Urban hydrological modelling requires precipitation data at a very high resolution, in both time and space. In time, durations down to 5-10 minutes need to be considered. In space, generally observations from a single gauge (or a few gauges within a very limited area) are used. The high resolutions required are problematic with respect to climate change assessment, as they are not matched by the resolutions used in Global Climate Models (GCMs) or either Regional Climate Models (RCMs). The calculation time step used in the latter (30 min, sometimes shorter) is rather close to the required temporal resolution, but the still relatively coarse spatial grid used in the calculations (typically 50*50 km, sometimes finer) makes RCM output unsuitable for direct application in urban hydrological climate change assessment.
Important differences in the extremes
The properties of short-term precipitation as observed in a single gauge and as averaged over a few thousand km², respectively, are very different. The most important difference concerns the extremes. The observed extremes can be very localised whereas a model gridbox represents an average of surface types and heights. Thus, the model extremes are expected to differ. Further, since the observed extremes occur over very short time periods, often shorter than the model time step, short term atmospheric fluctuations will not be fully captured by the model. In Sweden, local extremes are practically always generated in summer by localised short-term convective events. In the models, depending on the resolution and the convective parameterisation, these convective events can occur either on the resolved model grid box scale or on sub-grid scale. Thus, analyses of RCM extremes (and their changes) at the gridbox scale may not reflect the properties of local extremes.
Downscale extreme precipitation from the regional grid scale to local scale
The objective of the study is to assess the possibility to stochastically downscale extreme precipitation from the RCM grid scale to local scale. The main assumption behind the approach is that additional RCM variables, mainly precipitation type and cloud cover, can be used to rescale gridbox average precipitation to local precipitation. By assigning a probability that this local precipitation occurs in a specific point inside the gridbox, realisations of point value time series may be stochastically generated. The overall idea is to formulate such a stochastic scheme, calibrate it to reproduce observed statistical properties for the present climate, and then apply the calibrated scheme to estimate future point precipitation statistics.
A comparison of ERA40-driven RCA3-simulations with observations shows that the annual cycle of mean monthly total precipitation from RCA3 agrees well with observations, although there is some slight overestimation (Figure 1). The frequency is, however, overestimated, especially of large-scale (resolved) precipitation. Also the mean monthly RCA3 cloud cover agrees well with observations.
In future climate projections, no clear changes in the cloud cover are found. Large-scale and convective precipitation all increase gradually by some 20% by the period of 2071-2100. Total and extreme large-scale precipitation increases mainly in winter; convective precipitation in summer.
With the stochastic downscaling scheme is able to simulate point precipitation time series which satisfactorily reproduce observed extreme statistics. The downscaled series indicate a 5-10% increase in short-term extreme intensities in the period of 2011-2040 and a 10-20% increase in the period of 2071-2100. This is 5-10% more than the increase estimated from RCA3 gridbox data. Using more detailed cloud variables than ‘total clouds´ and ‘high clouds´ did not improve performance of the downscaling scheme.