteor Soc
112:677691

P. M. M. Soares et al.: ERA-Interim for Portugal


Betts AK, Miller MJ (1986) A new convective adjustment scheme.
Part II: Single column tests using GATE wave, BOMEX, and
arctic air-mass data sets. Quart J Roy Meteor Soc 112:693709
Brankovic C, Gregory D (2001) Impact of horizontal resolution on
seasonal integrations. Clim Dyn 18:123143
Bukovsky MS, Karoly DJ (2009) Precipitation simulations using
WRF as a nested regional climate model. J Appl Meteor
Climatol 48:21522159. doi:10.1175/2009JAMC2186.1
Caldwell PM, Chin H-NS, Bader DC, Bala G (2009) Evaluation of a
WRF based dynamical downscaling simulation over California.
Climatic Change 95:499521
Castro CL, Pielke RA Sr, Leoncini G (2005) Dynamical downscaling:
an assessment of value added using a regional climate model.
J Geophys Res (Atmospheres) 110:D05108. doi:10.1029/2004
JD004721
Chen F, Dudhia J (2001) Coupling an advanced land surface
hydrology model with the Penn StateNCAR MM5 modeling
system. Part I: model implementation and sensitivity. Mon
Weather Rev 129:569585
Christensen JH, Carter TR, Giorgi F (2002) PRUDENCE employs
new methods to assess European climate change. EOS 83:147
Christensen JH, Hewitson B, Busuioc A, Chen A, Gao X et al (2007)
Regional climate projections. In: Solomon S, Qin D, Manning
M, Chen Z, Marquis M, et al. (eds) Climate change 2007: the
physical science basis. Contribution of working group I to the
fourth assessment report of the intergovernmental panel on
climate change. Cambridge and New York: Cambridge University Press
Collins WD et al (2004) Description of the NCAR community
atmospheric model (CAM 3.0). NCAR tech. note, NCAR/TN4641STR, 226 pp
Compo GP, Whitaker JS, Sardeshmukh PD, Matsui N, Allan BJ, Yin
X, Gleason BE, Vose RS, Rutledge G et al (2011) The twentieth
century reanalysis project. Quarterly J Roy Meteorol Soc
137:128. doi:10.1002/qj.776
Derbyshire SH, Beau I, Bechtold P, Grandpeix J-Y, Piriou J-M,
Redelsperger J-L, Soares PMM (2004) Sensitivity of moist
convection to environmental humidity. Quart J Roy Meteor Soc
130:30553080
Dickinson RE, Errico RM, Giorgi F, Bates GT (1989) A regional
climate model for western United States. Clim Change
15:383422
Fischer E, Lawrence D, Sanderson B (2011) Quantifying uncertainties
in projections of extremesa perturbed land surface parameter
experiment. Clim Dyn 37:13811398. doi:10.1007/s00382-0100915-y
Flaounas E, Bastin S, Janicot S (2011) Regional climate modelling of
the 2006 West African monsoon: sensitivity to convection and
planetary boundary layer parameterisation using WRF. Clim
Dyn 36:10831105. doi:10.1007/s00382-010-0785-3
Fowler HJ, Blenkinshop S, Tebaldi C (2007) Linking climate change
modelling to impacts studies: recent advances in downscaling
techniques for hydrological modeling. Int J Climatol 27:1547
1578
Giorgi F (2002) Variability and trends of sub-continental scale surface
climate in the twentieth century. Obs Clim Dyn, Part I.
doi:10.1007/s00382-001-0204-x
Giorgi F, Bates GT (1989) The climatological skill of a regional
model over complex terrain. Mon Weather Rev 117:23252347
Giorgi F, Mearns LO (1991) Approaches to regional climate change
simulation: a review. Rev of Geophys 29:191216
Giorgi F, Mearns L (1999) Introduction to special section: regional
climate modeling revisited. J Geophys Res 104(D6):63356352
Giorgi F et al (2001) Regional climate information: Evaluation and
projections. In climate change 2001: the scientific basis:
contribution of working group I to the third assessment report

2521
of the intergovernmental panel on climate change. In: Houghton
JT et al. (ed).Cambridge and New York: Cambridge University
Press, pp 583638
Haylock MR, Hofstra N, Klein Tank AMG, Klok EJ, Jones PD, New
M (2008) A European daily high-resolution gridded data set of
surface temperature and precipitation for 19502006. J Geophys
Res 113:D20119. doi:10.1029/2008JD010201
Heikkila U, Sandvik A, Sorterberg A (2010) Dynamical downscaling
or ERA-40 in complex terrain using WRF regional Climate
model. Clim Dyn. doi:10.1007/s00382-010-0928-6
Herrera S, Fita L, Fernandez J, Gutierrez JM (2010) Evaluation of the
mean and extreme precipitation regimes from the ENSEMBLES
regional climate multimodel simulations over Spain. J Geophys
Res 115:D21117. doi:10.1029/2010JD013936
Hofstra N, HaylockM NewM, Jones PD (2009a) Testing EOBS
European high-resolution gridded dataset of daily precipitation
and surface temperature. J Geophys Res. doi:10.1029/2009
JD011799
Hofstra N, New M, McSweeney C (2009b) The influence of
interpolation and station network density on the distribution
and extreme trends of climate variables in gridded data. Clim
Dyn (in press)
Hong S-Y, Lim J-OJ (2006) The WRF single-moment 6-class microphysics scheme (WSM6). J Korean Meteorol Soc 42:129151
Jacob D, Barring L, Christensen OB, Christensen JH et al (2007) An
inter-comparison of regional climate models for Europe: model
performance in present-day climate. Clim Change 81:3152
Janjic ZI (1990) The step-mountain coordinate: physical package.
Mon Weather Rev 118:14291443
Janjic ZI (1994) The step-mountain eta coordinate model: further
developments of the convection, viscous sublayer and turbulence
closure schemes. Mon Weather Rev 122:927945
Janjic ZI (2000) Comments on development and evaluation of a
convection scheme for use in climate models. J Atmos Sci
57:3686
Janjic ZI (2001) Nonsingular implementation of the MellorYamada
level 2.5 scheme in the NCEP meso model. NCEP office note
437, 61 pp
Jiao Y, Caya D (2006) An investigation of summer precipitation
simulated by the Canadian regional climate model. Mon
Weather Rev 134:919932. doi:10.1175/MWR3103.1
Jones RG, Murphy JM, Noguer M (1995) Simulation of climate
change over Europe using a nested regional-climate model. I:
Assessment of control climate, including sensitivity to location
of lateral boundaries. Quart J Roy Meteor Soc 121:14131449
Kalnay E et al (1996) The NCEP/NCAR 40-year reanalysis project.
Bull Amer Meteor Soc 77:437471
Kanamaru H, Kanamitsu M (2007) Fifty-seven-year California
reanalysis downscaling at 10 km (CaRD10). part II: comparison
with North American regional reanalysis. J Clim 20:55725592.
doi:10.1175/2007JCLI1482.1
Kanamitsu M, Kanamaru H (2007) 57-year California reanalysis
downscaling at 10 km (CaRD10) Part I. System detail and
validation with observations. J Clim 20:55275552. doi:10.1175/
2007JCLI1482.1
Klein Tank AMG et al (2002) Daily dataset of 20th-century surface
air temperature and precipitation series for the European Climate
Assessment. Int J of Climatol 22:14411453
Klok EJ, Klein Tank AMG (2009) Updated and extended European
dataset of daily climate observations. Int J Climatol 29:1182
1191. doi:10.1002/joc.1779
Laprise R (2008) Regional climate modeling. J Comput Phys
227:36413666. doi:10.1016/j.jcp.2006.10.024
Leduc M, Laprise R (2009) Regional climate model sensitivity
to domain size. Clim Dyn 32:833854. doi:10.1007/s00382-0080400-z

123

2522
Leung LR, Qian Y (2009) Atmospheric rivers induced heavy
precipitation and flooding in the western U.S. simulated by the
WRF regional climate model. Geophys Res Lett 36:L03820. doi:
10.1029/2008GL036445
Leung LR, Mearns LO, Giorgi F, Wilby RL (2003) Regional climate
research: needs and opportunities. Bull Am Meteorol Soc
84:8995. doi:10.1175/BAMS-84-1-89
Liang XZ, Choi HI, Kunkel KE, Dai Y, Joseph E, Wang JXL (2005)
Surface boundary conditions for mesoscale regional climate
models. Earth Interactions 9
Lo JCF, Yang ZL, Pielke RA Sr (2008) Assessment of three dynamical
climate downscaling methods using the weather research and
forecasting (WRF) model. J Geophys Res 113:D09112. doi:
10.1029/2007JD009216
McGregor JL (1997) Regional climate modelling. Meteorol Atmos
Phys 63:105117
Meehl GA, Stocker TF, Collins WD, Friedlingstein P et al (2007)
Global climate projections. In: Solomon S, Qin D, Manning M,
Chen Z, Marquis M, Averyt KB, Tignor M, Miller HL (eds)
Climate change 2007: the physical science basis. Contribution of
working group I to the 4th assessment report of the IPCC.
Cambridge University Press, Cambridge, pp 747846
Miguez-Macho G, Stenchikov GL, Robock A (2004) Spectral
nudging to eliminate the effects of domain position and
geometry in regional climate model simulations. J Geophys
Res 109:D13104. doi:10.1029/2003JD004495
Miranda PMA, Coelho F, Tome AR, Valente MA, Carvalho A, Pires
C, Pires HO, Cabrinha VC, Ramalho C (2002) 20th century
Portuguese climate and climate scenarios. In: Santos FD, Forbes
K, Moita R (eds) Climate Change in Portugal: Scenarios,
Impacts and Adptation Measures 283 Gradiva
Mitchell TD, Jones PD (2005) An improved method of constructing a
database of monthly climate observations and associated highresolution grids. Int J Climatol 25:693712
Molders N, Kramm G (2010) A case study on wintertime inversions
in Interior Alaska with WRF. Atmos Res 95(23):314332
Paredes D, Trigo RM, Garca-Herrera R, Trigo IF (2006) Understanding precipitation changes in Iberia in early spring: Weather typing
and storm-tracking approaches. J Hydrometeorol 7:101113
Prommel K, Geyer B, Jones JM, Widmann M (2010) Evaluation of
the skill and added value of a reanalysis-driven regional
simulation for Alpine temperature. Int J Climatol 30:760773.
doi:10.1002/joc.1916
Randall DA, Wood RA et al (2007) Climate models and their
evaluation. In: climate change 2007: The physical science basis.
Contribution of working group I to the fourth assessment report
of the intergovernmental panel on climate change. In: Solomon
S, Qin D, Manning M, Chen Z, Marquis M, Averyt KB, Tignor
M, Miller HL (eds.) Cambridge University Press, Cambridge and
New York, NY
Rauscher SA, Coppola E, Piani C, Giorgi F (2010) Resolution effects
on regional climate model simulations of seasonal precipitation
over Europe. Part I: Seasonal. Clim Dynamics 35:685711

123

P. M. M. Soares et al.: ERA-Interim for Portugal


Sistema Nacional de Informacao de Recursos Hdricos (2010).
Available at http://snirh.pt/
Skamarock WC et al (2008) A description of the advanced research
WRF version 3. NCAR tech. note TN-475_STR, 113 pp
Smith RB, Barstad I (2004) A linear theory of orographic precipitation. J Atmos Sci 61:13771391
Solomon S, Qin D, Manning M, Chen Z, Marquis M, Averyt K,
Tignor M, Miller H (eds) (2007) Climate change 2007: the
physical science basis. Contribution of working group I to the
fourth assessment report of the intergovernmental panel on
climate change. Cambridge University Press, Cambridge
Stauffer DR, Seaman NL (1990) Use of four-dimensional data
assimilation in a limited area mesoscale model. Part I: experiments with synoptic-scale data. Mon Weather Rev 118:1250
1277
Sylla MB, Coppola E, Mariotti L, Giorgi F, Ruti PM, DellAquila A,
Bi X (2009) Multiyear simulation of the African climate using a
regional climate model (RegCM3) with the high resolution
ERA-interim reanalysis. Clim Dyn. doi:10.1007/s00382-0090613-9
Teixeira J et al (2011) Tropical and sub-tropical cloud transitions in
weather and climate prediction models: the GCSS/WGNE
Pacific crosssection intercomparison (GPCI). J Clim (in press)
Uppala SM et al (2005) The ERA-40 re-analysis. Q J R Meteorol Soc
131:29613012
van der Linden P, Mitchell JFB (eds) (2009) E ENSEMBLES: climate
change and its impacts: summary of research and results from
the ENSEMBLES project. Met Office Hadley Centre, FitzRoy
Road, Exeter EX1 3 PB, UK, 160 pp
Vannitsem SF, Chome F (2005) One-way nested regional climate
simulations and domain size. J Clim 18:229233
von Storch H, Langenberg H, Feser F (2000) A spectral nudging
technique for dynamical downscaling purposes. Mon Weather
Rev 128:36643673
Waldron KM, Peagle J, Horel JD (1996) Sensitivity of a spectrally
filtered and nudged limited area model to outer model options.
Mon Weather Rev 124:529547
Wang Y, Leung LR, McGregor JL, Lee D-K, Wang W-C et al (2004)
Regional climate modelling: progress, challenges, and prospects.
J Meteorol Soc Japan 82:15991628
Warner T, Peterson RA, Treadon RE (1997) A tutorial on lateral
boundary conditions as a basic and potentially serious limitation
to regional numerical weather prediction. Bull Am Meteor Soc
78:25992617
Wilks DS (2006) Statistical methods in the atmospheric sciences.
Elsevier, Amsterdam
Zhang Y, Dulie`re V, Mote P, Salathe EP Jr (2009) Evaluation of WRF
and HadRM mesoscale climate simulations over the United
States Pacific Northwest. J Clim 22:55115526
Zwiers FW, Kharin VV (1998) Changes in the extremes of the climate
simulated by CCC GCM2 under CO2 doubling. J Clim
11:22002222

Das könnte Ihnen auch gefallen