Hydrological Sciences Journal

ISSN: 0262-6667 (Print) 2150-3435 (Online) Journal homepage: https://www.tandfonline.com/loi/thsj20

Driving a lumped hydrological model with

precipitation output from weather generators of
different complexity

Korbinian Breinl

To cite this article: Korbinian Breinl (2016) Driving a lumped hydrological model with precipitation
output from weather generators of different complexity, Hydrological Sciences Journal, 61:8,
1395-1414, DOI: 10.1080/02626667.2015.1036755

To link to this article: https://doi.org/10.1080/02626667.2015.1036755

Published online: 20 Apr 2016.

Submit your article to this journal

Article views: 694

View related articles

View Crossmark data

Citing articles: 11 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=thsj20
HYDROLOGICAL SCIENCES JOURNAL – JOURNAL DES SCIENCES HYDROLOGIQUES, 2016
VOL. 61, NO. 8, 1395–1414
http://dx.doi.org/10.1080/02626667.2015.1036755

Driving a lumped hydrological model with precipitation output from weather

generators of different complexity
Korbinian Breinl
Department of Geoinformatics, Paris-Lodron University of Salzburg, Salzburg, Austria

ABSTRACT ARTICLE HISTORY

This paper deals with the question of whether a lumped hydrological model driven with lumped Received 9 August 2014
Accepted 25 March 2015
daily precipitation time series from a univariate single-site weather generator can produce equally
good results compared to using a multivariate multi-site weather generator, where synthetic EDITOR
precipitation is first generated at multiple sites and subsequently lumped. Three different D. Koutsoyiannis
weather generators were tested: a univariate “Richardson type” model, an adapted univariate
ASSOCIATE EDITOR
Richardson type model with an improved reproduction of the autocorrelation of precipitation
C. Perrin
amounts and a semi-parametric multi-site weather generator. The three modelling systems were
evaluated in two Alpine study areas by comparing the hydrological output with respect to KEYWORDS
monthly and daily statistics as well as extreme design flows. The application of a univariate single-site weather
Richardson type weather generator to lumped precipitation time series requires additional generator; multi-site
attention. Established parametric distribution functions for single-site precipitation turned out weather generator;
to be unsuitable for lumped precipitation time series and led to a large bias in the hydrological hydrological modelling; HBV;
parametric distributions
simulations. Combining a multi-site weather generator with a hydrological model produced the
least bias.

1 Introduction straightforward but ignore the serial correlation of

precipitation amounts. Another often cited single-
Coupling stochastic weather generation with hydrological
site model is the semi-parametric LARS-WG (Long
modelling is becoming increasingly important in
Ashton Research Station-Weather Generator)
hydrological impact assessment (Haberlandt et al.
(Racsko et al. 1991, Semenov and Barrow 1997,
2011). It is an alternative approach to obtaining
Semenov et al. 1998), where wet and dry spells as
design flows based on discharge observations (e.g.
well as precipitation amounts are sampled from
Grimaldi and Serinaldi 2006, Merz and Blöschl
semi-empirical distributions. Multi-site weather
2008, Keef et al. 2010). Moreover, the combination
generators are usually based on Markov processes
of a weather generator and a hydrological model
or resampling techniques. Wilks (1998), Brissette
allows us to study the impact and uncertainty of
et al. (2007) and Khalili et al. (2009) extended the
changing land use or climate conditions (e.g.
univariate Richardson type model to multiple sites,
Bergström et al. 2001a, Booij 2005, Raff et al.
which comes with technical challenges such as the
2009, Holzmann et al. 2010). To date, numerous
generation of correlated random numbers that are
concepts have been proposed for weather generation
more correlated than the observed occurrences.
models. The generation of weather variables at a
Furthermore, temporal correlations are underesti-
single site is less complex compared to multi-site
mated. Nearest neighbour resampling has been sug-
methods, which have to maintain not only single-
gested by various authors, including Beersma and
site but also inter-site statistics. The most common
Buishand (2003), Wilby et al. (2003), Leander and
univariate approach is the “Richardson type” model
Buishand (2009) and King et al. (2014). Resampling
that first simulates binary precipitation occurrences
is conceptually straightforward but there is a trade-
with a two-state Markov chain followed by the
off between the subjective specification of various
sampling of precipitation amounts from parametric
model variables and the duplication of observations
distribution functions (e.g. Katz 1977, Richardson
in the simulations (Buishand and Brandsma 2001,
1981). Richardson type models are conceptually
Mehrotra et al. 2006). Other concepts have been

CONTACT Korbinian Breinl korbinian.breinl@sbg.ac.at

© 2016 IAHS
1396 K. BREINL

proposed by Hughes et al. (1999), Ailliot et al. (2009), 2 Methodology

Bárdossy and Pegram (2009), Serinaldi (2009a),
Three different weather generators were applied: were a
Baigorria and Jones (2010) and Breinl et al. (2015).
univariate Richardson type model, an adapted univari-
For hydrological modelling in probabilistic analyses,
ate Richardson type model with an improved repro-
lumped or semi-distributed models may be preferable
duction of the autocorrelation of precipitation amounts
to distributed models, which can be computationally
and a semi-parametric multi-site weather generator.
costly and typically require high-resolution input data
These three models were called “Model A”
(Arnold et al. 1998, Romanowicz et al. 2005, Shrestha
(Richardson type), “Model A+” (adapted Richardson
et al. 2006). Conceptualized models are considerably
type) and “Model B” (multi-site). Goodness-of-fit tests
faster and allow multiple runs and uncertainty ana-
were used to find the optimum distribution functions
lyses at low computational costs.
for simulating precipitation amounts for both models
Only a few studies have focused on the impact of the
A/A+ and Model B, i.e. for site-specific and lumped
input precipitation from a weather generator on the
precipitation time series. Various studies examined the
hydrological model output. Harmel et al. (2000) exam-
applicability of different parametric distribution func-
ined the impact of precipitation on the hydrologic
tions for simulating daily precipitation in different
response of a small catchment in the USA (0.53 km2)
regions (e.g. Wilks 1999a, Serinaldi 2009b, Singh and
using a single-site Richardson type weather generator
Mishra 2011, Breinl et al. 2013, Li et al. 2013). Past
which was parameterized with four different precipita-
research has shown that the very common light-tailed
tion distribution functions. The weather generator was
gamma distribution is often not the best choice as it
fitted to records from a single centrally-located rain
underestimates extremes (Vlček and Huth 2009, Breinl
gauge. Li et al. (2013) examined the impact of six
et al. 2013). Often, heavy-tailed distributions such as
different precipitation distribution functions on the
the Generalized Pareto distribution (GPD) are prefer-
hydrological model output in two Canadian catch-
able (Papalexiou et al. 2013). In this research, four
ments with areas of 9700 and 3330 km2. They also
different distribution types were tested, including the
used a univariate Richardson type model but for gen-
Exponential, the Gamma, the Weibull and a heavy-
erating lumped catchment precipitation derived from
tailed hybrid distribution consisting of the empirical
24 raingauges instead of precipitation at a single site.
distribution of the observation data for low and the
Garavaglia et al. (2010) compared the Exponential and
GPD for high precipitation amounts (“Hybrid”). For
Generalized Pareto distributions with the so-called
the hydrological model, the lumped conceptual model
multi-exponential weather pattern distribution
HBV was chosen. Two different Alpine catchments
(MEWP), of which the latter was used by Paquet
(Salzach catchment in Austria with 4637 km2 and
et al. (2013) in combination with a hydrological
Ubaye catchment in France with 548 km2) were
model for extreme flood estimation in a French catch-
examined.
ment (2170 km2).
The structure of the paper is as follows: first, the
The present study deals with the question of
data and methods are described. Methods include the
whether a lumped hydrological model driven with
parametric distribution functions, the weather genera-
lumped daily precipitation averages from a univari-
tor types (Model A, Model A+ and Model B), the
ate weather generator, as, for example, suggested by
hydrological model as well as the evaluation proce-
Li et al. (2013), can produce equally good results
dure. Second, the results are presented for Model A,
compared to using a multivariate multi-site weather
Model A+ and Model B and each catchment, followed
generator, where synthetic precipitation is first gen-
by a discussion and a conclusion.
erated at multiple sites and subsequently lumped. In
the study by Li et al. (2013), “the averaged precipi-
tation distribution for each watershed shared the 2.1 Weather generator
same characteristics as those of the 24 stations” and
for this reason “using the stochastic weather genera- All weather generator models were set up on a monthly
tor based on the basin-averaged time series” was basis to capture the seasonality of precipitation. A pre-
“preferred to the much more complex option of gen- cipitation threshold of 0.1 mm for defining wet days
erating time series at each station using a multisite was applied.
approach and averaging in a second step”. However,
it remains unclear whether this is always possible, 2.1.1 Model A
which was one of the main motivations behind this For Model A (univariate), a straightforward model was
paper. chosen which is similar to the model by Richardson
 HYDROLOGICAL SCIENCES JOURNAL – JOURNAL DES SCIENCES HYDROLOGIQUES
(1981). Daily precipitation occurrences were simulated
with a Markov process where the transition probabil-
ities depend on m previous days:
PRfXtþ1 jXt ;Xt1 ;Xt2 ;...;X1 g
¼ PRfXtþ1 jXt ;Xt1 ;...;Xtm g with m < t 1 (1)
A two-state Markov chain was used for simulating
dry and wet days. Precipitation amounts on wet days
were sampled from one of the four distribution func-
tions (separate distribution parameters for each
month) as described in Section 2.2. Before fitting
the parametric distributions, the daily amounts of
all raingauges were weighted into lumped precipita-
tion time series by Thiessen polygons. Other meth-
ods including inverse distance weighting (IDW),
ordinary kriging (OK) and kriging with external
drift (KED) considering elevation were tested using
the “R” software package “hydroTSM” (Zambrano-
Bigiarini 2012). The computed distribution para-
meters did not differ significantly between methods,
except for KED in the Ubaye catchment. However,
for Ubaye, KED was considered less reliable as it
produced almost 5% negative precipitation values, a
known issue related to this method (Haberlandt
2007, Ly et al. 2011). Moreover, using the question-
able KED output for hydrological model calibration
likewise resulted in considerably lower Nash-Sutcliffe
efficiencies (Nash and Sutcliffe 1970) in both the
calibration and validation periods (−10.8%/–7.2%).
Mean daily temperature was simulated independently
by autoregressive moving average processes (ARMA).
Fourier time series were first fitted to represent the
annual cycles of temperature and standard deviation,
and the remaining residuals were simulated using the
ARMA process. See Breinl et al. (2015) for details.
The principle of Model A is schematized in Fig. 1.
2.1.2 Model A+
Model A+ is an adapted version of the univariate
Richardson type model as described in Section 2.1.1,
and has been developed for this study to reduce the
underestimation of the autocorrelation of precipita-
tion amounts in Model A. The latter is caused by the
random sampling of precipitation amounts on any
rainy day, ignoring the serial correlation of precipi-
tation amounts in the observations. Instead of simu-
lating precipitation occurrences with a two-state
Markov chain for dry and wet days, in Model A+,
the precipitation amounts are simulated with a six-
state Markov chain, representing dry days and five
different precipitation intensity classes (i.e. rainy
1397
days). Class one contains the lowest, class five the
highest observed precipitation amount. The six
intensity classes were estimated by k-means cluster-
ing (for details about k-means clustering see
Hartigan 1975 or Hartigan and Wong 1979). The
squared Euclidean distance was used to build the
classes. Once the occurrences have been generated,
precipitation amounts are first resampled from the
observations so that a randomly drawn observed day
of a specific class and month is assigned a synthetic
day that matches both class and month. Following
this, parametric precipitation amounts are generated
for each rainy day and reshuffled so that the ranks of
the resampled and parametric precipitation amounts
match. The reshuffling is applied over all simulated
years but separately for each month to make the
resampled and synthetic (i.e. parametric) time series
as similar as possible. Related reshuffling algorithms
have been described by Clark et al. (2004) and Breinl
et al. (2013). The choice of the number of precipita-
tion intensity classes is subjective. There is a trade-
off between the number of classes and the number of
observations in each class. In this study, the number
varies between 109 (highest intensity) and 1656
observations (lowest intensity) in the Salzach catch-
ment, and 61 and 3240 observations in the Ubaye
catchment. As in Model A, the daily amounts of all
raingauges were averaged/weighted into lumped pre-
cipitation time series by Thiessen polygons before
fitting the precipitation distributions. Temperature
was simulated in the same way as for Model A.
The principle of Model A+ is schematized in Fig. 1.
2.1.3 Model B
The multi-site Model B was likewise set up on a monthly
basis and is described in detail in Breinl et al. (2015).
They presented a model based on a precipitation algo-
rithm by Breinl et al. (2013). In this algorithm, daily
snapshots of precipitation fields (i.e. catchment-wide
precipitation patterns) called “amount vectors” are first
clustered into similar classes representing similar types
of patterns, which are then simulated using a Markov
process. The number of classes has direct impact on the
duplication of observations in the simulation output.
The number of classes was empirically chosen so that
the duplication of observations in the simulation was
below 1%. Once time series of classes have been simu-
lated, each single class is replaced by a randomly drawn
observed “amount vector” that matches the simulated
class and month. In the last two steps, precipitation
amounts are sampled from parametric distributions (at
each site separately) and reshuffled according to ranks of
the observed amount vectors to maintain the temporal
1398 K. BREINL

Figure 1. Schematics of the three weather generators, Model A, Model A+ and Model B. Model B is shown for an exemplary number
of three sites and 10 clusters.

and spatial correlation structures. Breinl et al. (2015) consisting of the empirical distribution for lower and
called this model the reduced-state-space model (RSS). the GPD for higher precipitation amounts.
The principle is schematized in Fig. 1 (right). The reader The probability density function (pdf) of the
is referred to Breinl et al. (2015) for more details on the Exponential distribution is given by:
algorithm. In Model B, the precipitation amounts from
all simulated sites were subsequently weighted by fE ðxÞ ¼ α expðαxÞ; α > 0 (2)
Thiessen polygons into lumped time series. As in the
cases of models A and A+, mean daily temperature was
simulated independently using an ARMA process. The
principles of the simulation procedures with Model A, where α is the single parameter of the distribution.
Model A+ and Model B are shown in Fig. 2. The orders The pdf of the Gamma distribution is given by:
of the Markov processes in models A/A+ and B were  
βα xα1 x
estimated using the Akaike information criterion, AIC fG ðxÞ ¼ exp ; α > 0; β > 0 (3)
(Katz 1981) and the Bayesian information criterion, BIC ΓðαÞ β
(Schwarz 1978). For all three models, first-order pro-
cesses are appropriate for all months in both study areas.
Daily mean temperature was simulated using time series where α is the shape and β is the scale parameter.
extracted from the centroid of an E-OBS grid cell in the The pdf of the Weibull distribution is given by:
Ubaye catchment and from three sites (15, 16 and 19) in
    α 
the Salzach catchment (Fig. 3 and Section 2.4). In the α x α1 x
Salzach catchment, the arithmetic mean of the three fW ð x Þ ¼ exp  ;α > 0;β > 0 (4)
β β β
temperature sites was used as input for the hydrological
model. The four parametric precipitation distributions
described in Section 2.2 were tested for all models.
where α is the shape and β is the scale parameter.
The pdf of the Hybrid distribution is defined by:

2.2 Parametric precipitation distributions fH ðx;PH Þ

Four different parametric precipitation functions were
tested in models A/A+ and Model B, which are (a) the
one-parameter Exponential distribution, (b) the two-
parameter Gamma distribution, (c) the two-parameter
Weibull distribution and (d) a Hybrid distribution
1  (5)
¼ fEmp ðxÞI ðx  uÞ þfGP ðx;κ;σ;uÞIðx > uÞ
Z
where fEmp ðxÞ is the pdf of the empirical distribution of
observation data and fGP ðx; κ; σ; uÞ is the GPD located
at u with shape parameter κ and scale parameter σ:
 HYDROLOGICAL SCIENCES JOURNAL – JOURNAL DES SCIENCES HYDROLOGIQUES 1399

The Hybrid distribution f H is not a continuous five-

parameter distribution but two distributions con-
nected at the threshold x ¼ u. The density of the
Hybrid function is integrated to one by Z. There
are a various publications examining the tail of dis-
tributions for precipitation amounts (e.g. Wilks
1999b, Vrac and Naveau 2007, Furrer and Katz
2008, Qian et al. 2008, Hundecha et al. 2009,
Hashmi et al. 2011). They all show that probability
distributions applied over all samples may not ade-
quately capture extreme amounts (Li et al. 2013,
Papalexiou et al. 2013). For this reason, compound
distributions were suggested, such as the mixed-
exponential distribution (Foufoula-Georgiou and
Lettenmaier 1987, Wilks 1998), or a hybrid distribu-
tion consisting of a light-tailed distribution for low
and a GPD for higher and extreme precipitation
amounts (e.g. Vrac and Naveau 2007, Furrer and
Katz 2008, Hundecha et al. 2009). Threshold selec-
tion for merging the two distributions is still an
open research topic (Coles 2001, Deidda 2010,
Figure 2. Simulation procedures using models A/A+ and Model B.
Lucio et al. 2010). Merging can be achieved dyna-
mically (e.g. Vrac and Naveau 2007, Hundecha et al.

Figure 3. Overview of the two study areas including location of the rain and river gauges. In the Ubaye study area, temperature
records were taken from grid-based observation data (E-OBS). In the Salzach study area, temperature records were available at sites
15, 16 and 19. Precipitation records of site 19 were not considered.

  1
σ 1 ðx uÞ 1κ 2009), or through peak-over-threshold (POT) analy-
fGP ðxÞ ¼ ¼ 1 þκ ; σ > 0;u  0 (6) sis (e.g. Furrer and Katz 2008). Dynamic merging of
κ σ σ
the distributions types is numerically unstable, data
1400 K. BREINL
sensitive and computationally expensive (Singh and
Mishra 2011). However, the POT method can be
subjective and (for large observation networks)
very time-consuming. In this paper, thresholds
were visually estimated with “mean residual life
plots” and “parameter stability plots” using the sta-
tistical software ExtRemes in “R” (see Gilleland and
Katz 2011).
2.3 Hydrological model
The present study requires a large number of mod-
elling runs for probabilistic analyses and to reliably
estimate uncertainty. Unlike conceptual hydrologi-
cal models, fully-distributed models can simulate
physical processes but they likewise come with
numerous challenges (Beven 1989, Refsgaard et al.
2010). Most importantly, though, they are compu-
tationally expensive and data demanding (Arnold
et al. 1998, Romanowicz et al. 2005, Shrestha et al.
2006). In addition, a higher-resolution model does
not always improve the model performance (Das
et al. 2008). The hydrological model HBV is a con-
ceptual model that has been applied in numerous
studies (e.g. Uhlenbrook et al. 1999, Bergström
et al. 2001b, Booij 2005, Hagg et al. 2006,
Bárdossy and Das 2008, Das et al. 2008, Steele-
Dunne et al. 2008, Gao et al. 2012, Beck et al.
2013). In this study, the HBV-light model was
used in a lumped set-up with 10 different elevation
zones for the distributed snow routine (to improve
the simulation of snow retention), and with three
groundwater boxes (Seibert 1997, 1999, 2000,
Seibert and Vis 2012). The 10 elevation zones were
computed in a GIS using a natural breaks classifica-
tion according to Jenks (1967). HBV-light generates
daily mean discharge using daily precipitation, daily
mean temperature and potential monthly evapora-
tion. The model essentially consists of four main
components: a snow routine, a soil moisture rou-
tine, a response routine and a routing routine. The
15 model parameters are calibrated to discharge
observations. The model was calibrated using a
GAP routine (Seibert 2000), consisting of a genetic
algorithm (GA), followed by Powell’s (P)
Table 1. Names, coordinates, altitudes and annual precipitation statistics of the gauge stations in the Ubaye study area.
ID Name Lat Long Altitude
(m .a.s.l.)
1 Barcelonnette 44.3915 6.6718 1155
2 Condamine 44.4598 6.7449 1325
3 Saint-Paul 44.5936 6.8424 1903
4 Uvernet 44.3186 6.6925 1660
quadratically-convergent method (Press et al.
2002). As different parameter sets in conceptual
models can lead to equally good simulation results
(“model equifinality”, see Beven 2001), 100 different
but suitable parameter sets were calibrated for each
catchment. In the simulations, the model was run
separately for each of the 100 parameter sets and
the output was subsequently averaged. A 10-year
calibration and validation period in both catch-
ments resulted in an average Nash-Sutcliffe effi-
ciency (Nash and Sutcliffe 1970) of 0.87/0.82 for
Salzach and 0.82/0.74 for Ubaye.
2.4 Study areas and data
Two catchments were chosen for this research: the
Salzach catchment in Germany and Austria down to
the city of Salzburg (4637 km2), and the Ubaye
Catchment in France down to the village of
Barcelonnette (548 km2) (Fig. 3). The discharge of
the Salzach River is dominated by snowmelt.
Extreme discharges and floods usually occur in
summer as a combination of snowmelt and long-
lasting rainfall, or rainfall caused by Vb weather
situations (Mudelsee et al. 2004). Extreme dis-
charges in the Ubaye River are generally associated
with snowmelt or high-intensity summer storms
(Turkington et al. 2014). For Salzach, daily precipi-
tation records were obtained for 1987–2010 at 18
sites. Temperature records were available at three
sites (15, 16, 19). For Ubaye, precipitation records
were available from 1971–2004 at four sites. As
gauge station based temperature records in the
area had multiple missing years and major gaps,
gridded observed temperature data from the
ENSEMBLES project (E-OBS dataset, 0.25 degree
regular grid) (Haylock et al. 2008) were used.
E-OBS data have been successfully used in other
flood hazard related studies (e.g. Photiadou et al.
2011, Schädler et al. 2012, Freudiger et al. 2014).
E-OBS temperature data show few inhomogeneities
in southeastern France (Hofstra et al. 2009).
Temperature time series were extracted from the
grid cell closest to the catchment centre (Fig. 3).
Tables 1 and 2 give details on the gauge stations.
Mean StDev 5th percentile 25th percentile 75th percentile 95th percentile
(mm) (mm) (mm) (mm) (mm) (mm)
711 140 481 629 795 984
663 143 472 569 782 874
924 207 569 765 1101 1206
987 214 688 857 1136 1332
 HYDROLOGICAL SCIENCES JOURNAL – JOURNAL DES SCIENCES HYDROLOGIQUES 1401

Table 2. Names, coordinates, altitudes and annual precipitation statistics of the gauge stations in the Salzach study area.
ID Name Lat Long Altitude Mean StDev 5th percentile 25th percentile 75th percentile 95th percentile
(m .a.s.l.) (mm) (mm) (mm) (mm) (mm) (mm)
1 Gerlos 47.2269 12.0444 1250 1278 105 1124 1180 1350 1443
2 Paß Thurn 47.2997 12.4222 1200 1206 131 1004 1096 1312 1402
3 Hochfilzen 47.4703 12.6217 960 1781 185 1479 1643 1925 2080
4 Felbertauerntunnel 47.1181 12.5056 1650 1403 133 1184 1287 1488 1624
5 Böckstein 47.0872 13.1158 1140 1342 132 1158 1268 1363 1622
6 Flachau 47.3472 13.3958 910 1154 93 1024 1087 1206 1304
7 Golling-Torren 47.5917 13.1575 473 1474 104 1293 1413 1533 1656
8 Gosau 47.5892 13.5431 765 1742 177 1460 1613 1895 2020
9 Hintersee 47.7611 13.2225 750 2035 246 1592 1850 2240 2388
10 Bischofswiesen-Loipl 47.6525 12.9315 845 1894 175 1575 1786 2044 2168
11 Hüttschlag 47.1772 13.2314 1030 1096 87 951 1038 1160 1239
12 Glanegg 47.7494 13.0175 450 1550 153 1284 1449 1641 1814
13 Enzingerboden 47.1667 12.6333 1768 1438 201 1120 1255 1573 1811
14 Schmittenhöhe 47.1786 12.7381 2102 1518 143 1315 1407 1652 1739
15 Zell am See 47.3267 12.7953 767 1175 98 1023 1114 1242 1351
16 St. Veit 47.3292 13.1550 747 1102 106 913 1052 1159 1301
17 Dienten 47.3908 13.0333 1265 1480 138 1282 1374 1640 1699
18 Hüttau 47.4131 13.3156 720 1138 102 984 1064 1201 1321
19 Salzburg Airport 47.8014 13.0017 424 1278 105 1124 1180 1350 1443

3 Model evaluation procedure procedure helps to evaluate the performance of the

3.1 Weather generation
Detailed validations of univariate Richardson type
models (Model A) can be found in the literature (e.g.
Semenov et al. 1998). The performance of Model A+ is
comparable to a classic Richardson type model but the
autocorrelation of precipitation amounts is better
reproduced (see below). The multi-site weather gen-
erator by Breinl (i.e. Model B) has been thoroughly
tested and validated in Breinl et al. (2015). For the
sake of comparison, the univariate catchment precipi-
tation time series generated by the three models were
evaluated with respect to three crucial parameters: wet
spells and the autocorrelation of precipitation occur-
rences and amounts. The four different parametric
distributions were evaluated with three different good-
ness-of-fit tests, the chi-square test at the 5% level, the
root mean square error (RMSE), and the Q-Q plot
correlation (Filliben 1975). The return periods of the
annual precipitation totals and annual maximum daily
precipitation were plotted for the observations and the
four different parametric distributions. The monthly
maxima of daily precipitation for each distribution
and for different percentiles were likewise plotted.
Finally, the monthly means of generated precipitation
amounts were directly compared with the hydrological
output (Section 3.2). In addition to this, the hydrolo-
gical simulation output was compared with bootstrap
simulations, meaning weather generation time series
using resampled observations instead of parametric
precipitation amounts. To achieve this, observations
were resampled but the parametric sampling and
reshuffling procedures were not conducted. This
different parametric distributions without the bias of
the HBV model. For a reliable evaluation of low fre-
quency flood events (see Section 3.2), 50 weather time
series of 120-year length (6000 years) were generated
for both study areas.
3.2 Hydrological output
The simulated hydrological output was compared with
the observations as well as the bootstrap time series
with respect to the mean monthly and mean daily
discharge. For extreme design flows, evaluations were
made for the Q10, Q20 and Q50 year return period of
annual, winter/spring and summer/autumn flows.
Higher return periods were not taken into account
due to the comparatively short length of observation
time series. As there were only 24 (34 in Ubaye) obser-
vation years, a parametric model was fitted to the
annual maximum flows to estimate the observed Q10,
Q20 and Q50 values. The GEV distribution turned out
to be the most suitable model for both study areas
(based on Q-Q plots), except for the winter/spring
season in Ubaye (log-logistic distribution).
Confidence intervals of the GEV and log-logistic dis-
tribution were estimated by parametric bootstrapping
(Reiss and Thomas 2007). Return periods and confi-
dence intervals of the simulations were estimated using
a “balanced resampling” approach by Burn (2003). The
evaluation was conducted for the two study areas, the
three weather generator models, the bootstrap time
series as well as the four parametric distribution
functions.
1402 K. BREINL
4 Model evaluation
4.1 Weather generation
The performance of all three models is comparable
with respect to the autocorrelation of precipitation
occurrences (Fig. 4). In the Salzach catchment, the
lag1 autocorrelation of occurrences is underesti-
mated by 3.8% in Model A, 3.9% in Model A+
and 4.8% in Model B. The difference is more pro-
nounced for amounts (−74.4%, −13.8% and
−17.7%). Model A+ performs best. For Ubaye, the
lag1 autocorrelation of occurrences is underesti-
mated by 4.2% in models A and A+ and 8.0% in
Model B. The same difference in the model perfor-
mances can be detected for amounts (−68.3%,
−13.1% and −25.8%). For higher lags (2–5), there
is a tendency to underestimate the autocorrelation
of occurrences in all models and both study areas
(on average −37.7%, −33.5% and −35.3% in Salzach
and −33.3%, −24.1% and −26.4% in Ubaye). For
amounts, there is a tendency to underestimate and
overestimate in Salzach (on average −21.9%, +32.0
and +23.7%). In Ubaye all models underestimate
(−66.3%, −32.1%, −46.7%). The analysis was con-
ducted with the bootstrap (i.e. weather generation
Figure 4. Autocorrelation of catchment precipitation occurrences (occ) and amounts (amt) for Model A, Model A+ and Model B in
both catchments.
only with resampled precipitation amounts, see
above). Results for mean (weighted) precipitation
amounts were comparable when simulating with
any of the four parametric distribution functions.
The differences between Model A, Model A+ and
Model B are less pronounced for generating wet
spells. The mean length of wet spells (7.2 days) in
the Salzach catchment is slightly underestimated by
the three models (−2.9%, −2.7% and −2.6%). The
95th percentile of wet spells is reproduced similarly
by all three models (observation: 23 days; simula-
tions: −11.0%, −6.3%, −8.8%). Results are similar for
Ubaye. Wet spells (mean of observations: 2.8 days)
are underestimated in Model A (−2.6%), Model A+
(−2.6%) and Model B (−9.1%). The 95th percentile
is reproduced well (observation: 7 days; simulations:
0.0%, 0.0%, −3.4%).
The three different goodness-of-fit tests for the
four parametric distributions are shown in Table 3.
For models A/A+ and the Salzach catchment, the
Hybrid distribution seems to be most appropriate,
although the Gamma distribution seems to be
equally valid. The results are similar for the Ubaye
catchment. As high values of the RMSE and lower
Q-Q correlations indicate, the Exponential and
 HYDROLOGICAL SCIENCES JOURNAL – JOURNAL DES SCIENCES HYDROLOGIQUES 1403

Table 3. Goodness-of-fit tests for the four distribution functions For Model B and the Salzach catchment, the
and for the three model set-ups and two study areas: chi-square Exponential distribution is the worst choice.
test at 5% level, RMSE and Q-Q correlation (average of all fits). Hybrid distributions seem to capture the observa-
Rejections chi- RMSE Q-Q
square test (mm) correlation tions well without any rejection produced by the
Catchment precipitation chi-square tests. Acceptable results are also achieved
Salzach models A/A+ Exponential 11 1.7803 0.9878 with Gamma and Weibull distributions, as the aver-
(12 fits) Gamma 0 0.8113 0.9930
Weibull 3 1.9223 0.9799 age RMSE and Q-Q correlations indicate. For the
Hybrid 0 0.7045 0.9948 Ubaye catchment, the performance of the Weibull
Ubaye models A/A+ Exponential 12 2.8977 0.9792
(12 fits) Gamma 0 0.9308 0.9948
and Hybrid distributions is similar. As in the case
Weibull 1 1.5221 0.9859 of Salzach, the Exponential distribution seems to be
Hybrid 0 1.0112 0.9941 the worst choice. The Gamma distribution performs
Site-specific precipitation better than the Exponential distribution but worse
Salzach Model B (216 Exponential 66 1.6825 0.9867
fits) Gamma 33 1.316 0.9881 than the Weibull and Hybrid distributions. When
Weibull 26 1.2396 0.9887 evaluating the annual maximum (catchment) preci-
Hybrid 0 1.127 0.9911
Ubaye Model B (48 Exponential 24 2.7651 0.9879 pitation (Fig. 5), a large spread in models A/A+
fits) Gamma 15 1.7284 0.9905 becomes obvious. For models A/A+ and both
Weibull 5 1.3826 0.9929
Hybrid 2 1.3845 0.9914 study areas, the Weibull distribution performs
worst. The Hybrid distribution performs consider-
ably better. In general, more bias can be detected
Weibull distributions appear to be less suitable for Salzach. The multi-site Model B is less sensitive
when fitting to lumped precipitation time series. to different parametric curves. In the Salzach area,
In particular the Exponential distribution causes a Hybrid and Weibull distributions outperform the
high number of rejections in the chi-square tests. Exponential and Gamma distributions. In Ubaye,

Figure 5. Return period of annual maximum daily (catchment) precipitation for the three model types, both study areas and the
four different parametric distribution functions implemented.
1404 K. BREINL

Figure 6. Return period of annual (catchment) precipitation totals for the three model types, both study areas and the four different
parametric distribution functions implemented.

all distributions except the Exponential appear to be
reasonable. The return periods of annual (catch-
ment) precipitation totals are plotted in Fig. 6.
4.2 Hydrological output
4.2.1 Monthly and daily statistics
For Salzach and models A/A+, the mean monthly
precipitation amounts are well reproduced by all
distributions except the Weibull distribution,
which tends to overestimate, leading to an overesti-
mation of daily and monthly mean flows (Fig. 7).
The overestimation is caused by the poor para-
metric fits for lumped precipitation time series as
discussed above. Model B performs well with all
distribution functions, reproducing the bootstrap
precipitation amounts and flows very well. Some
bias of the HBV model can be detected regardless
of the input precipitation. This bias is represented
by differences between the flows produced from
bootstrap input and the actual observations.
Models A+ and B better reproduce mean summer
flows compared to Model A, presumably due to a
better reproduction of the autocorrelation of
precipitation. In the Ubaye study area, differences
between types of models are generally less pro-
nounced (Fig. 8). In models A and A+, mean
monthly precipitation is well reproduced by all dis-
tribution functions. As in the Salzach study area,
the Weibull distribution tends to overestimate, at
least in the second half of the year. In Model B,
the hybrid distribution tends to overestimate preci-
pitation in the spring season. However, the impact
of the distribution functions on the resulting mean
monthly and daily flows is small. Model B performs
better than models A and A+ in the summer season.
In autumn, models A+ and B tend to slightly over-
estimate the flows. Model A performs better, but it
is possible that a poor reproduction of autocorrela-
tion compensates for too high precipitation
amounts from the parametric distributions. In gen-
eral, the bias from the HBV model (i.e. bias between
bootstrap and observations) is less pronounced
compared to the Salzach study area.
4.2.2 Extreme design flows
The performance of the different model types and
the four distributions was compared with respect to
 HYDROLOGICAL SCIENCES JOURNAL – JOURNAL DES SCIENCES HYDROLOGIQUES 1405

Figure 7. Observed and simulated monthly mean (catchment) precipitation and monthly and daily mean flows for the three model
types (Salzach catchment).

Figure 8. Observed and simulated monthly mean (catchment) precipitation and monthly and daily mean flows for the three model
types (Ubaye catchment).
1406 K. BREINL

Figure 9. Observed and simulated extreme design flows (annual, winter/spring, summer/autumn) for Q10, Q20 and Q50 (Model A,
Salzach). Bars denote the mean, lines the 95th confidence intervals. Results are shown for the observations (Obs), bootstrap
simulations (BS) and the four distribution functions. Black dots denote the distribution that best matches the mean of the
bootstrapping results.

the annual maximum flow, the annual maximum
winter/spring flow, and the annual maximum sum-
mer/autumn flow. As can be seen in Fig. 9, the
differences between the distribution functions are
pronounced. For annual flows, the Weibull distribu-
tion tends to overestimate for all return periods.
The most suitable distribution is the Hybrid distri-
bution. Results are similar for winter/spring and
summer/autumn extremes. Some bias between the
observations and bootstrap is obvious, especially for
the winter/spring season. However, this bias appears
to be small compared to the confidence intervals of
observations, in particular for Q20 and Q50. Results
are comparable for Model A+ (Fig. 10). The Hybrid
distribution appears to be most suitable. Simulated
values are generally higher, which is very likely
related to the better reproduction of autocorrelation
of precipitation compared to Model A. Figure 11
shows the results for Model B. Differences between
the types of distributions are considerably less pro-
nounced, which results in three different distribu-
tions (Exponential, Weibull, Hybrid) outperforming
the others. However, the Hybrid distribution per-
forms best in the majority of cases.
In the Ubaye study area, the Weibull distribution
likewise tends to overestimate extremes when using
models A and A+ (Figs 12 and 13), particularly for
higher return periods. The Hybrid distribution best
reproduces the bootstrap results. Differences between
types of distributions are obvious but less pronounced
compared to models A/A+ in the Salzach study area.
For Model B (Fig. 14), the differences are even less
pronounced than for the univariate model set-ups. All
distribution functions appear to be reasonable,
 HYDROLOGICAL SCIENCES JOURNAL – JOURNAL DES SCIENCES HYDROLOGIQUES 1407

Figure 10. Observed and simulated extreme design flows (annual, winter/spring, summer/autumn) for Q10, Q20 and Q50 (Model A+,
Salzach). Bars denote the mean, lines the 95th confidence intervals. Results are shown for the observations (Obs), bootstrap
simulations (BS) and the four distribution functions. Black dots denote the distribution that best matches the mean of the
bootstrapping results.

although the Hybrid distribution performs best in all
cases.
5 Discussion and conclusion
The study has shown that all three models (A, A+ and
B) have potential to be used for lumped hydrological
modelling. However, as could also be demonstrated,
the classic Richardson type Model A underestimates
the autocorrelation of precipitation amounts. Model
A+ and Model B, as well as other multi-site models,
perform better in this respect (e.g. Brandsma and
Buishand 1998, Buishand and Brandsma 2001,
Beersma and Buishand 2003). All model types show
similar performance in the simulation of wet spells.
The study has also shown that the averaging/weigh-
ing process to produce lumped time series obviously
changes the statistical behaviour of the site-specific
precipitation, leading to pronounced differences in
the fits of precipitation distributions. This is different
to the findings by Li et al. (2013), where distribution
parameters from areal precipitation and single sites
“shared the same characteristics”. Furthermore, in this
study, different weighting methods, including Thiessen
polygons, IDW, OK as well as KED, did not improve
the fits and led to similar curve parameters. Goodness-
of-fit tests as well as various precipitation related plots
revealed increased bias when fitting parametric distri-
butions to lumped precipitation time series. It is valid
to assume that this phenomenon is mainly related to
the mountainous character of the two catchments:
Time series from raingauges that are located at very
different altitudes in different micro-climatic condi-
tions, and which are exposed to orographic effects in
1408 K. BREINL

Figure 11. Observed and simulated extreme design flows (annual, winter/spring, summer/autumn) for Q10, Q20 and Q50 (Model B,
Salzach). Bars denote the mean, lines the 95th confidence intervals. Results are shown for the observations (Obs), bootstrap
simulations (BS) and the four distribution functions. Black dots denote the distribution that best matches the mean of the
bootstrapping results.

a complex topography, are lumped together. These
site- specific precipitation time series differ signifi-
cantly. In the Ubaye catchment, the difference in the
annual amount of precipitation between sites is up to
49%. The difference in altitude between the lowest and
highest rain gauge is 748 m. In the Salzach catchment,
the differences are even more pronounced (86% and
1652 m). In this study, the problem of lumped preci-
pitation time series particularly applies to the Weibull
distribution, which reproduces precipitation statistics
poorly in models A/A+ but well in Model B. In general,
for models A/A+, the Exponential and Weibull distri-
butions perform worse than the Gamma and Hybrid
distributions. The latter seems to be the most reason-
able choice. In Model B, the Weibull and Hybrid dis-
tributions tend to perform better than the Exponential
and Gamma distributions. Differences in the perfor-
mance of different parametric distributions are gener-
ally less pronounced in Model B. This becomes obvious
when simulating monthly and daily flow statistics
where all distributions produce reasonable results in
both study areas. The choice of the parametric distri-
bution in models A and A+ has a larger impact on the
monthly and daily flow modelling results. This parti-
cularly applies to the Salzach catchment. It also became
obvious that Model A tends to produce lower flows
than Model A+ and Model B (e.g. summer period
Salzach), which also affects the simulation of extremes.
As the performance of the three models is similar with
regard to the generation of wet spells, only the con-
siderably lower autocorrelation of the precipitation
amounts in Model A can explain this shortcoming
 HYDROLOGICAL SCIENCES JOURNAL – JOURNAL DES SCIENCES HYDROLOGIQUES 1409

Figure 12. Observed and simulated extreme design flows (annual, winter/spring, summer/autumn) for Q10, Q20 and Q50 (Model A,
Ubaye). Bars denote the mean, lines the 95th confidence intervals. Results are shown for the observations (Obs), bootstrap
simulations (BS) and the four distribution functions. Black dots denote the distribution that best matches the mean of the
bootstrapping results.

(see Section 4.1). The key conclusions of this study can
be summarized as follows:
● Fitting parametric distribution functions to
lumped precipitation in models A/A+ requires
additional attention. Poor fits can have strong
impact on the simulation results.
● Poor fits for lumped precipitation time series
are very likely caused by the mountainous
character of the two catchments and thus by
lumping time series that differ significantly
(due to different altitudes, different site-speci-
fic micro-climatic conditions and orographic
effects).
● The Gamma and Hybrid distributions seem to be
more suitable for fitting lumped precipitation
than the Exponential and Weibull distributions.
● Model B is less sensitive to the choice of para-
metric distribution functions. Potential poor fits
of single sites are averaged out in the subse-
quent weighting procedure. All parametric
curves produce reasonable mean and extreme
design flows, with the Hybrid distribution per-
forming best in the majority of cases.
● Model A tends to produce lower monthly and
daily flows as well as lower extremes. This
becomes obvious in the summer season of the
Salzach catchment and can be explained with
the lower autocorrelation of simulated precipi-
tation amounts (compared to models A+ and
B). Floods in the Salzach catchment often
occur in summer after long wet spells, i.e. the
reproduction of the serial autocorrelation is
important.
1410 K. BREINL

Figure 13. Observed and simulated extreme design flows (annual, winter/spring, summer/autumn) for Q10, Q20 and Q50
(Model A+, Ubaye). Bars denote the mean, lines the 95th confidence intervals. Results are shown for the observations (Obs),
bootstrap simulations (BS) and the four distribution functions. Black dots denote the distribution that best matches the mean
of the bootstrapping results.

● Independent modelling of precipitation and
temperature may introduce additional bias.
Some bias can be detected in the Salzach study
area (mean flows, overestimation of extreme
design flows in winter/spring). For this reason,
a sensitivity study (not shown in this paper)
with conditioned temperature and the bootstrap
model (using different Fourier series for tem-
perature on wet and dry days, see Semenov
et al. 1998) was conducted for the Salzach
study area, but did not noticeably improve the
results.
Altogether, all weather generators have the potential
to produce reasonable results, as long as the follow-
ing findings are taken into account: for the
univariate Richardson type weather generator (mod-
els A and A+), curve fitting requires particular atten-
tion. The wrong choice of parametric distribution
function can lead to a large bias in the simulations.
The major shortcoming of the Richardson type
model, which is the underestimation of the autocor-
relation of precipitation, can be tackled by an
improved Richardson model type (Model A+). The
multisite model (Model B) produces the least bias
and there is considerably less difference between the
different types of distribution functions. However,
there is a trade-off between the increased efforts of
setting up such a multi-site weather generator com-
pared to implementing comparatively straightforward
univariate models, and the potential improvements
of simulations. Additional studies would be required
 HYDROLOGICAL SCIENCES JOURNAL – JOURNAL DES SCIENCES HYDROLOGIQUES 1411

Figure 14. Observed and simulated extreme design flows (annual, winter/spring, summer/autumn) for Q10, Q20 and Q50 (Model B,
Ubaye). Bars denote the mean, lines the 95th confidence intervals. Results are shown for the observations (Obs), bootstrap
simulations (BS) and the four distribution functions. Black dots denote the distribution that best matches the mean of the
bootstrapping results.

to determine whether the findings of this study can Funding

be extrapolated elsewhere.
Acknowledgements
Data were obtained from the Central Institution for
Meteorology and Geodynamics (ZAMG), the Federal
Ministry of Agriculture, Forestry, Environment and Water
Management, the German Meteorological Service (DWD)
and the French Meterological Service (Météo-France).
Comments by Dirk Tiede are highly appreciated. Reviews
from the associate editor, Charles Perrin, and three anon-
ymous reviewers were gratefully received and helped to sig-
nificantly improve the paper.
Disclosure statement
No potential conflict of interest was reported by the author.
This research has been funded by the EU 7th framework
Marie Curie ITN project “CHANGES” (Grant Agreement
No. 263953).
References
Ailliot, P., Thompson, C., and Thomson, P., 2009. Space-
time modelling of precipitation by using a hidden Markov
model and censored Gaussian distributions. Journal of the
Royal Statistical Society: Series C (Applied Statistics), 58,
405–426. doi:10.1111/rssc.2009.58.issue-3
Arnold, J.G., et al., 1998. Large area hydrologic modeling and
assessment - Part 1: model development. Journal of the
American Water Resources Association, 34 (1), 73–89.
doi:10.1111/j.1752-1688.1998.tb05961.x
Baigorria, G.A. and Jones, J.W., 2010. GiST: a stochastic
model for generating spatially and temporally correlated
daily rainfall data. Journal of Climate, 23 (22), 5990–6008.
doi:10.1175/2010JCLI3537.1
1412 K. BREINL
Bárdossy, A. and Das, T., 2008. Influence of rainfall observa-
tion network on model calibration and application.
Hydrology and Earth System Sciences, 12 (1), 77–89.
doi:10.5194/hess-12-77-2008
Bárdossy, A. and Pegram, G.G.S., 2009. Copula based multi-
site model for daily precipitation simulation. Hydrology
and Earth System Sciences, 13 (12), 2299–2314.
doi:10.5194/hess-13-2299-2009
Beck, H.E., et al., 2013. The impact of forest regeneration on
streamflow in 12 mesoscale humid tropical catchments.
Hydrology and Earth System Sciences, 17 (7), 2613–2635.
doi:10.5194/hess-17-2613-2013
Beersma, J.J. and Buishand, T.A., 2003. Multi-site simulation
of daily precipitation and temperature conditional on the
atmospheric circulation. Climate Research, 25 (2), 121–
133. doi:10.3354/cr025121
Bergström, S., et al., 2001a. Climate change impacts on run-
off in Sweden - assessments by global climate models,
dynamical downscaling and hydrological modelling.
Climate Research, 16 (2), 101–112. doi:10.3354/cr016101
Bergström, S., et al., 2001b. Climate change impacts on run-
off in Sweden - assessments by global climate models,
dynamical downscaling and hydrological modelling.
Climate Research, 16 (2), 101–112. doi:10.3354/cr016101
Beven, K., 1989. Changing ideas in hydrology - the case of
physically-based models. Journal of Hydrology, 105 (1–2),
157–172. doi:10.1016/0022-1694(89)90101-7
Beven, K., 2001. How far can we go in distributed hydro-
logical modelling? Hydrology and Earth System Sciences, 5
(1), 1–12. doi:10.5194/hess-5-1-2001
Booij, M.J., 2005. Impact of climate change on river flooding
assessed with different spatial model resolutions. Journal
of Hydrology, 303 (1–4), 176–198. doi:10.1016/j.
jhydrol.2004.07.013
Brandsma, T. and Buishand, T.A., 1998. Simulation of
extreme precipitation in the Rhine basin by nearest-neigh-
bour resampling. Hydrology and Earth System Sciences, 2
(2/3), 195–209. doi:10.5194/hess-2-195-1998
Breinl, K., Turkington, T., and Stowasser, M., 2013.
Stochastic generation of multi-site daily precipitation for
applications in risk management. Journal of Hydrology,
498, 23–35. doi:10.1016/j.jhydrol.2013.06.015
Breinl, K., Turkington, T., and Stowasser, M., 2015.
Simulating daily precipitation and temperature: a weather
generation framework for assessing hydrometeorological
hazards. Meteorological Applications, 22 (3), 334–347.
doi:10.1002/met.1459.
Brissette, F.P., Khalili, M., and Leconte, R., 2007. Efficient
stochastic generation of multi-site synthetic precipitation
data. Journal of Hydrology, 345 (3–4), 121–133.
doi:10.1016/j.jhydrol.2007.06.035
Buishand, T.A. and Brandsma, T., 2001. Multisite simulation
of daily precipitation and temperature in the Rhine Basin
by nearest-neighbor resampling. Water Resources
Research, 37 (11), 2761–2776. doi:10.1029/2001WR000291
Burn, D.H., 2003. The use of resampling for estimating
confidence intervals for single site and pooled frequency
analysis. Hydrological Sciences Journal, 48 (1), 25–38.
doi:10.1623/hysj.48.1.25.43485
Clark, M.P., et al., 2004. A resampling procedure for gener-
ating conditioned daily weather sequences. Water
Resources Research, 40 (4). doi:10.1029/2003WR002747
Coles, S., 2001. An introduction to statistical modeling of
extreme values. New York: Springer.
Das, T., et al., 2008. Comparison of conceptual model per-
formance using different representations of spatial varia-
bility. Journal of Hydrology, 356 (1–2), 106–118.
doi:10.1016/j.jhydrol.2008.04.008
Deidda, R., 2010. A multiple threshold method for fitting the
generalized Pareto distribution to rainfall time series.
Hydrology and Earth System Sciences, 14 (12), 2559–2575.
doi:10.5194/hess-14-2559-2010
Filliben, J.J., 1975. The probability plot correlation coefficient
test for normality. Technometrics, 17 (1), 111–117.
doi:10.1080/00401706.1975.10489279
Foufoula-Georgiou, E. and Lettenmaier, D.P., 1987. A
Markov renewal model for rainfall occurrences. Water
Resources Research, 23 (5), 875–884. doi:10.1029/
WR023i005p00875
Freudiger, D., et al., 2014. Large-scale analysis of changing
frequencies of rain-on-snow events with flood-generation
potential. Hydrology and Earth System Sciences, 18 (7),
2695–2709. doi:10.5194/hess-18-2695-2014
Furrer, E.M. and Katz, R.W., 2008. Improving the simulation
of extreme precipitation events by stochastic weather gen-
erators. Water Resources Research, 44 (12), W12439.
doi:10.1029/2008WR007316
Gao, H.K., et al., 2012. Modeling the runoff and glacier mass
balance in a small watershed on the Central Tibetan
Plateau, China, from 1955 to 2008. Hydrological
Processes, 26 (11), 1593–1603. doi:10.1002/hyp.8256
Garavaglia, F., et al., 2010. Introducing a rainfall compound
distribution model based on weather patterns sub-sam-
pling. Hydrology and Earth System Sciences, 14 (6), 951–
964. doi:10.5194/hess-14-951-2010
Gilleland, E. and Katz, R.W., 2011. New software to analyze
how extremes change over time. Eos, Transactions
American Geophysical Union, 92 (2), 13–14.
Grimaldi, S. and Serinaldi, F., 2006. Asymmetric copula in
multivariate flood frequency analysis. Advances in Water
Resources, 29 (8), 1155–1167. doi:10.1016/j.
advwatres.2005.09.005
Haberlandt, U., 2007. Geostatistical interpolation of hourly
precipitation from rain gauges and radar for a large-scale
extreme rainfall event. Journal of Hydrology, 332 (1–2),
144–157. doi:10.1016/j.jhydrol.2006.06.028
Haberlandt, U., et al., 2011. Rainfall generators for application in
flood studies. In: A.H. Schumann, et al., eds. Flood risk assess-
ment and management. Dordrecht: Springer, 117–147.
Hagg, W., et al., 2006. Runoff modelling in glacierized
Central Asian catchments for present-day and future cli-
mate. Nordic Hydrology, 37 (2), 93–105.
Harmel, R.D., Richardson, C.W., and King, K.W., 2000.
Hydrologic response of a small watershed model to gen-
erated precipitation. Transactions of the ASAE 43 (6),
1483–1488. doi:10.13031/2013.3047
Hartigan, J.A., 1975. Clustering algorithms. New York: Wiley.
Hartigan, J.A. and Wong, M.A., 1979. Algorithm AS 136:
a K-Means clustering algorithm. Journal of the Royal
Statistical Society. Series C (Applied Statistics), 28 (1),
100–108.
Hashmi, M.Z., Shamseldin, A.Y., and Melville, B.W.,
2011. Comparison of SDSM and LARS-WG for simula-
tion and downscaling of extreme precipitation events
 HYDROLOGICAL SCIENCES JOURNAL – JOURNAL DES SCIENCES HYDROLOGIQUES
in a watershed. Stochastic Environmental Research and
Risk Assessment, 25 (4), 475–484. doi:10.1007/s00477-
010-0416-x
Haylock, M.R., et al., 2008. A European daily high-resolution
gridded data set of surface temperature and precipitation
for 1950–2006. Journal of Geophysical Research-
Atmospheres, 113 (D20), D20119. doi:10.1029/
2008JD010201
Hofstra, N., et al., 2009. Testing E-OBS European high-
resolution gridded data set of daily precipitation and surface
temperature. Journal of Geophysical Research-Atmospheres,
114 (D21), D2110. doi:10.1029/2009JD011799
Holzmann, H., et al., 2010. Auswirkungen möglicher
Klimaänderungen auf Hochwasser und Wasserhau-
shaltskomponenten ausgewählter Einzugsgebiete in
Österreich. Österr. Wasser- und Abfallwirtschaft, 56 (1–2/
10), 7–14. doi:10.1007/s00506-009-0154-9
Hughes, J.P., Guttorp, P., and Charles, S.P., 1999. A non-
homogeneous hidden Markov model for precipitation
occurrence. Journal of the Royal Statistical Society: Series
C (Applied Statistics), 48 (1), 15–30. doi:10.1111/1467-
9876.00136
Hundecha, Y., Pahlow, M., and Schumann, A., 2009.
Modeling of daily precipitation at multiple locations
using a mixture of distributions to characterize the
extremes. Water Resources Research, 45 (12), W12412.
Jenks, G., 1967. The data model concept in statistical map-
ping. International Yearbook of Cartography, 7, 186–190.
Katz, R.W., 1977. Precipitation as a chain-dependent process.
Journal of Applied Meteorology, 16 (7), 671–676. doi:10.1175/
1520-0450(1977)016<0671:PAACDP>2.0.CO;2
Katz, R.W., 1981. On some criteria for estimating the order
of a Markov-chain. Technometrics, 23 (3), 243–249.
Keef, C., et al., 2010. A multivariate model for the broad scale
spatial assessment of flood risk. In: BHS third international
symposium, managing consequences of a changing global
environment, July, Newcastle. British Hydrological Society.
Khalili, M., Brissette, F., and Leconte, R., 2009. Stochastic
multi-site generation of daily weather data. Stochastic
Environmental Research and Risk Assessment, 23 (6),
837–849. doi:10.1007/s00477-008-0275-x
King, L.M., McLeod, A.I., and Simonovic, S.P., 2014.
Simulation of historical temperatures using a multisite,
multivariate block resampling algorithm with perturba-
tion. Hydrological Processes, 28 (3), 905–912.
Leander, R. and Buishand, T.A., 2009. A daily weather gen-
erator based on a two-stage resampling algorithm. Journal
of Hydrology, 374 (3–4), 185–195. doi:10.1016/j.
jhydrol.2009.06.010
Li, Z., Brissette, F., and Chen, J., 2013. Finding the most
appropriate precipitation probability distribution for sto-
chastic weather generation and hydrological modelling in
Nordic watersheds. Hydrological Processes, 27, 3718–3729.
doi:10.1002/hyp.v27.25
Lucio, P.S., Silva, A.M., and Serrano, A.I., 2010. Changes in
occurrences of temperature extremes in continental
Portugal: a stochastic approach. Meteorological Applications,
17 (4), 404–418.
Ly, S., Charles, C., and Degré, A., 2011. Geostatistical
interpolation of daily rainfall at catchment scale: the
use of several variogram models in the Ourthe and
Ambleve catchments, Belgium. Hydrology and Earth
1413
System Sciences, 15 (7), 2259–2274. doi:10.5194/hess-
15-2259-2011
Mehrotra, R., Srikanthan, R., and Sharma, A., 2006. A
comparison of three stochastic multi-site precipitation
occurrence generators. Journal of Hydrology, 331 (1–2),
280–292. doi:10.1016/j.jhydrol.2006.05.016
Merz, R. and Blöschl, G., 2008. Flood frequency hydrology: 1.
Temporal, spatial, and causal expansion of information.
Water Resources Research, 44 (8), W08432. doi:10.1029/
2007WR006744
Mudelsee, M., et al., 2004. Extreme floods in central Europe
over the past 500 years: role of cyclone pathway
“Zugstrasse Vb’’. Journal of Geophysical Research-
Atmospheres, 109 (D23), D23101. doi:10.1029/
2004JD005034
Nash, J.E. and Sutcliffe, J.V., 1970. River flow forecasting
through conceptual models part I — a discussion of prin-
ciples. Journal of Hydrology, 10 (3), 282–290. doi:10.1016/
0022-1694(70)90255-6
Papalexiou, S.M., Koutsoyiannis, D., and Makropoulos, C.,
2013. How extreme is extreme? An assessment of daily
rainfall distribution tails. Hydrology and Earth System
Sciences, 17 (2), 851–862. doi:10.5194/hess-17-851-2013
Paquet, E., et al., 2013. The SCHADEX method: a semi-
continuous rainfall-runoff simulation for extreme flood
estimation. Journal of Hydrology, 495, 23–37.
doi:10.1016/j.jhydrol.2013.04.045
Photiadou, C.S., Weerts, A.H., and van den Hurk, B.J.J.M.,
2011. Evaluation of two precipitation data sets for the
Rhine River using streamflow simulations. Hydrology and
Earth System Sciences, 15 (11), 3355–3366. doi:10.5194/
hess-15-3355-2011
Press, W.H., et al., 2002. Numerical recipes in C++: the art of
scientific computing. 2nd ed. Cambridge: Cambridge
University Press.
Qian, B., Gameda, S., and Hayhoe, H., 2008. Performance of
stochastic weather generators LARS-WG and AAFC-WG
for reproducing daily extremes of diverse Canadian cli-
mates. Climate Research, 37 (1), 17–33. doi:10.3354/
cr00755
Racsko, P., Szeidl, L., and Semenov, M., 1991. A serial
approach to local stochastic weather models. Ecological
Modelling, 57 (1–2), 27–41. doi:10.1016/0304-3800(91)
90053-4
Raff, D.A., Pruitt, T., and Brekke, L.D., 2009. A framework
for assessing flood frequency based on climate projection
information. Hydrology and Earth System Sciences, 13 (11),
2119–2136. doi:10.5194/hess-13-2119-2009
Refsgaard, J.C., Storm, B., and Clausen, T., 2010. Systeme
Hydrologique Europeen (SHE): review and perspectives
after 30 years development in distributed physically-based
hydrological modelling. Hydrology Research, 41 (5), 355–377.
Reiss, R.-D. and Thomas, M., 2007. Statistical analysis of
extreme values. Basel: Birkhäuser.
Richardson, C.W., 1981. Stochastic simulation of daily pre-
cipitation, temperature, and solar-radiation. Water
Resources Research, 17 (1), 182–190. doi:10.1029/
WR017i001p00182
Romanowicz, A.A., et al., 2005. Sensitivity of the SWAT model
to the soil and land use data parametrisation: a case study in
the Thyle catchment, Belgium. Ecological Modelling, 187
(1), 27–39. doi:10.1016/j.ecolmodel.2005.01.025
1414 K. BREINL
Schädler, G., et al., 2012. Flood hazards in a changing climate.
Karlsruhe: Center for Disaster Management and Risk
Reduction Technology (CEDIM).
Schwarz, G., 1978. Estimating the dimension of a model. The
Annals of Statistics, 6 (2), 461–464. doi:10.1214/aos/
1176344136
Seibert, J., 1997. Estimation of parameter uncertainty in the
HBV model. Nordic Hydrology, 28 (4–5), 247–262.
Seibert, J., 1999. Regionalisation of parameters for a conceptual
rainfall-runoff model. Agricultural and Forest Meteorology,
98–99, 279–293. doi:10.1016/S0168-1923(99)00105-7
Seibert, J., 2000. Multi-criteria calibration of a conceptual
runoff model using a genetic algorithm. Hydrology and
Earth System Sciences, 4 (2), 215–224. doi:10.5194/hess-4-
215-2000
Seibert, J. and Vis, M.J.P., 2012. Teaching hydrological mod-
eling with a user-friendly catchment-runoff-model soft-
ware package. Hydrology and Earth System Sciences, 16
(9), 3315–3325. doi:10.5194/hess-16-3315-2012
Semenov, M.A. and Barrow, E.M., 1997. Use of a stochastic
weather generator in the development of climate change
scenarios. Climatic Change, 35 (4), 397–414. doi:10.1023/
A:1005342632279
Semenov, M.A., et al., 1998. Comparison of the WGEN and
LARS-WG stochastic weather generators for diverse climates.
Climate Research, 10 (2), 95–107. doi:10.3354/cr010095
Serinaldi, F., 2009a. Copula-based mixed models for bivariate
rainfall data: an empirical study in regression perspective.
Stochastic Environmental Research and Risk Assessment, 23
(5), 677–693. doi:10.1007/s00477-008-0249-z
Serinaldi, F., 2009b. A multisite daily rainfall generator
driven by bivariate copula-based mixed distributions.
Journal of Geophysical Research-Atmospheres, 114, 10.
doi:10.1029/2008JD011258
Shrestha, R., Tachikawa, Y., and Takara, K., 2006. Input data
resolution analysis for distributed hydrological modeling.
Journal of Hydrology, 319 (1–4), 36–50. doi:10.1016/j.
jhydrol.2005.04.025
Singh, V.P. and Mishra, A.K., 2011. Report as of FY2010 for
2009TX334G: “Hydrological Drought Characterization for
Texas under Climate Change, with Implications for Water
Resources Planning and Management”. Texas A & M
University, Department of Biological and Agricultural
Engineering.
Steele-Dunne, S., et al., 2008. The impacts of climate change
on hydrology in Ireland. Journal of Hydrology, 356 (1–2),
28–45. doi:10.1016/j.jhydrol.2008.03.025
Turkington, T., et al., 2014. Empirical atmospheric
thresholds for debris flows and flash floods in the
Southern French Alps. Natural Hazards Earth Systems
Sciences Discussions, 2 (1), 757–798. doi:10.5194/
nhessd-2-757-2014
Uhlenbrook, S., et al., 1999. Prediction uncertainty of con-
ceptual rainfall–runoff models caused by problems in
identifying model parameters and structure. Hydrological
Sciences Journal, 44 (5), 779–797. doi:10.1080/
02626669909492273
Vlček, O. and Huth, R., 2009. Is daily precipitation Gamma-
distributed? Adverse effects of an incorrect use of the
Kolmogorov-Smirnov test. Atmospheric Research, 93 (4),
759–766. doi:10.1016/j.atmosres.2009.03.005
Vrac, M. and Naveau, P., 2007. Stochastic downscaling of
precipitation: from dry events to heavy rainfalls. Water
Resources Research, 43, 7. doi:10.1029/2006WR005308
Wilby, R.L., Tomlinson, O.J., and Dawson, C.W., 2003.
Multi-site simulation of precipitation by conditional
resampling. Climate Research, 23 (3), 183–194.
doi:10.3354/cr023183
Wilks, D.S., 1998. Multisite generalization of a daily stochas-
tic precipitation generation model. Journal of Hydrology,
210 (1–4), 178–191. doi:10.1016/S0022-1694(98)00186-3
Wilks, D.S., 1999a. Interannual variability and extreme-value
characteristics of several stochastic daily precipitation
models. Agricultural and Forest Meteorology, 93 (3), 153–
169. doi:10.1016/S0168-1923(98)00125-7
Wilks, D.S., 1999b. Multisite downscaling of daily precipita-
tion with a stochastic weather generator. Climate Research,
11 (2), 125–136. doi:10.3354/cr011125
Zambrano-Bigiarini, M., 2012. “hydroTSM” - R package
version 0.4-2-1.