Both the ICHEC WRF and COSMO-CLM RCM simulations were performed utilising nested domains with 18, 6, and 2 km (WRF) or 1.5 km (COSMO-CLM) resolutions. Figure 1 illustrates the spatial coverage and topography of the three COSMO-CLM and WRF domains. The WRF 18 km domain is composed of a 176×183 grid with latitudinal extent 36.76 to 67.56∘ N and longitudinal extent 42.15∘ W to 24.15∘ E, the 6 km domain is on a 216×210 grid with latitudinal extent 47.74 to 59.52∘ N and longitudinal extent 17.78∘ W to 4.72∘ E and the 2 km domain is on a 216×273 grid with latitudinal extent 50.81 to 55.77∘ N and longitudinal extent 11.57 to 4.71∘ W. The three COSMO-CLM domains are only slightly larger than the WRF domains: 18 km on a 188×188 grid with latitudinal extent 35.51 to 68.36∘ N and longitudinal extent 45.54∘ W to 27.26∘ E; 6 km on a 245×245 grid with latitudinal extent 46.84 to 60.45∘ N and longitudinal extent 20.16∘ W to 5.76∘ E; and 1.5 km on a 328×398 grid with latitudinal extent 50.64 to 56.04∘ N and longitudinal extent 11.93 to 4.11∘ W. The 18 km simulations were driven at the boundaries by ERA-Interim reanalysis data, produced by ECMWF at 80 km resolution, with all outputs (Sect. 2.2) from each individual nested domain (for both COSMO-CLM and WRF) archived at hourly intervals.

The WRF model used here (v3.7.1) comes with topography data at four resolutions (10, 5, 2 and 0.5 arcmin) that can be used to construct terrain data for the model grid. Given that some climate variables (e.g. winds) are affected by nearby topography, it was realised that underlying data with much finer resolution was required. Therefore, a 1 arcsec topography dataset (The Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) Global Digital Elevation Model (GDEM)) was obtained and incorporated into the WRF simulations using the WRF Preprocessing System (WPS). By contrast, the COSMO-CLM model already includes the high-resolution ASTER topography dataset as part of the preprocessing stage (ExtPAR).

Both models have numerous parameter schemes that can potentially affect outputs. For instance, it is known that the choice of WRF sub-grid orographic, flow blocking and gravity wave drag schemes can influence bias in 10 m wind speeds, 2 m temperature and surface pressure (Koo et al., 2018). To ensure the most accurate options were employed, the results from several 1-month validation simulations previously performed at ICHEC (Nolan et al., 2017) were utilised. Summaries of the individual model settings, where different from the default option (or between different resolutions) are given in Tables 1 and 2.

All outputs from each of the 18 km, 6 km and highest resolution 2 km (WRF) and 1.5 km (COSMO-CLM) RCM simulations have been archived at hourly intervals by ICHEC. A full listing of all climate parameters (and their relevant units) archived at ICHEC is given in Tables 3–5 (COSMO-CLM) and 6–8 (WRF). The COSMO-CLM dataset is composed of 48 surface/near-surface parameters (Table 3), two sub-surface parameters at eight levels (Table 4) and three upper-air parameters at 10 heights (Table 5). The WRF dataset contains 23 surface/near-surface parameters (Table 6), two sub-surface parameters at four levels (Table 7) and two upper-air parameters at five levels (Table 8).

The climate parameters presented here constitute the highest (spatial) resolution, hourly dataset currently available for Ireland for the period 1981–2016 (data for 2017 is nearing completion and will soon be added to the dataset). Additionally, the datasets contain parameters that are currently not available elsewhere (MÉRA or observations). These new high-resolution parameters have potential for many applications and will be of use to researchers from different fields. For instance: hydrology (surface and subsurface runoff); wind energy (air density at turbine heights); extreme events (CAPE 3 km, Showalter index, surface lifted index); and agriculture (soil temperature and moisture content at four (WRF) and eight (COSMO-CLM) levels).

Hourly station observations of air temperature have been obtained from Met Éireann. In total, there are 25 of these stations with varying record lengths available. The COSMO-CLM and WRF datasets have been processed so that a comparison with these observations could be made – the 2 m temperature data is already stored in hourly files and data at the relevant station locations can be extracted immediately through (bilinear) interpolation. Overall values have been determined by treating all available station data as a single dataset. The (mean) error values found are -0.18 ∘C (COSMO-CLM) and -0.31 ∘C (WRF), whilst the overall standard deviations are 1.79 ∘C (COSMO-CLM) and 1.71 ∘C (WRF). Additionally, the mean absolute error (MAE) has been calculated (using the entire station dataset) – the values found are 1.34 ∘C (COSMO-CLM) and 1.31 ∘C (WRF). For comparison, an identical analysis of hourly MÉRA 2 m temperature yields the values -0.10 ∘C (mean error), 1.09 ∘ (standard deviation) and 0.81 ∘C (MAE).

In Fig. 4, the 2 m temperature error distributions for COSMO-CLM and WRF are shown to provide insight into the performance of each model. Overall, both COSMO-CLM and WRF display somewhat similar error, with WRF slightly more likely to be within 1 ∘C of observed values. Except for the error range [-2,-1), WRF is consistently more likely than COSMO-CLM to underestimate 2 m temperature. In turn, COSMO-CLM shows a higher frequency of large positive error. Additionally, whilst the mean errors and standard deviations of both models are similar for the left-most error range (-4.83 and 0.82 ∘C (COSMO-CLM) and -4.85 and 0.81 ∘C (WRF) respectively) COSMO-CLM produces larger positive outliers (mean error =5.38 ∘C and standard deviation =1.33 ∘C versus 5.10 and 1.10 ∘C respectively, for WRF).

Hourly 10 m wind speed synoptic station observations have also been obtained from Met Éireann. For both WRF and COSMO-CLM, the 10 m U and V wind components have been found at each station location through bilinear interpolation and used to calculate 10 m wind speeds (U2+V2) which are then compared to observations. Overall mean error values found are 0.85 m s-1 (COSMO-CLM) and 0.07 m s-1 (WRF), whilst the overall standard deviations are 2.30 m s-1 (COSMO-CLM) and 2.24 m s-1 (WRF). The overall MAEs found are 1.89 m s-1 (COSMO-CLM) and 1.67 m s-1 (WRF). For comparison, an identical analysis of hourly MÉRA 10 m wind speeds yields the values 0.29 m s-1 (mean error), 1.65 m s-1 (standard deviation) and 1.27 m s-1.

Figure 5 shows the 10 m wind speed error distributions found for COSMO-CLM and WRF. It can be readily seen that the COSMO-CLM distribution exhibits greater frequency of error at higher ranges. An opposite, but weaker, behaviour is evident for negative ranges where WRF shows higher frequencies. At the extremes (less than -4 m s-1 and greater than 4 m s-1) both COSMO-CLM and WRF exhibit similar mean errors (approximately -5.5 and 5.3 m s-1, respectively) and standard deviations (1.5 and 1.3 m s-1, respectively).

A similar analysis to that in Sect. 3.1 and 3.2 has been performed for sea-level pressures utilising synoptic station data from Met Éireann. The overall (mean) error values found are -0.87 hPa (COSMO-CLM) and -0.20 hPa (WRF), whilst the overall standard deviations are 2.56 hPa (COSMO-CLM) and 2.39 hPa (WRF). The overall MAEs found are 1.96 hPa (COSMO-CLM) and 1.69 hPa (WRF). By comparison, an identical analysis of hourly MÉRA sea-level pressures gives the values 0.03 hPa (mean error), 0.51 hPa (standard deviation) and 0.37 hPa (MAE).

The COSMO-CLM and WRF error distributions are shown in Fig. 6, where COSMO-CLM has greater frequency at all negative ranges, whilst an opposite but weaker effect occurs at positive ranges. At extreme ranges (less than -4 hPa and greater than 4 hPa) both models exhibit similar mean errors (-6 and 6 hPa, respectively) and standard deviations (2.2 and 2.1 hPa, respectively).

An analysis of hourly precipitation amounts has been performed, again using synoptic station data sourced from Met Éireann. Both COSMO-CLM and WRF show remarkably similar error and standard deviations – overall error values are less than 0.01 mm and overall standard deviations are 0.63 mm, for both models. Additionally, the MAEs found are 0.18 mm for both models. By comparison, an identical analysis performed for hourly MÉRA precipitation results in the values: <0.01 mm (error), 0.55 mm (standard deviation) and 0.16 mm (MAE).

Figure 7 presents the error distributions for both models for four different observed precipitation categories: dry (0 mm), light (up to 0.2 mm), moderate (0.2–2.5 mm) and heavy (>2.5 mm).

For no observed precipitation (Fig. 7, top-left panel), the count (3 063 887) is approximately 4.6 times higher than all other categories combined (671 887) – it does not rain as often as is commonly perceived. In this category, COSMO-CLM has much higher frequency of correct predictions. However, the overall error and standard deviation of both models are similar (0.07 and 0.4 mm, respectively). This is a result of COSMO-CLM producing larger (although fewer) errors than WRF – the means for the highest error range (>0.2 mm) are 1 and 0.9 mm, respectively.

For light rainfall (Fig. 7, top-right panel), both models display higher frequency of negative error, with WRF exhibiting higher frequency of low error. However, both show similar frequencies, means (1.3 mm) and standard deviations (1.3 mm) at the extreme error range (≥0.4 mm) which contribute to similar overall errors (>0.1 mm) and standard deviations (0.65 mm) for this category.

Both models exhibit similar error distributions for the moderate (Fig. 7, bottom-left panel) and heavy (Fig. 7, bottom-right panel) precipitation categories. For moderate precipitation, both models under-predict with the majority (∼0.85) of error values falling in the combined range [-1 mm, 0 mm). The overall errors and standard deviations for this category are similar: -0.38 and 1.05 mm (COSMO-CLM) and -0.35 and 1.01 mm (WRF). The frequencies of large negative and positive error are both less than 0.1. Both models exhibit similar mean errors and standard deviations for the error range <1.5 mm: -1.9 and 0.27 mm respectively. For the error range ≥1 mm, the corresponding values are slightly different: 2.35 and 1.8 mm (COSMO-CLM) and 2.23 and 1.6 mm (WRF).

Heavy precipitation is relatively rare in Ireland – the overall observed count (31 938) is an order of magnitude lower than that of the other two wet categories. For this category, both models typically under-predict, with WRF performing only marginally better than COSMO-CLM – the overall mean errors and standard deviations are -2.95 and 2.33 mm for COSMO-CLM and -2.84 and 2.32 mm for WRF. Indeed, the frequency of error in the range ≥0 mm is less than 0.1 for both models, with WRF performing slightly better – mean errors and standard deviations are 1.88 and 2.37 mm for COSMO-CLM and 1.63 and 1.98 mm for WRF. The frequencies, mean error (-7 mm) and standard deviation (2.5 mm) of large negative error are similar for both models, with WRF exhibiting only slightly lower frequency than COSMO-CLM.

Daily precipitation amounts have been obtained from Met Éireann's 484 station rainfall network and used to estimate mean 24 h accumulation errors for both COSMO-CLM (Fig. 8) and WRF (Fig. 9). Overall mean errors, standard deviations and MAEs found for COSMO-CLM are -0.23, 5.94 and 2.97 mm, respectively, whilst for WRF the respective values found are 0.10, 5.41 and 2.69 mm. By comparison, an identical analysis of MÉRA produces the values -0.09, 4.59 and 2.28 mm. For both COSMO-CLM and WRF, the largest errors occur over regions with complex topography (mountainous regions in the west) and during autumn and winter months when rainfall amounts tend to be largest. Both models tend to under-predict during drier spring and summer months – the mean errors found for WRF are -0.45 and -0.26 mm respectively, whilst for COSMO-CLM, the values found are -0.002 and -0.01 mm. Although COSMO-CLM shows better accuracy than WRF during drier seasons, the reverse is true during autumn and winter, when rainfall amounts are higher and both models over-predict – the mean values found for WRF are 0.29 and 1.20 mm respectively, whilst for COSMO-CLM the values found are 0.43 and 1.45 mm.

An analysis of daily mean 10 cm soil temperatures has been performed through comparison with daily values from 23 Met Éireann stations. Simple linear vertical interpolation (cdo command intlevel, Schulzweida, 2018) has been used to generate model data at this level from the archived levels (Tables 4 and 7). Horizontal bilinear interpolation to the station locations has then been applied for all stations surrounded by land points. Where a station is next to a sea point, a simple nearest-neighbour approach was taken. For those stations where bilinear interpolation was possible, the nearest-neighbour method was tested and compared – absolute differences in mean temperature values were small (typically less than 0.1 ∘C) and usually in favour of the bilinear method. The overall daily errors, standard deviations and MAEs found were 1.17, 1.26 and 1.37 ∘C (COSMO-CLM) and 1.01, 1.16 and 1.18 ∘C (WRF).

The mean daily errors per season for each station are shown in Figs. 10 (COSMO-CLM) and 11 (WRF). Both models exhibit positive mean error for each season (MAM, JJA, SON, DJF) with least error during colder seasons: (1.34, 1.79, 0.76, 0.78 ∘C) COSMO-CLM; (1.08, 1.62, 0.67, 0.68 ∘C) WRF. By comparison, Gleeson et al. (2017) (their Fig. 8a) show lower mean error over the equivalent time periods, albeit for 20 cm soil temperature, with consistent over- (under-) prediction during winter (summer) months.

The standard deviations found were (1.47, 1.92, 1.08, 1.01 ∘C) for COSMO-CLM and (1.25, 1.72, 0.85, 0.88 ∘C) for WRF whilst the MAEs found were (1.19, 1.35, 1.13, 1.07 ∘C) for COSMO-CLM and (1.20, 1.33, 0.87, 0.93 ∘C) for WRF. The two models show similar MAEs during spring and summer months, whereas WRF shows greater accuracy during autumn and winter months when temperatures are lower. From Figs. 10 and 11, there does not appear to be a pattern to the spatial distribution of errors. However, this could simply be due to the lack of observational data available – 23 stations here compared to 484 for the rainfall analysis in Sect. 3.4.

A preliminary analysis of COSMO-CLM parameters that are potentially of interest to researchers of weather extremes has been conducted. Observational data for these parameters are both rare and difficult to obtain. However, radiosonde data has been obtained for two locations in Ireland: Valentia and Castor Bay. The Valentia data covers the period 1981–present, whilst the Castor Bay data covers the period 2003–present. Typically, the soundings are recorded every 6 hours beginning at midnight on each day.

For CAPE 3 km, the overall errors, standard deviations and MAEs found for Valentia are 2.31, 47.5 and 2.31 J kg-1, respectively. For Castor Bay, the equivalent values are -0.68, 34.9 and 10.33 J kg-1. The Showalter index results at Valentia are -0.05, 3.51 and 2.58 whilst at Castor Bay the values are quite similar: 0.43, 3.36 and 2.52. The surface lifted index results are 0.06, 3.03 and 2.27 (Valentia) and 1.08, 3.09 and 2.58 (Castor Bay).

The error distributions for each parameter at Valentia and Castor Bay are shown in Figs. 12 and 13, respectively. For each parameter and location, the distributions typically have long tails – evidenced by the relatively large frequencies at the extremes of each as well as the means and standard deviations found for these ranges. For CAPE 3 km at Valentia, the respective values are -64.5 and 105.3 J kg-1 (<-15 range) and 55.3 and 36.6 J kg-1 (>15 range). At Castor Bay, the equivalent values found are -55.5 and 88.9 J kg-1, and 49.0 and 35.0 J kg-1 respectively. For Showalter index at Valentia, the respective values are -7.7 and 2.4 (<-5 range) and 6.7 and 1.9 (≥5 range). At Castor Bay, the equivalent values are similar: -7.7 and 2.5, and 6.6 and 1.6 respectively. Finally, for Valentia surface lifted index, the respective values are -7.2 and 2.2 (<-5 range) and 6.5 and 1.8 (≥5 range), whilst at Castor Bay, the equivalent values are (again) similar: -7.1 and 2.1, and 6.4 and 1.4 respectively.