Simulating surface wind over complex terrain is a challenge in
regional climate modelling. Therefore, this study aims at
identifying a setup of the WRF model that minimizes systematic
errors of surface winds in hindcast simulations. Major factors of
the model configuration are tested to find a suitable setup: the
horizontal resolution, the PBL parameterization scheme and the way
WRF is nested to the driving dataset. Hence, a number of sensitivity
simulations at a spatial resolution of 2
Prominent features of the North Atlantic and European climate are
cyclonic disturbances, which may be intensified and lead to severe
storms
The fundamental problem regarding surface wind is its intrinsically
complex nature, particularly over areas of complex terrain like the
Alps
RCMs, however, contain various sources of uncertainties, like
deviations in the driving dataset, numerical approximations, as well
as parametrizations of the sub-grid processes. A number of studies in
different locations assessed the sensitivity of the model performance
due to different model configurations.
Although the studies discussed above tackle the problem of the
uncertainties in the model configuration regarding wind, they do not
focus on areas of complex topography. As suggested by
The study is structured as follows: Sect. 2 presents the reanalysis product used to drive the RCM and the observational network. Section 3 describes the model setup including the different nesting options tested in this study. It further presents the set of sensitivity experiments carried out. The results are discussed in Sect. 4 focusing first on the role of the PBL scheme and the nesting method applied. Then, the role of the horizontal resolution is discussed, including how errors are spatially distributed over different areas of the Alps. Finally, Sect. 5 draws main conclusions.
The dataset providing the initial and boundary conditions for the RCM
is the ERA-Interim reanalysis
To evaluate the model's ability in dynamically downscaling wind storms, a reliable set of observations is required. In particular, this is the case in areas of complex terrain, where wind speed and direction can vary within distances of tens of meters. The Swiss Federal Office of Meteorology and Climatology (MeteoSwiss) provides such observations from a dense network of weather stations. This dataset contains 10 min mean values of wind speed and direction. The model simulations are evaluated hereafter by using hourly means of weather station wind measurements calculated from their 10 min mean.
Some basic data checks are carried out before using the data in the
evaluation. Following an approach similar to
The study is based on the Weather Research and Forecasting Model (WRF,
version 3.5), aired in September 2013
A first decision in regional climate modelling concerns the selection
of the domain to be simulated. Although this selection is susceptible
of introducing uncertainties, this study employs just one domain
setup, and hence the sensitivity of the performance to the model
domain is not investigated here. There are a number of reasons for
this. First, there is not much freedom, in the sense that the domain
is primary selected according to the area of interest, in this case
the Alpine area. The number of domains is conditioned by the
resolution of the driving data set and the final resolution of
2
Another source of uncertainty is related to the choice of the physical
parameterizations, such as microphysics, convection, radiation and the
formation of the PBL, among others
The PBL plays a major role in simulating surface winds
RCMs are nested in a global dataset, which drives the simulation by
providing the initial and boundary conditions. Dynamical downscaling
is hence mostly an initial value problem in the first days of the
simulation, which evolves into a boundary value problem when the
initial state has been “forgotten” by the atmosphere. However, how
to specify the lateral boundary conditions is a mathematically
ill-posed problem, since they become over-specified
The first approach basically consists of using Newtonian relaxation at the boundaries without any correction inside the domains. This is referred hereafter as “free simulations”. In favour of this approach, it is argued that simulations benefit from a better representation and the undisturbed development of regional processes. Another argument is that RCMs are often used to downscale climate change projections or paleo-simulations. Such simulations are performed with relatively coarse GCMs, so that modifications of the large-scale circulation maybe beneficial, as potential biases from the GCMs may be partly corrected by the RCMs.
In case of reanalysis data used at the boundaries, it may be desirable
that the RCM simulation stays close to the large-scale situation of
the driving data. A first method to achieve this is the so-called
reforecast simulation. The method consists of splitting a long
simulation in shorter simulation periods of one to few days, running
each period separately and finally merging them. This method
effectively minimises the impact of the boundaries, transforming the
problem into a mostly initial-value problem. The reforecast method is
regularly applied
A more sophisticated method is to force the RCM to follow the driving
large-scale conditions. This is implemented by additional terms in the
dynamic equations that restrict the degrees of freedom of the
simulation. This is the so-called nudging nesting, of which two
versions are available. The 3-D analysis nudging introduces
a Newtonian relaxation term in the prognostic equations of the model,
and was first introduced by
A variation of this method is spectral nudging, introduced by
This section summarises the set of simulations carried out to
investigate the sensitivities of the different settings. Following the
approach by
The comparison of observations and simulation results is performed for
hourly values at each observational site. For this, the simulation
result at the closest grid point to the observational site is
selected. Although this can lead to representativeness errors
To evaluate the sensitivity of the model result due to different PBL
schemes, the setups C1 to C4 are compared with each other
(Table
Figure
The surface winds over Switzerland during storm Lothar are presented
in Fig.
Figure
The temporal metrics (shown by boxplots) resemble the findings of the
time series in Fig.
The spatial metrics show that the biases behave similar to the ones of
the temporal scale (Fig.
Next, the sensitivity of the model to the PBL scheme is assessed with
respect to wind direction. Thereby, the wind rose of the storm Lothar
is shown (Fig.
Most of the conclusions drawn from the analysis of the storm Lothar
about the PBL schemes are consistent through the various storms
simulated. This is illustrated in a comprehensive although summarised
way in Fig.
Wind direction performance across all storms is analysed in a similar
fashion. However, this variable has to be treated differently, taking
into account the problems associated to its circularity. Thus,
similarly to
To assess whether the results of wind direction and the minor role of
the PBL scheme may depend on the storm selected, wind roses of three
additional storms are shown in Fig.
The analysis carried out in the former section indicates that the
YSU
Figure
The analysis of wind direction delivers similar results as in the
sensitivity to different PBL schemes (Fig.
As before the analysis is extended to all 24 storms. The mean temporal
correlation obtained for different storms is shown in
Fig.
As argued above, the horizontal resolution has a profound impact on the ability of the model to simulate wind speed. In particular this is the case if the closest grid point of the model to the weather station is used in the analysis. We note that this simple approach neglects the fact that the model averages subgrid terrain properties, and leads to so-called representativity errors. It is beyond the scope of this study to assess these errors and to address a method to minimise them, since they introduce systematic biases that only depend on the domain configuration, which is fixed across simulations, and thus play a secondary role in the evaluation of the relative skill of different model configurations. Still, such errors, and the model performance in general, depend on the model resolution, so the importance of model resolution and the type of station is discussed in more detail.
The representativity error is quantified by calculating the horizontal
distance
The influence of the horizontal resolution on the model performance is
investigated using the C6 configuration as an example
(Fig.
For temporal metrics, a somewhat unexpected behaviour is
found. Although the temporal correlation drops to a median value of
zero in the coarsest domain analysed, the model exhibits a remarkable
high correlation in the 6
The role of the representativity error is explored through the
separation of the observational sites in subcategories such as
stations in plains, mountain or valleys, as shown in
Fig.
The spatial correlation in the innermost domain shows a low value of 0.31 over the plains, which contrasts with the value of 0.78 obtained for mountains. This can be explained as a signal-to-noise artefact. The problem is that in the plains the mean wind is not as strongly modulated by height as it is in mountains, where there is a larger difference among stations. Thus, small variations in mean wind lead to large variations in the spatial correlation, since the mean wind speed is not a good predictor of the location of a station within plains. Additionally, the correlation is calculated according to only the 46 stations that corresponds to the plains in the Lothar storm. Such low number leads to a large variance of the estimator of correlation, which further contributes to the signal-to-noise problem. Thus, the spatial correlation of mean wind patterns over homogeneous terrain is not a meaningful measure of model skill and should be treated with care.
This paper analyses a number of sensitivity experiments aimed at
identifying a model setup for WRF that minimises systematic errors in
hindcast simulations of wind over areas of complex topography. The
simulations use the Era-Interim reanalysis for initial and boundary
conditions. These data are downscaled to a resolution of 2
The sensitivity tests designed to evaluate the role of the PBL
parameterization show that WRF systematically overestimates wind speed
compared to observations. The overestimation occurs in all types of
location (plains, valleys or mountains), and is exacerbated in
coarser domains. This result confirms previous studies pointing out
the overestimation of wind speed in simulations with WRF and its
relation with unresolved topograhy
The model is qualitatively able to reproduce the leading wind directions generated by very different synoptic conditions. However, the simulations still exhibit systematic biases in wind direction that cannot be improved through a suitable model configuration. Generally, the model performance in reproducing wind direction exhibits little sensitivity to all the evaluated model configurations. Thus, the model performance is dominated by other factors such as the driving conditions, insufficient resolution, or representativity errors.
Additionally, the sensitivity with respect to the nesting technique is
explored by comparing free simulations to analysis and spectral
nudging, as well as the so-called reforecast approach. The use of
nudging techniques slightly improves several aspects of the
simulation, like reducing the mean wind overestimation discussed above
and improving the spatial pattern of mean wind (in particular 3-D
analysis nudging). Further, the free simulations generally show
a lower temporal agreement with observations than nudged simulations,
a feature that is consistent across storms. Analysis nudging yields
a significant improvement for maximum wind speed, for which the
overestimation is reduced and leads to values closer to zero on
average than when no nudging is applied. These results indicate that
preserving the large-scale circulation via nudging slightly improves
the simulation of wind at regional scales, at least for hindcast
simulations where the driving dataset is generally reliable, and whose
aim is to be as close to the observations as possible. We note however
that for other scientific questions a free simulation setup could be
more appropriate, as atmospheric processes and their interactions with
regional scale features are able to develop desirable disturbances
that add value to RCM simulations. Typical examples are climate change
projections
Using the setup with analysis nudging and the YSU
In summary, this study suggests two setups depending on the scientific
question: (i) the configuration C6 with the YSU
The authors are grateful for the support provided by the Oeschger Centre for Climate Change Research and the Mobiliar lab for climate risk and natural hazards (Mobilab). Thanks are due to the Swiss National Supercomputing Centre (CSCS) for providing the supercomputing facilities and technical support required to perform the RCM simulations. The ERA-interim reanalysis was kindly provided by the ECMWF. Finally, we would like to thank Martina Messmer for sharing her great knowledge of the Swiss geography and the stimulating discussions.
List of 24 historical wind storms and the prevailing synoptic flow conditions. This list is adapted from Table 1 in
Summary of the eight model configurations used in the sensitivity studies.
Representativity error in different model domains. The mean
and standard deviation of the horizontal distance
Network of observational sites for wind speed and direction
run by MeteoSwiss. The orography of the area is illustrated by the
color shading, whereas each symbol indicates the location of an
observational site. The filling colour of the symbols indicates the
number of storms (Table
Configuration of the four two-way nested domains. The spatial
resolutions are 54, 18, 6 and 2
Synoptic situation of the storm Lothar: colour shading
depicts the sea level pressure, whereas blue contours indicate the
geopotential height at 500
Time series of wind speed for a 6-day period around storm
Lothar averaged for 109 stations. Thick black line depicts the
series corresponding to the observations, whereas the coloured lines
correspond to the simulation results with different model setups
(Table
Different skill metrics for the comparison of observations
and simulation results for storm Lothar. Each column represents the
results of one sensitivity simulation in Table
Wind roses corresponding to four different storm cases
selected from Table
Model performance for the comparison of observed and
simulated wind speed across the 24 storms defined in
Table
Performance metrics for wind direction for all storms. Left
(right) panel shows the model skill evaluated through the median
(RMSE) of
Influence of the grid size for the simulation skill, based on
the configuration C6 (Table
Site-averaged series of wind speed in the 6-day case study
containing the storm Lothar. Black and yellow lines correspond to
the observations and simulation in the domain D4 corresponding to C6
in Table