Accepted Manuscript A multi-state weather generator for daily precipitation for the Torne river basin, northern Sweden/western Finland David Rayner, Christine Achberger, Deliang Chen PII:

S1674-9278(15)30015-0

DOI:

10.1016/j.accre.2016.06.006

Reference:

ACCRE 60

To appear in:

Advances in Climate Change Research

Received Date: 10 December 2015 Revised Date:

22 June 2016

Accepted Date: 23 June 2016

Please cite this article as: Rayner, D., Achberger, C., Chen, D., A multi-state weather generator for daily precipitation for the Torne river basin, northern Sweden/western Finland, Advances in Climate Change Research (2016), doi: 10.1016/j.accre.2016.06.006. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

1

ACCEPTED MANUSCRIPT 1

A multi-state weather generator for daily precipitation for the Torne River basin,

2

northern Sweden/western Finland

3 David RAYNER *, Christine ACHBERGER, Deliang CHEN

5

Department of Earth Sciences, University of Gothenburg, 40530 Gothenburg, Sweden

6

* corresponding author: [email protected]

RI PT

4

7

SC

8

AC C

EP

TE D

M AN U

9

2

ACCEPTED MANUSCRIPT

Abstract

2

This paper describes a new weather generator – the 10-state empirical model – that combines

3

a 10-state, first-order Markov chain with a non-parametric precipitation amounts model.

4

Using a doubly-stochastic transition-matrix results in a weather generator for which the

5

overall precipitation distribution (including both wet and dry days) and the temporal-

6

correlation can be modified independently for climate change studies. This paper assesses the

7

ability of the 10-state empirical model to simulate daily area-average precipitation in the

8

Torne River catchment in northern Sweden/western Finland in the context of 3 other models:

9

a 10-state model with a parametric (Gamma) amounts model; a wet/dry chain with the

10

empirical amounts model; and a wet/dry chain with the parametric amounts model. The

11

ability to accurately simulate the distribution of multi-day precipitation in the catchment is

12

the primary consideration.

13

Results show that the 10-state empirical model represented accumulated 2- to 14-day

14

precipitation most realistically. Further, the distribution of precipitation on wet days in the

15

catchment is related to the placement of a wet day within a wet-spell, and the 10-state models

16

represented this realistically, while the wet/dry models did not.

17

Although all four models accurately reproduced the annual and monthly averages in the

18

training data, all models underestimated inter-annual and inter-seasonal variance. Even so,

19

the 10-state empirical model performed best.

20

We conclude that the multi-state model is a promising candidate for hydrological applications,

21

as it simulates multi-day precipitation well, but that further development is required to

22

improve the simulation of interannual variation.

23

Keywords: Weather generator; Multi-state; Torne River; Precipitation

AC C

EP

TE D

M AN U

SC

RI PT

1

24 25

26

1.

Introduction

27

A weather generator (WG) is a stochastic model that is designed to generate synthetic

28

weather time-series with the same statistical properties as observed data. WGs can provide

29

additional data when the observed climate record is insufficient with respect to completeness,

30

spatial coverage or length to reliably estimate of the probability of extreme events (Jones et

3

ACCEPTED MANUSCRIPT al., 2011; Kilsby et al., 2007; e.g. Wilks and Wilby, 1999). WGs can be used to simulate

2

short-term weather (e.g. at daily or sub-daily scales) for the past or future, and have become a

3

common tool for studying impacts of climate change on ecosystems and human settlements

4

(Jones et al., 2011; e.g. Wilks, 2010). One particular advantage of using WGs is that they can

5

generate time-series for an extended period without significant computational investment.

6

Within the broad family of WGs that have been developed for simulating daily precipitation,

7

two categories of models dominate the literature, which we call the “Richardson”

8

(Richardson and Wright, 1984; Richardson, 1981) and the “serial” (Racsko et al., 1991;

9

Semenov et al., 1998) types. The Richardson-type WG – applied in this study – simulates

RI PT

1

daily precipitation in two separate steps, the first to simulate rainfall occurrence and the

11

second to estimate the rainfall amount on wet days. With the classical Richardson model, the

12

first step is accomplished using a first-order, two-state Markov chain, which describes the

13

probability of a wet day following a dry day, a wet day following a wet day, etc. The

14

transition probabilities can be estimated from the observed data.

15

Once a certain day is modeled as wet, an “amounts model” simulates the precipitation

16

amount for that day. Parametric approaches (e.g. Chen et al., 2015; Chen and Brissette, 2014)

17

use pre-specified functions to approximate the observed precipitation distribution. With non-

18

parametric approaches, the observed precipitation distribution itself is used as the basis for

19

the amounts model. The simplest non-parametric approach is to resample directly from the

20

observed sequence, or from a sub-set of the sequence which represents the “nearest neighbors”

21

with-respect-to weather conditions (e.g. Sharma and Lall, 1999). Kernel-density smoothing

22

the observed distribution allows non-parametric methods to generate a continuous

23

distribution of precipitation, and also to generate values higher than observed historically (e.g.

24

Harrold et al., 2003; Mehrotra and Sharma, 2007a). Non-parametric methods allow a WG to

25

generate precipitation sequences that match the observed distribution with arbitrary-high

26

precision, at the cost of introducing arbitrarily-many parameters. Non-parametric methods

27

allow more flexibility for including new forms of conditional dependence; they have the

28

ability to reproduce features such as non-linearity, asymmetry, or multi-modality in observed

29

records (Mehrotra et al., 2006); and that they do not make any strong assumptions about the

30

precipitation distribution (Mehrotra and Sharma, 2007a).

31

Many WGs using Markov-approaches have been found to have 3 partially-related

32

deficiencies: they underestimate the frequency of extended drought periods (Mehrotra and

AC C

EP

TE D

M AN U

SC

10

4

ACCEPTED MANUSCRIPT Sharma, 2007a), they often ignore temporal correlations within wet-spells (Harrold et al.,

2

2003), and they underestimate low-frequency (usually described by inter-annual) variability

3

(e.g. Gregory et al., 1993; Katz and Zheng, 1999; Srikanthan et al., 2005). These issues

4

appear to be partially related, as a proportion of low-frequency variability in precipitation can

5

be accounted for by the short-lag correlation (Gregory et al., 1993) or by the rainfall

6

occurrence process (Mehrotra and Sharma, 2007a).

7

The tendency of purely Markov-based WGs to underestimate inter-annual variability has

8

been attributed to climatic non-stationarities, for example the influence of the El Niño

9

Southern Oscillation (Harrold et al., 2003). One way to increase the simulated inter-annual

RI PT

1

variability is to condition the WG parameters on a physical, slowly-varying index

11

representing atmospheric circulation or SST (Katz and Parlange, 1993; Wilby et al., 2002).

12

Another method is to use longer-period, aggregated precipitation as the conditioning index,

13

either explicitly (Harrold et al., 2003; Mehrotra and Sharma, 2007a, 2007b) or via wavelet

14

decomposition (Steinschneider and Brown, 2013). Finally, the low-frequency signal can be

15

increased by postulating dependence on a “hidden” index, whose variation must be estimated

16

using iterative methods (Katz and Zheng, 1999).

17

Consecutive days with similar rainfall amounts are often clustered in time, a property which

18

is not represented by two-state Markov chains that distinguish only between dry and wet.

19

This property can be represented using a multi-state Markov chain model that models

20

transitions between different precipitation bands (e.g. Boughton, 1999; Gregory et al., 1993;

21

Haan et al., 1976; Srikanthan and McMahon, 2001). The state boundaries can be defined

22

using geometric progression, resulting in increasing class widths (e.g. Haan et al., 1976;

23

Srikanthan et al., 2005) and a relatively even number observations in each state.

24

Simulation of catchment runoff often requires multiple precipitation time-series, each

25

representing a different sub-catchment or gauge, with realistic spatial correlations. There are

26

many approaches to generating such series. One is to drive a collection of individual models

27

(representing different locations) with a common “random” number series, which is in-turn

28

derived from an index of larger-scale atmospheric circulation. Another is to feed independent

29

models with serially-independent but spatially correlated random series (Wilks, 1998).

30

Finally, it is possible to simulate the catchment-average precipitation as a single time-series

31

(Chen et al., 2012), which is the approach used in this study.

AC C

EP

TE D

M AN U

SC

10

5

ACCEPTED MANUSCRIPT This paper introduces a new WG that combines a 10-state, first-order Markov chain and a

2

non-parametric precipitation amounts model. Our innovation is to adopt a doubly-stochastic

3

transition-matrix, rather than manually defining transition thresholds, which gives the model

4

the property that the overall precipitation distribution (defined as including both wet and dry

5

days) is independent of the transition-matrix. The paper first describes the new WG and its

6

implementation for the Torne River catchment in northern Sweden/western Finland. The

7

paper then quantitatively-assesses whether the model’s performance is an improvement over

8

simpler two-state approaches. The properties assessed are those that are important for

9

hydrological modeling: inter-annual, inter-seasonal and multi-day precipitation distributions;

RI PT

1

lengths of wet and dry spells; and the variations of precipitation within multi-day events.

11

2.

Site description and data 2.1.

M AN U

12

SC

10

The Torne River catchment

The Torne River catchment (Fig. 1) straddles the border of Sweden and Finland and covers

14

40,157 km². The catchment extends from the northern mountains of Sweden and north-

15

western Finnish Lapland, south-east down through marshes and lowlands to the Gulf of

16

Bothnia in the Baltic Sea. The lower reaches of the Torne River comprises the border

17

between Sweden and Finland, with the closely-connected towns of Haparanda (Sweden) and

18

Tornio (Finland) near the river mouth together having around 23,000 inhabitants. The Torne

19

River is essentially unregulated and the catchment is sparely populated. Around half of the

20

catchment area is covered by forest, a fifth is mountains and a tenth is marshland, with some

21

agriculture in the lowlands. For a comprehensive description of the catchment geography, see

22

Bowling et al. (2003).

23

Annual average precipitation is 550–600 mm in the lowland areas of the catchment,

24

increasing to 800 mm closer to the Scandinavian mountains and to over 1000 mm in the

25

western highland areas (Asp, 2011). The catchment is typically snow-covered October to

26

May, and peak river discharge occurs in May‒June (Carlsson, 1999).

27

Beginning in the mid-1980s, the social-disruption and damage caused by spring floods began

28

to receive increasing attention. Flood-levels were particularly high in 1990, but large floods

29

have also occurred in 1615, 1677 and in 1968 (MSB, 2011). The most critical situation

30

occurs when thick ice builds in the river mouth, and the ice behaves as a dam behind which

31

water-levels can rise rapidly. A system for forecasts and warnings has now been implemented

AC C

EP

TE D

13

6

ACCEPTED MANUSCRIPT 1

for Haparanda and Tornio. Haparanda is one of the 18 towns in Sweden designated as having

2

a “significant flood risk” (MSB, 2012).

2.2.

3

Precipitation time-series

The preparation of precipitation data for the Torne River basin is described in Asp (2011).

5

The daily precipitation data used in this paper were provided by the Swedish Meteorological

6

and Hyrdological Institute as 49 sub-catchment time-series covering the years 1961 to 2010.

7

Daily catchment-average precipitation was then calculated as the area-weighted average over

8

the sub-catchments. Annual-mean catchment precipitation is 627 mm (1.7 mm d-1), with a

9

standard deviation of 85 mm, with the highest precipitation in summer and the lowest in

RI PT

4

spring. Although the dry-day fraction (defined as precipitation < 0.1 mm, which is the

11

minimum non-zero value in the provided data) for individual sub-catchments is typically 0.35,

12

the fraction of days for which all 49 sub-catchments have < 0.1 mm is only 0.06.

13

There is significant temporal auto-correlation within the daily catchment-average time-series,

14

with lag-1 Spearman rank correlation coefficients (after removing the seasonal cycle) of 0.46

15

for all days and 0.32 for wet-days (precipitation ≥ 0.1 mm). For wet-days, lag-1 and lag-2

16

autocorrelations are statistically-significant for all months except October; in summer, lag-3

17

autocorrelations are also statistically-significant. Winter precipitation is closely correlated

18

with the North Atlantic Oscillation (Spearman rank correlation coefficient of 0.56; see also

19

Bartolini et al. 2009).

20

3.

3.1.

22

3.1.1.

Development of the WG

AC C

21

EP

Methods

TE D

M AN U

SC

10

Primary considerations

23

The WG was developed within the framework of the project “Future rainfall and

24

flooding in Sweden: a framework to support climate adaptation actions”. The aim was

25

to develop a compliment to the bias-corrected regional climate model outputs (Yang et

26

al., 2010) that are usually used for studies of future runoff in Sweden. The WG model

27

will be also be used to assess uncertainties related to downscaling methods.

28

We chose to simulate catchment-average precipitation as a single time-series, with a

29

view to subsequent spatial disaggregation to sub-catchment scale. Spatial

30

disaggregation is not covered in this paper. Our choice was partially motivated by a

7

ACCEPTED MANUSCRIPT preference for a downscaling framework that was based on GCM outputs: by avoiding

2

regional climate model outputs, we hope to obtain a more holistic assessment of

3

uncertainties in the downscaling process when comparing against methods based on

4

regional model outputs. Given this preference, modeling catchment-averages was a

5

natural choice, because the resolution of GCM outputs is more representative of

6

catchment-scales than sub-catchment scales, and thus modeling catchment-averages

7

avoids a large scale-mismatch. In addition, it is simpler to parameterize the effects of

8

climate change on a single time-series than on multiple, spatially-correlated time-series.

9

Initial experimentation with two-state WGs for catchment-average precipitation in the

10

Torne River catchment revealed, however, that the models were unable to realistically-

11

simulate multi-day precipitation totals (the two-state models are introduced below, and

12

their performance is documented in the Section 4). Thus, we decided to explore

13

whether we could develop a multi-state precipitation model that could generate realistic

14

multi-day precipitation totals whilst still being straightforward to modify to represent a

15

changed climate.

16

3.1.2.

17

All of the Richardson-type Markov models for daily occurrence of which we are aware

18

– both two- and multi-state models – have been constructed with “user-defined” states.

19

That is, for a two-state matrix, a threshold state for a wet day is chosen (e.g. 0.1mm),

20

and then the transition matrix P – defined such that the chance of transitioning from

21

state i to state j is pij – is estimated empirically. That is, suppose that the chance of

22

transitioning from state 1 (dry) to state 2 (wet) was 0.1, and the chance of transitioning

23

from state 2 to state 1 was 0.4, then the matrix would be defined by p11 = 0.9, p12 = 0.1,

24

p21 = 0.4 and p22 = 0.6. The matrix P is then said to be singly (or row) stochastic: the

25

sum of the elements in each row is 1, but the sums of the columns are not 1.

26

The transition matrix P also defines the overall fraction of wet and dry days. This

27

follows from the general result (e.g. Ching et al., 2013) that the steady-state distribution

28

π of an irreducible and aperiodic Markov chain can be calculated as the solution to the

29

equation sets:

TE D

EP

AC C

= ,

30 31

Precipitation state model

M AN U

SC

RI PT

1

and

(1)

8

ACCEPTED MANUSCRIPT = 1 .

(2)

For the two-states example given above, the solution to the above equations is

2

π = [0.8, 0.2]. That is, the fraction of dry days π1=0.8 and the fraction of wet days

3

π2=0.2.

4

An alternative strategy is to first specify the length-n vector π. If π is defined such that

5

all elements are equal (

6

steady-state distribution. The transition matrix P for a such a Markov chain must be

7

doubly-stochastic, i.e., both the rows and the columns of P sum to 1 (Sinkhorn and

8

Knopp, 1967).

9

To estimate the P for such a system, the observed data must be first classified into

= 1 ), then all states will be equally likely in the

SC

=

RI PT

1

groups of equal size, and then the transition matrix can be estimated empirically. An

11

obvious method for defining the groups is to split the data into quantiles, so then P

12

defines the transition probabilities between quantiles. For example, with n=10, P

13

defines transitions between precipitation deciles.

14

However, an issue arises immediately when attempting to estimate P using real

15

precipitation data: how to treat transitions to/from days of zero precipitation? For the

16

example above the dry day fraction is 0.8. If we choose n=10, we cannot say if a dry

17

day belongs to decile 1 or decile 8, and so we are unable to say if transition from a dry

18

day to a highest-decile wet day (j=10) should be classified as p1,10, p2,10,…or p8,10.

19

The solution is to realize that we do not actually need to know the quantile of each dry

20

day; all that is required to create matrix P is to divide the data into groups of equal size.

21

We are free to select any appropriate algorithm for classifying days into particular

22

states. In this study we model area-average precipitation, and the highest monthly dry-

23

day fraction is only 0.16. Thus, we employed the simple expedient of adding a tiny

24

random amount (from 0 to 10-5 mm) to all precipitation values to ensure unique

25

rankings across both wet and dry days. This process allocates dry days randomly and

26

uniformly amongst the lowest deciles; in the example above, with a dry-day fraction is

27

0.8, the dry days would be distributed evenly between deciles 1 to 8. More elaborate

28

possibilities exist; for example dry days could be ranked according to how long before

29

or after a wet-spell they occurred.

AC C

EP

TE D

M AN U

10

9

ACCEPTED MANUSCRIPT The advantage of using a Markov chain with a doubly-stochastic transition-matrix is

2

that the overall precipitation distribution (defined as including both wet and dry days) is

3

independent of the transition-matrix. This follows because all states are – by definition

4

– equally likely. This property is advantageous when modifying a WG to represent the

5

effects of climate change: changes to the transition matrix will not change the

6

precipitation distribution, provided that the modified matrix remains doubly-stochastic.

7

In contrast, modifying a classical two-state wet/dry transition matrix implicitly changes

8

the overall precipitation distribution, which may then necessitate further correction

9

factors. In addition, there are also (n−1)2 parameters in a doubly-stochastic matrix,

RI PT

1

slightly-less than the n·(n−1) parameters in a row-stochastic matrix of the same size.

11

In practice, transition matrices estimated from empirical data will not be doubly-

12

stochastic to within machine-precision because real data sets have finite length1.

13

However, it is simple to rescale the matrix to be double-stochastic using the Sinkhorn-

14

Knopp algorithm (Sinkhorn and Knopp, 1967; see Knight (2008) for a code fragment).

15

An n=10 transition matrix was used in this paper, a size selected so that most dry days

16

were allocated to one state. This setup is referred to as the 10-state model. Although

17

transition matrices were calculated independently for each month, they were all quite

18

similar, and the annual-average of the monthly transition matrices (shown in Fig. 2)

19

accounted for 71% of the variance in the monthly matrices, with no identifiable patterns

20

in the residuals.

21

The serial correlation in the precipitation time-series noted in Section 2.2 is evident in

22

the transition matrix average shown in Fig. 2. The transition probabilities are not

23

uniform; days with low (first-decile) precipitation are most likely to be followed by

24

first-decile precipitation days, and the same for high (tenth-decile) precipitation days. In

25

general, the probability of transition to a similar decile is larger than for transitions to

26

“more distant” deciles.

27

3.1.3.

28

Conversion from precipitation-state to precipitation-amount is straightforward with the

29

double-stochastic matrix. States are converted to a continuous cumulative-frequency value F

AC C

EP

TE D

M AN U

SC

10

1

Empirical precipitation amounts function

Dividing a dataset into n states, each with m observations, yields only (n×m–1) transitions.

10

ACCEPTED MANUSCRIPT by adding a uniform random value. Thus, for the 10-state model, state 1 covers the interval

2

[0, 0.1), state 2 covers [0.1, 0.2), etc.

3

The F values were then converted into precipitation amounts using a “mostly non-parametric”

4

function. A non-parametric approach was chosen as it provides a simple way to obtain a good

5

fit to the observed data. The disadvantages of our non-parametric approach, as compared to a

6

parametric approach, are that the model is over-parameterized and it is difficult to estimate

7

parameter uncertainty. However, these disadvantages are not detrimental to the aims of the

8

current paper.

9

The amounts-function is based on the observed cumulative frequency distribution for

SC

RI PT

1

0 < F ≤ 0.995, and an exponential fit to the high-tail of the observations for 0.995 < F < 1.

11

The later provides a convenient extrapolation of the observed cumulative frequency

12

distribution to high F values, and allows extreme precipitation to be represented as a

13

continuous distribution. We do not claim that the underlying distribution for the upper-tail

14

actually follows an exponential distribution rather than, for example, a Pareto distribution

15

(Lennartsson et al., 2008). Rather, it is simply an admission that 50 years of data does not

16

permit determining the shape of the tail of the distribution with any certainty. The choice of

17

tail distribution is not critical for the purposes of this paper, which focuses on how well the

18

new WG represents multi-day precipitation event statistics.

19

For computational expediency, the precipitation amounts model was implemented by

20

sampling the cumulative frequency distribution as 200 (F, prcp) pairs. This represents a

21

compromise between a precise representation the observed cumulative frequency

22

distribution and computational efficiency. The sampling covered both the observed and

23

fitted sections of the cumulative frequency distribution. The sampling-intervals were

24

non-linear in both F and prcp. The sampling was initially conducted with fixed

25

intervals in prcp (0.01 mm, 0.02 mm, 0.03 mm, … , 0.19 mm), and thereafter with

26

logarithmic intervals up to 10 times the maximum observed precipitation. In practice,

27

the WG generates F values that are continuous on the interval [0, 1), and these were

28

converted to continuous values of prcp by linearly-interpolating between the two

29

enclosing (F, prcp) pairs. F values less than F0.01mm were set to give prcp=0 mm.

30

This precipitation amounts model, shown graphically in Fig. 3, is referred to as the

31

“empirical model” in the remainder of this paper.

AC C

EP

TE D

M AN U

10

11

ACCEPTED MANUSCRIPT 1

3.2.

2

3.2.1.

Experimental setup Alternative models: wet/dry and parametric amounts

In this paper we investigate how well the new 10-state empirical model simulated

4

precipitation for the Torne River catchment, and also investigated whether it had

5

advantages over simpler models.

6

For the later purposes, a WG with a two-state (wet/dry) Markov chain with a parametric

7

(Gamma distribution) precipitation amounts model was developed. Treating all non-

8

zero precipitation values as “wet” gave a very poor fit to the observed cumulative

9

frequency distribution, and a wet-day threshold of prcp = 0.1 mm (corresponding to

SC

RI PT

3

F~0.15, dependent on month) was found to be appropriate by trial-and-error.

11

The experimental setup was designed to identify which component of the new WG – the

12

transition model or amounts model – was most important for accurately simulating multi-day

13

precipitation events. Thus, both of the precipitation state models (two-state and 10-state)

14

were paired with both of the precipitation amounts models, as shown in Table 1. The wet/dry

15

parametric model closely resembles the classical Richardson weather generator, and does not

16

generate precipitation values between 0 mm and 0.1 mm. Despite the name, the wet/dry

17

empirical model can actually generate precipitation values between 0 mm and 0.1mm, as the

18

F for dry days was selected from the interval [0, F0.1mm), and precipitation determined from

19

cumulative frequency distribution resampling. For the 10-state parametric model, a day is

20

determined to be dry if F< F0.1mm, and precipitation is set to zero, whereas for wet days, F

21

from the Markov chain is scaled to the cumulative frequency within the wet day fraction, and

22

then the precipitation calculated using the inverse-cumulative-frequency Gamma function.

TE D

EP

3.2.2.

Calibration and validation

AC C

23

M AN U

10

24

The 50-year observed data record was divided into interleaved calibration (1961, 1963,

25

1965… 2009) and validation (1962, 1964, 1966… 2010) datasets to provide a consistency-

26

check on the model formulation and calibration. The WGs’ parameters were determined

27

using the calibration dataset, and the tables and figures presented in this paper generally show

28

both calibration and validation datasets and the model results. Note that, with interleaved

29

calibration/validation data, the validation results cannot be interpreted as an estimate of how

30

accurately the model might represent an independent (future) period. This is especially so for

31

the Torne River catchment, because the observed dataset contains statistically-significant

32

precipitation trends. The issue of deriving a method to calibrate the model that incorporates

12

ACCEPTED MANUSCRIPT

such trends is inseparable from the issue of incorporating the effects of global climate change

2

into the model. We defer this issue to a future study. Note also that the observed annual

3

precipitation time-series do not show serial correlation.

4

The WG models each ran 200 times to quantify the range of random variability, with each

5

run simulating the entire 50-year record.

6

3.3.

Statistical measures

RI PT

1

The two-sample Kolmogorov-Smirnov test statistic was used to quantify differences between

8

observed and simulated precipitation distributions, both for single and multi-day distributions.

9

The test statistic is calculated following Conover (1999): = max

,

−

,

(3)

M AN U



SC

7

where n is the sample size, F1, n and F2, n are the empirical cumulative frequency

11

functions for the observed and simulated samples, respectively, over all values of

12

precipitation x. ks=0 would indicate a perfect match between the cumulative frequency

13

curve of a model and the observations; ks=1 would indicate that the ranges of modeled

14

and observed distributions do not overlap. Note that the ks statistic is always positive,

15

so the median ks for the model runs indicates the typical maximum differences between

16

the distributions; it does not indicate whether the models “over-estimate” or “under-

17

estimate” the distribution. When calculating the ks test statistics for the parametric

18

models (which cannot simulate daily precipitation between 0 mm and 0.1 mm, by

19

definition), observed precipitation values < 0.1 mm were also set to zero.

20

Throughout this paper, the 95% confidence intervals displayed for model simulations

21

were estimated from the 95th-percentile-range of the 200 model runs. The 95%

22

confidence intervals for the annual and seasonal means and standard deviations of the

23

calibration and validation datasets in Table 2 and Table 3 were obtained by fitting

24

normal distributions to the data.

AC C

EP

TE D

10

13

ACCEPTED MANUSCRIPT 1 2

4.

Results 4.1.

Annual averages and annual cycle

All models reproduced the observed annual-average precipitation from the calibration

4

dataset almost exactly (Table 2). The differences between the calibration and modeled

5

means were much less than between the calibration and validation means, and can

6

therefore be considered insignificant.

7

All the models also accurately-reproduced the annual cycle in precipitation of the

8

calibration dataset (Fig. 4). The results for the 10-state empirical model are shown in

9

Fig. 4, but results for the other models are very similar.

4.2.

SC

10

RI PT

3

Interannual variability

The interannual variability in the annual precipitation – as characterized by the standard

12

deviation in annual totals – is under-estimated by all the models (Table 2). The best-

13

performing model is the 10-state empirical model, for which the standard deviation is

14

71 mm per year, which compares reasonably with (76 ± 23) mm per year for the

15

calibration dataset. However, the standard deviation in the validation dataset is (92 ± 28)

16

mm per year. A test in which the models were calibrated with the validation years gave

17

modeled standard deviations very similar to Table 2, suggesting that the

18

underestimation of interannual variability is an inherent property of the model, rather

19

than a calibration effect. Further tests in which daily extreme precipitation (defined as F

20

values exceeding 0.9 to 0.999, depending on the test) were removed prior to model

21

calibration reduced the standard deviation for observed and modeled datasets by similar

22

fractions, demonstrating that this deficiency is not caused by approximating the tail of

23

the observed precipitation distribution with an exponential function.

24

The average seasonal totals and their interannual variability are shown in Table 3. Note

25

that the values for the observations were calculated over the entire (1961–2012) dataset,

26

as otherwise the calibration/validation scheme splits the winter season. The 10-state

27

empirical model underestimates the standard deviation in winter and summer

28

precipitation, but values for spring and fall seasons are reasonable. Again, interannual

29

variation in the other models is too low, and the wet/dry models’ performance is worst.

AC C

EP

TE D

M AN U

11

14

ACCEPTED MANUSCRIPT 1

4.3.

Cumulative multi-day precipitation distributions

The most important aspect of the current study was to determine whether the new WG

3

generated realistic multi-day (2- to 14-day) accumulated precipitation. Quantile-quantile

4

residual plots (see Cox, 2007) for block-sum 1-, 2-, 5- and 14-day precipitation are shown in

5

Fig. 5. Quantile-quantile residual plots show residuals after subtracting the 1:1 line from a

6

regular quantile-quantile plot: for two perfectly-matched distributions, the residuals would be

7

a horizontal line at 0 mm. The results are further summarized as Kolmogorov-Smirnov (ks)

8

statistics in Table 4.

9

Fig. 5 shows that the residuals for the 10-state empirical model do not deviate significantly

RI PT

2

from zero, and this is clearly the best-performing model in the experiment. Similarly, the

11

ks statistics (Table 4) show the 10-state empirical model has the lowest value for all

12

precipitation durations. All other models generally overestimate low-precipitation totals

13

(positive anomalies in Fig. 5) and underestimate high totals (negative anomalies), with two

14

exceptions. Of these, the wet-dry empirical model simulates 1-day precipitation precisely,

15

which occurs almost by definition. More surprisingly, the 10-state parametric model

16

simulates of the 14-day precipitation distribution relatively accurately.

17

Curiously, the 10-state model matches the validation observations more closely than the

18

calibration observations for higher-quantile 5-day and 14-day duration precipitation. A test in

19

which the model was calibrated using the validation observations yielded much smaller

20

residuals for higher-quantile accumulated precipitation, suggesting that this pattern is a

21

feature of the model construction, rather than of calibration.

M AN U

TE D

4.4.

EP

22

SC

10

Average lengths of wet and dry spells

The average lengths of wet and dry spells vary over the course of the year, with the longest

24

wet spells occurring in winter while the longest dry spells are in spring. The distributions of

25

winter wet-spell lengths and spring dry-spell lengths are shown in Fig. 6. Note that the

26

average durations of wet-spells (defined as catchment-average prcp > 0.1 mm) are longer for

27

the catchment as a whole than for individual stations within the catchment.

28

The observed distributions of wet- and dry-spells lay generally-within the 95% confidence

29

bands of the models. The 10-state models slightly overestimate the number of shorter wet-

30

spells, and both models slightly overestimate the number of longer dry-spells, but neither

31

model performs remarkably better than the other. The results for the 10-state empirical and

AC C

23

15

ACCEPTED MANUSCRIPT 1

10-state parametric models were indistinguishable, as were the results for the wet/dry

2

empirical and wet/dry parametric models.

3 4

4.5.

Distribution of precipitation amounts for isolated wet days and during wet spells

The distribution of observed precipitation on wet (prep > 0.1 mm) days in the catchment

6

depends on the placing of a wet day within a wet-spell. This is shown graphically in Fig. 7.

7

Isolated wet days (grey line) have lower precipitation than the first or last day of a multi-day

8

wet-spell (red line), which in-turn have lower precipitation than the second or second-last

9

days of wet-spells (green line) or the remaining days in the middle of 5-day-or-longer wet-

RI PT

5

spells (blue line). This result holds for both the observations and the 10-state empirical model

11

simulations. Note that further subdivision of the day 3 to day N‒2 (blue curve) data did not

12

yield a further distinct distribution. Although Fig. 7 contains data from all seasons, similar

13

results were obtained for each season separately. The results from the 10-state parametric

14

model (not shown) were similar to the 10-state empirical model, although the model curves

15

were further from the calibration data curves. The wet/dry models cannot (by definition)

16

generate different precipitation distributions for different placements within a wet-spell, and

17

so are unable to replicate this feature of the precipitation time-series.

AC C

EP

TE D

M AN U

SC

10

16

ACCEPTED MANUSCRIPT

5. Discussion

2

Our results showed that the new multi-state empirical WG could realistically represent both

3

accumulated 2- to 14-day precipitation distributions and the way that the distribution of daily

4

precipitation changes according to the placement of a wet day within a wet-spell in the Torne

5

River catchment. However, the model underestimated inter-annual and inter-seasonal

6

variance. This suggests that the multi-state model is a promising candidate for hydrological

7

applications in the Torne River catchment, but that further development will be required.

8

5.1.

RI PT

1

Simulation of multi-day precipitation and wet-spells

For the purposes of flood-simulation, the ability to accurately simulate the distribution of

10

accumulated, multi-day precipitation is a primary consideration. The empirical amounts-

11

model (paired with either a 10-state or wet/dry transition matrix) represents the observed 1-

12

day precipitation distribution better than the parametric amounts-model. This is not surprising,

13

however, as the empirical amounts-model uses 200 parameters and is intended as a proxy for

14

“perfect knowledge” of the 1-day precipitation distribution.

15

For multi-day precipitation, however, it is the transition model, rather than the amounts

16

model, which most affects the models’ performances. The precipitation distributions

17

simulated by wet/dry models increasingly deviate from the observed distribution as

18

accumulation period increases, such that they overestimate low-precipitation totals and

19

underestimating high totals. In contrast, the 10-state parametric model simulates 14-day totals

20

almost as well as the 10-state empirical model. That is, for accumulated precipitation, the 10-

21

state transition model paired with a very simple amounts model (a Gamma distribution)

22

performed better than a wet/dry transition model paired with an empirical amounts model,

23

even though the latter represents “perfect knowledge” of the daily precipitation distribution.

M AN U

TE D

EP

AC C

24

SC

9

5.2.

Distribution of wet- and dry-spell lengths

25

The analysis of wet- and dry-spell lengths found that neither the 10-state nor the wet‒dry

26

transition model gave obviously-better simulations. This is perhaps surprising, given that the

27

wet/dry model was calibrated using a threshold of 0.1 mm to define the transition matrix, and

28

the same threshold was used for calculating wet- and dry-spell lengths. In contrast, 0.1 mm

29

does not correspond to any specific transition-threshold in the 10-state model.

17

ACCEPTED MANUSCRIPT 1

5.3.

Precipitation distribution during wet-spells

That the distribution of precipitation is different for isolated wet-days, and differs within wet-

3

spells, has been observed before (see Section 1). Reproducing this behavior is certainly

4

desirable for hydrological modeling, and our results show that a multi-state model offers a

5

method for representing this phenomenon. What is surprising in this study is that model

6

matches the observations so well, given that this phenomenon was not explicitly included in

7

the model calibration.

8

The differences between precipitation on isolated wet-days, days at the beginning or end of

9

wet-spells, and days in the middle of wet-spells, have previously been attributed to

RI PT

2

observational procedures (Harrold et al. (2003) and references therein); because precipitation

11

is measured over fixed 24-hour periods, first/last days should be thought of as only “partially

12

wet”. But we also found differences between the second and third days of 5-day-or-longer

13

wet-spells. A physical explanation for these differences is less obvious. A direct causal

14

relationship (i.e. heavy precipitation on a given day physically contributes to heavy

15

precipitation on a subsequent day, presumably via enhanced evaporation) seems unlikely in

16

northern Sweden, especially during winter. A 5-day wet-spell usually represents the passage

17

of a series of frontal or convective systems, so the enhanced precipitation towards the middle

18

of the wet-spells is presumably related to larger-scale circulation features such as

19

atmospheric blocking systems.

M AN U

5.4.

TE D

20

SC

10

Interannual variability

High flows in the Torne River occur when heavy precipitation accompanies a rapid melt of

22

the winter snow-pack. Thus, the ability to simulate interannual variability (i.e. the

23

accumulation of above-average snow depth) is an important consideration. The 10-state

24

empirical model underestimates interannual variability, with the worst statistics for summer

25

and winter seasons.

26

The winter interannual variance in the observed data (32 ± 7) mm vs. (22 ± 4) mm in the 10-

27

state empirical model, is associated with the influence of the NAO (Bartolini et al., 2009). A

28

rough-estimate of the contribution of the NAO to interannual winter precipitation in the

29

catchment can be obtained using simple linear regression; after subtracting the influence of

30

the NAO, the standard-deviation of the residuals reduces to (26 ± 11) mm in winter, which is

31

within the 95% confidence interval of the model’s winter standard-deviation. We note,

32

however, that some of the interannual variability associated with the NAO has already been

AC C

EP

21

18

ACCEPTED MANUSCRIPT implicitly subsumed into the model’s transition matrices; a fair comparison would require

2

subtracting the influence of the NAO from both the observations and from the model. For the

3

latter, of course, this is not possible.

4

The drivers of inter-annual summer precipitation variability in Europe are not as well-

5

understood as for winter (Zveryaev and Allan, 2010). However, the Scandinavian (“Eurasia-

6

1”) index of Barnston and Livezey (1987) and the Summer NAO (e.g. Folland et al., 2009)

7

index have statistically-significant correlations with summer precipitation. The Scandinavian

8

teleconnection pattern, defined using 500 hPa heights, has one action center over Scandinavia,

9

and action-centers with opposite sign over the northeastern Atlantic and Siberia (see Bueh

RI PT

1

and Nakamurais, 2007). Both these indices are associated with changes in storm-track density

11

and blocking frequency (Dong et al., 2013). A statistically-significant correlation was found

12

between summer precipitation in the Torne River catchment and the Scandinavian index

13

(Spearman correlation coefficient = –0.32), but not with Summer NAO (correlation

14

coefficient = –0.23). Removing the influence of the Scandinavian index on summer

15

precipitation by linear regression, as above, yielded a series with standard deviation of (54 ±

16

11) mm. This is still higher than the 10-state empirical model’s standard deviation of (49 ± 10)

17

mm in summer, but is within the 95% confidence interval.

18

6. Further work

19

We identify 3 lines of work that we believe need to be undertaken to confirm that the 10-state

20

model is suitable for hydrological simulations, and one important extension.

TE D

M AN U

6.1.

EP

21

SC

10

Improve interannual variance

The multi-state empirical WG model clearly needs to be further developed so it can

23

reproduce observed interannual variability. As discussed in the Introduction, many methods

24

have been proposed in the literature for conditioning WGs to show realistic low-frequency

25

variability. However, these might not be suitable. The most attractive feature of the multi-

26

state empirical model, for climate change impact studies, is that the daily precipitation

27

distribution and transition behaviors can be modeled and modified independently. Thus, the

28

challenge is to introduce low-frequency variability in such a way that it, too, can be modeled

29

and modified independently. This is not merely a desire for computational elegance:

30

Steinschneider and Brown (2013) found, in their study of the Connecticut River catchment,

AC C

22

19

ACCEPTED MANUSCRIPT that GCMs simulated an increase in annual mean precipitation with very little change in

2

interannual standard deviation.

3

Within the current modeling framework, a pragmatic approach could be to modulate the

4

random sequence used to convert precipitation state (i.e. decile) into a cumulative-frequency

5

value F. We have conducted a simple test to demonstrate that this technique can inflate the

6

modeled interannual variability sufficiently to match the observed. In the test, the WG

7

algorithm was modified so that F values for odd years were selected from the lower half of

8

each decile, and F values for even years were selected from the upper half. The result

9

doubled the standard-deviation in modeled annual-average precipitation. Even so, the overall

10

precipitation distribution and transition matrices in the WG output match those of the original

11

dataset.

12

6.2.

SC

RI PT

1

M AN U

Reducing the number of model parameters

Another line of future work would be to reduce number of model parameters. Thus far, we

14

have not made any systematic study of how many states a transition matrix needs in order to

15

give an acceptable simulation of multi-day precipitation. But there are other options for

16

reducing the number of model parameters, too.

17

The results from this study suggest that even a simple Gamma distribution, when paired with

18

the 10-state transition model, can represent multi-day accumulated precipitation well. This

19

suggests a simple way to reduce the number of model parameters would be to adopt a

20

parametric amounts-model, presumably a mixed-model so that extremes are accurately

21

represented. Alternatively, the non-parametric amounts-model could be retained, but the 200-

22

point sampling scheme used to represent the daily precipitation distribution replaced with a

23

series of splines.

24

In the present study, the transition matrices were derived completely-independently for each

25

month, but the inter-monthly differences were relatively small. Thus, another option for

26

reducing the number of model parameters would be to use the average annual transition

27

matrix, or at least to approximate the monthly transition matrices as linear combinations of

28

the annual matrix and a small set of modifier matrices. We note that some scheme for

29

allowing seasonal variation should be retained to facilitate incorporating seasonally-distinct

30

climate-change signals into the transition probabilities in future studies.

AC C

EP

TE D

13

20

ACCEPTED MANUSCRIPT 1

6.3.

Compare with dynamical downscaling approaches

Global climate models provide information about how the future climate will evolve under

3

enhanced-greenhouse conditions. Because GCM-outputs typically have a horizontal

4

resolution of 100−200 km, their outputs are downscaled in regional climate models (RCMs)

5

to produce climate time-series with higher spatial-resolution (12−50 km) over a limited area.

6

Even the outputs from RCMs, however, show biases compared to present-day local climate

7

conditions (Kotlarski et al., 2005). The Swedish Meteorological and Hydrological Institute

8

has developed bias-correction techniques that adjust RCM outputs so that they have a similar

9

statistical distribution to observed climate time-series (Yang et al. 2010). Such adjustments

10

are then applied to RCM outputs for future periods, using climate change simulations from

11

the ENSEMBLES project (van der Linden and Mitchell, 2009). These methods will be used

12

to generate local climate time-series that can be compared with the stochastic model results

13

presented in this paper.

6.4.

SC

M AN U

14

RI PT

2

Evaluate the model using other river basins

This study is exclusively concerned with the Torne River catchment in northern

16

Sweden/western Finland. In order to be useful to the wider community, the model will need

17

to be calibrated and evaluated for other river basins.

18

6.5.

TE D

15

Generation of complimentary temperature time-series

Hydrological models usually require more meteorological inputs than simply precipitation.

20

Temperature is particularly important, both for simulating snow-accumulation in high-

21

latitude or alpine catchments, and as a component of simulating evapotranspiration during the

22

growing season. Thus, the model will need to be extended to simulate temperature time-series

23

that are consistent with the precipitation time-series before it can be useful for hydrological

24

purposes.

25

7. Conclusions

26

Modeling catchment-average precipitation is a reasonable choice for studies that will

27

characterize the effect of climate change using global climate model outputs, as it avoids the

28

large scale-mismatch between Global Climate Model and sub-catchment scale. However,

29

modeling catchment-average precipitation in the Torne River catchment using a two-state

AC C

EP

19

21

ACCEPTED MANUSCRIPT

(wet/dry) Markov chain did not simulate the multi-day precipitation distribution well, nor the

2

variation in the distribution of daily precipitation within wet-spells.

3

In this paper, we introduced a new multi-state WG for catchment-average daily precipitation.

4

The new WG uses a doubly-stochastic transition-matrix, which has the property that the

5

overall precipitation distribution (including both wet and dry days) and the temporal-

6

correlation can be modified independently in climate change studies.

7

The new WG successfully simulated the distribution of accumulated, multi-day precipitation

8

in the catchment, which is a primary consideration for the purposes of flood-simulation.

9

However, the ability to simulate interannual and seasonal variability is also an important

RI PT

1

consideration, and the model does not perform well in this regard.

11

Thus, we conclude that the multi-state model is a promising candidate for hydrological

12

applications. However, further development is required to improve the simulation of

13

interannual variation, and evaluation using other river basins is also required.

14

Acknowledgements

15

Financial support for this study by the Swedish Civil Contingencies Agency, though the

16

project “Future rainfall and flooding in Sweden: a framework to support climate adaptation

17

actions”, is gratefully acknowledged.

18

References

19 20

Asp, S.-S., 2011. Sammanställning kalibrering och dimensionering av Torneälven samt Tärendöälven 2009-2011. Norrköping (in Swedish).

21 22

Barnston, A., Livezey, R., 1987. Classification, seasonality and persistence of low-frequency atmospheric circulation patterns. Mon. Weather Rev. 115, 1083–1126.

23 24 25

Bartolini, E., Claps, P., D’Odorico, P., 2009. Interannual variability of winter precipitation in the European Alps: relations with the North Atlantic Oscillation. Hydrol. Earth Syst. Sci. 13, 17–25.

26 27

Boughton, W.C., 1999. A daily rainfall generating model for water yield and flood studies 99/9. Cooperative Research Centre for Catchment Hydrology, Canberra.

28 29 30

Bowling, L.C., Lettenmaier, D.P., Nijssen, B., el al., 2003. Simulation of high-latitude hydrological processes in the Torne–Kalix basin: PILPS Phase 2(e). Glob. Planet. Change 38, 1–30.

31

AC C

EP

TE D

M AN U

SC

10

22

ACCEPTED MANUSCRIPT Bueh, C., Nakamura, H., 2007. Scandinavian pattern and its climatic impact. Q. J. R. Meteorol. Soc. 133, 2117–2131.

3 4

Carlsson, B., 1999. Some facts about the Torne and Kalix River basins: a contribution to the NEWBALTIC II workshop in Abisko, June 1999. SMHI Hydrol. 80.

5 6

Chen, D., Achberger, C., Ou, T., et al., 2015. Projecting future local precipitation and its extremes for Sweden. Geogr. Ann. Ser. A, Phys. Geogr. 97, 25–39.

7 8

Chen, J., Brissette, F., 2014. Stochastic generation of daily precipitation amounts: review and evaluation of different models. Clim. Res. 59, 189–206.

RI PT

1 2

Chen, J., Brissette, F., Leconte, R., 2012. Downscaling of weather generator parameters to quantify hydrological impacts of climate change. Clim. Res. 51, 185–200.

11 12 13

Ching, W.-K., Huang, X., Ng, M.K., et al., 2013. Markov Chains, International Series in Operations Research & Management Science. Springer US, Boston, MA. doi:10.1007/978-1-4614-6312-2

14

Conover, W.J., 1999. Practical Nonparametric Statistics, 3rd ed. Wiley, New York.

15 16

Cox, N.J., 2007. Stata tip 47: quantile-quantile plots without programming. Stata J. 7, 275– 279.

17 18 19

Dong, B., Sutton, R.T., Woollings, T., et al., 2013. Variability of the North Atlantic summer storm track: mechanisms and impacts on European climate. Environ. Res. Lett. 8, 034037.

20 21

Folland, C.K., Knight, J., Linderholm, H.W., et al., 2009. The summer North Atlantic Oscillation: past, present, and future. J. Clim. 22, 1082–1103.

22 23

Gregory, J., Wigley, T., Jones, P., 1993. Application of Markov models to area-average daily precipitation series and interannual variability in seasonal totals. Clim. Dyn. 8, 299–310.

24 25

Haan, C., Allen, D., Street, J., 1976. A Markov chain model of daily rainfall. Water Resour. Res. 12, 443–449.

26 27

Harrold, T.I., Sharma, A., Sheather, S.J., 2003. A nonparametric model for stochastic generation of daily rainfall amounts. Water Resour. Res. 39, SWC 8-1–8-12.

28 29 30

Jones, P.D., Harpham, C., Goodess, et al., 2011. Perturbing a Weather Generator using change factors derived from Regional Climate Model simulations. Nonlinear Process. Geophys. 18, 503–511.

31 32

Katz, R.W., Parlange, M.B., 1993. Effects of an index of atmospheric circulation on stochastic properties of precipitation. Water Resour. Res. 29, 2335–2344.

33 34

Katz, R.W., Zheng, X., 1999. Mixture model for overdispersion of precipitation. J. Clim. 12, 2528–2537.

AC C

EP

TE D

M AN U

SC

9 10

23

ACCEPTED MANUSCRIPT Kilsby, C.G., Jones, P.D., Burton, A., et al., 2007. A daily weather generator for use in climate change studies. Environ. Model. Softw. 22, 1705–1719.

3 4

Knight, P. A., 2008. The Sinkhorn–Knopp algorithm: convergence and applications. SIAM J. Matrix Anal. Appl. 30, 261–275.

5 6

Kotlarski, S., Block, A., Böhm, U., et al., 2005. Regional climate model simulations as input for hydrological applications: evaluation of uncertainties. Adv. Geosci. 5, 119–125.

7 8

Lennartsson, J., Baxevani, A., Chen, D., 2008. Modelling precipitation in Sweden using multiple step markov chains and a composite model. J. Hydrol. 363, 42–59.

RI PT

1 2

Mehrotra, R., Sharma, A., 2007a. A semi-parametric model for stochastic generation of multi-site daily rainfall exhibiting low-frequency variability. J. Hydrol. 335, 180–193.

11 12

Mehrotra, R., Sharma, A., 2007b. Preserving low-frequency variability in generated daily rainfall sequences. J. Hydrol. 345, 102–120.

13 14

Mehrotra, R., Srikanthan, R., Sharma, A., 2006. A comparison of three stochastic multi-site precipitation occurrence generators. J. Hydrol. 331, 280–292.

15

MSB, 2011. Seminarium och beredskapsövning Torne älv: Utvärderingsrapport (in Swedish).

16 17

MSB, 2012. Översvämningsförordningen - Sveriges genomförande av översvämningsdirektivet. Karlstad (in Swedish).

18 19

Racsko, P., Szeidl, L., Semenov, M., 1991. A serial approach to local stochastic weather models. Ecol. Modell. 57, 27–41.

20 21

Richardson, C.W., 1981. Stochastic simulation of daily precipitation, temperature, and solar radiation. Water Resour. Res. 17, 182–190.

22 23

Richardson, C.W., Wright, D.A., 1984. WGEN : A Model for Generating Daily Weather Variables. United States Department of Agriculture ARS-8.

24 25

Sharma, A., Lall, U., 1999. A nonparametric approach for daily rainfall simulation. Math. Comput. Simul. 48, 361–371.

26 27

Sinkhorn, R., Knopp, P., 1967. Concerning nonnegative matrices and doubly stochastic matrices. Pacific J. Math. 21, 343–348.

28 29

Srikanthan, R., Harrold, T.I., Sharma, A., et al., 2005. Comparison of two approaches for generation of daily rainfall data. Stoch. Environ. Res. Risk Assess. 19, 215–226.

30 31

Srikanthan, R., McMahon, T. A., 2001. Stochastic generation of annual, monthly and daily climate data: a review. Hydrol. Earth Syst. Sci. 5, 653–670.

32 33 34

Steinschneider, S., Brown, C., 2013. A semiparametric multivariate, multisite weather generator with low-frequency variability for use in climate risk assessments. Water Resour. Res. 49, 7205–7220.

AC C

EP

TE D

M AN U

SC

9 10

24

ACCEPTED MANUSCRIPT van der Linden, P., Mitchell, J.F.B., 2009. ENSEMBLES: Climate change and its impacts: summary of research and results from the ENSEMBLES project. Met Office Hadley Centre, Exeter, UK.

4 5 6

Wilby, R.L., Conway, D., Jones, P.D., 2002. Prospects for downscaling seasonal precipitation variability using conditioned weather generator parameters. Hydrol. Process. 16, 1215– 1234.

7 8

Wilks, D.S., 1998. Multisite generalization of a daily stochastic precipitation generation model. J. Hydrol. 210, 178–191.

RI PT

1 2 3

Wilks, D.S., Wilby, R.L., 1999. The weather generation game: a review of stochastic weather models. Prog. Phys. Geogr. 23, 329–357.

11 12

Wilks, D.S., 2010. Use of stochastic weather generators for precipitation downscaling. Wiley Interdiscip. Rev. Clim. Chang. 1, 898–907. doi:10.1002/wcc.85

13 14 15

Yang, W., Andréasson, J., Phil Graham, L., et al., 2010. Distribution-based scaling to improve usability of regional climate model projections for hydrological climate change impacts studies. Hydrol. Res. 41, 211–229.

16 17

Zveryaev, I.I., Allan, R.P., 2010. Summertime precipitation variability over Europe and its links to atmospheric dynamics and evaporation. J. Geophys. Res. 115, D12102.

M AN U

SC

9 10

18

AC C

EP

TE D

19

25

ACCEPTED MANUSCRIPT 1

Table 1. Summary of weather generator model configurations. Model name

Cfd resampling, 0 ≤ F ≤ 0.995 Exponential, 0.995 < F < 1 Dry (prcp≤0.1 mm), Cfd resampling, 0 ≤ F ≤ 0.995 Wet/dry empirical wet (prcp>0.1 mm) Exponential, 0.995 < F < 1 Gamma for prcp > F0.1mm, 10-state parametric Ten equally-likely zero otherwise. Gamma for prcp > 0.1 mm, Dry (prcp≤0.1 mm), Wet/dry parametric zero otherwise. wet (prcp>0.1 mm) Note: Cfd represents cumulative frequency distribution Ten equally-likely

AC C

EP

TE D

M AN U

SC

RI PT

10-state empirical

2

Precipitation amounts function

States

26

ACCEPTED MANUSCRIPT 1 2

Table 2. Mean and standard deviation in annual-average precipitation for observations

3

and models. Uncertainties are 95% confidence intervals. Standard deviation (mm per year) 76 ± 23 92 ± 28 71 ± 13 54 ± 10 64 ± 12 48 ± 10

SC

Observations, calibration Observations, validation 10-state empirical Wet/dry empirical 10-state parametric Wet/dry parametric

Mean (mm per year) 643 ± 31 611 ± 38 646 ± 19 645 ± 14 642 ± 17 641 ± 14

RI PT

Precipitation

AC C

EP

TE D

M AN U

4 5

ACCEPTED MANUSCRIPT

27 1

Precipitation mean (mm) Winter Spring Summer Fall Observations (1961‒2012) 109 ± 8 219 ± 16 170 ± 9 129 ± 9 10-state empirical 117 ± 8 223 ± 11 174 ± 10 131 ± 6 Wet/dry empirical 117 ± 6 222 ± 10 174 ± 7 131 ± 5 10-state parametric 116 ± 6 222 ± 13 173 ± 8 130 ± 5 Wet/dry parametric 117 ± 6 221 ± 10 174 ± 6 131 ± 4

Precipitation standard-deviation (mm) Winter Spring Summer Fall 32 ± 7 28 ± 6 58 ± 12 32 ± 7 23 ± 5 28 ± 5 50 ± 9 34 ± 8 18 ± 4 22 ± 5 37 ± 7 27 ± 5 21 ± 5 25 ± 6 45 ± 9 30 ± 5 16 ± 3 19 ± 4 33 ± 6 23 ± 4

M AN U

Observation/model

RI PT

Table 3: Mean and standard deviation in seasonal-average precipitation for observations and models.

SC

2

AC C

EP

TE D

3

28

ACCEPTED MANUSCRIPT 1 2

Table 4. Median Kolmogorov-Smirnov statistic between calibration observations and model

3

outputs for 1-, 2-, 5- and 14-day precipitation. Duration 2-day

5-day

14-day

Validation observations

0.02

0.03

0.04

0.07

10-state empirical model

0.01

0.01

0.02

Wet-dry empirical model

0.01

0.06

0.10

10-state parametric model

0.04

0.04

0.05

Wet-dry parametric model

0.04

0.10

0.13

AC C

EP

TE D

M AN U

5

0.03

0.10

0.05

0.12

SC

4

RI PT

1-day

29

RI PT

ACCEPTED MANUSCRIPT

SC

1

Fig. 1. Location and topography of the Torne River catchment (red outline), showing

3

major rivers and settlements.

M AN U

2

AC C

EP

TE D

4

30

RI PT

ACCEPTED MANUSCRIPT

SC

1

Fig. 2. Annual-average of the monthly catchment-average precipitation transition matrices.

3

The distribution of precipitation amounts is shown in Fig. 3

M AN U

2

AC C

EP

TE D

4

31

RI PT

ACCEPTED MANUSCRIPT

SC

1

Fig. 3. The “empirical” precipitation amounts model for the Torne River catchment. The

3

model consists of 200 (F, precipitation) pairs for each month, but here months with similar

4

values have been averaged into seasonal curves for clarity. The first 30 individual (F,

5

precipitation) points are shown for each season, after which the data are plotted as continuous

6

curves.

M AN U

2

AC C

EP

TE D

7

32

ACCEPTED MANUSCRIPT 100

60

40 Calibratrion obs. Validation obs. 10-state empirical 10-state empirical 95% CI

20

0 Jan

Feb Mar Apr May Jun

Jul

Aug Sep Oct Nov Dec

Month

SC

1

RI PT

Precipitation (mm)

80

Fig. 4. Annual cycle for precipitation simulated by the 10-state empirical model. Note the

3

calibration observations (dashed black line) and the mean of the 200 model runs (solid red

4

line) are essentially identical. Shading indicates 95th percentile range of 200 model runs.

M AN U

2

AC C

EP

TE D

5

33

M AN U

SC

RI PT

ACCEPTED MANUSCRIPT

1

Fig. 5. Quantile-quantile residuals for (a) 1-day, (b) block-sum 2-day, (c) 5-day, and (d) 14-

3

day precipitation. The horizontal axes describe quantiles 0.01–0.99 of the calibration

4

observations, and the vertical axis the difference with the matching quantile from the model

5

simulations or validation observations. Shading shows the 95th percentile range from the 10-

6

state empirical model runs; the ranges for the other models are similar, and are omitted for

7

clarity. Note the changes in vertical and horizontal scales from (a) to (d).

EP

AC C

8

TE D

2

34

1.0

1.0 a) 0.8

0.5 0.6

10

100

0.4 1

Cumulative frequency

1.0

10

Spring wet-spell length (days)

RI PT

0 1

Winter wet-spell length (days) 1.0 c)

Wet/dry empirical Calibration obs. Validation obs. model 95% CI

d) 0.8

0.5 0.6 0 1

10

100

Winter wet-spell length (days)

0.4 1

10

Spring wet-spell length (days)

M AN U

1

10-bin empirical Calibration obs. Validation obs. model 95% CI

b)

SC

Cumulative frequency

ACCEPTED MANUSCRIPT

2

Fig. 6. Cumulative probability that a winter wet-spell (a, c) or a spring dry-spell (b, d) has a

3

given duration. Top row (a, b) the 10-bin empirical model, and bottom row (c, d) the wet/dry

4

empirical model. Shading indicates 95th-percentile-ranges calculated from 200 model runs.

AC C

EP

TE D

5

35

SC

RI PT

ACCEPTED MANUSCRIPT

1

Fig. 7. Cumulative frequency distributions for days classed according to their position in a

3

wet-spell of length N. Days are classed as: (grey) isolated wet days; (red) the first (day 1) or

4

last (day N) day of a wet-spell; (green) second (day 2) or second-last (N–1) day; (blue) all

5

other days (day 3 to N–2). Solid lines show median of the 10-state empirical model runs;

6

shading indicates the 95th percentile range; calibration observations shown as dashed lines.

M AN U

2

AC C

EP

TE D

7