WRF high resolution simulation of Iberian mean and extreme precipitation climate



In this study precipitation from a high resolution WRF climate simulation is presented and evaluated against daily gridded observations in the Iberian Peninsula. The simulation corresponds to a dynamical downscaling of ERA-Interim, in the period 1989–2009, performed with two nested grids, at 27 and 9 km horizontal resolution. The higher resolution simulation indicates a significantly improved representation of Iberian precipitation fields, at all timescales, with emphasis on the representation of variability and of extreme weather statistics. Results compare well with recent studies with other models and/or for other regions, further supporting the use of WRF as a regional climate model.

1. Introduction

The precipitation regime in the Iberian Peninsula is characterized by large interannual and spatial variability (Esteban-Parra et al., 1998; Muñoz-Díaz and Rodrigo, 2004). This variability intrinsically related to the Mediterranean climate is substantially enhanced by complex topography and coastal processes (Serrano et al., 1999; Miranda et al., 2002).

The highlands of the northwest are among the wettest regions in Europe, with the highest mean annual precipitation above 2800 mm, whereas near the eastern coast values below 200 mm are observed (Couto et al. 2011). Furthermore, especially in the latter region, Iberia is occasionally affected by flash floods. More than 300 mm have been recorded in 24 h (Couto et al., 2011), resulting from a combination of ingredients like high Mediterranean Sea surface temperature and topographic enhancement of convective systems (Romero et al., 1998; Martín et al., 2007). In some places, annual precipitation is the result of less than 60 d of rainfall, and periods of more than 120 d without precipitation have been registered (Martin-Vide and Gomez, 1999).

The Mediterranean region has been identified as one of the most vulnerable to climate change, due to consistent losses of precipitation found in CMIP3 ensemble of climate simulations (Giorgi, 2002) and to an increase of interannual variability, leading to higher frequency of precipitation extremes (Solomon et al., 2007). Iberia, in the western border of the European Mediterranean sector, is expected to experience large changes in its precipitation climate, and, being characterized by large gradients of rainfall, constitutes a challenge for regional climate modelling.

The Iberian Peninsula precipitation regimes and climate change impact has been studied in the framework of European projects PRUDENCE (Christensen and Christensen, 2007; Christensen et al., 2007) and ENSEMBLES (van der Linden and Mitchell, 2009), both based on the analysis of results from ensembles of Regional Climate Models (RCM). In PRUDENCE, RCMs were run at a horizontal resolution of 50 km, while ENSEMBLES compared results at both 50 and 25 km. Most studies with those data focused on European scale features, assessed results through comparisons with datasets of gridded observations at comparable resolutions, namely monthly data from the Climate Research Unit (CRU) (Mitchell and Jones, 2005), at 0.5 × 0.5°, and the ENSEMBLES daily observational gridded dataset for Europe (E-OBS) (Haylock et al., 2008; Klok and Klein Tank, 2009), at 0.25 × 0.25°. Different analyses have been done for different European regions, including Iberia (Rauscher et al., 2010).

While the CRU dataset, with only monthly precipitation, is unusable for high frequency statistics, the E-OBS dataset, in Iberia, was until very recently the only available choice for daily precipitation assessment. However, it relies only on 26 + 52 stations (Portugal + Spain) in this region, implying that its real resolution is much coarser than the nominal value of 0.25°. Herrera et al. (2010a, 2010b) used an extensive compilation of about 2000 daily observations in Spain to build a new daily gridded dataset, at 0.2° × 0.2° of horizontal resolution (latitude–longitude), for the period 1950 to March 2008. Belo-Pereira et al. (2011) compiled around 400 stations and used the similar technical specifications to extend the dataset into Portugal. The merged dataset contains daily precipitation for the whole Iberian Peninsula, from 1950 to 2003, based on a dense network of rain gauges, of about 2400. This grid, based on 30 times more data than E-OBS, has enough information to look into finer details of precipitation features down to the daily scale, even with some limitations which are unavoidable in any gridded dataset, in particular problems coming from very localized rainfall events in complex mountainous regions impossible to represent by a feasible regular grid (Caldwell et al., 2009).

Herrera et al. (2010a) used the new Spanish gridded dataset to evaluate the results from ENSEMBLES RCMs. In general, they found that results were good, capturing the main spatial patterns, but with large intermodel spread, and with most models overpredicting mean annual Spanish precipitation. ENSEMBLES RCMs, mostly developed for IPCC AR4, used a common European domain and similar resolutions, down to 25 km, with up to 40 vertical levels. Some of those models have been developed as climate models, whereas others are adaptations of numerical weather prediction (NWP) models.

The increased focus on temporal and spatial variability of climate fields, and the need to assess the impact of global warming in the extremes of the statistical distributions of climate variables, are leading to a change in the climate modeling paradigm towards a seamless prediction approach (Palmer et al., 2008; Hazeleger et al., 2010). A good representation of synoptic weather systems, as done by state of the art NWPs and by the new reanalysis products, such as ERA-Interim (Berrisford et al., 2009; Dee et al., 2011), is crucial to drive RCMs into producing realistic regional climates. For the same reasons, state of the art RCMs, at higher horizontal and vertical resolutions, are needed to represent local circulations and low level flow.

The Weather Research and Forecast model (WRF, Skamarock et al., 2008) model is being developed as a global community NWP and atmospheric research model, but is increasingly being used also as an RCM. While many studies have looked into the performance of WRF in different weather conditions, only a few assess it as an RCM, namely Leung and Qian (2009), Qian et al. (2009), Caldwell et al. (2009), Zhang et al. (2009), Salathé et al. (2008), mostly within the USA. Heikkilä et al. (2011) discusses results from WRF, at 10 and 30 km horizontal resolution, in Norway, concluding that it is an improvement from the 25 km ENSEMBLES results and that the refinement of the resolution represents added value and should be considered in when addressing extreme precipitation.

In spite of the crucial role of resolution in regional climate simulations, one must keep in mind that there is always a complex interplay between resolution and parametrizations. Fernández et al. (2007), using MM5 at 45 km resolution in Iberia, found sensitivity of the results to parametrization options, with different climate variables responding in their own way to those changes. As a consequence, a choice of model options is not straightforward even for a confined region as Iberia.

In this study, a high resolution WRF simulation, at 9 km in the horizontal and 49 vertical levels, forced by the full ERA-Interim reanalysis (1989–2009), is compared with the new Iberian daily gridded precipitation dataset. Analysis will focus on detailed error statistics, down to the daily scale, emphasizing the ability of WRF to represent individual transient weather systems and to reproduce observed spatial variability and the statistics of weather extremes.

In the following section, the observational dataset and the WRF model setup are summarized. A detailed analysis of the model performance against observations is presented in the third section. And, finally, the main conclusions are in the fourth section of this study.

2. Data and methods

2.1. WRF simulation

The WRF model, version 3.1.1, was used for a climate simulation of precipitation over Iberia. A high regional resolution is achieved by using two nested grids (Figure 1), with 27 and 9 km resolution, respectively, and one-way nesting. The outer integration domain covers 4320 km × 3240 km encompassing a region of the North Atlantic up to the Azores Islands, in an attempt to capture the large scale systems as they cross the Atlantic, and the western half of the Mediterranean Sea, to guarantee that mesoscale systems that affect the Mediterranean coast of Spain are resolved. The inner grid, with 1467 km × 1278 km, covers the entire Iberian Peninsula as well as the Balearic Sea, the Alboran Sea, the Gulf of Cadiz, the Bay of Biscay and a significant portion of the nearby Atlantic Ocean, enough to capture sea-breeze circulations and other coastal processes that control the Iberian climate, especially in summer.

Figure 1.

(a) WRF model domains, at dx = 27 km (full map), and dx = 9 km (black rectangle), and (b) the main Iberian basins.

Iberian topography, a major forcing of its regional climate, is represented in Figure 2, for three different resolutions, from about 1 km to the ERA-Interim grid (about 80 km). Unlike ERA-Interim, the WRF grid, at 9 km, captures the main mountain chains and river basins, although with some smoothing.

Figure 2.

Iberian topography (m) according to (a) Gtopo 30 dataset (30″ resolution, approximately 1 km), (b) WRF model grid at 9 km resolution, and (c) ERA-Interim reanalysis (0.7° resolution).

In this simulation, 49 vertical levels are used, the model top is fixed as 50 hPa, the first level is set at approximately 10 m from the ground, the second level at 30 m, and the following distances, between levels, increase by 10% in the first 1000 m where the grid merges with a standard 41 level grid. The planetary boundary layer encompasses roughly 20 vertical levels, in an attempt of better capture the convective boundary layer. This results in a rather high resolution for climate purposes. The physical parameterizations used include the microphysics Double-Moment six class (mp6) scheme of Hong and Lim (2006), the boundary layer scheme of Mellor–Yamada–Janjic (Janjic, 2001) and the Betts–Miller–Janjic (BMJ) cumulus scheme (Betts, 1986; Betts and Miller 1986; Janjic, 1990, 1994, 2000). The simulation was performed with varying sea surface temperature, the NCAR Community Atmospheric Model (CAM) shortwave and longwave radiation scheme (Collins et al., 2004) and the Noah LSM four-layer soil temperature, moisture model (Chen and Dudhia, 2001).

WRF was integrated continuously for 21 years, 1989–2009, using initial and boundary conditions from ERA-Interim reanalysis. To keep the interanual memory of the state of the soil, the model is only initialized on the 1 January 1989, and afterwards ERA-Interim data is only given at the spatial boundaries and as SST fields. To reduce phase propagation errors in WRF, which may lead to desynchronization and increase error statistics, nudging is performed on the outer (27 km) grid, every 6 h, at all levels above the planetary boundary layer, unless the latter is bellow grid level 10 (∼425 m). In this case, nudging is only performed above that level. The lateral boundary conditions for the outer grid are provided at the 6 hourly intervals of ERA-Interim. SST are supplied every 6 h from ERA-Interim for both grids. In both domains, 11 grid points are used as lateral relaxation areas. Output from WRF is archived on an hourly basis, for all 3D fields, allowing for direct computation of high frequency variability, and for integration of daily precipitation in the different daily periods of the Spanish (6-6 UTC) and Portuguese (9-9 UTC) climatological networks.

2.2. Observations

The observations in Iberia rely on two gridded precipitation datasets recently proposed. Herrera et al. (2010b) presented a new precipitation dataset for Spain, based on approximately 2000 good quality rain gauges. Belo-Pereira et al. (2011) complemented that work, gridding more than 400 rain gauges in Portugal. Both datasets include daily precipitation on a regular grid at resolution of 0.2°, about 20 km. The two datasets have been merged for the common period 1950–2003, although the Spanish grid is available until March 2008. Therefore, it should be kept in mind that the WRF evaluation performed in this study for the Spanish territory has a wider temporal range (1989–2008) than the one used for mainland Portugal (1989–2003).

2.3. Comparison methods

Due to the different resolutions of observations and model data, there is a need to perform spatial averaging prior to the computation of error statistics. To simplify the process, WRF output is averaged to the nearest observational grid box, always using 4 × 4 WRF grid points for each 0.2° of the merged Iberian grid. For each 0.2° grid box, observed and simulated precipitation are accumulated over different time intervals: daily, 3, 5, 8, 10, 15, 20, 30, 40, 50, 60 and 90 d, monthly, seasonally and yearly. The different standards of daily accumulation used in Portugal (9-9) and Spain (6-6) are taken into account in the comparisons. For each grid box and accumulation period (merging time and space), the following standard error statistics are computed: bias (1), normalized bias (2), mean absolute error (3), mean absolute percentage error (4), root mean square error (5), correlation coefficient (6) and standard deviation (7), defined as:

display math(1)
display math(2)
display math(3)
display math(4)
display math(5)
display math(6)
display math(7)

where N is the number of observed/predicted days and ō and math formula stand for the mean of observed and simulated values. A simple bootstrapping technique (Wilks, 2006, pg. 166ff) using 10 000 samples was used to estimate the 95% confidence interval of the different error statistics.

The use of different time aggregation intervals is a way to look at different time scales in the Iberian climatology. At the 1-d time scale results are still penalized by phase errors in the propagation of individual storms, a problem that tends to be enhanced in high resolution simulations (Mass et al., 2002). At the longer aggregation times, results are relevant for the assessment of intra and interannual variability and such phase errors are reduced.

For comparison, ERA-Interim forecasted precipitation is also used to compute error statistics. However, in this case, the comparison is made against the nearest grid point, e.g. without averaging of the 0.2° observational grid into the 0.7° ERA-Interim grid, as this would impede the discussion of the high-resolution climate features, which are our main goal. ERA-Interim precipitation fields cannot be taken with the same level of confidence as other (analysed) fields since they constitute a global model forecast. However, they are an indication of what a very good global model, at a good horizontal resolution, could produce and they have some degree of synchronization with the real world through the pressure field. For those reasons, a comparison with ERA-Interim is a relevant indication on the added value of the downscaling by high resolution simulations.

Spatial aggregation is also applied to the data, using a simplified river basin system (Figure 1(b)), which includes the main Iberian rivers (Tejo, Douro, Guadiana, Ebro, Segura and Guadalquivir) and combinations of smaller river basins. The assessment of model performance at this scale is relevant for hydrological modelling and water resources management, and offers a relevant, but smoother, view of the spatial patterns of precipitation. Basins are used to compute basin-averaged precipitation fields, which can be used to assess the ability of the model to represent observed temporal variability. Furthermore, grid boxes belonging to each basin are pooled together to investigate climate extremes, e.g. high-frequency/small scale processes, through the analysis of high-rank percentiles.

Finally, eight standard precipitation climate indices from ECA&D (European Climate Assessment and Dataset) are computed from gridded observations, WRF and ERA-Interim forecast. WRF results are evaluated by spatial correlation against the observations.

3. Results

3.1. General evaluation

The annual mean precipitation from the observational grid and models (WRF 9 km and ERA-Interim) is illustrated in Figure 3. A significant southeast-northwest gradient with high precipitation (more than 2000 mm year–1) in the north-northwest Atlantic coast associated with the path of the winter baroclinic synoptic-scale systems (Zorita et al., 1992) and an extremely dry (less than 200 mm year–1) southeastern coast, is observed (Figure 3(a)). The incursion of the winter frontal systems in the Tejo river basin transports moist air into the center of the peninsula, leading to increased precipitation due to enhancement by the underlying topography (Figure 4(a)). The frontal systems do not reach the southwest of Portugal as often as the northwest and combined with lower topography conduces to lower precipitation in this area (less than 700 mm year–1). As in the Tejo river basin, the Guadiana and the Guadalquivir river basins also convey moisture inland. Apart from these three areas, the central part of continental Spain displays low precipitation (less than 500 mm year–1). In the north, the Cantabrian Mountains prevent the northern Atlantic circulations from reaching far inland (Herrera et al., 2010a, 2010b), in the west the Estrela, Montemuro, Marão and Morena Mountains perform the same role. The Sierra Nevada Mountains, in the southeast, also prevent moist Mediterranean air from travelling inland and also show higher precipitation due to local orographic enhancement. The eastern coast presents an annual maximum of less than 700 mm year–1 of rainfall (Herrera et al., 2010a, 2010b) and is characterized by large temporal variability, which is often linked to severe events connected to high sea surface temperatures in the Mediterranean and orographic enhancement (Romero et al., 1998; Martín et al., 2007).

Figure 3.

Annual mean precipitation from (a) observational grid dataset with 0.2° resolution, (b) WRF 9 km and (c) ERA-Interim. Relative differences computed as (d) (WRF-Obs)/Obs, and (e) (ERA-Interim-Obs)/Obs.

Figure 4.

Seasonal mean precipitation from observational grid dataset with 0.2° resolution, WRF 9 km and ERA-Interim (correlation between 0.2° maps and WRF or ERA-Interim in brackets) (a) winter (0.8, 0.8), (b) spring (0.69, 0.69), (c) summer (0.68, 0.8) and (d) autumn (0.77, 0.75). Note the different scale in spring.

Figure 3(b) shows the annual mean precipitation from the WRF climate simulation. The southeast-northwest precipitation gradient is immediately recognized, as well as the Tejo, Guadiana and Guadalquivir river basins and Sierra Nevada local precipitation maximums. Overall, WRF overestimates precipitation in the central part of Iberia (Figure 3(d)) where the overestimation pattern depicts quite remarkably the highest orographic features (compare with Figure 2(b)). In some measure, these topography related overestimation by WRF may be due to the smoothness of the observation grid, since, for example, the Portuguese NW mountainous area has recorded, in some weather stations, annual mean precipitations above 2500 mm, which are absent in the grid dataset. Similar reasoning could be applied to the Pyrenees. Caldwell et al. (2009) points out that there is substantial uncertainty inherent to interpolating station data to a grid and that station measurements are not unbiased, particularly at high altitudes. Thus, the bias in high orography should be carefully valorized. In general, WRF reveals an underestimation of annual precipitation of the coastal areas, and an overestimation in the interior.

The ERA-Interim annual precipitation (Figure 3(c)) represents the southeast-northwest gradient, but the major influences of the topography are absent, which is not surprising due to its smooth orography. Accordingly, the complex spatial pattern of precipitation is oversmoothed in ERA-Interim. The relative difference of ERA-Interim to the observed grid (Figure 3(e)) shows a wet bias associated to the model higher topography, and its mismatch with the real topography. Strong dry bias can be observed in the Tejo and Guadalquivir basins as well as on the upwind side of the mountain ranges.

The seasonal geographical distribution of precipitation, observational and model results, are depicted in Figure 4. Precipitation in the north and western coasts occurs mostly in winter, while in the eastern coasts the maximum seasonal precipitation occurs in autumn. In these three seasons the northwest-southeast gradient is unmistakable, only in summer does the gradient change to north–south. In the summer, half of Iberia has less than 50 mm of seasonal accumulated precipitation, were Algarve, Guadalquivir river basin and the southern coast of Spain have even less than 10 mm. Apart from the northern coasts, the summer precipitation in most parts of Iberia is typically associated to convective systems produced by the combination of strong soil heating and instability. The precipitation patterns for winter, spring and autumn are similar to the yearly precipitation with the lowest intensity occurring in the spring. In March, according to Paredes et al. (2006), the spring cyclones typically cross the North Atlantic at latitudes higher than the Iberian Peninsula and since these are the major sources of precipitation there is a reduction in precipitation during this month.

WRF is able to reproduce the seasonal distribution of rainfall. In all wet seasons, it underestimates precipitation in the Guadalquivir basin and near the Gibraltar strait and overestimates in regions with high topography, while ERA-Interim is only able to reproduce the northwest-southeast gradient in the wet seasons and the north–south gradient in the summer.

Rauscher et al. (2010) investigated the resolution effect on ENSEMBLES RCM simulations of seasonal precipitation over Europe. One of their main findings was that the majority of models over-predicted precipitation. The higher resolution models, 25 km, showed a smaller ratio of convective to total precipitation when compared to the 50 km resolution. More importantly, and related to the latter fact, the finer resolution models revealed a better representation of both spatial patterns and temporal evolution of precipitation in summer, although not in winter months or in the annual mean. Here, somewhat the contrary occurs. The global seasonal correlation of WRF with observations shows (Figure 5(a)) a clear annual cycle, being maximum in winter and minimum in summer. In fact this correlation cycle is in line with the global seasonal precipitation. WRF and ERA-Interim have similar seasonal correlation coefficients showing a good ability to globally describe the precipitation phase in Iberia. WRF correlation is slightly higher than ERA-Interim in winter and autumn, around 0.86 (0.84) and 0.80 (0.77), respectively. However, in summer, when the correlation values are smaller, ERA-Interim outperforms WRF (0.81 and 0.73, respectively). While in winter the local underestimation and overestimation in WRF cancel each other, prompting a −0.7 mm season–1 bias (0% normalized), in spring the overestimation associated to the high topography is greater, originating a 11% normalized bias (19 mm season–1 bias) and in autumn the opposite occurs (−10% and −22 mm season–1). The amplitude of the errors is comparable in the three seasons, since MAPE is 29% in winter and autumn and 32% in spring (Figure 5(c)). The RMSE (not shown), due to the squared error being very sensitive to large occasional deviations, reveals that spring (75 mm season–1) has less large deviations than winter and autumn (87 and 89 mm season–1, respectively). In the summer, WRF overestimates precipitation in most of Iberia, except in the north and northwest coasts as well as in the eastern coast, hence a 21% normalized bias and a 53% MAPE. The highest values are found in Sierra Nevada.

Figure 5.

Iberian seasonal precipitation global error measures of WRF9km and ERA-Interim against the observational grid (Obs_0.2 grid). The presented errors, computed for seasonal accumulation of precipitation, are correlation coefficient, normalized BIAS and MAPE. Horizontal lines indicate the limits of the 95% confidence interval of the corresponding variable, computed by 10 000 bootstrapping samples.

ERA-Interims' statistical errors, have a similar seasonal behavior as WRF's. While WRF presents smaller MAPE and RMSE than ERA-Interim in winter and autumn the opposite is verified in summer.

Figure 6 shows temporal error measures for increasing accumulating time intervals. For each time scale, the corresponding running accumulation of the observed and simulated precipitation time series, at every individual grid point, were calculated, and then pooled together, forming two independent time series. Finally, the errors were computed between the two series. The correlation (Figure 6(a)) improves significantly with increasing accumulating period from 1 to 20 d but saturates after 30 d, as in Zhang et al. (2009). Note that all values are significant at significance level 0.01 based on the Fisher transformation. The correlation of the WRF domain is always higher than ERA-Interim for periods shorter than 60 d and from this point they become equal. For example, for 1 d accumulation the correlation coefficients are 0.73 for WRF and 0.71 for ERA-Interim. The high correlation coefficients suggest that both models capture well the weather systems accountable for most of the precipitation. This can be confirmed by the higher correlation distribution on the Portuguese west coast and Galicia (Figure 7) where the Atlantic fronts are responsible for the majority of the rain bearing storms. The higher correlation coefficients for the WRF domain, denotes that the high-resolution terrain improves precipitation estimates. This is particularly evident in the higher correlations of WRF in the Cantabrian Mountains, Tejo and Guadalquivir river basins. On the East coast, both WRF and ERA-Interim have lower correlations which are associated to the convective nature of precipitation in this area. WRFs' underperformance in the eastern side of Iberia may be related to the proximity of the lateral relaxation grid points to the region. As pointed out by the study of Herrera et al. (2010a), where the performance of ENSEMBLES RCMs were analysed for the Spanish territory, the model using wider boundary relaxation area for the wind, the KNMI model, gave the best spatial correlation for precipitation of the participating models (0.82 monthly correlation, see Table 1). On the other hand, with the present high resolution, 9 km, to increase further the domain would imply a prohibitively high computational cost. Nevertheless, the higher resolution orography in WRF increases the correlation in the high topography areas. Figures 5-7 show correlations computed in rather different ways. In Figure 5, results were aggregated by season, implying that the assessment comprises both interannual variability of seasonally aggregated fields and their spatial variability. In Figure 6, results are pooled together for all grid points, in both space and time, offering a global measure of the local fit between model and observations, and its response to time aggregation. In Figure 7, correlations are performed between time series in the same spatial location, for different aggregation intervals, focusing the analysis on the spatial distribution of model performance across Iberia. The latter reveals the large heterogeneity found in Iberian climate and as in Figure 6, the correlation increases for all areas with increasing accumulating period, although the east–west divide never disappears.

Figure 6.

Global error measures of WRF 9 km and ERA-Interim precipitation against the observational grid (Obs_0.2 grid) for Iberia. The errors are correlation coefficients (correlation), mean absolute percentage error (MAPE) and standard deviations. The presented errors are computed for increasing accumulation periods of precipitation, from daily to 90-d. Horizontal lines indicate the limits of the 95% confidence interval of the corresponding variable, computed by 10 000 bootstrapping samples.

Figure 7.

Correlation maps of the observational grid and WRF9km, and ERA-Interim. (a) monthly, (b) 5 d and (c) daily.

Table 1. WRF Iberian results compared with ENSEMBLES. Results for both WRF resolutions and the two ENSEMBLES models selected for their especially good performance in Iberia, KNMI and ETHZ. Best results italicised
Model CorrelationN Bias (%)MAPERMSE (mm/d)N σ
ENSEMBLES5 d(0.44, 0.7)(2.9, 28.5)(0.73, 1.16)(13.2, 18.7)(0.83, 1.12)
 Month(0.54, 0.82)(2.9, 26.0)(0.45, 0.78)(37.7, 60.3)(0.84, 1.17)
ETHZ5 d0.704.10.7514.10.97
KNMI5 d0.702.90.7313.20.84
WRF 27 km5 d0.70−4.70.7213.50.94
WRF 9 km5 d0.801.40.5910.80.96

In Figure 6(b), it is visible that MAPE decreases significantly from 82% for a daily accumulation to 32% for 90 d accumulation. MAPE improves significantly for 3 d accumulation with a reduction of 15–65%. The lower MAPE is observed in the western coast, particularly in Galicia and northern Portugal. With increasing accumulating period, MAPE is reduced in all areas, with higher values remaining over the high topography in central Spain (not shown). The biases in WRF and ERA-Interim have opposing values, while WRF has a wet bias; ERA-Interim has a dry bias. WRF's and the ERAs' wet bias are located in the centre of the Peninsula, but the latter has strong dry bias in the Tejo and Guadalquivir basins (not shown). When the bias is normalized by the observed average precipitation WRF performs better than ERA-Interim, it is 1.4% for WRF, while it is 6% for the later for all accumulation periods.

The amplitude of the various precipitation patterns can be summarized by the standard deviation (Figure 6(c)), which, however, mixes spatial and temporal variability in a single indicator. Both WRF and the ERA-Interim follow the observations in the increase of standard deviation with increasing accumulating period and their ratio to the observed is kept constant, 0.9 and 0.7, respectively. This, along with a 0.56 mean daily spatial correlation, is indicative that the ERA-Interim fields are too smooth and underestimate the amplitude of the precipitation events. WRF has a 0.62 mean daily spatial correlation and is able to represent the amplitude of each storm reasonably well.

3.2. Basin analysis

In the last section, the whole domain was analysed, but since orography plays such a significant role in the spatial distribution of precipitation, a river basin analysis is now explored. In Figure 8, the spatial variability and the seasonal cycle are both presented. The monthly mean of the daily precipitation is spatially averaged for each individual basin. Note that merely continental basins are considered and that the precipitation scale is maintained throughout the plots. Both WRF and ERA-Interim capture well the seasonal cycle. In all basins, the minimum daily precipitation occurs in July where, with the exception of the North, Catalana and Ebro, less than 0.5 mm d−1 is observed. WRF overestimates summer precipitation except in the first two.

Figure 8.

Seasonal cycle of monthly mean daily precipitation for each basin, results from observational grid (black), WRF 9 km (red) and ERA-Interim (blue – dashed).

The winter and autumn maximum and local March minimum of the western river basins is correctly simulated (Figure 8(a)–(g)). In these basins precipitation is mostly due to the frontal systems that cross the North Atlantic during these seasons. In the northeastern basins, the two peaks of precipitation occur in spring and autumn, which according to Romero et al. (1998) is a feature of this region's climatology and is captured by both WRF and ERA-Interim.

To evaluate the performance of the simulations, similar error statistics for all time periods were calculated for each river basin, with the daily values shown in Figure 9. For each time period, the precipitation in each basin was spatially averaged. The ability of WRF to represent the timing of precipitation is mirrored in the high correlation coefficient for total daily precipitation. In the NW basins, were the influence of frontal systems that cross the North Atlantic is mostly felt, the correlation is above 0.9. In accordance with Figure 7, the correlation for the eastern basins is lower, with the lowest value observed in the Segura basin (0.77) indicating that the onset of cumulus convection is deficient since this is the major source of precipitation in these regions. The remaining basins have correlations above 0.86. In contrast, ERA-Interim does not perform as well, except in Estremadura and SW & Algarve. In the eastern basins its correlation is lower than 0.43.

Figure 9.

Error measures of WRF 9 km and ERA-Interim daily basin precipitation against the observational grid (Obs_0.2 grid), for the river basins considered. The errors are correlation coefficients (correlation), root mean square errors (RMSE), mean absolute percentage error (MAPE), normalized bias (Bias%). Horizontal lines indicate the limits of the 95% confidence interval of the corresponding variable, computed by 10 000 bootstrapping samples.

Although the phase is well simulated, it is patent from Figure 8 that WRF does not perform in the same manner for all basins. In the North, precipitation is underestimated all year round, which is reflected in −16% normalized bias (Figure 9(b)) which is comparable to the values found in Estremadura and SW & Algarve. This might be attributable to the smoother topography of the model relative to the real world, which is not high enough to force higher precipitation rates. In fact, in the latter the small mountain ranges that characterize these two regions are smoothed out in WRF's topography (compare Figure 2(a) and (b)). Since there is very little precipitation in summer in these latter regions, and topographic enhancement does not have a preponderant role, precipitation is well simulated (Figure 8(c) and (f)). However, WRF overestimates in the Douro and Ebro basins all year round but specially in late spring when the source of precipitation transitions from cyclonic frontal systems to convective. In fact, apart from the basins already referenced, WRF overestimates late spring precipitation. This is particularly significant in the eastern basins where convection is one of the main sources of precipitation. On the other hand, autumn precipitation is underestimated in many basins. The overall basin bias (Figure 9(b)) shows that WRF is wetter than ERA-Interim in all basins except Guadalquivir where they match, a good results considering the dryness of ERA-Interim results.

The RMSE in WRF is less than 3 mm d−1 in all basins, while ERA-Interim goes up to 5 mm d−1. WRF has, RMSE lower than the ERA-Interim, indicating that the latter has higher deviations from the observations which were already apparent in Figure 8. The exception is Estremadura and SW & Algarve, where WRF and ERA-Interim have similar performances. The higher RMSE values in the eastern basins are consistent with the low correlations found in these regions. MAPE in WRF, with exception of Estremadura, is lower than 46% in the western basins and less than 73% in the remaining (Figure 9(d)). In the eastern basins, ERA-Interim has MAPEs higher than 100%.

To analyse the intensity and frequency of precipitation events, the 2.5, 10, 20, 25, 30, 40, 50, 60, 70, 75, 80, 90, 95, 97.5, 99 and 99.9 percentiles were computed and q–q plots were constructed for each basin. As in ERA-Interim, WRF output is compared against the nearest grid point since the grid box averaging smoothes the extremes. The grid points in each basin are pooled together and the percentiles are determined considering only wet days (defined as having precipitation greater than 0.1 mm). Both models underestimate significantly the lower daily precipitation quantiles, as also found by Herrera et al. (2010a) and by Soares et al. (2012b), where all the RCMs and both the ensembles created from them, underestimate daily precipitation quantiles. The larger quantiles are also underestimated by ERA-Interim. For the later, the basin with the best results is Estremadura, where the top 10% are only underestimated by less than 32% and the top 0.1% is underestimated by 23%. The worst results are obtained for the SSE and Levante basins which are in line with Herrera et al. (2010a, 2010b). These are mostly due to the large temporal and spatial variability of precipitation in these regions. In the western basins, WRF performs similarly to ERA-Interim in the lower quantiles, although its difference is about 10% smaller. Only in the SW & Algarve is WRF worse than ERA-Interim for the lower quantiles. On the contrary, for the higher quantiles, WRF outperforms ERA-Interim. In the latter, the precipitation is underestimated by more than 40%, while WRF overestimates the highest quantiles in the northern basins, i.e. in the wettest basins. This is in line with Kjellström et al. (2010), who established that in Europe overestimation of precipitation worsens with increasing precipitation in most RCMs. In the dryer basins, WRF underestimates precipitation, but there is clearly an improvement in relation to ERA-Interim. ERA's poorer performance is probably due to the low resolution which smoothes the orography and thus reducing its enhancing effect on precipitation (Figure 10). Similar results were found in Norway, by Heikkilä et al. (2011).

Figure 10.

Quantiles of daily precipitation for each basin. The scales are different to add legibility to the plots.

A basin average analysis was also performed and a considerable improvement was found in the agreement between the simulated and observed values (not shown). Similar results were found by Herrera et al. (2010a, 2010b).

3.3. Extreme analysis

The high precipitation associated to the higher quantiles is often related to short lived and extreme events, thus its correct evaluation is paramount to an accurate hazard assessment.

The histogram of daily precipitation as a measure of the skill of WRF and ERA-Interim to simulate the intensity and frequency of events is shown in Figure 11(a). In the low end of the spectrum, ERA-Interim overpredicts the frequency of events while in the high end of the spectrum it is unable to correctly reproduce the distribution, where the largest values are completely missing. That is emphasized in Figure 11(b) where the highest percentiles deviate from the observed by more than 44%. Conversely WRF is able to reproduce much better the precipitation distribution and is able to accurately simulate the extremes. Results in Figure 11(b) indicate that WRF27km compares better with the 0.2° grid than WRF9km, in what concerns the distribution of high-rank quantiles. This is consistent with the resolution represented by the observational grid and could be used to question the usefulness of the higher resolution run. However, the WRF9km simulation performs better in most other statistics, most notably in what concerns bias and correlation, and it was also found to compare better against (ungridded) station observations (Soares et al., 2012a).

Figure 11.

(a) Histogram, and (b) quantiles of the observational grid, WRF 9 km, WRF 27 km, ERA-Interim daily precipitation for the whole time series and full range of results from the ENSEMBLES models as a grey shadow.

The excessive frequency of precipitation, in ERA-Interim, in the low end of the spectrum is also illustrated in Figure 12(a) where the ratio of simulated to observed wet days with precipitation greater than 0.1 mm is between 150 and 278%, except in Estremadura and SW and Algarve. The greatest discrepancies (more than 200% of wet days) occur in the dryer regions where the prevalent precipitation is associated to convective local systems. When the wet day standard index (r > 1 mm) is analysed the proportion of days is considerably reduced. It fluctuates between basins by 50% and is never greater than 173%. The reduction in the number of wet days between the two thresholds (r > 0.1 mm and r > 1 mm) is especially significant in the dryer basins (about 100%), which indicates that in these basins it rains very frequently but by small amounts.

Figure 12.

The river basins percentage of wet days above (a) 0.1 mm and 1 mm, and percentiles (b) 95, (c) 99 and (d) 99.9, compared with the basin observational data.

In the lower threshold, the percentage of overestimation is between 25 and 107% (except in Estremadura where WRF predicts the correct number of rainy days). At the higher limit, the number of wet days in the majority of basins is between 125 and 150%. In WRF, there is also an excess of very light rain, but not as significant as in ERA-Interim. In Estremadura and SW and Algarve WRF underestimates by no more than 13%, but overall overestimates by 26% the number of wet days (40% for ERA-Interim).

The basin distribution of fourth, second and highest percentiles is depicted in Figures 12(b)–(d). As shown before, it is evident that ERA-Interim is completely unable to reproduce the high end quantiles, by underpredicting by more than 40%. Conversely, WRF reasonably reproduces the spatial variability of the extremes, although it underestimates in most regions. The highest differences occur in the Levante basin where high precipitation is associated to short lived high intense storms.

For an assessment of the spatial distribution of extreme events, the 15 standard indicators from ECA&D were calculated for the three datasets as well as the frequency of wet days (r > 1 mm). Here we present a subset of seven indices – moderate wet days (days with precipitation above 75th percentile–r75p), consecutive dry days (maximum number of consecutive days with precipitation below 1 mm – cdd), consecutive wet days (maximum number of consecutive days with precipitation above 1 mm – cwd), very heavy precipitation days (days with precipitation above 20 mm – r20mm), highest precipitation in one day (r × 1 d), highest precipitation in 5 d (r × 5 d) and percentage of precipitation above 95th percentile (r95ptot) (Figures 13 and 14).

Figure 13.

Geographical distribution of yearly mean values of (a) frequency of wet days, with daily rainfall above 1 mm, (b) maximum number of consecutive days with precipitation, (c) maximum number of consecutive days without precipitation and (d) moderate wet days with precipitation above the percentile 75. Values on the top of each map, show yearly average spatial correlation values of each individual map with 0.2 grid.

Figure 14.

As in figure 13 but (a) number of days with a maximum precipitation over 20 mm, (b) maximum precipitation, (c) maximum of precipitation during 5 d and (d) percentage of precipitation above the 95th percentile. Values on the top of each map, show yearly average spatial correlation values of each individual map with 0.2 grid.

The frequency of wet days (Figure 13(a)) mirrors northwest-southeast distribution of precipitation; a very wet N-NW, with more than 25% percent of days of precipitation and an extremely dry SE with less than 10% of days of rain. In the first region there are significant areas where it rains for more than 35% of days. This also mirrored in the number of consecutive wet days (Figure 13(b)) where there are more than 14 consecutive days in the NW and less than 6 in the east. The number of consecutive dry days reflects more of a NW-S divide, where there are less than 40 consecutive days without precipitation in the summer in the north and more than 90 dry days in the south. As before, the topography plays an important role in these distributions. WRF is able to reproduce the frequency of precipitation quite well, with some overestimation associated to the topography, especially in the mountain ranges around the Submeseta Norte, and to a lesser extent it also overestimates in the south and east coasts. Nevertheless, the spatial correlation is of 0.75, which is in line with the results from the RCMs in the ENSEMBLES project (Herrera et al., 2010a). As seen before, ERA-Interim considerably overestimates the frequency of wet days over all of Iberia which is in agreement with Belo-Pereira et al. (2011).

The northwest-southeast gradient in the number of consecutive wet days is captured by WRF, but in this case there is underestimation in the NW. Herrera et al. (2010a) also obtained similar results, wherein RCMs have difficulty in maintaining long periods of precipitation. Nevertheless the spatial correlation is also 0.75. All central and eastern Spain are overpredicted in ERA-Interim. The N–S distribution of the maximum uninterrupted dry days is patent in both ERA-Interim and WRF. Both underestimate the duration of dry spells but correlate very well with the observations, 0.93 and 0.84, respectively.

The moderate wet days where precipitation is above the 75th percentile varies between 40 in the NW and 7 days in the SE with a spatial distribution similar to the frequency of wet days. WRF has a similar behavior as in the frequency index, but now the spatial correlation is 0.97. Once again ERA-Interim shows a tendency to overestimate.

Very high precipitation occurs mostly in N-NW, Tejo and Guadalquivir river basins, in the Gibraltar strait (Figure 14). This is well captured by WRF which has a 0.78 correlation. The results for the maximum daily and 5-d precipitation are similar to the previous one. In addition to the mentioned regions, in these two indices, the Mediterranean short lived but strong storms connected to high sea surface temperatures and orographic enhancement have a significant signal. Note the maximums in the Levante coast. Since the 5-d accumulation reduces small spatial and temporal desynchronization, the latter has as spatial correlation 0.78, while the former has 0.7. The 95th has also a significant signal along the Mediterranean coast. WRF captures the overall distribution with a 0.87 correlation, yet its underestimates precipitation in this region (not shown). WRF also overestimates the 95th percentile in the NW (Gerês mountain). The fact that gridded observations tend to smooth out extremes can partially contribute to this result. The percentage of precipitation above the 95th percentile is less than 22% for the majority of Iberia, except in the Mediterranean coast where it can be 30%. WRF overestimates in these areas, but overall has a spatial correlation of 0.95.

4. Discussion and conclusions

A common approach in assessing RCM results is to compare them against other state of the art RCMs'. The ENSEMBLES project offers a good set of results for such a task, which have been comprehensively used. However, the comparison has to be taken with caution, since the ENSEMBLES models used a different reanalysis, namely ERA-40 instead of ERA-Interim, and refer to a different time interval, i.e. 1961–2000 instead of 1989–2008. A further problem comes from the daily accumulation period for precipitation, defined in ENSEMBLES as 0–24 UTC, whereas in the gridded observations it is defined as 6-6 UTC in Spain and 9-9 UTC in Portugal. The latter problem may be addressed by restricting the comparisons to accumulation periods above 5 d. Since some climate statistics may be considered not stationary in the full reanalysis period 1961–2010, the results are not a fair comparison, but they still provide a relevant qualitative scale for the present analysis. Finally, one must keep in mind that some of the differences observed in the comparison may be due to changes in the reanalysis system.

Table 1 shows a selection of results from ENSEMBLES, including the full range of the ENSEMBLE statistical parameters, and two ENSEMBLES models selected for their especially good performance in Iberia, from KNMI and ETHZ. The KNMI model has the best performance in four of five indices, whereas ETHZ's was the second best in the five indices but has the better representation of observed variability including the distribution of quantiles of precipitation. Those results are to be compared with the same parameters computed for WRF at both 9 and 27 km resolution (the outer domain), since its resolution is nearer to the 25 km ENSEMBLES resolution and also closer to the resolution of the gridded observations.

From Table 1, one may conclude that WRF 9km has a better agreement with observations for all parameters, with the exception of the normalized 5-d standard deviation, for which ETHZ has a best performance. In what concerns correlation, normalized bias, MAPE and RMSE, the results of WRF 9km are not only better than any of the ENSEMBLES models but they are outside the ENSEMBLES range. Improvements are present in both 5-d and monthly accumulation periods, but are more striking in the 5-d case. Results from WRF 27km, however, are mostly within the range of ENSEMBLES, although with negative bias inherited from ERA-Interim (outside the large range of positive biases found in ENSEMBLES), very near the low end in MAPE and RMSE and at the high end in 5-d correlation.

Heikkilä et al. (2011) did a similar comparison between WRF simulations at 10 and 30 km resolution against ENSEMBLES results for Norway. While their results are not easily compared with the present ones, considering the very different climates, a clear improvement from the higher resolution simulation was also found. In spite of the very different climate settings Norway and Iberia share a very rugged topography, a condition justifying the need for significant resolution. It is important to mention that study of Heikkilä et al. (2011) used the same ERA-40 boundary conditions as ENSEMBLES, making the comparison straightforward.

The representation of extreme weather by RCMs is an increasingly important issue for impact assessment. Some models may produce very good mean statistics, namely bias, MAPE or correlation coefficients, and perform poorly in the representation of high frequency variability. The normalized standard deviation, presented in Table 1, gives an indication of model variability. A better description was presented in Figure 11, showing the distribution of quantiles of precipitation. In that Figure, the full range of results from the ENSEMBLES models for Iberia was also shown, indicating that WRF9km outperforms any individual ENSEMBLES model, being better in almost all quantiles taken from any model in the ensemble. Surprisingly, WRF27km does even better, with a really good match with observations. The better results from WRF27km, unlike what was found in the mean statistics, may be due to the fact that the gridded data has a resolution of 0.25°, similar to WRF27km, and higher ranking percentiles (e.g. >P90) are very sensitive to horizontal averaging. While the WRF results are generally good, there is some indication of difficulties in the representation of summer precipitation everywhere in Iberia, and of heavy precipitation events, especially at the Mediterranean coast. The poor representation of the latter events may be inherited from ERA-Interim boundary conditions. In both cases, mesoscale processes are probably to blame. Like in many other regional models, WRF is found to produce too much light rain, and to underrepresent heavy precipitation (e.g. above P95 or above P99), although with a very significant gain from ERA-Interim.

The WRF model, originally developed for NWP and for atmospheric research, was here shown to be a good performer in regional climate studies. Previous studies in other regions, e.g. by Heikkilä et al. (2011) for Norway, by Zhang et al. (2009) in the United States, also arrived to similar conclusions. Improvements gained with high resolution runs in complex terrain areas seem to justify the increased computational cost of those simulations, especially if one is concerned with the representation of fast variability, associated with weather extremes. Higher resolution also leads to gains in all mean statistics, being able to cancel a negative bias found in ERA-Interim, and to largely improve the values of correlation coefficients, MAPE and RMSE, at the level of the regional basins, indicating a much better representation of the geographical distribution of the precipitation field, a result that was also found for the ENSEMBLES models in Iberia.


This study was partially funded by the Portuguese Science Foundation (FCT) under the Project REWRITE – PTDC/CLI/73814/2006, co-financed by the European Union under Program FEDER. The ENSEMBLES data used in this work were funded by the EU FP6 Integrated Project ENSEMBLES (contract GOCE-CT-2003-505539) whose support is acknowledged; we also acknowledge DMI for hosting the RCM repository. R.M.C. also acknowledges the financial support of FCT under the grant SFRH/BPD/47221/2008.