Downscaling of Regional Climate Model Precipitation for Urban ...

11 th International Conference on Urban Drainage, Edinburgh, Scotland, UK, 2008 

Downscaling of Regional Climate Model Precipitation for 

Urban Hydrology 

J. Olsson 1 *, C.B. Uvo 2 , E. Kjellström 1 

1 Swedish Meteorological and Hydrological Institute, 601 76 Norrköping, Sweden 

2 Dept. Water Resources Engineering, Lund University, Box 112, 221 00 Lund, Sweden 

*Corresponding author, e-mail jonas.olsson@smhi.se 

ABSTRACT 

The demand from society for assessments of the climate change impact on also regional and 

even local scales is rapidly increasing. In an urban hydrological context, such assessment 

requires estimation of future precipitation intensities as observed by a single gauge (i.e. in a 

point). Due to the coarse spatial resolution of climate models, their precipitation output can 

not be directly used to make any inference concerning point value intensities. In this paper, a 

simple stochastic method to downscale precipitation from a climate model grid box to an 

arbitrary point inside it is proposed and tested. The method is based on a separation between 

the convective and large-scale components of the total precipitation and assumptions about 

the typical extensions of the associated synoptic systems. Evaluation for the city of Kalmar, 

Sweden, indicated that the downscaling method is able to generate point value series with 

properties very similar to point value observations, both in terms of extremes and general 

statistics (a similar performance is indicated in ongoing testing also in two other locations). 

One possible application of the method is to downscale model simulated climate change 

scenarios to assess future changes in point precipitation. 

KEYWORDS 

Regional climate model; precipitation; downscaling 

INTRODUCTION 

There is today a widespread consensus that global warming is a real threat to the future 

climate (IPCC, 2007). General Circulation Models (GCM) are used to predict the associated 

climate impacts on a global scale. However, not least in light of the recent strong medial focus 

on the climate change issue, the demand from society for assessments of the climate change 

impact on also regional and even local scales is rapidly increasing. This necessitates some 

form of downscaling of the GCM results from a resolution of typically 1-3º down to much 

finer spatial grids, and even virtually point values if changes in local processes are to be 

assessed. There are today two main downscaling strategies: dynamical and statistical. The 

former implies that a Regional Climate Model (RCM) is nested inside the GCM to increase 

the spatial resolution (to typically 20-50 km); the latter is based on statistical relationships 

between GCM results and regional or local observations. 

The main objective of this preliminary study is to assess the possibility to relate the result in 

terms of spatially averaged precipitation from an RCM grid box to point value observations 

from a nearby located precipitation gauge. The key issue is with which accuracy RCM results 

can be statistically downscaled to represent the precipitation process as manifested in a single 

point. The underlying motivation for the study is urban hydrological assessment, i.e. what is 

Olsson et al. 1


the expected impact of climate change in terms of flooding and other problems (e.g. pollution 

transport) in the storm drainage network of large cities. To be accurately described by 

computer models, this urban runoff process requires point value precipitation data, i.e. even 

the spatial resolution of RCM:s is far too coarse. Also the temporal resolution required for 

urban hydrological assessment is very high (~5-10 min), but recent RCM output is at least 

approaching these time scales. 

DATA AND STUDY REGION 

The regional climate model data consist of results from an RCM named RCA3, which stands 

for the third version of the Rossby Centre Atmospheric model (Kjellström et al., 2005). This 

model has been developed at the Swedish Meteorological and Hydrological Institute. The 

model domain covers Europe and is shown in Figure 1a. The spatial resolution of the results 

used here was 49×49 km and the time step 30 min. Data from grid boxes close to the 

observation stations (Figure 1b) were used. 

A general difficulty with GCM results, whether further downscaled or not, concerns their 

comparability with observations representing the present climate. As (1) climate exhibits lowfrequency 

oscillations (on decadal time scales or even longer, e.g. NAO phases) and (2) GCM 

runs start from arbitrary initial conditions, it is not certain that GCM results in a given time 

period are “in phase” with the real climate. Averaging over several decades may reduce the 

problem, but to which degree is difficult to assess, as is the relative contributions from “model 

errors” and “phase errors”. Comparing GCM-driven RCM results with observations is 

therefore associated with large uncertainty and requires long time series, with seldom exist for 

short-term point precipitation observations. An alternative strategy, used in this experiment, is 

to drive the RCA3 model not by GCM results but by data from the ERA-40 reanalysis 

(Uppala et al., 2005). ERA-40 data from 60 vertical levels with resolutions 2º in space and 6 h 

in time were interpolated to the RCA 3 grid and then used to specify the boundary conditions 

for an RCA3 run. Thus the model boundary conditions were “perfect” (within the accuracy of 

the ERA40 data set), which provides the most suitable RCA3 results for comparison with 

observations and reduces the need for long time periods. In this experiment, the time period 

used extends from the mid-80’s (when high-resolution observations started) to year 2000 

(when the available ERA-40-driven RCA3 data ended), i.e. covering ~15 years. 

Precipitation observations from single gauges in three Swedish cities (Kalmar (1), Jönköping 

(2), Stockholm (3); Figure 1b), representing climate regions of both maritime and more inland 

character, were used. The observations were obtained by tipping-bucket gauges, recording the 

time when 0.2 mm had accumulated, and were aggregated to 30-min values to conform with 

the RCA3 output time step (Hernebring, 2006). 

DOWNSCALING METHODOLOGY 

A simple stochastic scheme was formulated and tested for downscaling the RCA3 

precipitation from a grid box to possible realisations of the precipitation in a point within the 

box (Figure 2). The method is based on not only total precipitation from the RCA3 model, but 

also on its two components: large-scale and convective precipitation (which when summed 

adds up to the total precipitation). For each of these components, a typical spatial coverage (% 

of grid box) needs to be estimated. For example, as illustrated in Figure 2, the rainfall 

typically produced by a convective system may perhaps occupy 1/9=11% (=cc) of the grid 

box (~250 km² with a grid size of 49×49 km). Thus the actual precipitation intensity is 9/1=9 

2 Downscaling of climate model precipitation


a b 

Figure 1. RCA3 domain (Europe; a) and location of gauges (Sweden; b). 

times the grid box mean intensity and it has 11% probability of occurring in a certain point 

within the grid box. Rainfall produced by a large-scale system (e.g. stratiform), on the other 

hand, may typically cover 89% (=cl) of the grid box, i.e. with 89% probability producing 

precipitation with an intensity of 9/8=1.125 the grid box mean in a certain point (Figure 2). 

In this experiment the overall applicability of the proposed simple scheme was assessed and 

the parameters roughly optimised for the three Swedish cities. As urban hydrological 

modelling is critically dependent upon a proper representation of precipitation extremes, in 

the optimisation of the scheme the main focus was on reproducing the observed 30-min 

precipitation extremes as expressed in Intensity-Duration-Frequency (IDF) curves. As shortterm 

precipitation extremes in the region are generally produced by convective systems, the 

key parameter in this case is cc. The large-scale precipitation, on the other hand, governs 

general statistical properties of 30-min time series such as the mean 30-min volume and the 

wet fraction (i.e. probability of a non-zero volume). It was however found difficult to 

reproduce the observed general statistics by optimising cl. The reason for this was that the 

RCA model too frequently generates small amounts of large-scale precipitation (Räisänen et 

al., 2004). To compensate for this, a relatively low value of cl had to be used to get a proper 

wet fraction, but as a consequence the generated intensities became overestimated. Therefore 

a different strategy was adopted, in which cl was omitted (or, equivalently, kept fixed at 1) 

and replaced by a threshold value tl below which all large-scale precipitation was ignored. 

Large-scale 

P 

Convective 

P 

Assumed coverage Actual intensity Probability in a point 

Figure 2. Schematic of the downscaling methodology. 

RCA/(8/9)=RCA*1.125 8/9 = 89% = c l 

RCA/(1/9)=RCA*9 1/9 = 11% = c c 

Olsson et al. 3 

2 

1 

3


RESULTS 

Figures 3-5 show some examples of results from the simple downscaling scheme, for the 

Kalmar gauge (nr. 1 in Figure 1b). The parameters were roughly optimised by trial-and-error 

and the final values obtained were cl=8% and tl=0.18 mm. All results have been averaged over 

10 realizations from the stochastic scheme, to obtain stable estimates 

Figure 3 shows the result in terms of IDF curves, return periods 15 (top left), 5 (top right), 3 

(bottom left) and 1 (bottom right) year(s). As expected, the short-term extremes in the grid 

box averages are substantially lower than in the point observations for short durations. For the 

maximum duration considered (1 day), however, the grid box values are in fact rather close to 

the observed IDF curves. Thus, at a daily scale the local (i.e. grid box) precipitation extremes 

simulated by the RCA3 model appear to well agree with point observations, which indicates 

that the RCA3 precipitation is realistic with respect to local extremes. Overall, the IDF curves 

obtained from the downscaled time series very well reproduce the observed curves, for all 

durations and return periods considered. There is a weak tendency towards underestimation, 

but only with a small amount. 

Figure 3. IDF curves for Kalmar from observed point series (solid), RCA3 grid box average 

series (dashed) and downscaled point series (dotted). 



Figure 4a shows the result in terms of four general statistical properties of the time series: (1) 

mean 30-min volume (mm), (2) standard deviation (mm), (3) mean non-zero 30-min volume 

(mm) and (4) probability of non-zero 30-min volume (-). For the grid box averages, the nonzero 

probability is substantially higher than in the observations. A higher value is expected for 

spatial averages than for point values, but the difference is likely exacerbated by the 

overestimated frequency of wet 30-min periods in the RCA3 data discussed above. The mean 

30-min volume is also overestimated in the grid box averages but to a lesser degree. In the 

downscaled series, all properties are close to the observed. Figure 4b shows the 

autocorrelation function for lags 1-6. In the observations, the autocorrelation is ~0.5 for lag 1 

and then decreases to ~0.25 for lag 6. For the time series of grid box averages, the value at lag 

1 is ~0.8 and at lag 6 ~0.6. It is expected to have a higher autocorrelation in spatial averages, 

but also in this case the difference compared with the point observations is probably amplified 

by the inaccuracies in the RCA3 data. The erroneously generated precipitation in the RCA3 

model often occurs in the form of long, uninterrupted sequences of similar (small) values, 

which increases autocorrelation. The autocorrelations in the downscaled series does not 

perfectly reproduce the shape of the observed function, but overall the level is very 

reasonable. 

a b 

Figure 4. General descriptive statistics (a) and autocorrelation functions (b) for Kalmar from 

observed point series (black bar, solid line), RCA3 grid box average series (grey bar, dashed 

line) and downscaled point series (white bar, dotted line). 

Figure 5 shows the result in terms of some event-related properties, mean and standard 

deviation of event duration (Figure 5a), event volume (Figure 5b) and length of dry period 

between events (Figure 5c). As expected from the results shown in Figures 3 and 4, and the 

associated discussion, as compared with point observations the events are longer and the dry 

periods shorter in the time series of grid box averages. The mean event volume is, however, 

well reproduces. After downscaling, also the properties of event durations are well 

reproduced. The dry periods lengths become somewhat overestimated though, possibly 

because the thresholding removes also some minor precipitation events that should have been 

regarded as “true”. 

a b c 

Figure 5. Statistical properties of event duration (a), event volume (b) and dry period length 

(c) for Kalmar from observed point series (black), RCA3 grid box average series (grey) and 

downscaled point series (white). 

Olsson et al. 5


CONCLUSIONS AND FUTURE WORK 

From this preliminary evaluation, it appears possible to combine RCM precipitation output, 

separated into components representing convective and large-scale precipitation, with a 

simple stochastic scheme reflecting spatial coverage, to create reasonably accurate realisations 

of the associated point rainfall process. The downscaled point precipitation series compare 

well with observations both in terms of general descriptive statistics and extreme values. The 

results support the possibility to statistically relate RCM results to short-term point 

observations. One possible application discussed above is to use dynamically downscaled 

GCM output to assess future changes in point precipitation, which however will require a 

very careful treatment of the calibration for present climate. However, also other applications 

may be envisioned, e.g. simulating point precipitation properties in ungauged areas. 

We believe the present methodology may constitute a step forward in the search for tools to 

estimate future point precipitation. Most efforts to date have involved different versions of socalled 

delta change, where an observed rainfall time series is modified in line with changes 

found in RCM output. The proposed, process-based methodology is likely more flexible than 

delta change, which is heavily constrained by the properties of the available observations. It 

should however be emphasised that it is associated with uncertainty, e..g. because of the 

simplified descriptions of rainfall generating emchanisms in climate models. 

Future work will include the following activities: (1) optimisation of the coverage parameters 

(cc and cl) for the different cities, sensitivity testing, split sample calibration, (2) estimation of 

coverage not as fixed areas but as given by additional RCA3 output parameters, e.g. 

cloudiness, (3) evaluation and optimisation for not only ERA-40-driven RCA3 results but for 

GCM-driven RCA3 results representing present climate, (4) application of optimised 

downscaling scheme to GCM-driven RCA3 results representing future climate. 

ACKNOWLEDGEMENT 

The study was funded by The Swedish Research Council for Environment, Agricultural 

Sciences and Spatial Planning (FORMAS) and The Foundation for Strategic Environmental 

Research (MISTRA). 

REFERENCES 

Hernebring C. (2006). 10-års regnets återkomst förr och nu (Design storms in Sweden then and now). VA-forsk 

publ. 2006-04, Svenskt Vatten, Stockholm (in Swedish). 

IPCC (2007). Climate Change 2007: The Physical Science Basis, Summary for Policymakers. Contribution of 

Working Group I to the Fourth Assessment Report of the IPCC 

Kjellström E., Bärring L., Gollvik S., Hansson U., Jones C., Samuelsson P., Rummukainen M., Ullerstig A., 

Willén U. and Wyser K. (2005). A 140-year simulation of European climate with the new version of the 

Rossby Centre regional atmospheric climate model (RCA3). Reports Meteorology and Climatology 

108, SMHI, Sweden, 54 pp. 

Räisänen J., Hansson U., Ullerstig A., Döscher R., Graham L.P., Jones C., Meier H.E.M., Samuelsson P. and 

Willén U. (2004). European climate in the late twenty-first century: regional simulations with two 

driving global models and two forcing scenarios. Clim. Dyn., 22, 13-31. 

Uppala S.M., Kållberg P.W., Simmons A.J., Andrae U., da Costa Bechtold V., Fiorino M., Gibson J.K., Haseler 

J., Hernandez A., Kelly G.A., Li X., Onogi K., Saarinen S., Sokka N., Allan R.P., Andersson E., Arpe 

K., Balmaseda M.A., Beljaars A.C.M., van de Berg L., Bidlot J., Bormann N., Caires S., Chevallier F., 

Dethof A., Dragosavac M., Fisher M., Fuentes M., Hagemann S., Hólm E., Hoskins B.J., Isaksen L., 

Janssen P.A.E.M., Jenne R., McNally A.P., Mahfouf J.-F., Morcrette J.-J., Rayner N.A., Saunders 

R.W., Simon P., Sterl A., Trenberth K.E., Untch A., Vasiljevic D., Viterbo P. and Woollen J. (2005). 

The ERA-40 re-analysis. Quart. J. R. Meteorol. Soc., 131, 2961-3012.

Downscaling of Regional Climate Model Precipitation for Urban ...

Create successful ePaper yourself

Delete template?

Save as template?