Convection-resolving precipitation forecasting and its predictability in Alpine river catchments

17
Convection-resolving precipitation forecasting and its predictability in Alpine river catchments Andre ´ Walser * , Christoph Scha ¨r Institute for Atmospheric and Climate Science, ETH, Zurich, Switzerland Received 14 December 2002; accepted 20 November 2003 Abstract Predictability limitations in quantitative precipitation forecasting arising from small-scale uncertainties in the initial conditions are investigated for Alpine river catchments, with particular consideration of their implications on hydrological runoff forecasting. To this end, convection-resolving ensembles of limited-area simulations are performed using a nonhydrostatic numerical weather prediction (NWP) model, and results are analysed in terms of catchment-averaged precipitation. The applied ensemble strategy uses slightly modified initial conditions representing observational uncertainties, but identical lateral boundary conditions representing a perfectly predictable synoptic-scale forcing. A total of four case studies is carried out for different synoptic conditions leading to heavy precipitation. Ensemble integrations of 12 members are analysed for 24-h forecasting periods, with particular attention paid to precipitation in the Po basin and in its sub-catchments in the Lago Maggiore area. The simulations exhibit a large variability in the predictability of precipitation amounts, both from case to case and from catchment to catchment. It is demonstrated for an episode of thermal convection, that the predictability may be very low even in large-scale catchments of , 50,000 km 2 . In more synoptically dominated cases, predictability limitations appear to be restricted to catchments smaller than , 10,000 km 2 , while in one case predictability is found to be high in catchments as small as 200 km 2 . Overall, the simulations show that precipitation forecasts for alpine river catchments may on occasions be critically affected by predictability limitations, even though the NWP model and the synoptic-scale forcing are assumed to be prefect. It is demonstrated that a substantial fraction of the predictability limitations is due to the scattered and unpredictable occurrence of convective cells, but the presence of convective precipitation alone does not necessarily limit predictability. It is also shown that the predictability is systematically higher in mountainous catchments. q 2003 Elsevier B.V. All rights reserved. Keywords: Ensemble simulation; Probabilistic forecast; Quantitative precipitation forecasting; High-resolution numerical weather prediction 1. Introduction The Alpine region is frequently affected by flood events due to exceptional precipitation amounts and intensities. Some of these events are devastating and inflict loss of life and large property damage 0022-1694/$ - see front matter q 2003 Elsevier B.V. All rights reserved. doi:10.1016/j.jhydrol.2003.11.035 Journal of Hydrology 288 (2004) 57–73 www.elsevier.com/locate/jhydrol * Corresponding author. Present Address: MeteoSwiss, Zurich, Switzerland. Tel.: þ41-1-256-94-43; fax: þ 41-1-256-92-78. E-mail address: [email protected] (A. Walser).

Transcript of Convection-resolving precipitation forecasting and its predictability in Alpine river catchments

Convection-resolving precipitation forecasting and its

predictability in Alpine river catchments

Andre Walser*, Christoph Schar

Institute for Atmospheric and Climate Science, ETH, Zurich, Switzerland

Received 14 December 2002; accepted 20 November 2003

Abstract

Predictability limitations in quantitative precipitation forecasting arising from small-scale uncertainties in the initial

conditions are investigated for Alpine river catchments, with particular consideration of their implications on hydrological

runoff forecasting. To this end, convection-resolving ensembles of limited-area simulations are performed using a

nonhydrostatic numerical weather prediction (NWP) model, and results are analysed in terms of catchment-averaged

precipitation. The applied ensemble strategy uses slightly modified initial conditions representing observational uncertainties,

but identical lateral boundary conditions representing a perfectly predictable synoptic-scale forcing.

A total of four case studies is carried out for different synoptic conditions leading to heavy precipitation. Ensemble

integrations of 12 members are analysed for 24-h forecasting periods, with particular attention paid to precipitation in the Po

basin and in its sub-catchments in the Lago Maggiore area. The simulations exhibit a large variability in the predictability of

precipitation amounts, both from case to case and from catchment to catchment. It is demonstrated for an episode of thermal

convection, that the predictability may be very low even in large-scale catchments of ,50,000 km2. In more synoptically

dominated cases, predictability limitations appear to be restricted to catchments smaller than ,10,000 km2, while in one case

predictability is found to be high in catchments as small as 200 km2. Overall, the simulations show that precipitation forecasts

for alpine river catchments may on occasions be critically affected by predictability limitations, even though the NWP model

and the synoptic-scale forcing are assumed to be prefect. It is demonstrated that a substantial fraction of the predictability

limitations is due to the scattered and unpredictable occurrence of convective cells, but the presence of convective precipitation

alone does not necessarily limit predictability. It is also shown that the predictability is systematically higher in mountainous

catchments.

q 2003 Elsevier B.V. All rights reserved.

Keywords: Ensemble simulation; Probabilistic forecast; Quantitative precipitation forecasting; High-resolution numerical weather prediction

1. Introduction

The Alpine region is frequently affected by flood

events due to exceptional precipitation amounts and

intensities. Some of these events are devastating

and inflict loss of life and large property damage

0022-1694/$ - see front matter q 2003 Elsevier B.V. All rights reserved.

doi:10.1016/j.jhydrol.2003.11.035

Journal of Hydrology 288 (2004) 57–73

www.elsevier.com/locate/jhydrol

* Corresponding author. Present Address: MeteoSwiss, Zurich,

Switzerland. Tel.: þ41-1-256-94-43; fax: þ41-1-256-92-78.

E-mail address: [email protected] (A. Walser).

(Frei et al., 2000). Experience suggest that appropriate

warnings could substantially mitigate the conse-

quences of such events. Flood forecasting and

warning systems in intermediate-scale catchments

(,50,000 km2) best employ both an atmospheric and

a hydrological numerical model. In order to improve

such systems, much research has been undertaken in

the last years towards the use of coupled numerical

weather prediction (NWP) and hydrologic models

(Miller and Kim, 1996; Yu et al., 1999; Benoit et al.,

2000; Droegemeier et al., 2000; Jasper and Kauf-

mann, 2003).

For the prediction of critical river stages, quanti-

tative precipitation is the most decisive variable

among the external inputs of hydrologic models.

However, forecasts of precipitation have an inherent

uncertainty, since deterministic NWPs are intrinsi-

cally limited by the chaotic nature of the atmospheric

dynamics. Already in the 1960s, Lorenz (1963)

demonstrated in a seminal study that small errors in

the initial conditions of a weather forecast can grow

rapidly, leading to highly diverging solutions. The

degree of this divergence, and hence the uncertainty,

varies from one occasion to the next. In order to take

this uncertainty into account, probabilistic river stage

forecasts (PRSF) have been proposed in the past

decade (Fritsch et al., 1998; Krzysztofowicz, 1998).

However, most hydrological forecasting systems still

produce deterministic forecasts, finding a single

estimate without quantifying the uncertainty. The

currently available short-term PRSF are based on a

statistical quantification of both the precipitation

uncertainty (Kelly and Krzysztofowicz, 2000) and

the hydrologic uncertainty (Krzysztofowicz and Herr,

2001). It is expected that in future the input of

precipitation for PRSF will be derived from high-

resolution ensemble simulations of NWP models

which provide a quantitative dynamical approach for

probabilistic precipitation forecasts.

In the last decades, much research has been

devoted to estimate the prediction of the uncertainty

in numerical weather forecasts (see the reviews by

Ehrendorfer, 1997 and Palmer, 2000). Only in the

1990s, computing resources have become sufficient to

implement probabilistic atmospheric forecasts into

operational practice (Molteni et al., 1996; Houteka-

mer et al., 1996; Toth and Kalnay, 1997), using global

ensemble prediction systems (EPSs). Such systems

are based on many parallel NWPs (ensemble

members), starting from slightly different initial

conditions. In addition, uncertainties in the model

formulations are taken into account by the use of

stochastic perturbations to physical tendencies

(Buizza et al., 1999) or multi-model approaches

(Grimit and Mass, 2002). Ideally, ensemble predic-

tions should span the entire space of possible

solutions. In practice, however, the atmosphere has

too many degrees of freedom to fully cover its

probabilistic behaviour. EPSs were originally devel-

oped for the evolution of baroclinic perturbations and

designed for medium-range weather forecasts. In the

last decade, studies have been devoted also to

mesoscale predictability of precipitation using lim-

ited-area models (Stensrud et al., 2000; Marsigli et al.,

2001; Frogner and Iversen, 2002) and have accounted

for moist physics in detecting strong error growth

(Ehrendorfer et al., 1999). Nevertheless, the step

towards short-range high-resolution EPS forecasts for

hydrological applications still requires a further

increase in spatial resolution.

Runoff forecasts for typical alpine river catchments

are often affected by the spatiotemporal occurrence as

well as the intensity of convective systems and squall

lines. Such meso-b scale precipitation systems have

short time-scales and their evolution is highly non-

linear. It is still unknown at what catchment scale a

prediction of such systems is skilful. In a recent study,

we found that the range of scale affected by small-

scale predictability limitations may extend up to the

scale of several hundred kilometres (Walser et al.,

2004), even when the synoptic-scale forcing is

perfectly known. This result implies that, at least on

occasions, NWPs may be unable to provide a useful

area-mean representation of convective precipitation

even in major alpine river basins.

The main purpose of this study is to examine the

uncertainty in mean-catchment precipitation arising

from small-scale observational uncertainties. Predict-

ability is thus investigated in the absence of larger-

scale uncertainties, assuming perfectly predictable

synoptic-scale conditions. Following earlier studies in

short-range ensemble forecasting (Du et al., 1997;

Stensrud et al., 1999), we assume a perfect model, and

the resulting estimates will thus be referred to as

potential predictability. Our study quantifies the

potential predictability of precipitation in selected

A. Walser, C. Schar / Journal of Hydrology 288 (2004) 57–7358

Alpine river catchments in four case studies, using a

variant of the ensemble methodology. In order to

simulate the small-scale processes as reliable as

possible, the underlying simulations were performed

in a convection-resolving setup with 3 km horizontal

grid spacing.

The paper is structured as follows: the relevant

features of the model and the experimental setup are

described in Section 2, complemented by background

information on the process of deep convection in the

simulations. Section 3 presents the results of our

ensemble experiments on the basis of four case

studies, including an intercomparison with radar

observations. Finally, conclusions of the study are

presented in Section 4.

2. Experimental design

2.1. Model chain

The model chain and simulation strategy used in

this study has been introduced in Walser et al. (2004).

The numerical weather simulations were undertaken

with a model chain that includes two limited-area

NWP models and that is driven by the operational

European Centre for Medium-Range Weather Fore-

casts (ECMWF) analysis. This analysis assimilates all

available observations in a spatially and temporally

consistent fashion with the help of a global NWP

model. The driving analyses of the year 1999 have a

horizontal mesh size of about 0.58 ðTL ¼ 319Þ; 50

vertical levels (60 since November) and a six hourly

temporal resolution.

The High Resolution Model (HRM) is a hydro-

static limited-area model developed at the German

Weather Service. For this study, it is used with a

horizontal grid spacing of 0.1258 over Central Europe

(Fig. 1), and 31 vertical levels in a hybrid sigma

coordinate system. In the present model chain, the

operational ECMWF analysis provides the initial and

lateral boundary conditions for the HRM simulation.

The Canadian Mesoscale Compressible Commu-

nity (MC2) limited-area model is based on the

compressible set of nonhydrostatic equations which

govern the atmospheric dynamics. The model is

thus suited for the simulation of atmospheric flow at

very high resolution. The prognostic equations are

formulated in a semi-implicit, semi-Lagrangian

scheme (Tanguay et al., 1990; Benoit et al., 1997).

The model was used during the special observing

period of the Mesoscale Alpine Programme (MAP

SOP; Bougeault et al., 2001) in a quasi-operational

setup (Benoit et al., 2002; Schar et al., 2003). The

same physical setup (MC2 version 4.9) is used for this

study. Nested in the HRM run, a MC2 simulation with

14 km horizontal grid spacing and 35 vertical levels is

first conducted for an integration period of 34 h. This

MC2 simulation on a slightly smaller domain and with

basically the same resolution as the HRM simulation

allows a smooth transition to the height-based Gal-

Chen Somerville coordinates (Gal-Chen and Somer-

ville, 1975) and the different physical schemes used in

the MC2. It provides the initial and lateral boundary

conditions for MC2 ensemble simulations over the

European Alps (innermost domain in Fig. 1). In order

to resolve convection explicitly (discussed later in

Section 2.4), a 3 km grid spacing and 50 vertical

levels are used for these ensemble simulations,

yielding a horizontal domain of 343 £ 293 grid points.

2.2. Ensemble strategy

Ensemble forecasting makes use of a series of

model integrations that slightly differ from one

Fig. 1. Computational domains of the HRM (outer box), MC2 with

14 km grid spacing (middle box), and MC2 with 3 km grid spacing

(innermost box) superimposed on the HRM topography.

A. Walser, C. Schar / Journal of Hydrology 288 (2004) 57–73 59

another in terms of initial conditions. The spread of

the ensemble members during the integration period

then represents a measure of the predictability. As

typical atmospheric models have a huge number of

degrees of freedom (for our model setup about

5 £ 107), the ensemble strategy can only partly

represent the full spectrum of uncertainty. The

choice of the initial conditions of the individual

ensemble members is thus of key importance. To this

end, a large number of techniques have been

developed (Section 1). The aim of these is to find

initial perturbations that (i) grow quickly in time, and

(ii) to realistically represent typical observational and

analysis uncertainties. For systems with an approxi-

mately linear behaviour, rapidly growing pertur-

bations may be derived using singular vector

analysis. This analysis delivers linear structures that

exhibit a maximum linear growth over some

specified time period, and the initial amplitude of

these perturbations is specified such as to match

typical analysis uncertainties. Singular vector tech-

niques have been highly successful in global NWP

(Buizza and Palmer, 1995) to identify large-scale

baroclinic perturbations that may grow into forecast-

ing errors. However, on the small scales considered

in the present study, alternative growth mechanisms

need to be addressed, such as moist dynamical

processes and convective instabilities. In addition,

due to the large growth rates of small-scale

perturbations (which often possess doubling times

below an hour), the assumption of a linear system

appears poorly justified.

The present study uses an ensemble method-

ology based on a lagged initialisation technique.

Each of the ensemble members uses identical

lateral boundary conditions but slightly modified

initial conditions realised by shifting the initialisa-

tion time of six simulations. Member 1 is initialised

at 2100 UTC, member 2 at 2000 UTC, etc. and

finally member 6 at 1600 UTC. This procedure is

shown in Fig. 2. Within this framework, differences

between ensemble members essentially derive from

model differences between the driving low-resol-

ution simulation (MC2 14 km) and the target high-

resolution simulations (MC2 3 km). As all ensem-

ble members entail the same observational data (as

contained in the driving model simulation), it is not

surprising that the generated perturbations have

small amplitudes (Walser et al., 2004), and this will

require amplification after some initial integration

time, such as to represent initial analysis uncer-

tainty. Such a procedure is straightforward, pro-

vided that the amplitude of the perturbations is in

the range of validity of linear theory.

In order to amplify the initial perturbations, the

deviations from the ensemble mean of each

member are amplified at 0000 UTC by some factor

a: This is performed uniformly for temperature,

humidity, horizontal wind and pressure with a ¼ 3;

leading to six atmospheric states for 0000 UTC

Fig. 2. Model chain and setup for the ensemble simulations. The MC2 ensemble members (horizontal thin arrows) utilise a high-resolution grid

(3 km grid spacing, 50 vertical levels) and are generated using a shifted initialisation strategy followed by an amplification of the perturbations

and a change of the perturbations’ sign at 0000 UTC (see text).

A. Walser, C. Schar / Journal of Hydrology 288 (2004) 57–7360

with enlarged perturbations. Six additional atmos-

pheric states are obtained by setting a ¼ 23; that

is, by inverting the sign of the amplified pertur-

bations. Together, the 12 different atmospheric

states are used as initial conditions to perform 12

members with a 24-h integration period from 0000

to 2400 UTC. Since the perturbations are compara-

tively small, the integrations are continued without

using an additional initialisation procedure after the

amplification/inversion of the perturbations with the

a-factor. Negative values for humidity resulting

from the amplification are set to zero. Super-

saturation is not corrected but is removed in the

first few time steps of the model integration.

The proposed ensemble strategy is not meant as a

setup for an operational forecasting system. However,

it allows isolating predictability issues related to

meso-b scale, since the identical lateral boundary

conditions for the ensemble members prevent synop-

tic-scale perturbations. The method thus reveals the

inherent uncertainty in high-resolution weather fore-

casts due to nonlinear error growth of small-scale

errors in the initial conditions. With the proposed

ensemble strategy, the initial perturbations are

generated by design choices of the driving NWP

model chain. Such small-scale perturbations have a

tendency to grow within the forecasting system

(Walser et al., 2004), such that presumably fewer

members are needed for a representation of the

uncertainty as compared to randomly generated

perturbations. However, our EPS strategy is ad hoc

in many regards and does hardly produce those

perturbations that would grow faster. In this respect,

our ensemble simulations may suggest rather too

optimistic (i.e. low) uncertainty levels.

2.3. Analysis strategy

Ensemble simulations are performed for four case

studies. These were chosen from the year 1999 and

include a strong convective summer day, two heavy

precipitation events with embedded convection in

September, and a late autumn frontal passage. An

overview of these cases is presented together with the

discussion of the results in Section 3.

For the analysis of the ensemble simulations

attention is given to precipitation in six sub-catch-

ments of the Po basin, located to the south of the main

Alpine crest: the river catchments Verzasca, Maggia,

Ticino, Toce, Lago Maggiore as well as the Po

catchment upstream of Piacenza (see Fig. 3 and

Table 1). In order to include the entire eastern flat part

of the Po Valley (from which a large fraction does not

belong to the Po river basin), a hypothetical extended

Po catchment is defined which encompasses the whole

Po plain between Venezia in the North and Ravenna

in the South. The main idea behind selecting these

catchments is to span a wide range of scales from

186 km2 (Verzasca) to 120,000 km2 (extended Po).

2.4. Convection in a cloud-resolving model

In this section, we present some background

information on the process of deep convection in

our simulations. The mesh size of 3 km allows

Table 1

Catchments used for the analysis of the ensemble simulations

Catchment Gauge/gorge Area (km2)

Extended Po ,120,000

Po Piacenza 42,030

Lago Maggiore Miorina 6599

Toce Candoglia 1531

Ticino Bellinzona 1515

Maggia Locarno–Solduno 926

Verzasca Lavertezzo 186

Fig. 3. Catchment definitions used for the analysis: extended Po

catchment (blue line), Po catchment upstream of Piacenza (red),

Lago Maggiore (black), Toce (green), Maggia (violet), Verzasca

(white), and Ticino (gold). The background field shows the MC2

model topography of the innermost computational domain.

A. Walser, C. Schar / Journal of Hydrology 288 (2004) 57–73 61

an explicit simulation of deep convection, i.e. the

vertical motions and condensation/evaporation in up-

and downdrafts are explicitly represented. The cloud

physics is treated with an advanced version of Kong

and Yau (1997) explicit cloud microphysics scheme.

It includes a bulk representation of the five water

species water vapour, cloud water, rain water, ice

crystals and graupel (Misra et al., 2000). Ice

microphysics is of particular importance for the

simulation of Alpine heavy precipitation (Richard

et al., 2003; Yuter and Houze, 2003). At present, it is

still somewhat unclear whether a horizontal grid

spacing of 3 km resolves convection sufficiently,

despite the promising experience from the MAP

experiment with the same setup as used in our study

(Benoit et al., 2002).

Moist deep convection requires three ingredients:

instability, moisture, and some vertical lifting. In a dry

atmosphere, the criteria for static instability can be

deduced from the first law of thermodynamics

(Holton, 1992) and requires that the potential

temperature

u ¼ Tðps=pÞR=cp ð1Þ

decreases with height, i.e. ð›u=›zÞ , 0: Here, T is the

temperature; R; the gas constant for dry air

(287 J K21 kg21); cp is the specific heat of dry air at

constant pressure (1004 J K21 kg21). In a moist

atmosphere the situation is more complicated since

the release of condensational energy has to be taken

into account. This leads to the equivalent potential

temperature, ue; defined as the potential temperature

that an air parcel would have if all its moisture is

condensed and the resulting latent heat is used to

warm the parcel. For a saturated parcel, ue is defined

as (Holton, 1992, p. 289)

ue < u expðLcqs=cpTÞ ð2Þ

where Lc is the latent heat of condensation

(2.5 £ 106 J kg21 at 0 8C), and qs the saturation

mixing ratio (mass of vapour per unit mass of dry

air in a saturated parcel). If ue decreases with height,

the atmosphere is referred to as potentially unstable.

The term potential here signals that the unstable layer

must first undergo some finite vertical displacement

and reach saturation in order to release the instability

(Schultz and Schumacher, 1999; Houze, 1993).

Fig. 4 shows an example of modeled convective

activity due to a potentially unstable lower tropo-

sphere in the Lago Maggiore area. In the evening of

20 September 1999 (for a case description see later in

Section 3.1) a strong southerly flow impinges upon the

Alps and the orographically forced lifting releases the

instability. Two convective cells produce strong

updrafts and precipitation rates exceeding 25 mm/h.

The characteristic time-scale of such convective cells

is typically only a few hours, for the two cells shown

in Fig. 4 even less.

The temporal evolution of the convective cell to

the north is shown in the vertical section in Fig. 5. The

baseline of this section is represented by the black line

in Fig. 4. The main flow is almost along the cross-

section with a strong vertical wind shear (the

horizontal wind increases from 10 m/s in the bound-

ary layer to about 40 m/s at 300 hPa). Panel (a) shows

the updraft roughly 1 h after the initial triggering of

the cell, while it is still growing.

The extent of the convective cloud centred at a

height of 6 km is represented by the cloud water

concentration (liquid and solid phases). Panel (b)

Fig. 4. MC2 simulated 12 min accumulated precipitation (grey-

scales) in millimetre per hour for 2000 UTC 20 Sep 1999,

superimposed by wind vectors at 700 hPa and by vertical velocity at

500 hPa (contours 1 m/s, with zero-contour omitted) showing

updrafts (solid contours) and downdrafts (dashed contours) in a

strong southerly flow. The diagram zooms into the Swiss/Italian

Lago Maggiore area. The black line indicates the vertical section

used in Fig. 5. Wind arrows are shown at every second grid point.

A. Walser, C. Schar / Journal of Hydrology 288 (2004) 57–7362

shows the cell 12 min later at the time of maximum

updraft (9 m/s) and with a vertical extension up to the

tropopause. In panel (c), again 12 min later, the

updraft becomes weaker and a secondary updraft

evolves upstream. The horizontal displacement of the

primary deep cell is about 21 km in 24 min,

corresponding to a propagation speed of 15 m/s. In

comparison, the horizontal wind velocity amounts to

25 m/s at the level of the maximum updraft.

Although state-of-the-art NWP models are capable

to resolve large-scale convective clouds quite realis-

tically, one cannot expect a skilful prediction of the

spatial and temporal occurrence of convective cells.

Convection is an intrinsically chaotic process which

implies that the evolution of convection depends upon

the initial conditions in a highly nonlinear way.

3. Results

In this section, we first present a brief overview of

the weather conditions for the four case studies. In the

following, we then analyse the model’s ability to

simulate precipitation by intercomparison against

radar data, and assess the potential predictability of

precipitation in Alpine river catchments.

3.1. Description of the case studies

The four day–long case studies are chosen from

the year 1999 such as to involve heavy precipitation

under different atmospheric stratifications and differ-

ent synoptic conditions. Three of these cases are in the

MAP SOP. All cases have already been discussed in

Walser et al. (2004).

The 29 July 1999 was characterised by a flat

pressure distribution over Central Europe and thus by

a weak synoptic forcing (Fig. 6a). In the entire Alpine

area strong thermal convection developed in the

potentially unstable stratified atmosphere, producing

locally heavy precipitation. Daily accumulated pre-

cipitation exceeding 100 mm are simulated by the

ensemble simulations over some Alpine peaks (see

later in Fig. 8b).

The 20 September 1999 belongs to the MAP

intensive observing period (IOP) 2b. It represents the

strongest precipitation event during the whole MAP

SOP and is the subject of several recent studies

(Medina and Houze, 2003; Rotunno and Ferretti,

2003). A trough was approaching the Alps from west,

leading to a strong persistent low-level moist flow

from the Mediterranean Sea towards the south side of

the Alpine barrier (Fig. 6b). The flow, strong at all

levels, was warm, moist and potentially unstable in

the lower troposphere. Hence, the air rose easily over

the rising terrain, releasing the instability and

favouring the development of convective cells.

Radar observations show moderate convection during

the day and deep convection in the afternoon

(Rotunno and Ferretti, 2003). Our simulations suggest

sporadic deep convection in the afternoon after the

passage of the cold front (as discussed in Section 2.4).

Fig. 5. Time series of cloud water concentration (condensed water and ice crystals) (g/kg) of a convective cell along the south–north vertical

section indicated by the black line in Fig. 4. The vertical velocity (contours 1 m/s, with zero-contour omitted) is overlaid and shows the strong

updraft (solid contours) and weak downdrafts (dashed contours). The time interval between the panels is 12 min, and the first panel is at 1948

UTC 20 Sep 1999. The axes are labeled in kilometres.

A. Walser, C. Schar / Journal of Hydrology 288 (2004) 57–73 63

According to the dataset of Frei and Haller (2001), the

heavy rainfall intensity led to more than 100 mm

accumulated daily precipitation in the Lago Maggiore

(LM) area and on the Alpine slopes of the Friuli-

Veneto region.

The 25 September 1999 was part of MAP IOP 3,

which is characterised by a similar synoptic

situation as IOP 2b. A trough extending from the

British Isles to Spain propagated slowly eastwards,

leading to a persistent south-westerly flow over the

Alpine region (Fig. 6c). South of the Alps, this

moist potentially unstable flow impinges on the

Alpine slopes, leading to heavy prefrontal precipi-

tation enhanced by embedded convection in the LM

area. In contrast to the 20 September, the heavy

rainfall is confined to the LM area and to south-

eastern France where strong frontal precipitation

occurred.

On 6 November 1999 (MAP IOP 15), a cutoff

low originating from the British Isles moved

quickly towards the Alps, advecting very cold air

towards the Mediterranean Sea (Fig. 6d). This

evolution was followed by a fast lee cyclogenesis

event centred over the Gulf of Genova. In the LM

area, this led to a southerly, later south-easterly

flow towards the Alpine barrier, before a northerly

flow established. The maximum rainfall amounts

were observed in the eastern Po Valley, where the

daily sum exceeded 60 mm (dataset of Frei and

Haller, 2001).

3.2. Comparison with radar observations

An objective validation of the MC2’s simulation of

precipitation was performed by Jasper and Kaufmann

(2003) for the Ticino–Verzasca–Maggia basin. They

Fig. 6. Overviews of the synoptic situation in the four case studies in terms of geopotential height (m) at 850 hPa (contour interval is 40 m) at

1200 UTC from ECMWF analysis for (a) 29 Jul 1999, (b) 20 Sep 1999, (c) 25 Sep 1999, and (d) 6 Nov 1999.

A. Walser, C. Schar / Journal of Hydrology 288 (2004) 57–7364

compared the operational real-time MC2 forecasts

during the MAP SOP (Benoit et al., 2002) with an

extensive surface observation network. They reported

that the temporal variability of the predicted precipi-

tation sequences was generally in good agreement

with the observations, while the MC2 precipitation

amounts were substantially and persistently under-

estimated by as much as 43%. Similarly, Benoit et al.

(2002) reported a systematic underestimation for the

entire model domain evaluating also the MAP SOP

forecasts.

In this section a comparison of radar derived and

modeled precipitation is presented. During the MAP

SOP a radar composite combining the data of

operational radar networks from France, Germany,

Switzerland and Austria has routinely been generated

in quasi-real time (Hagen, 1999). Unfortunately, such

a product is not available for the 29 July. Thus, the

comparison is restricted to the three MAP cases

considered in this study. Radar derived precipitation

is clearly less accurate than rain gauge measurements

(Joss et al., 1998; Germann and Joss, 2002), and

thus less suited for a validation of accumulated

precipitation amounts. However, radar scans provide

high spatially and temporally resolved observations.

These allow evaluating how well the model performs

in reproducing the spatiotemporal occurrence of

precipitation. Due to a lack of radar scans in the

eastern Po Valley, the comparison does focus on the

western part of the valley, in particular on the LM

area, and on Switzerland.

Subsequently, the radar derived precipitation and

the simulated precipitation of member 2 of the

ensembles are compared at the time of the strongest

intensity in the LM area (Fig. 7). The choice of

ensemble member 2 is arbitrary. Since the ensemble

spread is small for 20 September and 6 November

(see later), the comparison depends not on the chosen

member in these cases. The spread is larger,

however, for the 25 September and thus the choice

of the member does somewhat influence the

comparison.

On 20 September 1999, precipitation was persist-

ently strong in the LM area until 1500 UTC with

a reenhancement in the evening. The model simu-

lates this evolution quite accurately (Fig. 7a and b).

Fig. 7. Comparison of modeled and radar derived precipitation intensity (mm/h). The panels show the MC2 3-km simulation of ensemble

member 2 (top) and the MAP Alpine composite (bottom) for (a,b) 0800 UTC 20 Sep 1999, (c,d) 1930 UTC 25 Sep 1999, and (e,f) 1330 UTC 6

Nov 1999, zoomed into Switzerland and the northwestern Po Valley.

A. Walser, C. Schar / Journal of Hydrology 288 (2004) 57–73 65

At 0800 UTC when the strongest intensity occurred,

the model captures well the precipitation pattern in

the western Po Valley including the heavy precipi-

tation in the LM area and the shielded region to the

southeast. The model appears to capture the squall

line extending from south-western Switzerland to the

French Vosges, but its amplitude is considerably

underestimated.

On 25 September 1999, the prefrontal precipitation

in the LM area started at around 0400 UTC, enhancing

slowly and reaching the maximum only at 1930 UTC.

This maximum as well as the evolution is quite well

captured by the model (Fig. 7c and d). However, north

of the Alps the precipitation pattern of the approach-

ing cold front is very poorly simulated, showing a

serious underestimation of the extension of the

rainfall area at this time.

On 6 November 1999, the prefrontal precipitation

in the LM area, which started on the previous evening,

persisted until about 1700 UTC. It reached its

maximum at 1330 UTC. The model simulates the

precipitation related to the passing cold front well in

this case (Fig. 7e and f). The simulation misses,

however, the latest 3 h of the precipitation in the LM

area. In addition, the model represents the postfrontal

cellular precipitation in the west near the French/

Swiss borderline quite realistically.

Overall, the model simulates the evolution of the

precipitation in the western Po Valley qualitatively

remarkably well, even though it underestimates the

persistency of the rainfall in the LM area in one of

the cases. The model has also demonstrated its ability

to simulate small-scale shower-like cellular precipi-

tation behind the cold front. The systematic under-

estimation of precipitation by the MC2 is evident

from this comparison, despite the rather uncertain

accuracy of the radar-derived precipitation amounts.

At least in the cases considered, the model bias can

be related to a misrepresentation of the spatial

extension of the precipitation bands, while the

simulated precipitation intensities appear realistic.

In addition, the model shows a distinct shortage of

precipitation near the in flow boundary, affecting up

to 100 km in the flow direction (see later in Fig. 8).

This feature must be associated with the treatment in

the lateral relaxation zone. In particular, the con-

densed and frozen water is not nested from the

driving simulation.

3.3. Modeled precipitation

In this section, the potential predictability of

precipitation in the four case studies is investigated.

A measure for predictability is the spread of an

ensemble, which can be determined, for example, by

the standard deviation of the ensemble members with

respect to the ensemble mean. In order to evaluate the

potential predictability of accumulated daily precipi-

tation p from our ensemble simulations, a normalised

spread is defined as

Sp ¼1

p

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1

M 2 1

XM

m¼1

ðpm 2 �pÞ2

vuut ; ð3Þ

where pm denotes the daily precipitation of member

m; �p; the ensemble mean daily precipitation; M is the

ensemble size (equals to 12). Because of the normal-

isation by �p; the analysis is restricted to areas with

�p $ 1 mm. Fig. 8 provides normalised spread Sp and

ensemble mean accumulated daily precipitation,

respectively, for the four case studies. The panels

reveal large differences in Sp among the case studies,

but also in the spatial distribution. These aspects are

discussed below.

For 29 July the highest Sp values exceed 2 in

some regions, implying that the standard deviation

of the daily precipitation amounts to more than

twice the ensemble mean (Fig. 8a). Such high

values occur in particular over sea and over the flat

eastern Po region, while in the Alps and Appen-

nines the values are smaller.

In contrast, the heavy precipitation event on 20

September is characterised by a small normalised

spread (Fig. 8c), similar as the late autumn case on

6 November (Fig. 8g). In the Po basin, where the

highest precipitation amounts occurred in both these

cases, the Sp values for the 20 September amount to

less than 0.1 in a large part of that region, while

they are substantially larger for 6 November in the

eastern Po Valley and over the Ligurian Sea.

The ensemble for 25 September shows a higher

normalised spread in the LM area and north of the

Alps, as compared to the other two MAP cases

(Fig. 8e). The large spread in the western part of the

model domain is likely associated with convective

activity at the leading edge of the cold front, while

A. Walser, C. Schar / Journal of Hydrology 288 (2004) 57–7366

Fig. 8. Spatial distribution of left normalised ensemble spread Sp of accumulated daily precipitation (see text), and (right) ensemble mean of

accumulated daily precipitation (mm) for (a,b) 29 Jul 1999, (c,d) 20 Sep 1999, (e,f) 25 Sep 1999, (g,h) 6 Nov 1999. The panels show the MC2

3 km model domain without the relaxation zone.

A. Walser, C. Schar / Journal of Hydrology 288 (2004) 57–73 67

the high values to the north occur with ensemble

mean precipitation of only a few millimetres.

Overall, the four case studies considered reveal

remarkable differences in the normalised ensemble

spread of daily precipitation Sp: Convective activity,

above all thermal convection, enhances Sp and hence

reduces the potential predictability. However, as will

be seen below the predictability is not a mere function

of convective activity.

3.4. Predictability in typical Alpine river catchments

The simulated precipitation of the four ensembles

are evaluated for the seven river catchments Verzasca,

Maggia, Ticino, Toce, LM, Po Piacenza and extended

Po, defined in Section 2.3. Fig. 9 provides

the accumulated daily precipitation of the 12 members

for the four case studies. It should be noted that the

mean precipitation amounts in the smaller catchments

differ by up to an order of magnitude between the case

studies.

For 29 July, the members show notable differences

at all catchments considered, even in the extended Po

basin. In the Po Piacenza basin, member 1 suggests

more than twice the value of member 8. In the other,

smaller catchments the differences are similar. In

contrast, for 20 September the members yield very

similar amounts in all catchments, despite the presence

of convective activity in this case. On 25 September,

even though the synoptic conditions and the intensity

of the convection are qualitatively similar as on 20

September, the LM catchment, and in particular its

Fig. 9. Accumulated daily precipitation (mm) in seven river catchments (Section 2.3) from four case studies. Shown are catchment-mean

amounts from ensemble simulations with 12 members. Note the different ordinate in the four panels.

A. Walser, C. Schar / Journal of Hydrology 288 (2004) 57–7368

sub-catchments Verzasca, Maggia and Ticino, show

substantial differences between the members. On

6 November, the differences between the members

are small, except in the Verzasca catchment.

In the following, attention is given to the temporal

evolution of the accumulated precipitation in the river

catchments. Fig. 10 displays the evolution in the

Verzasca, Maggia, Toce, Po Piacenza and extended

Po catchment. As expected, the 20 September and the

6 November do show little spread (except in the

Verzasca catchment on 6 November) pointing out that

not only the daily precipitation sum of the members

are close to each other, but also the temporal evolution

of precipitation.

On 29 July, the precipitation intensities averaged

over the different catchments are rather small, and

precipitation is mostly due to scattered convective

cells. The evolution of the precipitation intensity

varies remarkably within the members, even in the Po

Piacenza catchment (first column in Fig. 10). This

behaviour is very pronounced in the Toce catchment

where some members show higher precipitation

intensities in the earlier hours of the day, while others

have maximum precipitation in the afternoon or are

characterised by relatively constant intensities during

most of the day. However, the comparatively high

precipitation intensity in the first few hours in some

members, in particular in the Verzasca and Maggia

catchment, appears to be related to model spinup

caused by the amplification of the perturbations.

Nevertheless, even when disregarding the first few

hours, the remaining period is clearly unpredictable.

The differences highlight the unpredictable spatial

and temporal occurrence of convective cells and the

respective impact upon accumulated precipitation

within a typical Alpine river catchment.

On 25 September, when precipitation was much

stronger, the ensemble reveals also remarkably

varying precipitation intensities between the members

in the small Verzasca catchment. In this case, some

members suggest a notable increase in precipitation

intensity at 1600 UTC, whereas other members show

it later and weaker.

In order to calculate the normalised spread of daily

precipitation Sp for the seven river catchments, Eq.

(3) is used with pm and �p representing averages over

the catchments instead of grid point values. This

information is shown in Fig. 11 as a function of

catchment area (see symbols).

In addition, a location-independent background

normalised spread SpðAÞ is evaluated as a function of

the area A: SpðAÞ is derived from the simulation

output in a 768 £ 768 km2 box centred in the 3 km

model domain (equals to almost the entire 3 km model

domain). To this end, SpðAÞ was calculated according

to Eq. (3) with pm and �p; representing averages for all

possible subdomains with the areas

A ¼ ð3 £ 2xÞ2 km2; x ¼ 0;…; 8 ð4Þ

within this box (see Walser et al., 2004 for details).

The results are summarised in terms of the solid lines

in Fig. 11, quantifying the predictability of precipi-

tation amounts for the four case studies and for areas

between 10 km2 (grid point scale) and 106 km2. In

essence, the background spread SpðAÞ represents a

mean location-independent estimate of normalised

spread as a function of area A:

A general trend towards high predictability with

increasing area is obvious in both the catchments and

background values. For 29 July, the background

SpðAÞ exceeds 1 for the smallest areas considered.

Even for larger areas, SpðAÞ does not appear to

converge to 0, in contrast to the other three cases. For

most of the smaller catchments and in all the cases, the

Sp values are substantially lower than the back-

ground values. In the larger catchments, however, the

Sp values are mostly higher than the background

values, in particular in the Po Piacenza catchment.

The transition between the two characteristics occurs

in the LM catchment, which is the largest of the

catchments with entirely mountainous character,

located directly to the south of the main Alpine

ridge. Hence, complex topography seems to increase

the predictability of quantitative precipitation. This is

also indicated, at grid point scale, in Fig. 8a. It is likely

that this property is related to the topographic control

of both the convection triggering mechanism and the

larger-scale uplift on the Alpine slopes.

Additionally, the Sp values for the river catch-

ments exhibit a large case-to-case variability. As an

extreme example, consider the Maggia catchment. It

shows values between 0.015 (20 September) and 0.33

(29 July) indicating excellent and poor predictability,

respectively. Note also the large differences among

A. Walser, C. Schar / Journal of Hydrology 288 (2004) 57–73 69

Fig. 10. Evolution of accumulated precipitation (mm) from ensemble simulations for four cases studies in the river catchments Verzasca,

Maggia, Toce, Po Piacenza, and extended Po. The x-axis in the panels indicates time in UTC.

A. Walser, C. Schar / Journal of Hydrology 288 (2004) 57–7370

the catchments within a single case. For instance on

29 July, the Po Piacenza catchment reveals a Sp value

of 0.29, while for the smaller LM catchment it is only

0.15. For a single event, a simple dependence of

predictability upon the size of a specific catchment is

thus not generally deducible.

4. Summary and conclusions

The predictability of quantitative precipitation in

typical Alpine river catchments has been investigated

using ensembles of numerical weather simulations.

These simulations were performed with the Canadian

NWP model MC2 in a convection-resolving setup.

The applied variant of the ensemble methodology

assumes perfectly predictable synoptic conditions and

a prefect model. Since the uncertainty arises only from

small-scale perturbations in the initial conditions, this

setup allows to isolate predictability limitations

related to the meso-b scale, e.g. related to convection

and gravity-wave propagation. In reality, uncertainties

due to synoptic-scale atmospheric evolution as well as

model errors would contribute towards further redu-

cing the predictability. The ensemble strategy applied

in this study is thus not meant as an operational

limited-area EPS.

The study is based on four case studies and includes

a summer day with strong thermal convection, two

early autumn day with embedded convection, and a

late autumn frontal passage with stable stratification.

The results were analysed with respect to the potential

predictability of precipitation in selected Alpine river

catchments spanning a wide range of scales from 200

to 120,000 km2. The key results are:

† The potential predictability strongly depends upon

the weather conditions. In particular during

episodes of thermal convection, precipitation

forecasts can be critically affected by predictability

limitations, even in intermediate-scale river catch-

ments (,50,000 km2).

† However, the presence of convection alone does

not necessarily limit predictability, at least in

mountainous regions. More specifically, in one

case with moderate convection, we found the

ensemble members to be virtually identical. The

dynamical and physical reason for this peculiar

behaviour is beyond the scope of the present study,

but it clearly deserves further investigation.

† The potential predictability of precipitation shows

large case-to-case and catchment-to-catchment

variability. A relationship between catchment

area and predictability can be noted in the mean,

but is not deducible in a single case.

† Precipitation amounts in mountainous catchments

appear to be more predictable than in the

foreland. This is most likely due to the

topographic control of precipitation, both through

the triggering of convection and the larger-scale

uplift due to the underlying topography.

The present study has some notable limitations.

The initial perturbations are generated using an ad hoc

technique that does at best qualitatively match the

observational uncertainties. Furthermore, only four

cases were considered, and each of them investigated

with only 12 ensemble members. In addition, the MC2

is characterised by a systematic underestimation of

the precipitation. Nevertheless, we believe that

the main conclusion are not affected by these model

and setup shortcomings.

Fig. 11. Scale dependence of predictability of precipitation in

doubly logarithmic display. Normalised spread Sp for 29 Jul 1999

(red), 20 Sep 1999 (green), 25 Sep 1999 (brown), 6 Nov 1999 (blue)

is shown. Symbols and solid lines refers to the seven river

catchments and to background values for the model domain,

respectively (see text).

A. Walser, C. Schar / Journal of Hydrology 288 (2004) 57–73 71

In particular, our results suggest that short-range

quantitative precipitation forecasting may at least on

occasion be seriously affected by intrinsic predict-

ability limitations, even with a perfect model and a

perfect synoptic forcing. This result underlines the

practical importance of probabilistic forecasting

strategies. The trend in NWP towards finer meshes

will unquestionably continue, favouring further the

one-way coupling of atmospheric and hydrologic

models. Hence, the combination of an atmospheric

short-range high-resolution EPS with a hydrologic

model and uncertainty processor seems to be an

attractive strategy towards flood forecasting and

warning systems.

Acknowledgements

We are indebted to Robert Benoit and his research

group at Recherche en Prevision Numerique (RPN),

Canada, for providing access to and support of the

MC2 model. The authors acknowledge the support of

Karsten Jasper and Massimiliano Zappa for providing

the coordinates of the river catchments. The numeri-

cal simulations have been conducted on the NEC SX-

5 at the Swiss Center for Scientific Computing (SCSC,

Manno). The research has been funded by ETH Zurich

under research contract TH02134.

References

Benoit, R., Desgagne, M., Pellerin, P., Chartier, Y., Desjardins,

S., 1997. The Canadian MC2: a semi-Lagrangian, semi-

implicit wideband atmospheric model suited for finescale

process studies and simulations. Mon. Wea. Rev. 125,

2382–2415.

Benoit, R., Pellerin, P., Kouwen, N., Ritchie, H., Soulis, E.D., 2000.

Toward the use of coupled atmospheric and hydrologic models

at regional scale. Mon. Wea. Rev. 128, 1681–1706.

Benoit, R., SchaR, C., Binder, P., Chamberland, S., Davies, H.C.,

Desgagne, M., et. al., 2002. The real-time ultrafinescale forecast

support during the special observing period of the MAP. Bull.

Am. Meteorol. Soc. 83, 85–109.

Bougeault, P., Binder, P., Buzzi, A., Dirks, R., Houze, R., Kuettner,

J., et al., 2001. The MAP special observing period. Bull. Am.

Meteorol. Soc. 82, 433–462.

Buizza, R., Palmer, T.N., 1995. The singular-vector structure

of the atmospheric general circulation. J. Atmos. Sci. 52,

1434–1456.

Buizza, R., Miller, M., Palmer, T.N., 1999. Stochastic represen-

tation of model uncertainties in the ECMWF ensemble

prediction system. Q. J. R. Meteorol. Soc. 125, 2887–2908.

Droegemeier, K.K., Smith, J.D., Businger, S., Doswell III, C.,

Doyle, J., Duffy, C., et al., 2000. Hydrological aspects of

weather prediction and flood warnings: report of the ninth

prospectus development team of the US weather research

program. Bull. Am. Meteorol. Soc. 81, 2665–2680.

Du, J., Mullen, S.L., Sanders, F., 1997. Short range ensemble

forecasting of quantitative precipitation. Mon. Wea. Rev. 125,

2427–2459.

Ehrendorfer, M., 1997. Predicting the uncertainty of numerical

weather forecasts: a review. Meteorol. Zeitschrift, N.F. 6,

147–183.

Ehrendorfer, M., Errico, R.M., Raeder, K.D., 1999. Singular-vector

perturbation growth in a primitive equation model with moist

physics. J. Atmos. Sci. 56, 1627–1648.

Frei, C., Haller, E., 2001. Mesoscale precipitation analysis from

MAP SOP rain-gauge data, MAP Newsletter No. 15, MeteoS-

wiss, Zurich, pp. 257–260.

Frei, C., Davies, H.C., Gurtz, J., Schar, C., 2000. Climate dynamics

and extreme precipitation and flood events in Central Europe.

Integrated Assessment 1, 281–299.

Fritsch, J.M., Houze Jr., R.A., Adler, R., Bluestein, H., Bosart,

L., Brown, J., et al., 1998. Quantitative precipitation

forecasting: report of the eight prospectus development

team, US weather research program. Bull. Am. Meteorol.

Soc. 79, 285–299.

Forgner, I.-L., Iversen, T., 2002. High-resolution limited-area

ensemble predictions based on low-resolution targeted singular

vectors. Q. J. R. Meteorol. Soc. 128, 1321–1341.

Gal-Chen, T., Somerville, R., 1975. On the use of a coordinate

transformation for the solution of the Navier–Stokes equations.

J. Comput. Phys. 17, 209–228.

Germann, U., Joss, J., 2002. Mesobeta profiles to extrapolate radar

precipitation measurements above the Alps to the ground-level.

J. Appl. Meteorol. 41, 542–557.

Grimit, E.P., Mass, C.F., 2002. Initial results of a mesoscale short-

range ensemble forecasting system over the Pacific northwest.

Wea. Forecast. 17, 192–205.

Hagen, M., 1999. The Alpine radar composite, MAP Newsletter No.

11, MeteoSwiss, Zurich, pp. 20–21.

Holton, J.R., 1992. An Introduction to Dynamic Meteorology.

Academic Press, New York, 511 pp.

Houtekamer, P.L., Lefaivre, L., Derome, J., Ritchie, J., Houteka-

mer, H.L., 1996. A system simulation approach to ensemble

prediction. Mon. Wea. Rev. 124, 1225–1242.

Houze, R.A., 1993. Cloud Dynamics. Academic Press, New York,

573 pp.

Jasper, K., Kaufmann, P., 2003. Coupled runoff simulations as

validation tools for atmospheric models at the regional scale.

Q. J. R. Meteorol. Soc. 129, 673–692.

Joss, J., Schadler, B., Galli, G., Cavalli, R., Boscacci, M., Held, E.,

et al., 1998. Operational Use of Radar for Precipitation

Measurements in Switzerland. VDF Hochschulverlag AG an

der ETH, Zurich, 108 pp; ISBN 3 7281 2501 6.

A. Walser, C. Schar / Journal of Hydrology 288 (2004) 57–7372

Kelly, K.S., Krzysztofowicz, R., 2000. Precipitation uncertainty

processor for probabilistic river stage forecasting. Water

Resour. Res. 36, 2643–2653.

Kong, F., Yau, M.K., 1997. An explicit approach to microphysics in

MC2. Atmos. Ocean 35, 257–291.

Krzysztofowicz, R., 1998. Probabilistic hydrometeorological fore-

casts: toward a new area in operational forecasting. Bull. Am.

Meteorol. Soc. 79, 243–251.

Krzysztofowicz, R., Herr, H.D., 2001. Hydrologic uncertainty

processor for probabilistic river stage forecasting: precipitation-

dependent model. J. Hydrol. 249, 46–68.

Lorenz, E.N., 1963. Deterministic nonperiodic flow. J. Atmos. Sci.

20, 130–141.

Marsigli, C., Montani, A., Nerozzi, F., Paccagnella, T., Tibaldi, S.,

Molteni, F., Buizza, R., 2001. A strategy for high-resolution

ensemble prediction. Part II: Limited-area experiments in

four Alpine flood events. Q. J. R. Meteorol. Soc. 127, 2095–2115.

Medina, S., Houze, R.A., 2003. Air motions and precipitation

growth in Alpine storms. Q. J. R. Meteorol. Soc. 129, 345–371.

Miller, N.L., Kim, J., 1996. Numerical prediction of precipitation

and river flow over the Russian river watershed during the

January 1995 California storms. Bull. Am. Meteorol. Soc. 77,

101–105.

Misra, V., Yau, M.K., Badrinath, N., 2000. Atmospheric water

species budget in mesoscale simulations of lee cyclones over the

Mackenzie river basin. Tellus 52A, 140–161.

Molteni, F., Buizza, R., Palmer, T.N., Petroliagis, T., 1996. The

ECMWF ensemble prediction system: methodology and

validation. Q. J. R. Meteorol. Soc. 122, 73–119.

Palmer, T.N., 2000. Predicting uncertainty in forecasts of weather

and climate. Rep. Prog. Phys. 63, 71–116.

Richard, E., Cosma, S., Tabary, P., Pinty, J.-P., Hagen, M., 2003.

High-resolution numerical simulations of the convective system

observed in the Lago Maggiore area on 17 September 1999

(MAP IOP2a). Q. J. R. Meteorol. Soc. 129, 543–563.

Rotunno, R., Ferretti, R., 2003. Comparative analysis of rainfall in

MAP cases IOP2b and IOP8. Q. J. R. Meteorol. Soc. 129,

373–390.

Schar, C., Sprenger, M., Luthi, D., Jiang, Q., Smith, R.B., Benoit,

R., 2003. Structure and dynamics of an Alpine potential

vorticity banner. Q. J. R. Meteorol. Soc. 129, 825–855.

Schultz, D.M., Schumacher, P.N., 1999. The use and misuse of

conditional symmetric instability. Mon. Wea. Rev. 127,

2709–2732.

Stensrud, D.J., Brooks, H.E., Du, J., Tracton, M.S., Rogers, E.,

1999. Using ensembles for short-range forecasting. Mon. Wea.

Rev. 127, 433–466.

Stensrud, D.J., Bao, J.W., Warner, T.T., 2000. Using initial

conditions and model physics in short-range ensemble simu-

lations of mesoscale convective systems. Mon. Wea. Rev. 128,

2077–2107.

Tanguay, M., Robert, A., Laprise, R., 1990. A semi-implicit semi-

Lagrangian fully compressible regional forecast model. Mon.

Wea. Rev. 118, 1970–1980.

Toth, Z., Kalnay, E., 1997. Ensemble forecasting at NCEP and the

breeding method. Mon. Wea. Rev. 125, 3297–3319.

Walser, A., Luthi, D., Schar, C., 2004. Predictability of precipitation

in a cloud-resolving model. Mon. Wea. Rev. 132, 560–577.

Yu, Z., Lakhtakia, M.N., Yarnal, B., White, R.A., Miller, D.A.,

Frakes, B., Barron, E.J., Duffy, C., Schwartz, F.W., 1999.

Simulating the river-basin response to atmospheric forcing by

linking a mesoscale meteorological model and hydrologic

model system. J. Hydrol. 218, 72–91.

Yuter, S.E., Houze, R.A., 2003. Microphysical modes of precipi-

tation growth determined by vertically pointing radar at

Locarno–Monti during MAP. Q. J. R. Meteorol. Soc. 129,

455–476.

A. Walser, C. Schar / Journal of Hydrology 288 (2004) 57–73 73