Intensification of summer precipitation with shorter time-scales in Europe

We have applied the Weather Research and Forecasting model (WRF) (Skamarock and Klemp 2008) at convection-permitting 3 km horizontal resolution for four different regions across Europe (see section 2.3). An overview of domain setups and parameterization schemes is given in tables S1, S2. For each of the four regions, the 3 km horizontal resolution domain was one-way nested from an intermediate 15 km domain, which was again nested from a 45 km domain. 50 vertical levels were used, and spectral nudging was applied for temperature, horizontal winds, humidity and geopotential height in the outermost (45 km) domains. Two 20 year time periods were simulated to represent recent (1986–2005) and future (2081–2100) climate for the summer and early fall. Each year was initialized on May 1 and the run ended on September 30, but the first month was excluded from the analysis and considered as spin-up.

The initial and 6 hourly boundary conditions and sea-surface temperatures to WRF were from a coupled atmosphere-ocean simulation using the National Center for Atmospheric Research (NCAR) Community Earth System Model (CESM1.0.4) (Gent et al 2011). The CESM was set up with the CAM4 atmosphere module (Neale et al 2010) at 1.9°× 2.5° horizontal resolution and 26 vertical levels, the CLM4 land module (Lawrence et al 2011), the CICE4 sea-ice model (Hunke and Lipscomb 2008), and a full ocean model (Danabasoglu et al 2012), which is based on the Parallel Ocean Program version 2 (Smith et al 2010). Soil temperature and moisture was initialized in WRF by interpolating the 15 CLM4 soil levels to the four soil levels used in WRF. The CESM model was run from 1850 to 2100 using the high emission Representative Concentration Pathway 8.5 (RCP8.5) (van Vuuren et al 2011), and necessary variables to drive WRF were stored. Additional WRF experiments were run for the 1996–2005 (1 May–30 Sep) period with initial and boundary conditions from the 2.5° × 2.5° resolution NCEP/NCAR reanalysis (Kalnay et al 1996).

For comparison with the CESM and WRF results, further GCM output was retrieved from the Coupled Model Intercomparison Project phase 5 (CMIP5) archive (Taylor et al 2011) and regridded to 1°×1° horizontal resolution for the analysis (see table S3). Further RCM output has been taken from two large ensemble datasets and their driving GCMs. First, from the 21-member COSMO-CLM model ensemble over Europe (Addor and Fischer 2015), at approximately 50 km × 50 km horizontal resolution for 1950–2100 (RCP8.5), downscaled from an ensemble of CESM1-CAM4 simulations (Fischer et al 2013) (hereafter 'COSMO-LE' and 'CESM1-LE', respectively). Second, from the 50-member ensemble with the Canadian Regional Climate Model (CRCM5) downscaled over Europe for 1950–2099 (RCP8.5) at 12 km × 12 km resolution from an ensemble of CanESM2 GCM simulations (Leduc et al 2019, von Trentini et al 2019) (hereafter 'CRCM5-LE' and 'CanESM2-LE', respectively).

This study makes use of hourly and daily precipitation from rain gauges at observational sites for validation of model results. The hourly precipitation data were from the Norwegian Meteorological Institute through the eKlima data server (eKlima 2017), the Swedish Meteorological and Hydrological Institute through their portal for weather observations (SMHI 2017), the Royal Netherlands Meteorological Institute (KNMI 2017) and the MIDAS UK Hourly Rainfall Data catalogue (Met Office 2006). Daily precipitation data are taken from the non-blend version European Climate Assessment & Dataset (Klein Tank et al 2002). For the United Kingdom, daily precipitation was calculated based on the observed hourly precipitation.

The four high-resolution domains simulated by WRF cover southern Norway (hereafter 'S Norway'), Denmark incl. Southern Sweden and Northern Germany ('Denmark'), Benelux incl. Northern France and Southeast England ('Benelux'), and the Balkans focused on Albania ('Albania'). The extent of the domains, minus a boundary of 10 grid cells, is shown in figure 1 (figure S1 is available online at stacks.iop.org/ERL/14/124050/mmedia for Albania) and corner points of the domains are given in table S4.

Zoom In Zoom Out Reset image size

For all model datasets, we have analyzed future changes in mean and extreme summer precipitation per °C global- and annual-mean near-surface temperature change from each GCM simulation (or driving GCM simulation in the case of the RCMs). Here, we define the extreme indices we have used, Rx1day/Rx1hour/Rx10min, as 20 year averages of the day/hour/10 min with highest precipitation amount during each summer season (JJAS).

Figure 1 shows comparisons between station observations and WRF model results for three of the four regions simulated at convection-permitting scales. Observed Rx1day is generally very well reproduced by WRF in terms of magnitude and distribution, except for an overestimation of Rx1day along the west coast of S Norway and parts of Northern Germany (figures 1(a), (c)). A WRF simulation driven with reanalysis data instead of free-running climate data improves the comparison against observations for the S Norway region (figure S2). The overestimation is larger for mean summer precipitation than for Rx1day (figure S3). Further inspection shows that the altitude of the WRF grid box closest to each observation station is in many cases several hundred meters higher than the actual altitude of the station (figure S4). Stations are often located in valleys where there is less precipitation. This indicates that an even higher resolution than 3 km may be needed to properly resolve the complex terrain in S Norway. For the Albania domain, the very few stations that exist indicate an underestimation of Rx1day and mean precipitation by WRF (figures S1 and S3).

Moving from Rx1day to Rx1hour, we see that the spatial distribution is quite different, especially in S Norway (figure 1(b)). This is because maximum hourly compared to daily precipitation amounts during summer are more often a result of intense convective events. Comparison with station observations for Rx1hour shows that WRF is again slightly overestimated for S Norway, but compares very well for the other two regions (no hourly observations could be obtained for the Albania region) (figures 1(b), (d)).

Considering the strong regional differences in projections of mean summer precipitation in Europe (see section 1), analysis of the CMIP5 model projections, their water budget and how they agree over Europe, could give some more insight and be useful to have in mind before analyzing future precipitation projections from WRF. Figure 2(a) shows that future projections of mean summer precipitation are divided into three zones: an area of robust increase in the north, robust decrease in the south, and a transition zone in central Europe with large model diversity. Compared with annual mean precipitation changes (figure S5), this transition zone is located further north, and the region of precipitation decrease is more prominent and covers a larger area in Europe during the summer. The large inter-model standard deviation over central Europe seems greatest over the Alps (figure S6), indicating perhaps that the models' representation of orographic precipitation leads to larger uncertainty. In terms of extreme Rx1day precipitation, models agree on an annual increase in most of Europe, while uncertainties are large for Central and Southern Europe for the summer season (figure S5).

Zoom In Zoom Out Reset image size

The net surface water budget, or precipitation minus evaporation (P–E) (e.g. Byrne and O'Gorman 2015), shows to a large extent the same pattern as for precipitation, but is modulated by evaporation changes, which show a strong land-ocean contrast over southern Europe (figure 2(a)). This leads to less drying over land for P–E compared with the pure precipitation changes, and the opposite over the Mediterranean Sea. Contrasted with present-day P–E conditions (figure S6), we see a strong tendency of the so-called 'wet-get-wetter, dry-get-drier' conditions, except over land in central Europe, and this is also seen for S Norway ('wet-get-wetter') and Albania ('dry-get-drier') in figure 2(b). In general, present-day and future change in evaporation shows higher agreement between models compared to precipitation and P–E (figure 2(b)). Except for S Norway, uncertainties are much larger for the summer season compared to the annual mean (figures 2(b); S7). The drying over the Albania region is striking—it is the only one among the four regions where the change in summer precipitation is robustly different from zero, and the model-median decrease is large with around 25% of the present-day value. Drying in the Mediterranean region has already been observed over the last decades, and has mainly been attributed to changes in horizontal energy transport and a shift in the jet stream, due to increased concentrations of well-mixed greenhouse gases (Tang et al 2018).

In figure 3 we show averages over the four regions for mean and extreme precipitation, and for several models and resolutions (results for JJA and September separately are shown in figure S8). Most models show a future decrease in mean precipitation, but with some diversity in the sign of the change between CMIP5 models for the regions located in northern/central Europe. In terms of dry summers, all model medians show clear indications of more dry summers in the future, with stronger precipitation decreases in the 1-in-20-year driest (JJA) summer compared with the 20 year summer average precipitation (figure S8).

Zoom In Zoom Out Reset image size

When going from changes in mean to daily (Rx1day) extreme precipitation (figure 3), there is an intensification, or less reduction, for all models and regions. The signals of reduced future mean and increased Rx1day precipitation have been found for Western Europe in a single-model large-ensemble RCM study (Aalbers et al 2018). Further analysis of CMIP5 results shows that this intensification is also very clear for mid-century (2041–2060) but that the model spread is larger due to a weaker climate change signal compared to the end-of-the-century (figure S9). For nearly all regions and models, there is an intensification (or less reduction in regions with negative Rx1day change) when going from daily to hourly (Rx1hour) extreme precipitation (figure 3), and this is much more evident for CRCM5-LE than CESM1-CAM4 and WRF, including the WRF results at comparable resolution (15 km) as CRCM5-LE (12 km). However, the ability to simulate sub-daily precipitation in models with parameterized convection is questionable, and, for a given realization, more confidence should be given to WRF 3 km. When going from hourly to sub-hourly (Rx10min) extremes, available in WRF only, there is a further intensification for all regions except Benelux.

Uncertainties are in most cases dominated by the variation between different models (CMIP5), but natural variability, represented in figure 3 by the two large ensemble studies CESM1-LE/COSMO-LE and CanESM2-LE/CRCM5-LE, can also be large, e.g. spanning from −17 to 0%/°C for Rx1day over Albania for COSMO-LE. To investigate whether 2 × 20 summer seasons are enough to get robust results and get an indication of the influence of natural variability, which may be especially important for WRF since it is based on a single realization, we have analyzed the first and last ten summer seasons separately (see alternative versions of figure 3 in figure S10). We still find a consistent upward trend in the precipitation change versus duration, and this strongly indicates that our result of an intensification of precipitation with shorter time-scales, based on 20 summer seasons, is robust.

An interesting question is whether the convection-permitting model results (WRF 3 km) are notably different from the simulations in which convection is parameterized (WRF 15 km). With only a few exceptions (e.g. Rx1hour for Albania), the 3 km resolution results seem to have slightly lower values, i.e. smaller future increases and larger reductions, than the 15 km results (figure 3). However, the relatively coarse resolution driving model (i.e. CESM1-CAM4) shows in most cases even lower values than WRF 3 km, and this makes it difficult to conclude that there is a clear dependence on model resolution. Other factors than resolution can also lead to differences between CESM1-CAM4 and WRF results. Nevertheless, the convection-permitting WRF 3 km results do not differ greatly from its driving GCM, and the same is true for the two other RCMs, COSMO-LE and CRCM5-LE, and their driving models.

Separating the JJAS summer season into JJA and September, which is normally an autumn month in northern Europe, shows that the level of intensification when going from changes in daily to sub-daily extreme precipitation in WRF 3 km is more evident for September than JJA (figure S8). Previous research has shown that Rx1day will occur later in the year (Marelle et al 2018), and a subject for future research would be to investigate the robustness of the intensification across different seasons, and to see whether the close correspondence between the convection-permitting simulations and their driving data applies to other seasons.

Figure 4 shows that the future decrease in summer (JJAS) mean precipitation from WRF is evident throughout all four regions, except for the west coast of S Norway and parts of the Balkan Peninsula. It also shows that the projected decrease is less strong for the regions located furthest to the north (S Norway and Denmark). As for CMIP5 (figure S5), this decrease is not so evident for Rx1day, and an increase is projected in many areas. In WRF, some of the strongest increases in Rx1day seem to occur on mountain lee sides (north in the S Norway and Albania domains) (figure 4), where Rx1day in the historical period is lower than in its surroundings (figures 1; S1). The further intensification for shorter duration extremes, Rx1hour and Rx10min, is most pronounced for the two northernmost regions (S Norway and Denmark) (figure 4). Interestingly, there does not seem to be major differences in the intensification with shorter time-scales between regions of different weather regimes, such as between the low-altitude Denmark region and the S Norway region, where the western part is dominated by orographic precipitation and the eastern part is on the mountain lee side. Increases in short-duration extremes, which are more dominated by convection than for mean precipitation, would be consistent with the already observed increase in convective precipitation (and decrease in non-convective precipitation) over Eurasia (Ye et al 2017).

Zoom In Zoom Out Reset image size

Soil moisture can be an important driver of convection and how this variable is initialized in the WRF model is important. The effect of running WRF in seasonal time slices with one month of spin-up (see section 2.1) rather than simulations with a longer spin-up period has been investigated at 15 km resolution (table S5), and these results indicate that our main results and conclusions are not likely to be altered if a longer spin-up period would be used.

Our results show future increases in Rx1day, Rx1hour and Rx10min extreme indices from WRF 3 km of 5%/°C or less for all four regions (figure 3). The corresponding number is 6%/°C or less when the extreme indices are calculated as 1-in-20 year maximums rather than 20 year averages of these extreme indices (figure S8). Spatially, the projected changes in extremes are also in general below 7%/°C and only exceed this value locally over small areas, most notably in the S Norway and Denmark regions (figure 4). Thus, there is no direct indication of superadiabatic (>7%/°C) scaling, a finding that is in agreement with Ban et al (2015) and in contrast to Kendon et al (2014, 2019). Part of the reason could be that both Ban et al (2015) and this study have investigated summer seasons for regions with a relatively strong decrease projected for mean precipitation, and this implies the need for further investigating mechanisms of any relationship between seasonal mean and extreme precipitation changes.

In figure S8, the convection-permitting model results (WRF 3 km) show that the 1-in-20-year maximum for all extreme indices (Rx1day, Rx1hour and Rx10min) and nearly all regions (both JJA and September) increase more than the 20 year average of these extreme indices and are only rarely below zero. This strongly indicates that the extremes of the extremes increase more than the mean of extremes.

An interesting metric that has not been given much attention before, but could be of importance for society, is how the heaviest local precipitation event in a region may change in a future climate. The WRF results show that the precipitation amount in the land gridbox with the maximum summer Rx1day event throughout the entire 20 year period increases at around 10%–20%/°C depending on the region (figure 5). While this increase is based on only one value from each of the 20-year simulations, results for N = 10 (the 10 most extreme local precipitation events) are similar, indicating that the increase of 10%–20%/°C is a fairly robust result. For higher values of N, the change in precipitation is lower and approaches zero, consistent with the projected decrease in mean precipitation.

Zoom In Zoom Out Reset image size

Previous modelling studies have often used the local or regional (dew point) temperature when scaling with precipitation (Kendon et al 2014, Ban et al 2015, Kendon et al 2019) and observational studies typically use local daily mean (dew point) temperature to derive precipitation scaling (Lenderink and van Meijgaard 2008, 2010, Shaw et al 2011, Loriaux et al 2013). We have scaled precipitation with the global and annual mean temperature because precipitation is constrained by the global energy budget (e.g. O'Gorman et al 2012) and because local temperature does not necessarily represent the origin of the air masses leading to extreme precipitation. For the WRF results, we have tested the effect of using the seasonal mean temperature for each region, instead of the global and annual mean, and find only small differences. The magnitude of precipitation changes is larger in S Norway, smaller in Albania, and almost unchanged in the two other regions, but Rx10min for S Norway still does not show superadiabatic scaling (now 6.4%/°C compared to 4.8%/°C when scaling with global and annual mean temperature change) (not shown).

While the Rx1day extreme index is widely used, many previous studies (including Kendon et al 2014, Ban et al 2015, Kendon et al 2019) have calculated future changes in extremes using percentiles. We have tested the effect of analyzing percentile indices instead of Rx1day, Rx1hour and Rx10min in the WRF results and find only small differences when choosing percentiles that correspond to extreme precipitation events occurring, on average, once every summer (JJAS) (figure S11). In fact, all four regions show a further intensification of extremes for sub-daily and sub-hourly extremes when using percentiles.