全球变化知识资源中心

Few studies have examined changes in rainfall on hourly timescales due to sparse sub-daily observations and the inability of climate models to reliably simulate sub-daily rainfall. The studies so far suggest greater increases in hourly compared to daily rainfall extremes^{5, 8}, but as a result of model deficiencies we have low confidence in these projections. This is of concern as it is short-duration convective extremes which tend to be responsible for flash flooding events, such as the Boscastle flood in August 2004 (ref. 9), particularly important in urban environments and for small or steep river catchments.

The Clausius–Clapeyron (CC) relation describes the rate of change of saturated water vapour pressure with temperature as approximately 7% °C⁻¹, and sets a scale for change in precipitation extremes¹. Increasing evidence from observational studies suggests intensities of sub-daily precipitation extremes increase more rapidly with temperature than for daily extremes; above the CC rate, at least in some regions^{8, 10}. This seems to be a property of convective precipitation¹⁰ and may be explained by latent heat released within storms invigorating vertical motion, leading to greater increases in rainfall intensity. However, the extent to which this scaling may apply over the longer-term with global warming is uncertain.

Global and regional climate models (with typical grid spacings of 60–300 km and 10–50 km respectively) rely on a convective parameterization scheme to represent the average effects of convection. This simplification is a known source of model error, and leads to deficiencies in the diurnal cycle of convection¹¹ and the inability (by design) to produce hourly precipitation extremes^{5, 6, 8}. Very high resolution models (order 1 km grid spacing), on the other hand, can represent deep convection explicitly without the need for such a parameterization scheme^{3, 12}. Such models are termed ‘convection-permitting’ because larger storms and meso-scale convective organization are permitted (largely resolved) but convective plumes and small showers are still not represented.

Convection-permitting models are commonly used in short-range weather forecasting. They give a much more realistic representation of convection and are able to forecast localized extreme events not captured at coarser resolutions¹³. However, there are few examples of convection-permitting resolutions being applied in climate studies, owing to their high computational cost. Previous studies have been limited to small domains and often just a single season^{12, 14, 15} or selected events^{16, 17}. Some studies have built up multi-year climatologies through a sequence of seasonal^{18, 19} or shorter²⁰ simulations. However, long continuous simulations are needed to represent long-term memory in the soil and its feedbacks with precipitation²¹. We recently³ carried out the first extended (20-year) length climate simulation with a convection-permitting (1.5 km) model over a region of the UK. Here we use the same model to examine future changes. To our knowledge this is the first time that continuous multi-year simulations at such high resolutions have been carried out to study rainfall change for a future climate scenario. Climate change experiments have been carried out at 4 km resolution over the western US (ref. 22), but this resolution is not high enough to adequately represent typical convection over the UK (ref. 13).

We compare future changes in hourly rainfall in the 1.5 km model with results from a 12 km regional climate model (RCM) over the southern UK. The models are run for 13-year present-day (1996–2009) and 13-year future (~2100, under the Intergovernmental Panel on Climate Change RCP 8.5 scenario) periods, driven by a 60 km global climate model (GCM). Model biases for the present-day have been assessed by comparison with gridded hourly observations from radar, available for 2003–2012 (ref. 23). Because radar tends to systematically underestimate heavy rain²⁴, we apply a bias correction using daily gauge observations.

On hourly timescales, rainfall is heavier over the southern UK in summer than in winter (Figs 1 and 2). Model biases compared to radar data are also larger in summer. In particular, the 12 km-RCM significantly underestimates heavy rainfall in summer, whereas the 1.5 km model tends to provide an overestimate, particularly in the south-east. The tendency for heavy rain to be too intense in small convective cores in the 1.5 km model is understood and is a current inherent weakness of a ‘convection-permitting’ model¹³. Smaller showers are not properly resolved, with some showers having updrafts on the wrong scale with insufficient turbulent sub-grid mixing. Nevertheless, the 1.5 km model gives a much better representation of hourly rainfall characteristics, including extremes^{3, 4}, than the 12 km model, and extensive testing within numerical weather prediction trials at the Met Office has shown considerable benefit from the 1.5 km model, leading to its operational implementation as a replacement for the previous 4 km and 12 km models.

a, Observed radar. b,c, Difference between model and observed radar for 12 km and 1.5 km models, respectively. d,e, Difference between 2100 and present-day for 12 km and 1.5 km models, respectively. Heavy rainfall is defined as the mean of the upper 5% of wet values (>0.1 mm h⁻¹). White indicates differences or future changes not significant at the 1% level compared to year-to-year variability. The radar data has been bias corrected using daily rain gauge data.

The models used here are all configurations of the Met Office Unified Model (MetUM; ref. 25). The 1.5 km model spans the southern UK and is driven by the 12 km-RCM, which spans Europe and is in turn driven by the 60 km-GCM. The 1.5 km model is as described previously³, with some upgrades to the model physics, particularly an improved microphysical parameterization of drizzle and fog (ref. 26). The 12 km-RCM and 60 km-GCM both have the UM Global Atmosphere 3.0 configuration²⁵, and have similar model physics to that in the 1.5 km model except that, at 1.5 km resolution, the convection scheme has been switched off and Smagorinsky–Lilly turbulence diffusion is applied.

For the present-day control runs, monthly sea surface temperature (SST) and sea-ice forcings from the Program for Climate Model Diagnosis and Intercomparison were used. Other forcings follow the Atmospheric Model Intercomparison Project II (AMIP-II) protocols, excepting that Shine–Li ozone climatology²⁷ was used for the 1.5 km and 12 km RCMs. For the future runs, SST was configured as time-varying monthly SST from the control run plus the (multi-year) mean SST change for each month between 1990–2010 and 2090–2110 in the HadGEM2-ES runs²⁸ under the IPCC RCP8.5 scenario. Carbon dioxide, methane, nitrous oxide, CFC and HFC concentrations were adjusted accordingly, but do not vary with time. Sea-ice comes directly from the HadGEM2-ES integration, as a repeat monthly cycle. Ozone and aerosol forcings were not changed between the present-day and future runs. In the 60 km-GCM and 12 km-RCM, aerosol mass mixing ratios provide the cloud droplet number for autoconversion. In the case of the 1.5 km model, however, autoconversion limits are based on droplet number assumptions²⁹. Because aerosol forcings are not changed for the future simulations, this is not expected to have a large impact on the climate change results.

Soil moisture evolves freely using the Joint UK Land Environment Simulator (JULES; ref. 30). Soil moisture in the 1.5 km model is initialized from the 12 km-RCM, and takes a few months to ‘spin-up’ (except potentially in the very deepest layer, where it can take several years to fully reach equilibrium). Thus the first few months of the simulation were discarded (the simulations were actually 13 years 7 months), and the analysis here only uses 13 years of model data from December 1996 to November 2009 for the present-day and similarly for the future runs. We note that a key benefit of long-continuous simulations (rather than seasonal slices) is that long-memory land-surface feedbacks can be represented.

The radar data used here are 5 km hourly data from the Nimrod database²³. Radar data offer good spatial coverage and are available for an extended period (2003–2012). However, the radar tends to underestimate heavy rain because of beam attenuation²⁴, and so we apply a bias correction using daily rain gauge observations (further details of the observational datasets are provided in the Supplementary Methods). In particular, at times when hourly rainfall is a major contributor to the daily total, it is possible to estimate an upward correction to the hourly radar intensity by comparing daily radar totals with daily gauge totals. Specifically we identify heavy hourly rainfall amounts in the radar data, and identify when these are > 0.3× daily radar total. If this criterion is met, and the daily radar total < the daily gauge total for that grid point on that day, then we upscale the radar hourly amount as follows:

If the daily radar total > the daily gauge total, no correction is applied. This corresponds to the situation of a heavy localized shower missed by the gauges, for which the radar provides the best (although potentially biased) estimate. The sensitivity of the bias correction to the selection criterion and the impact of the correction on the results are discussed in the Supplementary Methods (Supplementary Figs 2 and 3).

All analysis here is carried out at the 12 km scale, with the hourly precipitation fields for the 1.5 km model and 5 km radar being first aggregated onto the 12 km-RCM grid. Bootstrap resampling is used to assess the significance of model biases and future changes with respect to year-to-year variability. A total of 1,000 bootstrap samples are produced for the model (radar) data by selecting 13 (10) years from the full dataset randomly with replacement. These are used to produce 1,000 estimates of the difference between the model and radar, or the future and present-day model runs, allowing a confidence interval for the difference to be calculated.

1. Trenberth, K. E., Dai, A., Rasmussen, R. M. & Parsons, D. B. The changing character of precipitation. Bull. Am. Meteorol. Soc. 84, 1205–1217 (2003).
   • ADS
   • ISI
   • Article
2. Min, S-K., Zhang, X., Zwiers, F. W. & Hegerl, G. C. Human contribution to more-intense precipitation extremes. Nature 470, 376–379 (2011).
   • CAS
   • ADS
   • Article
3. Kendon, E. J., Roberts, N. M., Senior, C. A. & Roberts, M. J. Realism of rainfall in a very high resolution regional climate model. J. Clim. 25, 5791–5806 (2012).
   • Article
4. Chan, S. C. et al. The value of high-resolution Met Office regional climate models in the simulation of multi-hourly precipitation extremes. J. Clim. http://dx.doi.org/10.1175/JCLI-D-13-00723.1 (in the press).
5. Hanel, M. & Buishand, T. A. On the value of hourly precipitation extremes in regional climate model simulations. J. Hydrol. 393, 265–273 (2010).
   • Article
6. Gregersen, I. B. et al. Assessing future climatic changes of rainfall extremes at small spatio-temporal scales. Climatic Change 118, 783–797 (2013).
   • Article
   URL: