The first study area, the WegenerNet FBR (indicated by the orange-framed white rectangle in the middle panel of Fig. , enlarged in a) lies in the Alpine foreland of southeastern Styria, Austria, centered near the town of Feldbach (46.93∘ N, 15.90∘ E). It covers a dense grid of 155 meteorological stations within an area of about 22 km × 16 km, in a hilly terrain, characterized by small differences in altitude . The typical difference in altitude between the valleys and the crests is about 100 m and the highest peak is the Gleichenberger Kogel, with an elevation of 598 m.

This region, with a more Alpine climate at the valley floors and more Mediterranean climate along hillsides, is quite sensitive to climate change . Furthermore, it exhibits rich weather variability, especially through strong convective activity and severe weather in summer . The wind fields in this study area are characterized by thermally induced local flows and influenced by thermally driven regional wind systems with weak wind speeds, caused by a dynamical process called Alpine pumping . Furthermore, nocturnal drainage winds, which lead to cold air pockets, are relevant for this region, which is dominated by agriculture. Especially in fall and winter, the nocturnal cold air production is amplified by temperature inversions in relation to high-pressure weather conditions. In the WegenerNet FBR, hillside locations are thermally preferred to valley locations at night. Results related to the WPG-diagnosed empirical wind fields in the WegenerNet FBR can be found in .

The second study area, the WegenerNet JBT (indicated by the gray-framed white rectangle in the middle panel of Fig. , enlarged in b), is named after the Johnsbachtal river basin (location of the village of Johnsbach: 47.54∘ N, 14.58∘ E) and situated in the eastern Alpine region, in the Ennstaler Alps and the Gesäuse National Park, in northern Styria, Austria. The terrain of this mountainous region is characterized by large differences in elevation. The Hochtor, with an elevation of 2369 m, is the highest summit, and the valleys lie roughly at a height of 600 to 800 m . This region spans an area of about 16 km × 17 km and comprises 11 irregularly distributed meteorological stations including two summit stations at altitudes of 2.191 and 1.969 m .

The climate is Alpine with mean annual temperatures of around 8 to 0 ∘C and an annual precipitation of about 1.500 to 1.800 mm from the valley to the summit regions . Typical for this region are thermally induced local flows and westerly flow synoptic weather conditions. Details related to initial studies and their results as well as to the cooperation and partnerships can be found in and in their most up-to-date form in . Recently, computed and evaluated WPG-generated empirical wind fields in this region.

The data acquired from the two WegenerNet regions FBR and JBT are automatically quality controlled and processed by the WegenerNet Processing System (WPS), consisting of four subsystems : the Command Receive Archiving System transfers raw measurement data via wireless transmission to the WegenerNet database in Graz, the Quality Control System checks the data quality, the Data Product Generator (DPG) generates regular station time series and gridded fields of weather and climate products, and the Visualization and Information System offers the data to users via the WegenerNet data portal (https://wegenernet.org/portal/, last access: 2 July 2019).

Besides weather and climate time series, based on a spatial interpolation of the station observations, the DPG generates gridded fields of the variables temperature, precipitation, and relative humidity for the WegenerNet FBR. These gridded products of the WegenerNet FBR are available to users in near-real time with a latency of about 1–2 h. and provide detailed information about the subsystems of the WPG.

The DPG furthermore includes a newly developed wind field application, the WPG, as briefly introduced in Sect. . The WPG provides high-resolution wind fields for the WegenerNet FBR as well as for the WegenerNet JBT. The WPG uses the freely available empirical California Meteorological Model (CALMET) as core tool and generates, based on meteorological observations, terrain elevations, and information about land use and mean wind fields at 10 and 50 m height levels with a spatial resolution of 100 m × 100 m and a temporal resolution of 30 min, again with a maximum latency of about 1–2 h. In order to keep the meteorological input data of the WPG independent from the data pertaining to the other operational station networks, observations from the Central Institute for Meteorology and Geodynamics (ZAMG) stations (violet stars in Fig. a and b) and other external stations are not used as WPG input. For the WegenerNet FBR, the gridded wind fields are available starting in 2007 and for the WegenerNet JBT starting in 2012. The wind fields at 10 m height level are used for the model intercomparisons.

The CALMET model is a diagnostic model that omits time-consuming integrations of nonlinear equations, such as the governing equations of dynamical models . It is hence not capable of simulating dynamic processes such as flow splitting and grid-resolved turbulence or of delivering prognostic information. Specific parameterizations allow the model to empirically take into account conditions such as the kinematic effects of terrain, slope flows, and terrain-blocking effects . We enhanced the model by implementing methods developed by to as well take into account topographic shading through relief, topographic slope and aspect, and the sun position for the estimation of solar radiation. In addition, the modeling of temperature fields is now based on vertical temperature gradients, calculated from meteorological station observations located at different altitudes, and the influence of vegetation cover is taken into account. Details about these advanced algorithms can be found in .

The quality of the generated wind fields depends above all on the quality and the spatial and temporal resolution of the meteorological observations and surface-related datasets which are used as model input . A detailed description of the WPG application and the statistical results for the WegenerNet FBR can be found in . More information regarding statistical results related to the WegenerNet JBT as well as information regarding evaluation results from 5-year climate data of the WegenerNet JBT in comparison to 9-year climate data from the WegenerNet FBR can be found in .

The INCA system has been developed at the ZAMG in Vienna, Austria, to provide realistic analyses and nowcasts of quantities of several meteorological variables for the highly mountainous and the overall complex terrain of Austria. In the case of the variable wind, the system operationally generates spatially distributed analysis wind fields in 3-D and for 10 m height above ground with a horizontal grid spacing of 1 km × 1 km and a temporal resolution of 1 h.

The basic idea of the INCA wind module is to statistically correct a numerical weather prediction (NWP) model first guess (i.e., in operational mode the latest available NWP model run) with the latest observational data, which do not enter the NWP data assimilation. Thus, the skill of the INCA analysis depends on the station density, their representativeness, and the spatial distribution of station observations, as well as on the skill of the NWP model providing the first guess. The impact of the NWP model on the analysis skill is further discussed in .

In this study, NWP model outputs used as a first guess for INCA were generated by the spectral ARPEGE-ALADIN (ALARO) model in a revised version . ALARO has a horizontal grid spacing of 4.8 km × 4.8 km (600×540 grid points) and includes 60 vertical layers up to the 2 hPa level (about 43 km altitude), covering central Europe, eastern France, and the northern part of the Mediterranean Sea. It is run with a temporal resolution of 180 s using a hydrostatic semi-implicit semi-Lagrangian dynamical solver and the ALARO-0 physics package, which includes the 3MT microphysics-convection scheme , the Interaction Soil Biosphere Atmosphere (ISBA) force restore 2 L soil scheme , and the Actif Calcul de Rayonnement et Nebulosité (ACRANEB) radiation scheme . Soil temperature and moisture are initialized by a 6 h-cycle optimal interpolation data analysis taking into account the latest ALARO forecast as a first guess and 2 m relative humidity and temperature observations from synoptic (SYNOP) and national stations. The 2 m values are transferred to soil variables via empirical relations . To reduce initial spin-up, a digital filter initialization is applied.

The model gets its lateral boundary and atmospheric initial conditions from the high-resolution deterministic operational global integrated forecast system (IFS) of the European Centre for Medium-range Weather Forecasts (ECMWF) model in lagged mode (i.e., ALARO 00:00 UTC is linked to IFS 18:00 UTC of the day before, ALARO 06:00 UTC to IFS 00:00 UTC, etc.). This is due to the rather late availability of the IFS data. Coupling is achieved by one-way nesting via Davies relaxation . Sea surface temperature is interpolated from the deterministic IFS model to the ALARO grid. More details about ALARO development and configurations can be found in .

The INCA wind fields have already been evaluated for a moderately hilly region in the north of Austria (47.78–49∘ N, 13.8–17∘ E), where the wind analyses show significantly higher errors compared to the statistical results from other meteorological variables . These higher errors mainly root in the limited representativeness of station data, as well as on the low station density, which can be only partly compensated for by INCA’s analysis algorithms .

In the WegenerNet FBR area, INCA assimilates observations from the ZAMG Feldbach and Bad Gleichenberg stations (violet stars in Fig. a) to the NWP’s first guess, and in the WegenerNet JBT region, observations from the ZAMG Admont station are used. However, data from WegenerNet FBR and JBT stations are not used in INCA data assimilation and hence the WPG fields can be used for independent evaluation .

The coordinate system of the INCA datasets is transformed into WGS84/UTM zone 33∘ N coordinates. Furthermore, we resampled the wind fields at 10 m height levels from the INCA gridding onto the WegenerNet FBR and WegenerNet JBT grids, using a bilinear interpolation method. Based on extensive sensitivity tests regarding different interpolation methods, we concluded that the statistical results do not significantly depend on the interpolation method.

Regarding the CCLM , we use available wind fields generated with the model version 5.0. These wind fields were generated during the course of a previous study and are available for the period 2008 to 2010. The data have a comparatively coarse horizontal resolution of 3 km × 3 km on an hourly basis that is nevertheless the highest resolution available to this study. This limited resolution leads to a smoothed orography, which may result in different wind patterns with errors in wind speed or direction. Furthermore, the winds may be displaced by an incorrect position of the topographic slopes . The CCLM fields are provided for 40 vertical levels. The first level is simulated for 10 m above ground and the last level corresponds to the 100 hPa level (about 16 km altitude), whereby the vertical resolution is higher in the boundary layer and decreases towards the top level.

CCLM is a non-hydrostatic model with a Runge–Kutta dynamical core, which makes use of a third-order scheme with diffusion damping to discretize the advection term in the compressible Euler equations . In order to avoid numerical instability, the model's orography is additionally smoothed via a 10th-order filter. The vertical coordinate system is a terrain-following, time-invariant Gal-Chen pressure-based sigma coordinate . Deep and shallow convection are parameterized following , and turbulence is parameterized based on Mellor and Yamada . Vertical mixing comes from a prognostic formulation of the turbulent kinetic energy (TKE) with a 2.5 closure that accounts for grid- and subgrid-scale water and ice clouds. It uses a statistical cloud scheme for cloud cover and cloud water content (so-called Gaussian closure scheme). Horizontal diffusion follows the Smagorinsky approach.

Land cover data for CCLM are based on the Global Land Cover 2000 project from SPOT4 satellite products . In the model setup used (3 km resolution), deep convection is resolved explicitly, which means that parameterization for deep convection was switched off. Shallow convection is still parameterized. In climate research, such simulations are referred to as convection-permitting climate simulations (CPCSs) . To minimize decoupling effects from model-internal variability, which usually occurs in large model domains and if nudging techniques are not used , CCLM is operated in a small domain encompassing the greater Alpine region, and it is also driven by ECMWF's IFS. The data assimilation system IFS includes a wide range of observations and is assumed to provide perfect boundary conditions with a horizontal grid spacing of about 25 km at midlatitudes and at 91 vertical levels . Every 6 h (00:00, 06:00, 12:00, 18:00 UTC) of the IFS data is an analysis field from the assimilation system, and every alternate 6 h (03:00, 09:00, 15:00, 21:00 UTC) is a short-range forecast field. This procedure has already been used by and keeps the modeled synoptic patterns in agreement with the observed ones.

In the course of the data preparation for the study, CCLM data at 10 m height level were also transformed into WGS84/UTM zone 33N coordinates, resampled and mapped onto the high-resolution WPG grid, and checked for sensitivity with respect to the interpolation method which was also found to be weak (see Sect.   above).

The WegenerNet, INCA, and CCLM wind fields are intercompared for two representative types of wind events: thermally induced wind events and strong wind events. For this purpose, we defined eight evaluation cases, four for each of the two study areas (Table ). For the cases shown in Table  we use the WegenerNet data as a reference, except for evaluating the CCLM wind fields for the WegenerNet JBT. The reason for this is that the CCLM data used in this study are available from January 2008 to December 2009, but the WegenerNet JBT data are available only as of January 2012, since this latter network has been sufficiently completed for long-term monitoring only since 2012 (Table , cases CCLMvsINCA_therm_JBT and CCLMvsINCA_strong_JBT).

In both study areas, autochthonous weather conditions mainly lead to thermally induced wind systems, meaning that the wind fields are controlled by small-scale temperature and pressure gradients. These small-scale gradients lead to characteristic interacting systems of air motion, like slope winds and mountain-valley winds, and create complex everyday flow patterns. The autochthonous days are characterized by small synoptic influences, cloudless or nearly cloudless skies, low relative humidity and increased radiation fluxes between the Earth surface and the atmosphere . Due to frequently occurring temperature inversions in relation to clear-sky and high-pressure weather conditions in winter, which often lead to a stable atmospheric stratification in the whole WegenerNet FBR and in the valley regions of the WegenerNet JBT, autochthonous days are only selected from spring, summer, and fall (March to October).

The automatic selection of thermally induced and strong wind events is done based on thresholds, which we defined based on sound physical and careful sensitivity checks summarized in Table . For the estimation of autochthonous days, we compared the observed daytime mean values of wind speed (v) and relative humidity (rh) as well as the nighttime mean values of net radiation (Qn) from selected stations with the respective thresholds. A further criterion for the selection of such days is high daily global radiation, which indicates fair weather conditions. For this purpose, we compared the daily mean modeled global radiation (Qg,m) for clear-sky conditions with the observed daily mean net radiation (Qn,o) for the WegenerNet FBR and with the observed daily mean global radiation (Qg,o) for the WegenerNet JBT at defined station locations (Table , reference data). The reason for the comparison of Qg,m with Qn,o for the WegenerNet FBR is that this station network includes no global radiation sensors. Due to the almost linear relationship between the daytime Qg,o and Qn,o for clear-sky conditions, we find that the same selection method can robustly be applied to both study areas by defining different thresholds (Table , Qg,m–Qn(g),o). If all criteria are fulfilled for a given day, the data from the entire day are added to the thermally induced wind event dataset, leading to 24 h events (i.e., 24 h mean wind speed values).

The modeling of the global radiation is done based on ESRI's ArcGIS Area Solar Radiation Tool. This tool is designed for local landscape scales and derives the incoming solar radiation based on a digital elevation model. Small differences of daily mean values between Qg,m and Qn(g),o indicate fair weather conditions and high global radiations during the day. If all criteria are fulfilled for a given day, the data from that day are included in the thermally induced wind events dataset.

The strong wind events, caused by synoptic weather conditions such as cyclones and frontal system at a larger scale, are selected on an hourly basis from preselected days, by comparing hourly mean values from gridded reference datasets with defined minimum wind speeds. These preselected days were estimated by comparing the daily average wind speed from the gridded datasets with a defined minimum average wind speed (Table , v‾ and v for strong wind speed cases). If the hourly mean value of this reference dataset is larger than the defined minimum wind speed, the data of this reference dataset and the corresponding model dataset are used as part of the hourly event data for evaluating strong wind events.

In order to account for spatial differences and displacements between the model and the reference data and to analyze wind speed and direction in a combined way, we apply a novel wind verification methodology. This methodology extends the fractions skill score (FSS), a spatial verification metric developed by , which is classified as a neighborhood-based approach and originally used for verifying precipitation. The FSS is based on the assumption that a model is useful when the model data and the corresponding reference data show a similar spatial frequency of precipitation events, which alleviates the requirement of the models to predict the events at exactly correct positions, which is an unduly strict assumption. Furthermore, this metric avoids the double-penalty problem and provides scale-dependent information about the level of model skill .

In order to obtain information on how the skill of a model varies with spatial scale, the FSS is calculated for different neighborhood sizes. A neighborhood size of n defines a square consisting of n×n grid points; i.e., it denotes the side length of the square (e.g., for n=5, the square contains 25 grid points). These squares of defined neighborhood sizes are moved as sliding windows over the datasets and centered successively at each grid point, whereby the area outside the domain is assumed to contain no wind class. In terms of the FSS value, it is a 0-to-1 normalized metric; i.e., the lower limit of the FSS is 0 and the upper limit +1, with values approaching +1 representing a better degree of model performance.

The extended version of the FSS is called the wind fractions skill score, denoted WFSS, which has been developed by . The score is calculated based on user-defined wind classes. The definition of these classes is partly subjective and can significantly affect the WFSS. The wind vector field should be classified in such a way, that the definition reflects what a user wants to analyze. For example, a complex terrain leads to strong changes in wind directions; therefore, it is reasonable to define smaller class intervals regarding the wind directions. For upper-level winds the focus could be more on the magnitude of wind speed.

We defined eight wind direction classes with an interval size of 45∘ for a range of three wind speed categories, as shown in the wind roses of Fig. . Wind speeds <0.5ms-1 were classified as calm, independently of the wind direction. The small interval size of 45∘ was chosen to be able to capture the varying wind directions in the study areas, especially for thermally induced winds. Because of the generally much lower wind speeds in WegenerNet FBR, we defined a smaller interval size of the wind speed categories for this region (Fig. a and b, lower panels) than for the WegenerNet JBT (Fig. c and d, lower panels).

Table  includes the equations for the calculation of the WFSS and the asymptotic WFSS (AWFSS). This AWFSS will always be reached for a neighborhood size ≥2N-1, where N is the number of grid points of the largest domain size. At such a large neighborhood size, the estimated fractions within the domain are the same at all locations, and further enlarging this size will not affect the WFSS. A bias always leads to an AWFSS value <1, which indicates systematic differences in the frequency of wind classes between the model and the reference wind classification.

The WFSS is calculated for each hourly field for the selected events. The event-averaged score values are calculated based on averaging these 1 h event WFSS values over all the hourly events within the analyzed multiyear period, for each evaluation case listed in Table .

As briefly explained above, the chosen thresholds of the classification and the number of classes influence the score. We found, in particular, that the wind direction thresholds can have a strong impact on the score values. For example, a small change in the wind direction value from prevailing northwesterly winds, which are close to a threshold to distinguish between W and N, could dramatically change the WFSS value. Such a small error in the model data could indicate a poor modeling performance, whereas a human analyst would assess the forecast as reasonably good. To avoid this problem we calculated the hourly WFSS for every rotation between 0 and 45∘ with an interval size of 5∘ (nine trail classes), in addition to the original class definition. As a next step, the maximum values of the hourly score values at each neighborhood size are always used for computing the final case-averaged score values. We applied this approach for 7597 selected events in total (Table , number of events) and estimated the case-averaged score value for each of the eight evaluation cases. A more detailed description of the wind fractions skill score metric can be found in .

Furthermore, we applied traditional grid-point-based statistical performance measures such as bias, root-mean-square error, and others, to each evaluation case. All statistical performance metrics used in this study are summarized in Table .

Figure a–d illustrate typical examples of modeled wind fields (upper rows of panels) and the corresponding wind roses of relative frequency of wind directions divided by wind speed categories (lower rows of panels) from selected representative evaluation events. Each panel depicts the modeled and the associated reference data for the WegenerNet FBR (Fig. a and b) and the WegenerNet JBT (Fig. c and d), for thermally induced (Fig. a and c) and strong wind events (Fig. b and d). Figure e shows the WFSS values of these examples, estimated as explained in Sect.  above.

The thermally induced wind event on the 29 July 2009 from 16:00 to 17:00 UTC for the WegenerNet FBR (Fig. a) shows thermally driven regional flows caused by Alpine pumping. This flow is called “Antirandgebirgswind”, which arises usually in the afternoon as southerly wind with maximum wind speeds of about 2.5ms-1. The Antirandgebirgswind is a compensating flow between the bordering mountains of the eastern Alps, and the hilly countryside region of southeastern Styria (called Riedelland), which comprises the WegenerNet FBR . The INCA (upper-left panel of Fig. a) and the WegenerNet wind fields (upper-right panel of Fig. a) show a similar distribution with generally low wind speeds and prevailing southerly wind directions. The intercomparison of these INCA data with WegenerNet data for this event shows the largest WFSS values for all neighborhood sizes, which indicates a good overlap of the wind classes (Fig. e, INCAvsWN_therm_FBR). Furthermore, it shows a nearly perfect asymptotic value of about 0.99. This large AWFSS indicates a very small bias, which is also reflected by the similar wind classification results (lower-left and lower-right panel of Fig. a).

The CCLM wind field shows similarly low wind speeds, compared to the WegenerNet wind field, but a shift in wind directions from the S sector mainly to the SE and partly to the E and NE sectors between the CCLM and the WegenerNet data can be observed (lower-middle and lower-right panel of Fig. a). This shift is reflected by small WFSS values at all neighborhood sizes, especially below a scale of 10 km. The AWFSS shows the largest value of all CCLM intercomparisons but is still low compared to INCA evaluation cases, which indicates a large bias (Fig. e, CCLMvsWN_therm_FBR). Evidently, this regional climate wind field modeling at 3 km grid spacing is not adequately representative for the given challenging hilly terrain.

The strong wind speed event in the WegenerNet FBR on the 30 October 2008 from 10:00 to 11:00 UTC led to southwesterly to southerly winds (Fig. b). The INCA model data and the WegenerNet reference data show maximum 1 h vector-mean wind speeds of around 9–10ms-1 (upper-left and right panel of Fig. b). Regarding the wind directions, differences in the wind sectors can be observed (lower-left and right panel of Fig. b). The INCA data show wind directions mainly from the SW sector (lower-left panel), while the WegenerNet data show wind directions from the S and SW sectors (lower-right panel). The WFSS for this case shows small values at small neighborhood sizes and increases with increasing neighborhood size (Fig. e, INCAvsWN_strong_FBR). These low WFSS values are mainly caused by the differences in wind direction classes, especially in the southern part of the domain and through some spatial displacements in wind speed classes. Despite low WFSS values at small neighborhood sizes caused by differences in wind sectors, the AWFSS shows a high asymptotic value (AWFSS>0.97). This high value is caused by the prevailing wind directions in the WegenerNet data, which are close to the threshold values for distinguishing between S and the SW. Hence, in this case, the 5∘ azimuthal class rotation procedure avoids lower score values.

Regarding the CCLM data (lower-middle panel of Fig. b), the whole wind field shows wind speeds from about 6.5 to 7.5ms-1 and is therefore assigned to the wind class with wind speeds higher than 6ms-1, whereas, for the WegenerNet wind fields, a large proportion is assigned to the class with wind speeds from 3 to 6ms-1 of this region (Fig. 2e, CCLMvsWN_strong_FBR) and indicates that the dynamically modeled CCLM wind speeds are systematically overestimated relative to the empirically diagnosed wind speeds.

This discrepancy leads to the smallest WFSS values at all neighborhood sizes for this region (Fig. e, CCLMvsWN_strong_FBR) and indicates that the dynamically modeled CCLM wind speeds are systematically underestimated relative to the empirically diagnosed wind speeds for this hourly event.

On the 1 August 2012 the winds were thermally driven and the local pressure and temperature gradients were causing varying wind speeds and wind directions in the WegenerNet JBT. This is illustrated for the late afternoon INCA and WegenerNet wind fields in the upper-left panels of Fig. c. The WFSS for the evaluation of the INCA wind field shows the second largest value at the 1 km neighborhood size, which indicates overlapping areas at this neighborhood size, equal to the horizontal resolution of the INCA analysis (Fig. e, INCAvsWN_therm_JBT). The large AWFSS value again indicates a small bias, which is also reflected by the similar wind classification shown in the wind roses of the corresponding lower-left panels of Fig. c. The high asymptotic value (AWFSS>0.9) indicates a small bias and that the WFSS is mostly influenced by the spatial displacement.

In a further example of a thermally induced wind event on the 31 May 2008, we intercompare the CCLM with INCA wind fields (right panels of Fig. c). Especially in the CCLM, the smoothed terrain leads to uniform wind speeds and directions. Regarding the INCA wind fields, some variability in wind speed can be observed, with higher values in the summit regions and lower values at lower altitudes in the valleys of this region. Furthermore, a valley wind in the Enns valley is simulated by INCA. Probably the analysis part of the INCA model with its higher-resolved digital elevation model (DEM) and assimilated ZAMG observations leads to a somewhat better representation of the wind field. Comparing wind directions, the largest part of the CCLM modeled flow is from the N and NE sectors, while the INCA system estimated wind directions mainly from the E sector and partly from the NE and SE sectors (bottom right panels of Fig. c). This simplistic pattern of wind directions in CCLM leads to low WFSS values for all neighborhood sizes, including the lowest asymptotic value of all examples, indicating a very poor representation of the wind field by the dynamical modeling of the CCLM in this challenging mountainous terrain.

The strong wind speed event for the WegenerNet JBT on the 7 December 2013 is caused by northwesterly weather conditions. These synoptic-scale flow conditions led to strong wind speeds with maximum 1 h mean wind speeds of around 20ms-1 from 17:00 to 18:00 UTC. Both the INCA and the WegenerNet wind fields show wind directions mainly from the NW, with some proportions from the N and the W sectors, caused by a channeling of the air flow through the pronounced valleys of this study area. The INCA wind fields show much lower wind speeds in the valley regions compared to the WegenerNet wind fields, resulting from the observations of the ZAMG Admont (ADM) station that flow into the INCA analysis but are considered far off the area and not used by the diagnostic modeling (upper-left panels of Fig. d). As the neighborhood size increases, the WFSS also increases, but due to spatial displacements, the values are generally low (Fig. e, INCAvsWN_strong_JBT). The low AWFSS value is caused by the differences in wind speed categories (lower-left panels of Fig. d).

For the 5 November 2008 we intercompare the CCLM wind fields with INCA wind fields from 01:00 to 02:00 UTC, for a strong wind speed event (right panels of Fig. d). In this example, the influence of the smoothed terrain caused by the coarse horizontal resolution of the CCLM becomes obvious. This smoothed topography results in systematically lower wind speeds compared to the INCA wind fields. The WFSS shows similar results to the previous INCAvsWN_strong_JBT evaluation, with small values at all neighborhood sizes (Fig. e, CCLMvsINCA_strong_JBT), indicating the clear limits of the CCLM dynamical modeling fields also for strong wind events.

The statistical event-averaged WFSS values from the large ensemble of events over multiple years are represented for each evaluation case in Fig. . Overall, the WFSS values show a monotonic increase with neighborhood size for all cases so that the AWFSS is the largest value, indicating relatively the best performance at large scales.

For the WegenerNet FBR, the statistical WFSS values, calculated for the INCA wind fields compared to the WegenerNet wind fields, shows nearly the same behavior for both the thermally induced and strong wind events (Fig. a, INCAvsWN_therm_FBR and INCAvsWN_strong_FBR). The WFSS values for these cases indicate a reasonably good spatial matching at all neighborhood sizes. Furthermore, the AWFSS values are higher than 0.8, reflecting generally small INCA biases of wind classes.

The statistical WFSS estimated for the CCLMvsWN_therm_FBR case indicates that the CCLM clearly and systematically underperforms in the case of thermally induced wind events for the WegenerNet FBR. Evidently, due to the coarse horizontal resolution of the wind fields, the CCLM wind fields appear fundamentally unable to capture the varying wind directions for such events in this region. For the CCLMvsWN_strong_FBR case, however the results indicate a similar spatial matching as for the INCAvsWN_therm_FBR and INCAvsWN_strong_FBR cases, just with a somewhat higher bias of wind class differences. This similar performance despite the coarser horizontal resolution of the CCLM model is explained through a weaker influence of the terrain on the wind fields under strong wind conditions in this region.

Because of the challenging terrain of the WegenerNet JBT, the statistical WFSS values are generally low for this region, signalling large biases (Fig. b). These biases are indicated by low asymptotic values, which tend to be between 0.61 and 0.64, except for the INCAvsWN_strong_JBT case, which shows an even lower value (AWFSS=0.39).

The spatial displacement and the biases for the INCAvsWN_therm_JBT case are mainly caused by the differences in wind directions for these thermally induced wind events. Especially at small neighborhood sizes at the 1 km scale, WFSS values indicate large spatial displacements.

The INCAvsWN_strong_JBT case shows the lowest values at all neighborhood sizes, but this time caused by the differences in the wind speed categories. These low values are caused by the INCA-analyzed wind speeds, which, in the case of strong winds, are overestimated in the summit regions and underestimated in the valley regions. Slightly overestimated WegenerNet wind speeds in the Enns valley somewhat reinforces the difference between the INCA and the WegenerNet wind speeds. These differences in wind speed especially in the valley and the summit regions become obvious in Fig. c and d and are discussed in further detail below.

The intercomparison of the CCLM wind fields with INCA delivers nearly the same (low) WFSS values for both types of wind events. In the case of thermally induced events (CCLMvsINCA_therm_JBT), the spatial displacements and biases are mainly caused due to differences in wind directions. For strong wind events (CCLMvsINCA_strong_JBT), the smoothed terrain caused by the coarse resolution of the CCLM leads to systematically underestimated wind speeds.

Table  summarizes, in addition to the AWFSS values, the results estimated with traditional statistical methods of the INCA analysis and CCLM dynamical fields. Due to the less challenging region of the WegenerNet FBR, all traditional statistical parameters show better performance for this region compared to the WegenerNet JBT. The absolute-value statistical metrics (bias B, standard deviation SDo, root-mean-square error RMSE) applied to the hourly vector-mean wind speeds show higher values for the WegenerNet JBT, resulting from the generally higher wind speeds in addition to the effects of the complex mountainous terrain on the wind fields in this region. The B values are slightly positive for the WegenerNet FBR and negative for the WegenerNet JBT. The substantially negative B value for the INCAvsWN_strong_JBT case again reflects the underestimation of wind speed in the valleys, as explained above. Furthermore, the CCLMvsINCA_strong_JBT intercomparison also shows a negative bias, caused by the coarse resolution of the CCLM, which leads to lower wind speeds for strong wind events.

The RMSE values range from 0.79 to 1.85 m s-1 for the WegenerNet FBR and from 1.3 to 8.6 m s-1 for the WegenerNet JBT. The high value of 8.6 m s-1 for the INCAvsWN_strong_JBT case is caused by the underestimation of wind speed in the valleys as well as the overestimation in the summit regions by the INCA model. The mean R values show a better correlation for the WegenerNet FBR than for the WegenerNet JBT. The mean absolute error of wind direction (MAEdir) applied to hourly vector-mean wind directions also shows better performance (INCA and CCLM fields) for the WegenerNet FBR. Due to the varying wind directions caused by thermally induced circulations, the MAEdir is higher for such events for both study areas, with the highest value of 68∘ for the INCAvsWN_therm_JBT case.

Figure  shows the mean wind speed bias spatial distributions for all evaluation cases, for the WegenerNet FBR (Fig. a and b) and the WegenerNet JBT (Fig. c and d). The distribution for the INCAvsWN_therm_FBR case, the case for thermally induced wind events for the WegenerNet FBR, shows large areas with nearly no B values (left panel of Fig. a). The maximum B value for this case can be observed in the area of the Gleichenberger Kogel, north of the ZAMG Bad Gleichenberg station with a value of around 1.4 m s-1. The evaluation of the CCLM for the same thermally induced events shows an overestimation of wind speeds for the whole study area, with B values from 0.5 to 1 m s-1 in the western part and from 1 to about 1.4 m s-1 in the eastern part of the study area (right panel of Fig. a).

The overestimation of the wind speeds for the WegenerNet FBR can be explained by the too frequent flow-over patterns simulated for this region, which lead to a more dominant orographic speedup effect. Due to orographic smoothing, flow-over patterns are generally more frequent than flow-around patterns, especially for the WegenerNet FBR with its small differences in altitude .

The evaluation of the INCA model for strong wind speeds illustrates the strong influence of the terrain on this model. The results show a good agreement in the valleys of the study area, with partly small negative B values (left panel of Fig. b). The hilltop regions exhibit positive B values, with maximum values of around 5 m s-1 again in the area of the Gleichenberger Kogel. Overall positive B values of CCLM dynamical wind speed fields for strong wind events are seen in the right panel of Fig. b, showing the systematic overestimating by CCLM fields in this case. These large B values are probably also due to the speedup effect explained for the above case CCLMvsWN_therm_FBR.

For the WegenerNet JBT, the strong influence of the terrain on the INCA-analyzed wind speeds can be observed in all evaluation cases in this region. The evaluation of the INCA model for thermally induced wind events exhibits negative B values in the valleys, whereby positive values partly occur in the summit regions (left panel of Fig. c). At lower elevations, the intercomparison of the CCLM with the INCA model shows nearly no B values for thermally induced events (right panel of Fig. c). Furthermore, small negative B values in the summit regions and some spots with positive values can be observed for this case. Due to these small bias values, similar results to these ones can be expected for a comparison of CCLM with WegenerNet data.

Similar bias distribution patterns as for the INCA evaluation for thermally induced wind events are present for strong wind events in the INCAvsWN_strong_JBT case, but this time with strong negative and positive B values ranging from -14.4 to 4.9ms-1 (left panel of Fig. d). These strong negative B values are again caused by the severely underestimated INCA wind speeds and the somewhat overestimated WegenerNet wind speeds in the valley regions of the WegenerNet JBT.

Opposite patterns can be seen for the intercomparison of the CCLM with the INCA model. This intercomparison exhibits small positive values in the valley regions and strong negative values in the summit regions (right panel of Fig. d). The main reason for these strong negative B values is the coarse resolution of the CCLM data and the resulting underestimation of wind speeds for strong wind events, as explained above. The negative B values are likely caused by negative orographic speedup effects, which are preferred in flow-around patterns and flow-splitting patterns that occur especially when the differences in the altitude of ridges of mountains are large .