Dynamical downscaling of ERA-40 in complex terrain using the WRF regional climate model
- First Online:
- 2.3k Downloads
Results from a first-time employment of the WRF regional climate model to climatological simulations in Europe are presented. The ERA-40 reanalysis (resolution 1°) has been downscaled to a horizontal resolution of 30 and 10 km for the period of 1961–1990. This model setup includes the whole North Atlantic in the 30 km domain and spectral nudging is used to keep the large scales consistent with the driving ERA-40 reanalysis. The model results are compared against an extensive observational network of surface variables in complex terrain in Norway. The comparison shows that the WRF model is able to add significant detail to the representation of precipitation and 2-m temperature of the ERA-40 reanalysis. Especially the geographical distribution, wet day frequency and extreme values of precipitation are highly improved due to the better representation of the orography. Refining the resolution from 30 to 10 km further increases the skill of the model, especially in case of precipitation. Our results indicate that the use of 10-km resolution is advantageous for producing regional future climate projections. Use of a large domain and spectral nudging seems to be useful in reproducing the extreme precipitation events due to the better resolved synoptic scale features over the North Atlantic, and also helps to reduce the large regional temperature biases over Norway. This study presents a high-resolution, high-quality climatological data set useful for reference climate impact studies.
KeywordsWRF Downscaling ERA-40 ENSEMBLES Regional climate modelling
A number of model studies have addressed the regional effects of future climate change in Europe (ENSEMBLES members 2009). These studies point to increased precipitation in Northern Europe in the future which can have important impacts on hydrology, vegetation and infrastructure in the influenced areas. In order to deliver reliable information for the society on these issues we need to focus on a local level. This study is a first step towards a higher resolution assessment of future climate prediction in Norway. We provide high-resolution climate parameters for Norway which are increasingly required and of crucial importance for driving various climate impact models. This study employs a new model for Europe (WRF: Weather Research and Forecasting, http://www.wrf-model). It provides a first-time comprehensive model evaluation for new users wanting to apply the WRF model for climatological simulations over Europe, and also a qualitative comparison of the performance of the WRF model against other state-of-the-art regional climate models.
A common approach used in regional climate simulations for this region has been to include only the continent of Europe with little ocean into the high-resolved regional model domain. In this study, we aim to improve the representation of climate in Europe by increasing the size of the regional model domain to cover the whole North Atlantic. Such a setup will increase the independency of the regional climate model from the driving data. In this way also the synoptic scale features on open water will be better resolved on our high-resolution domain (30 km) than in the driving ERA-40 reanalysis (1.1°) before they reach the coast of Europe. We apply the spectral nudging procedure to force the regional model to keep its large scale circulation consistent with the driving reanalysis data.
The focus of this study is on the validation of the WRF model in a period (1961–1990) for which many observations and model runs exist and on finding an optimal setup for future prediction simulations. The main questions to address are (1) how well does the WRF model agree with observations when run with “ideal” boundary conditions (ERA-40 reanalysis) and spectral nudging on the 30 and 10 km horizontal resolution? (2) Does the model simulation improve the driving ERA-40 reanalysis? Is the 10 km resolution adding significant value to the 30 km resolution? (3) How well does the WRF model capture the regional differences of climate in Norway? and (4) is the model able to reproduce the observed extreme values of precipitation and temperature? In order to put our model simulation into a larger context of the state-of-the-art of regional climate modelling we perform a comparison with 12 European and Canadian models which participated in the recently finished ENSEMBLES project (see Sect. 3.2 for more details on the project). The WRF and the ENSEMBLES model simulations are not directly comparable because of the different setup used. Keeping this in mind, the comparison presented in this study should be understood as qualitative.
2 Model description and setup
The model employed for this study is the WRF regional climate model (version 3.1.1). The WRF model has a rapidly growing user community and has been used for climatological studies, various case studies and operational weather forecasting among other purposes in the recent years. For the experiments of this study we used a large model domain of a size of 9,090 km (W–E) × 5,490 km (S–N) with a horizontal grid resolution of 30 km (Fig. 1). We included one nest inside the model domain with a horizontal resolution of 10 km. This nest has a size of 880 km (W–E) × 1,840 km (S–N). Both domains have 40 vertical levels reaching up to 50 hPa. The first reason for choosing such a large domain was that the precipitation in the western coast of Norway is mainly large scale and the moisture can have its origin far in the southwestern North Atlantic (Stohl et al. 2008). Another reason was that we wanted to find an optimal setup for subsequent future climate predictions. A larger model domain will give the regional model more freedom to develop its own synoptic and mesoscale circulation. This may be an advantage in regions where the climate change signal is strongly influenced by advective processes.
It has been noted that using a large domain may lead to deviation of the large scale features from the driving fields creating problems close to the boundaries (Jones et al. 1995; Koltzow et al. 2008). To reduce this risk, we use spectral nudging. Spectral nudging is a method which allows the passing of the driving global model information not only onto the lateral boundaries but also into the interior of the regional model domain (Waldron et al. 1996). The value of spectral nudging has been discussed in the literature (e.g. Alexandru et al. 2008; Miguez-Macho et al. 2004, 2005; Radu et al. 2008; Von Storch et al. 2000; Zahn et al. 2008) and there is some controversy. Most studies agree that nudging too strongly will not allow the regional model to deviate much from the driving fields. While spectral nudging seems to reduce the sensitivity to the chosen model domain or grid size (Alexandru et al. 2008; Miguez-Macho et al. 2004) other studies show that it can affect extreme precipitation or high frequency dynamical phenomena (Alexandru et al. 2008; Radu et al. 2008). We conducted several tests to evaluate the sensitivity of the modelled surface variables to nudging. We found that spectral nudging has an important effect in keeping the large scale circulation of the regional model in phase with the global model, but does not constrain the model’s ability to develop small scale features. The extreme precipitation events were actually better reproduced by the nudged run than the free run.
We applied the spectral nudging technique following previous studies by Miguez-Macho et al. 2005 and Radu et al. 2008. We nudged only in the outer domain in order to let the regional model create its own structures in the high-resolution nest. For the same reason we applied nudging only on vertical levels above the boundary layer. The threshold for wavelengths over which the waves were nudged was 1,000 km. Following Miguez-Macho et al. 2005 and Radu et al. 2008, the nudging was applied to u and v winds, temperature and geopotential height but not to humidity. The sensitivity to the strength of the nudging was tested but no significant differences were found between stronger (every 6 h) and weaker (every 24 h) nudging. We chose the weaker nudging approach in order to maximize the freedom of the regional model to deviate from the driving global fields.
We simulated the years from 1960 to 1990 because many climatology simulations exist for this period, such as the regional model runs of the EU-project ENSEMBLES (see Sect. 3.2). The first year was used to spin up the soil moisture and not included in the analysis. The driving global data used was the ERA-40 reanalysis (Uppala et al. 2005) with 1.1-degree horizontal resolution and 24 vertical pressure levels. The experiment was performed using the default setup of the WRF model for the physical parameterizations as much as possible to keep the runtime low. The cloud microphysical scheme used was the 3-class scheme (Hong et al. 2004), the Kain–Fritsch scheme (Kain 2004) for the convective parameterization, the Yonsei University (YSU) (Hong et al. 2006) planetary boundary layer scheme, the Monin–Obukhov scheme for surface layer processes and the 4-layer Noah land-surface model (Ek et al. 2003) for the land-surface and soil processes. We used the new MODIS land use data set to describe the vegetation and land use classes in Norway (http://modis.gsfc.nasa.gov/). The Community Atmosphere Model (CAM) schemes were used for short-wave and long-wave radiation (Collins et al. 2006). We tested the sensitivity of the model to different microphysical schemes and found no significant differences between the simpler and the more sophisticated schemes on the spacial scales (10 km) or time scales (daily) of this study. We used the so called 1-way nesting procedure which passes information only from the outer domain to the inner nest. This is a common approach in climatological studies because of possible stability problems introduced by 2-way nesting.
The results obtained within this study are evaluated against daily surface observations (precipitation, 2-m temperature and 10-m wind speed) from the Norwegian meteorological office in a similar manner with Barstad et al. 2009. The observational network consists of several hundred meteorological stations covering the whole country and provides the best data available for Norway. The data was checked for continuity and consistency and only stations which contained a continuous 30-year data set were taken into account in this comparison. This left us with 316 stations of precipitation, 66 stations of 2-m temperature and 67 stations of 10-m wind speed data. The comparison was made using the nearest gridpoint of the model to the observations. Although the horizontal resolution of the model is quite high (30 and 10 km) the error of the elevation of the model gridpoint to the actual elevation can be large at some points, especially on the coast and in the mountain slopes. For temperature we used a simple lapse-rate correction assuming that the temperature drops 6 K each 1,000 m, as has been used in several studies (e.g. Barstad et al. 2009; Kostopoulou 2009). Assuming a constant negative lapse rate neglects many effects, such as the complexity of the temperature profile in a boundary layer. In a case of a winter-time inversion, for example, this correction actually increases the error. Still, without this correction the temperature bias will reflect mostly the smoothed topography and not the correctness of the model dynamics or the physical parameterizations. Moreover, comparison of the temperatures of the ERA-40 and both WRF simulations with very different resolutions would not be fair without such a correction.
In the case of wind the issue is more complicated as there is no standard procedure to correct for the altitude error. We know that the stations measuring wind in the mountains are located in small valleys which are not resolved by the model topography. Therefore, the wind observations are not necessarily representative for the areas they are located in. The wind observations are made optically which can introduce an error in some cases. In order to use only quality-checked data representative for the location in question we use the ten coastal stations chosen by Barstad et al. 2009. These stations were chosen as recommended by the meteorological office responsible for the observations. Data was written out from the model every 6 h for the 30-km domain and every 3 h for the 10-km domain and daily means were calculated from these values.
3.2 Comparison with models participating in the ENSEMBLES project
There has been large regional climate modelling and model inter-comparison activity in Europe during the recent years. The ENSEMBLES project (Ensembles-based predictions of climate changes and their impacts) was finished in the end of 2009 (ENSEMBLES members 2009). Its aim was to produce an ensemble of downscaled global future climate projections in order to provide the European society and economy with more detailed information on the future climate. Some ten state-of-the-art European and Canadian climate models took part in the project and several experiments were performed with different combinations of global model, greenhouse gas emission scenarios and horizontal resolution of the regional models. One part of it was, similar to the goal of this study, to validate the models driven with the ERA-40 reanalysis data for the period of 1958–2002. The results of this project give us an excellent opportunity to put our model results in a larger perspective and investigate how well the WRF model is performing within the spread of the ENSEMBLES models. We chose a set of 12 simulations with different models for comparison and performed the same analysis as with our simulations for the period of 1961–1990. These models are listed in Fig. 12.
We chose the 25-km resolution of the ENSEMBLES model runs to allow for a comparison as accurate as possible with our 30 and 10-km simulations. The number of vertical levels in the ENSEMBLES runs was lower than in our runs and varied from 19 to 32. No spectral nudging was used in these runs. Their domain size was smaller, covering Europe including the Mediterranean in the south but just only including the northernmost part of Norway in the north and not the whole Atlantic ocean. The analyzed precipitation, 2-m temperature and 10-m wind speed are daily means. The ENSEMBLES means shown are calculated as simple averages in each case and are not weighted based on model performance or any other way.
3.3.1 Geographical distribution of precipitation bias
3.3.2 Histogram of the daily mean precipitation
In order to ignore the few exaggerated extremes we look at the quantiles of the daily mean precipitation (lower panel of Fig. 4). The ERA-40 reanalysis again lacks the highest values of the spectrum and a few of the ENSEMBLES models perform even worse than the driving reanalysis in reproducing the quantiles from 0.6 to 0.99. The ENSEMBLES mean performs very well. The WRF model shows good skill and the refined resolution of 10 km adds more value to the 30 km simulation producing nearly a perfect agreement between the observed and modelled quantiles. We also see that these results are consistent during all seasons.
3.3.3 Regional differences of precipitation
This is also the case for the extreme precipitation which is defined as 0.95 quantile. The figure did not change if we changed the 0.95 quantile to 0.9, 0.99 or 0.999. In relative numbers the models mostly overestimate the extreme precipitation in the driest areas (3 and 10–12) and perform better along the wet west coast of Norway. Both WRF simulations are comparable with the ENSEMBLES mean.
The bias in the number of wet days (Fig. 5) shows that all models overestimate the frequency of precipitation. The error is largest again in the dry areas, 3 and 12, but otherwise the difference between the regions is smaller than in the case of total precipitation of extreme precipitation. The WRF model seems to be performing very well giving a low bias compared with the ENSEMBLES models.
3.3.4 Extreme values of precipitation
3.4 2-m temperature
3.4.1 Geographical distribution of 2-m temperature bias
3.4.2 Histogram of the daily mean temperature
The same is shown by the modelled quantiles in the lower part of Fig. 8, plotted against the observed quantiles. Both WRF simulations reproduce well the higher quantiles but overestimate the lower quantiles during the DJF season. The results of the 10-km nest are slightly improving the 30-km results of the WRF model. There is a large spread between the individual ENSEMBLES models in the lower end of the quantiles but the ENSEMBLES mean is performing quite well. In the upper end of the quantiles all ENSEMBLES models underestimate the observed temperatures leading to a larger general cold bias than the bias of the WRF simulations.
3.4.3 Regional differences of temperature
The lower two panels of Fig. 9 show the 0.05 and 0.95 quantile temperatures describing the extremely low and extremely high temperatures. Both WRF simulations are in good agreement with the observed extreme temperatures and outperform the ENSEMBLES mean or the ERA-40 reanalysis. There are no large differences between the regions in the extremely high temperatures and the 10-km WRF simulation is giving the best results. The models vary more in reproducing the extreme low temperatures. All models and the ENSEMBLES mean perform well in the southern regions 4–6 but fail to reproduce the extremely cold temperatures of below −30°C observed in the northern regions, as discussed in the previous section.
3.4.4 Extreme values of temperature
Generally the agreement between the modelled and observed excesses is satisfying. The WRF simulations fail to reproduce the extremely low temperatures as we have seen in Fig. 8. The 10-km simulation is improved from the 30-km simulation but still lacking the extremely cold temperatures. The ERA-40 and all of the ENSEMBLES models have a better agreement with the observed distribution than the WRF simulations. A few of the ENSEMBLES models produce too cold extreme temperatures and give the distribution of the ENSEMBLES mean too long a tail but the shape of the distribution is correct.
The situation is changed in the case of the extremely high temperatures. The WRF simulations are reproducing the distribution of the observed temperatures reasonably well. There is almost no difference between the 30 and 10-km results of the WRF model. The WRF model is overestimating the 0.95 quantile temperatures whereas the ENSEMBLES models are underestimating them. The error on both sides is approximately as large. These differences reflect the overall shift towards cold temperatures of the ENSEMBLES models compared with the WRF simulations.
3.5 10-m wind speed
The winds are generally well simulated or slightly too low (in the order of 1–2 m/s) on the coast and overestimated (up to >50%) in the inland stations in all models (not shown). The mean statistics show that all models are very similar and that refining the horizontal resolution from 30 to 10 km does not make a significant difference. This is likely to be due to the land use data used in the models, which generally does not describe the Norwegian vegetation in high detail. In studies which concentrate on surface winds a higher horizontal resolution as well as use of a more detailed description of land use would be important.
The quantiles in the lower panel of Fig. 11 show that despite of the over- (under-) estimation of the low (high) winds the form of the histogram is reasonable. Only the lower quantiles (from 0 to 0.5) are significantly overestimated in the 10-km WRF simulation. Here we see a clear improvement of the 10-km nest in the WRF simulation. The spread of the ENSEMBLES models is large but the ENSEMBLES mean agrees very well with the observed quantiles.
4 Summary and discussion
Results are presented from a dynamical downscaling of the ERA-40 reanalysis, with the WRFV3.1.1 regional climate model, to 30 and 10 km resolutions for 1961–1990 in Norway. The results of 12 different regional climate model simulations from the ENSEMBLES project are also presented as a reference. We concentrate the analysis on surface variables on complex terrain: precipitation, 2-m temperature and 10-m wind speed and compare the model results with a large number of observations within Norway.
Figure 12 summarizes the general behaviour of all experiments analyzed within this study. The biases shown are deviations of the daily mean modeled values of the observations, averaged over all stations over the whole period of 1961–1990. Precipitation and wind biases are shown in percent and the temperature biases in degrees. We focus on the “mean” (0.5 quantile) and “extreme” (0.95 quantile for extremely high and 0.05 quantile for extremely low) values. The figure shows that there is large spread in the quality of the modeled precipitation and wind between the individual models. The WRF simulations perform comparably well and the value added by the refinement of the resolution to 10 km is obvious. The ENSEMBLES mean has low biases and only a few of the models are performing better. In case of temperature the WRF simulations have clearly lower biases than the individual ENSEMBLES models or the ENSEMBLES mean. Again, the 10-km simulation reduces the bias compared with the 30-km simulation.
The precipitation on the Norwegian coast is largely driven by advective systems. As opposed to the traditional setup for regional climate models downscaling the European climate we included the whole North Atlantic into the larger model domain and applied spectral nudging to keep the large scale circulation consistent with the driving data. This turned out to be advantageous in several ways. First, the phase of the precipitation events was improved from the ERA-40 indicating that the synoptic scale features were better resolved by the 30 km grid than in the reanalysis. Also the representation of extreme precipitation on the coast was much improved from the reanalysis, probably due to sharper gradients and better resolved fronts. Another advantage seemed to be the larger independence of the regional model compared with the driving data. The WRF simulations were able to reduce the large regional biases of surface temperature in the ERA-40 reanalysis which had been largely inherited by the ENSEMBLES simulations.
A relatively high horizontal resolution turned out to be important in complex terrain, such as the Norwegian coast and the mountains. The precipitation has a large orographic enhancement which was largely improved from the reanalysis by the WRF simulations. The orographic lifting in the 10-km simulation was stronger and better resolved than in the 30-km simulation which also lead to an improvement of the representation of the extreme precipitation events, especially in the mountains. We conclude that the use of a horizontal resolution of 10 km, or higher, is preferable for producing climate projections, especially for impact studies dealing with extreme precipitation.
The fact that the precipitation and coastal winds are improved on a higher resolution grid is a consequence of a better representation of topography and coastline. This is in accordance with the general findings from several regional climate model studies (Rummukainen 2010). It also has to be kept in mind that precipitation of the ERA-40 reanalysis is a pure model product but temperature and winds are more constrained by the observations which improves the agreement. Also the fact that temperature and winds from the ERA-40 reanalysis are input fields for the WRF model, but precipitation not, explains why the differences between the temperature and winds of WRF runs and ERA-40 are smaller than for precipitation.
This study was the first application of the WRF model to climatological simulations in Europe. Generally the WRF model performed very well in reproducing the observed climate in Norway. The default setup of physical schemes in the WRF model turned out to be a suitable approach in climatological studies keeping the runtime low but producing results similar to the more sophisticated schemes. Spectral nudging proved to be a very useful method in these simulations where the outer model domain was large. The phase of precipitation and temperature was significantly improved in the nudged runs compared with the free runs (not discussed in this paper) and the simulated extreme values of precipitation were more realistic. This model configuration is useful for downscaling of GCM future predictions and the high-resolution data set created provides input for further downscaling and impact studies.
This work has been financed by the Norwegian Research Council through the SFF grant to the Bjerknes Centre for Climate Research and by the Municipality of Bergen through the MARE project. The authors acknowledge the computer resources and support by the Parallab at the Uni Research in Bergen, Norway. The ENSEMBLES data used in this work was funded by the EU FP6 integrated project ENSEMBLES (contract number 505539) whose support is gratefully acknowledged. This is publication no. A-307 from the Bjerknes Centre for Climate Research.
This article is distributed under the terms of the Creative Commons Attribution Noncommercial License which permits any noncommercial use, distribution, and reproduction in any medium, provided the original author(s) and source are credited.