Theor Appl ClimatolDOI 10.1007/s00704-015-1551-6

ORIGINAL PAPER

Extreme precipitation in WRF during the Newcastle EastCoast Low of 2007

James B. Gilmore1 · Jason P. Evans1 · Steven C. Sherwood1 ·Marie Ekstrom2 · Fei Ji3

Received: 21 December 2014 / Accepted: 29 June 2015© Springer-Verlag Wien 2015

Abstract In the context of regional downscaling, we studythe representation of extreme precipitation in the WeatherResearch and Forecasting (WRF) model, focusing on amajor event that occurred on the 8th of June 2007 alongthe coast of eastern Australia (abbreviated “Newy”). Thiswas one of the strongest extra-tropical low-pressure systemsoff eastern Australia in the last 30 years and was one ofseveral storms comprising a test bed for the WRF ensem-ble that underpins the regional climate change projectionsfor eastern Australia (New South Wales/Australian CapitalTerritory Regional Climate Modelling Project, NARCliM).Newy provides an informative case study for examining pre-cipitation extremes as simulated by WRF set up for regionaldownscaling. Here, simulations from the NARCliM physicsensemble of Newy available at ∼ 10 km grid spacingare used. Extremes and spatio-temporal characteristics areexamined using land-based daily and hourly precipitationtotals, with a particular focus on hourly accumulations. Ofthe different physics schemes assessed, the cumulus and

� James B. Gilmorejames.gilmore@unsw.edu.au

1 Climate Change Research Centre and ARC Centreof Excellence for Climate System Science,University of New South Wales, Sydney, Australia

2 CSIRO Land and Water, PO Box 1666, Canberra,ACT 2601, Australia

3 Office of Environment and Heritage, NSW Departmentof Premier and Cabinet, PO Box 733, Queanbeyan,NSW 2620, Australia

the boundary layer schemes cause the largest differences.Although the Betts-Miller-Janjic cumulus scheme producesbetter rainfall totals over the entire storm, the Kain-Fritschcumulus scheme promotes higher and more realistic hourlyextreme precipitation totals. Analysis indicates the Kain-Fritsch runs are correlated with larger resolved grid-scalevertical moisture fluxes, which are produced through theinfluence of parameterized convection on the larger-scalecirculation and the subsequent convergence and ascent ofmoisture. Results show that WRF qualitatively reproducesspatial precipitation patterns during the storm, albeit withsome errors in timing. This case study indicates that whilstregional climate simulations of an extreme event such asNewy in WRF may be well represented at daily scales irre-spective of the physics scheme used, the representation athourly scales is likely to be physics scheme dependent.

1 Introduction

Precipitation extremes remain one of the most challengingquantities to simulate in climate models (Stephens et al.2010), regional climate studies (Sunyer et al. 2012) and innumerical weather prediction (Lavers and Villarini 2013).Here, we investigate how the choice of physical scheme ina regional climate inspired model configuration can influ-ence the simulation of precipitation extremes (Fita et al.2010; Liang et al. 2012). With the number of regionalclimate simulations being performed at ∼ 10 km resolu-tion increasing (e.g. Evans et al. 2014; Jacob et al. 2014)and growing interest in the representation of precipitationextremes within these models (particularly for specific eventstudies Ji et al. 2015), a better understanding of the sensitiv-ity of extreme precipitation to physics schemes is required.

J. B. Gilmore et al.

For regional climate modelling, it is important that the quan-titative precipitation distribution can be captured locally atdifferent time scales, while the timing of these events isless important (in contrast to a numerical weather predictioncontext).

To investigate the uncertainty associated with the selec-tion of different, but commonly used, physical parameteri-zation schemes, we use a case study approach to assess theirimpact on model skill in terms of capturing observed tem-poral and spatial structures of an extreme rainfall event. Thespecific event studied here is an extreme extra-tropical low-pressure system (storms known as East Coast Lows (ECLs)Speer et al. 2009) that formed off eastern Australia in June2007 and caused significant damage to the city of Newcas-tle (here the storm is referenced as “Newy”). Newy was oneof several ECLs used in studies of different configurationsof the Weather Research and Forecasting (WRF) model forthe production of regional projections for eastern Australiain the New South Wales/ACT Regional Climate Modelling(NARCliM) project (Evans et al. 2012, 2014). Each config-uration reflects a different combination of physics param-eterization schemes for microphysics, long and shortwaveradiation, cumulus and the planetary boundary layer.

Evans et al. (2012) examined the performance of theensemble over four ECL events including Newy. They con-sidered modelled and observed patterns for a range ofvariables and found that whilst no single ensemble memberperformed best, a small number of combinations consis-tently showed low skill in simulating a range of ECL events.They also found that the spread of performance amongst themembers was greater in more intense events, suggesting thatextreme precipitation events provide good test environmentsto differentiate the impact of different physics parameteriza-tions. Recently, Ji et al. (2014) examined the spatial patternsof precipitation for eight such ECL events. Both studiesfound that parameterizations of cumulus convection and theplanetary boundary layer significantly influenced the spa-tial patterns of precipitation produced. We build on thiswork, focusing on the role played by the physical param-eterizations in the quantitative simulation of precipitationextremes for hourly and daily totals. The extreme nature ofNewy as one of the largest storms in the region for the past30 years, means it is well suited to an in depth analysis ofthe influence of physics parameterization on precipitationextremes.

A number of previous studies have examined multi-physics ensembles of WRF (Gallus and Bresch 2006;Bukovsky and Karoly 2009; Argueso et al. 2011; Flaounaset al. 2011; Schumacher et al. 2013) but few havefocused on short-term precipitation during an extreme event.Jankov et al. (2005) used a WRF multi-physics ensemble at12 km grid spacing, created using three cumulus schemes,three microphysics schemes and two planetary boundary

layer (PBL) schemes, to simulate a series of warm seasonmesoscale convective systems that included some extremeprecipitation. While they found no single physics combi-nation performed best, the systems were most sensitive tothe cumulus scheme followed by the PBL scheme and thenthe microphysics scheme. When examining rain rates, thecumulus scheme was the dominant factor, while the micro-physics scheme had a stronger influence on total rain vol-ume. They also found that interactions of different schemescould influence the results as much as changing a singlescheme though this effect varied between events.

Lowrey and Yang (2008) examined the ability of amulti-physics WRF ensemble to simulate daily extremeprecipitation in Texas, USA. Their ensemble was madeusing four different microphysics schemes, three differ-ent cumulus schemes and two different radiation schemes.They found that the simulation of an extreme precipitationevent was most sensitive to the cumulus parameterization,slightly affected by the microphysics scheme and largelyunaffected by the radiation schemes. They found that theBetts-Miller-Janjic (BMJ) cumulus scheme coupled to theLin microphysics scheme produced the best precipitationestimate, while other cumulus schemes (including Kain-Fritsch) tended to underestimate the precipitation intensity.They did not consider different PBL schemes. They alsofound the cumulus parameterization improved the precipi-tation simulation even at 4 km grid spacing which is oftenconsidered sufficient grid spacing to turn off the cumu-lus parameterization. The present study addresses howshorter timescale precipitation extremes respond to differ-ent physics schemes in the context of regional climatesimulations.

This paper is outlined as follows. In the Methods section,we present the case study, and then discuss the model sim-ulations and the observational precipitation datasets usedin this paper. The Results section presents a comparisonbetween the observations and the 36 simulations and theirensemble averages at daily and hourly resolution. We thendiscuss these results in light of regional climate modellingand close with recommendations for the use of physicsschemes in regional climate ensembles, where precipitationextremes are of interest.

2 Methods

The work presented here consists of a comparative analysisbetween precipitation extremes observed in the Newy eventoccurring in June 2007 and 36 model simulations using dif-ferent combinations of physics parameterization schemes.After introducing the case study, we discuss the WRF simu-lations, and then we describe the observational data used forthe precipitation comparisons.

Extreme precipitation in WRF during the Newcastle East Coast Low

2.1 Case study

Newy was an extreme extra-tropical low-pressure system(storms known as East Coast Lows (ECLs) Speer et al.2009) that formed off eastern Australia in June 2007. Newyproduced highly localized extreme precipitation near New-castle, Australia (32.9167o S, 151.7500o E) and was oneof the strongest ECL events observed in the last 30 years.Newy flooded the Hunter River to levels higher than in theprevious 36 years, breached the Pasha Bulker carrier andproduced flash flooding in the Newcastle, Australia region.Nine fatalities were recorded, and an estimated 20,000calls to emergency services were made (Mills et al. 2010;Verdon-Kidd et al. 2010). Wind gusts to 135 km/h were alsoreported. The storm developed in an existing low-pressuretrough located in the northern Tasman Sea, resulting inhumid air being funnelled towards the coast of eastern Aus-tralia from the north. The synoptic situation is shown inFig. 1. Beyond its formation, the major contributing fac-tors to the extreme nature of this event were the warmerthan usual Tasman Sea, the high-pressure system to thesouth creating large pressure gradients and high atmospherictemperature gradients over eastern Australia. The synopticsituation and dynamics leading up to the event are furtherdiscussed in Mills et al. (2010).

2.2 Atmospheric model simulations

The atmospheric model used in this study is WRF ver-sion 3.2.1, and its primary use lies in mesoscale numerical

Fig. 1 Synoptic situation of the Newcastle East Coast Low at 12 UTC07 June 2007. The major event at Newcastle occurs 24–36 h later.Data from ERA Interim and chart adapted from the Bureau of Meteo-rology, Australia (see http://www.bom.gov.au/nsw/sevwx/facts/events/june-07-ecl/e1-msl-loop.shtml where an animation is available)

weather simulations, for both research and operational pur-poses. Due to its versatile configuration, WRF contains anincreasing number of physics parameterizations that can beused almost interchangeably (Skamarock et al. 2008). Thereare now such a large number of parameterization schemesfor each physical process available within the WRF mod-elling system that it is only feasible to examine a smallsubset within a single study. Here, we wish to examinemodel configurations relevant for climate length simula-tions and hence focus on simple to medium complexityschemes that have been used in previous regional climatesimulations. We briefly describe the model setup used forsimulating Newy, noting that full details of the experimentsare documented in Evans et al. (2012).

The model domains used in this study are the same asEvans et al. (2012) and are shown in Fig. 2. The outerdomain is the Australasian Coordinated Regional climateDownscaling Experiment domain and the inner domain cov-ers eastern Australia and a significant portion of the Tasmansea. The grid spacing of the two domains is approximately0.44◦ and 0.088◦ (48.92 km and 9.78 km, respectively)with dimensions 216×145 and 311×201, respectively. Theatmosphere comprises 30 levels in both domains with foursoil layers and the sea surface temperature is updated every6 h. Gravity wave damping at the model top is used in bothdomains with a damping layer depth of 5 km. The modelwas run from an initial condition starting at 00 UTC 01 June2007 to 00 UTC 15 June 2007 with the event peak occurringfrom June 7 to 9.

Thirty-six different combinations of physics options aresimulated starting from the same initial condition, bound-ary and sea surface temperature forcing. The employedschemes, although not exhaustive, provide a selection ofthose typically employed in WRF simulations and, there-fore, allow us to probe typical uncertainty in schemechoice. Two cumulus schemes are used, the Kain-Fritsch(KF) (Kain and Fritsch 1990; Kain 2004) and BMJ(Betts 1986; Betts and Miller 1986; Janjic 1994) cumu-lus schemes, and two PBL schemes are used, the Yon-sei University (YSU)/MM5 similarity (Hong et al. 2006;Paulson 1970) and Mellor-Yamada-Janjic (MYJ)/Eta simi-larity (Janjic 1994) schemes. Three microphysics schemesare employed: WRF Single Moment 3-class (WSM 3),WSM 5 (Hong et al. 2004) and WRF Double Moment5-class (WDM 5) (Lim and Hong 2010), and threeshortwave/longwave radiation combinations are simulated:Dudhia/Rapid Radiative Transfer (RRTM) model (Dudhia1989; Mlawer et al. 1997), Community Atmosphere Model(CAM)/CAM (Collins et al. 2004) and the global ver-sion of RRTM: RRTMG/RRTMG. The run number of eachcombination is shown in Table 1.

The initial and boundary conditions to the WRFsimulations are provided by the ERA-Interim reanalysis

J. B. Gilmore et al.

Fig. 2 Outer and inner WRFdomains used in the study of theNewcastle East Coast Lowstorm. Elevation in meters abovesea level is also show

(Dee et al. 2011). The outer domain employs spectralnudging of the wind and geopotential fields in the upperatmosphere. One way nesting is used to set the innerdomain boundary conditions using the spectrally nudgedouter domain fields.

2.3 Precipitation data

The WRF simulations are compared with measures ofobserved extreme precipitation totals during the Newyevent, and this is undertaken for daily and hourly totals. Forboth analyses, we attempt to consider uncertainty encom-passed in the gridded observed data; a process describedin the following sections. We now discuss the precipita-tion data used in this work and detail how the hourlyprecipitation totals are derived.

2.3.1 AWAP daily data

For the daily precipitation analysis, we use the AustralianWater Availability Project (AWAP) (Jones et al. 2009).1

AWAP uses a climatological anomaly based interpolationemploying daily station data from across Australia and hasbeen found to be suitable for studies of extremes despitesome underestimation (King et al. 2012). We use AWAPdata to perform comparisons with WRF from 9.00 a.m. EST02 June 2007 to 9.00 a.m. EST 15 June 2007 (EST is 10 hahead of UTC). To estimate the errors in the interpolatedAWAP precipitation fields, we employ the root mean square(RMS) error product accompanying the standard interpola-tion. In the AWAP interpolation, the size of the RMS erroris positively correlated to the precipitation totals and theAWAP precipitation error is largest where the extreme totalsoccur. This necessitates a careful consideration of errors inAWAP for extreme precipitation. To estimate the errors, a

1See also http://www.bom.gov.au/jsp/awap/.

Monte Carlo bootstrap procedure is used over the AWAPRMS error product. The bootstrap is performed by assumingeach precipitation value in AWAP has an associated Gaus-sian error distribution with standard deviation approximatedby the RMS error at that point. We bootstrap AWAP on itsnative grid before interpolating to the WRF grid (whose gridspacing is about 1/2 that of AWAP) for model-observationcomparison. We find about 200 bootstraps are sufficient toestimate the error.

2.3.2 Hourly precipitation data

In addition to a model-observation evaluation on dailytimescales, we also investigate the model on the sub-dailytimescale using hourly precipitation. Hourly precipitationdata provides much greater temporal sensitivity than thedaily AWAP product, and this allows us to understand theinfluence of the various physics options on precipitationover a range of time scales. To assess the agreement withobservations, and the influence of the physics parameteriza-tions on precipitation extremes, we undertake a grid pointanalysis and construct a simple hourly interpolation fieldusing inverse distance interpolation.

The hourly precipitation data consists of data from Aus-tralia’s pluviograph network. We use 88 Australian Bureauof Meteorology (BOM) stations with hourly totals withina 500 km radius of Newcastle from 9.00 a.m. EST 01–15June 2007, shown in Fig. 3. All daily pluviograph stationtotals have been verified with the published daily totals forthe same station and different (daily only) stations within40 km using BOM’s climate data online service.2 Only onepluviograph was excluded as it showed erroneous readingscompared to the daily totals, after storm peak. The tem-poral coverage of all stations during the aforementionedperiod is not continuous and some stations drop out after the

2Station verification performed at http://www.bom.gov.au/climate/data/.

Extreme precipitation in WRF during the Newcastle East Coast Low

Table 1 Physics options usedin the ensemble runs. Thecumulus scheme (CU),planetary boundary layer(PBL), microphysics (MP) andradiation schemes (RAD) areshown

Run CU PBL MP RAD

N1 KF YSU WSM3 Dudhia/RRTM

N2 KF YSU WSM3 CAM/CAM

N3 KF YSU WSM3 RRTMG/RRTMG

N4 KF YSU WSM5 Dudhia/RRTM

N5 KF YSU WSM5 CAM/CAM

N6 KF YSU WSM5 RRTMG/RRTMG

N7 KF YSU WDM5 Dudhia/RRTM

N8 KF YSU WDM5 CAM/CAM

N9 KF YSU WDM5 RRTMG/RRTMG

N10 BMJ YSU WSM3 Dudhia/RRTM

N11 BMJ YSU WSM3 CAM/CAM

N12 BMJ YSU WSM3 RRTMG/RRTMG

N13 BMJ YSU WSM5 Dudhia/RRTM

N14 BMJ YSU WSM5 CAM/CAM

N15 BMJ YSU WSM5 RRTMG/RRTMG

N16 BMJ YSU WDM5 Dudhia/RRTM

N17 BMJ YSU WDM5 CAM/CAM

N18 BMJ YSU WDM5 RRTMG/RRTMG

N19 KF MYJ WSM3 Dudhia/RRTM

N20 KF MYJ WSM3 CAM/CAM

N21 KF MYJ WSM3 RRTMG/RRTMG

N22 KF MYJ WSM5 Dudhia/RRTM

N23 KF MYJ WSM5 CAM/CAM

N24 KF MYJ WSM5 RRTMG/RRTMG

N25 KF MYJ WDM5 Dudhia/RRTM

N26 KF MYJ WDM5 CAM/CAM

N27 KF MYJ WDM5 RRTMG/RRTMG

N28 BMJ MYJ WSM3 Dudhia/RRTM

N29 BMJ MYJ WSM3 CAM/CAM

N30 BMJ MYJ WSM3 RRTMG/RRTMG

N31 BMJ MYJ WSM5 Dudhia/RRTM

N32 BMJ MYJ WSM5 CAM/CAM

N33 BMJ MYJ WSM5 RRTMG/RRTMG

N34 BMJ MYJ WDM5 Dudhia/RRTM

N35 BMJ MYJ WDM5 CAM/CAM

N36 BMJ MYJ WDM5 RRTMG/RRTMG

event, with some of these being in the Hunter Valley wheresubstantial flooding occurred.

Our approach to the hourly distribution analysis consistsof undertaking a grid point analysis and an interpolatedhourly precipitation analysis. We undertake both approachesto probe the robustness of our results for hourly totals. Thegrid point analysis involves comparing the observations andmodel at (1) the model point nearest to the station and (2) atall model points within 0.25◦ of the nearest model point tothe station.

For the interpolation, we first conduct a variogram anal-ysis of the 88 stations. This shows that the decorrelation

length of the hourly totals is ∼ 0.5◦. As we discussbelow, a cut-off of 1.0◦ is favoured in the inverse distanceinterpolation. This introduces some uncertainty in the inter-polation procedure, which we quantify below by consid-ering the parametric errors in the interpolation procedure.Comparison across the grid point and interpolation methodsshows our results for the hourly precipitation distributionsare consistent (see Results Section 3.2). We also note thatNewy was a heavy precipitation winter event, which waswide spread, large scale and very fast moving on hourlytimes scales, meaning that the precipitation interpolationmethod should be much better in this situation than in a

J. B. Gilmore et al.

Fig. 3 Station locations with hourly rain gauge data. These stationsprovide the basis for the hourly precipitation analysis. The land maskfor the hourly interpolation encompasses all stations and a circle ofradius 1◦ surrounding each station. The eastern Australian states aremarked, along with Newcastle and Sydney for reference

summertime convective situation. Since we are only inter-ested in the statistics (precipitation distribution and totals)of the hourly precipitation totals, and we find broad con-sistency with AWAP and the other hourly comparisons, wejudge this interpolation method to be satisfactory.

2.3.3 Interpolation of hourly station data

To interpolate the hourly station data for comparison withthe precipitation produced by the WRF model, we use aninverse distance interpolation procedure. This allows us toestimate the hourly precipitation totals P between the sta-tion locations. The inverse distance interpolation is given by

P(x, t) =(

n∑i=1

Dαi

) (n∑

i=1

p(t)i

Dαi

)(1)

where p(t)i is the ith station precipitation time series, Di =‖x − yi‖ is the distance between the station location yi andthe current grid point x and α is the weighting exponentof the inverse distance interpolation. The sum is performedover the n stations within the distance Di < Dmax, whereDmax is the threshold after which stations are not includedfor estimating the precipitation at the current grid pointx. Typically, n ranges from 1 to 10 in the interpolationdepending on station density. The two parameters of theinterpolation method, α and Dmax, are unknown a priori andmust be chosen to reproduce a best interpolation.

To determine the values of α and Dmax for the hourlyinterpolation, the daily AWAP product is used to constrainthe hourly interpolated product. Using the AWAP dailyproduct to estimate the values of α and Dmax representsa “matching” approach across time scales, since we areselecting α and Dmax such that the daily AWAP totals areoptimally reproduced when summing the hourly interpola-tion fields over the interpolation grid. This approach allowsus to incorporate the precipitation totals from the entiredaily station network (which are used in the AWAP con-struction), of which there are many more stations whencompared to the pluviograph station network with hourlytotals within 500 km of Newcastle.

The approach used to constrain the interpolation param-eters is based on a χ2 measure using AWAP as a ‘true’reconstruction of the daily precipitation. The hourly inter-polation is summed from 9.00 a.m. to 9.00 a.m. the next dayand compared with the corresponding AWAP value. The dif-ference between the daily AWAP value and summed hourlyinterpolation is weighted by the corresponding RMS prod-uct from AWAP at each grid point. The measure can bewritten schematically as

χ2 =∑m,j

⎛⎝P

m,jAWAP − P

m,j

Hourly

Pm,j

AWAP,RMS

⎞⎠

2

(2)

where the sum is taken over all days m and grid points j

with daily precipitation values falling within the land maskfrom the hourly interpolation procedure. Here, PAWAP arethe daily AWAP totals, PHourly are the summed daily totalsfrom the hourly interpolation and PAWAP,RMS is the dailyRMS error from the AWAP interpolation. We also restrictthe sum to precipitation values P

m,jAWAP > 2.0 mm day−1 to

avoid weighting drizzle in the parameter estimation.To determine the interpolation parameters α and Dmax,

we minimize the χ2 by standard optimization techniques.The search range used is α = 0.5 to 5.0 and Dmax = 0.25◦to 2.0◦. We find the interpolation parameters α and Dmax

that minimize the χ2 are (α, Dmax) = (2.2, 1.0), and theseparameters are used here. Figure 4 shows the daily pre-cipitation from AWAP versus summed daily precipitationtotals from the hourly interpolation at the same grid point.Good agreement is seen over the precipitation range—withdrizzle and extremes > 200 mm day−1 showing satisfac-tory agreement. Due to the difference between the best-fitcut-off of 1.0◦ and the variogram decorrelation length of∼ 0.5◦ for the station data, we also estimate the parametricerrors in this reconstruction. We examine the χ2 at the 3σ

limit for the α parameter and provide these curves as the 3σ

errors in the Results section for the interpolated quantities.It is difficult to estimate the exact bias of the interpola-tion for observational data, however, given the rarity of high

Extreme precipitation in WRF during the Newcastle East Coast Low

Fig. 4 Daily precipitation for AWAP versus the summed daily totalsfor the hourly precipitation interpolation. All precipitation values con-tained in the station land mask are shown as well as a one-to-one line.This comparison takes place on the AWAP grid, before interpolation tothe grid of the Newcastle WRF simulations. Good agreement is seenfor the entire precipitation range

percentile extremes; we would expect the interpolation tounderestimate the true precipitation extremes.

3 Results

Newy generated substantial precipitation from 7–9 June2007, which lead to flooding in the Hunter Valley andaround Newcastle. To understand how WRF simulated thisevent, we examine both the daily and hourly precipitationmeasures using the datasets discussed above. Using thedaily AWAP data, we first consider totals, spatial and dis-tributional properties as well as event timing and structure.Similar characteristics are then assessed for hourly totals.

3.1 Daily evaluation

3.1.1 Total areal precipitation

The total precipitation produced over land from 9 a.m. EST02 to 15 June 2007 in the inner domain provides a generalmeasure of the precipitation activity of the storm (Fig. 5).Compared to the AWAP precipitation of 5.19 × 107 mm,the total amount of precipitation is overestimated by themodel across all physics runs, except N32, which has a15 % deficit in total precipitation. The excess is on average50 %, with a maximum overestimation of 82 % by N9.

1 6 12 18 24 30 36

100

150

200

Run Number

TotalPrecipitatio

nRelativeto

AWAP

KF YSU

BMJ YSU

KF MYJ

BMJ MYJ

Fig. 5 Total land-based precipitation relative to AWAP. Runs are num-bered N1 through N36 and color-coded. The runs can be split intothe cumulus-boundary layer scheme combinations: KF-YSU (N1-9),BMJ-YSU (N10-18), KF-MYJ (N19-27) and BMJ-MYJ (N28-36).The averages of these combinations are also shown with color-coding:KF-YSU red, KF-MYJ dashed dark red, BMJ-YSU blue, BMJ-MYJdashed dark blue. The same colours are used throughout for theindividual runs and the cumulus-boundary layer ensemble averages

The choice of boundary layer and cumulus scheme has thelargest effect, with average precipitation excesses of 69,46, 70 and 19 %, for the KF-YSU, BMJ-YSU, KF-MYJand BMJ-MYJ combination respectively. Ensemble aver-ages over the radiation and microphysics schemes showsonly small changes in the precipitation excesses betweendifferent schemes. This indicates the cumulus and bound-ary layer schemes play an important role in determining thetotal precipitation making landfall. We find the effect of thecumulus and boundary layer schemes to be important formost quantities considered below and will show averagesfor the nine runs belonging to each cumulus and boundarylayer combination, in addition to the behaviour of the fullensemble.

3.1.2 Spatial measures

We now examine the spatial precipitation properties of thesimulations. Figure 6 shows total WRF precipitation from9.00 a.m. EST 02 to 15 June 2007 for both grid resolvedand parameterized precipitation. Figure 6 supports our state-ments above regarding total precipitation, with the KFscheme (N1-9 and N19-27) generally promoting larger spa-tial simulation precipitation than the BMJ cumulus scheme(N10-18 and N28-36). Even though the cumulus schemeinfluences total rainfall, analysis of the WRF runs hereshows that for each specific run, total rainfall predomi-nately comes from the microphysical scheme at the gridspacing used in simulations of this event. For example,the cumulus scheme precipitation can only achieve rates of∼ 10 mm hr−1, where microphysical rates can be as large as∼ 100 mm hr−1.

J. B. Gilmore et al.

Fig. 6 Total precipitation from 1 to 15 June 2007 from AWAP andall 36 simulations. Each panel is labelled according to the run num-ber. Excluding AWAP, the first and last three columns are the KF and

BMJ cumulus schemes respectively. The first and last three rows arethe YSU and MYJ boundary layer schemes, respectively

Extreme precipitation in WRF during the Newcastle East Coast Low

To understand Fig. 6 more thoroughly, we decomposeprecipitation by latitude and coastal distance. Figure 7ashows total land only precipitation converted to a con-stant latitude projection and then summed over the storm.This shows that Newy produced a substantial precipitationpeak at ∼ 32◦S, with ∼5 and ∼3.5 times less precipita-tion south and north of the main peak, respectively. TheWRF simulations have a peak in precipitation at ∼ 32◦Sin good agreement with the observation, except for someruns which show a more northerly position of the peakprecipitation. The size of the peak is also well captured,with the largest model-observation discrepancy being lessthan 50 %.

Mixed results are found on either side of the mainpeak in Fig. 7a. South of 32◦ S, there is good agreementbetween model and observed data with no systematic overor underestimation of precipitation. North of 32◦ S, thereis a clear excess of precipitation in the model, with theaverage excess of all physics runs from 30◦ S to 24◦ Sbeing twice the AWAP precipitation which has good spatialsampling of daily stations in this region. The precipitationnorth of the storm peak occurs before the main event inboth WRF and the observations, and this indicates greaterprecipitation activity before the storm peak over southernQueensland in WRF relative to AWAP. Figure 7b shows

the cumulus and boundary layer scheme ensembles as anaverage over each of its nine members. The storm peakis generally well captured by all combinations. An excep-tion is the simulations using the BMJ-MYJ combination.For these runs, the ensemble average shows a northwardshift of the peak, and an overall smoothed appearance dueto the variability in peak location. The precipitation to thesouth of the main peak agrees well with AWAP in all cases.To the north, all four averages overestimate the total pre-cipitation, indicating a shared origin of this effect in thesimulations.

This storm produced extreme totals localized near thecoast and we would like to know if the WRF ensembleadequately simulates this coastal precipitation. We exam-ined precipitation as a function of linear westward distancefrom the eastern coast of Australia (Fig. 7c) using the con-stant latitude projection. Total coastal WRF precipitationranges from 10 × 103 mm (N32) to 25 × 103 mm (N9)over the 36 different model physics runs. The AWAP obser-vations indicate 14.9 × 103 mm, clearly falling within thelower end of the models. The cumulus and the boundarylayer schemes have the largest influence on the coastal pre-cipitation (Fig. 7d). Compared to AWAP, the BMJ schemeproduces coastal precipitation that more closely mirrors theobservations. The PBL influence is not as pronounced as

40 35 30 250

2000

4000

6000

8000

10000

12000

Latitude

Summed

Precipitatio

nmm

40 35 30 250

2000

4000

6000

8000

10000

12000

Latitude

Summed

Precipitatio

nmm

0 200 400 600 800 1000 1200 14000

5000

10000

15000

20000

25000

Distance From Coast km

Summed

Precipita

tion

mm

0 200 400 600 800 1000 1200 14000

5000

10000

15000

20000

25000

Distance From Coast km

Summed

Precipita

tion

mm

a b

c d

KF-YSU

BMJ-YSU

KF-MYJ

BMJ-MYJ

Fig. 7 Spatial precipitation from WRF and AWAP. Top row: Total pre-cipitation as a function of latitude for a all model configurations and baverages for each combination of cumulus and boundary layer scheme.Bottom row: Total precipitation over the domain as a function of

distance from the coastline, c all model runs and d scheme combina-tion averages. AWAP is the black curve in all panels. The run coloursand ensemble averages are the same as Fig. 5. In this figure, we haveused a 0.125o lat-lon regridding

J. B. Gilmore et al.

the cumulus scheme, but the YSU scheme does producehigher coastal precipitation than MYJ. Of interest are thedifferences in the precipitation less than 40 km from thecoast for the four cumulus and boundary layer combinations(Fig. 7d). The KF cumulus scheme has more than twice theprecipitation compared to the BMJ scheme, whereas inlandthis difference is considerably reduced.

3.1.3 Precipitation structure and timing

Figure 8 shows the daily precipitation totals around stormpeak starting from 9.00 a.m. EST 05 to 10 June 2007.The two model runs shown are those with the largest dailyextreme totals over land (∼ 400 mm, model N2), andthe smallest daily extreme precipitation total (∼ 160 mm,model N28), along with AWAP. Clearly, the storm in WRFis late by ∼1 day in both the maximal and minimal extremeevent simulations. This delay is common amongst all sim-ulations and is discussed later. Even though the events aredelayed, both N2 and N28 have strong qualitative agreementin precipitation structure. For example, both have similarprecipitation totals in Queensland on 06–07 June around 24◦S, becoming coastal precipitation about the same latitudeon 07–08 June, then moving down to 30◦ S on 08–09 June,and finally both events on 09–10 June produce substantialprecipitation around Newcastle at 32◦ S. Throughout thisdevelopment, the precipitation in the Tasman sea is alsosimilar. One clear distinction occurs for extreme precipita-tion totals greater than 200 mm day−1, which are commonlyseen in N2 with the KF-YSU scheme. While in N28 withthe BMJ-MYJ scheme, these “cells” of extreme precipita-tion greater than 200 mm day−1 are absent. This shows thatfor this event and WRF configuration, the KF scheme pro-motes localized precipitation extremes relative to BMJ. Wealso note that the cumulus scheme produces only a minorcomponent of the rain in these “cells”, with the micro-physical scheme producing the majority. At least for thisevent and WRF configuration, this indicates that KF is moreefficient at promoting large-scale accent of moisture, wherethe microphysics can then convert the moisture into surfacerainfall.

3.1.4 Event precipitation distributions

We now examine daily precipitation distributions for theentire period over land. Distributional measures can be use-ful to identify certain precipitation characteristics, enablinga quantifiable comparison of precipitation events (Karl andKnight 1998; Adler et al. 2000; Lonfat et al. 2004). Here,we use metrics of precipitation distributions to tell us if theoverall precipitation extremes are reproduced by the differ-ent physics choices. We are aware that by using different

physics combinations, storm dynamics will be slightly dif-ferent for each run and this will play a role in determiningthe extremes. However, we would like to know if the land-based precipitation distribution from the Newcastle eventcould be reasonably modelled by the different physics com-binations, and when evaluated with other measures, suchas spatial properties, timing and totals, this provides valu-able information. This situation has direct applicability toregional climate downscaling, where limited computingresources are typically allocated for long simulations atsmall grid spacing rather than simulating a large range ofdifferent physics combinations as undertaken for Newy here(Fita et al. 2010; Liang et al. 2012).

Figure 9a shows the daily precipitation distribution forthe model (coloured dots) and AWAP (large black dots)along with ±3σ (gray lines) from the AWAP Monte Carlodiscussed in the Methods section. All physics combinationsrun with WRF reproduce the overall distribution shape sur-prisingly well—especially the extreme daily totals. Closerexamination of Fig. 9a shows the simulations produce anexcess over the AWAP distribution in the intermediate pre-cipitation range, 40 − 140 mm day−1. This excess resultsfrom the additional precipitation over southern Queenslandbefore the event maximum, as discussed above. The numberof extreme events from 150–250 mm day−1 is well repro-duced. A number of simulations produce daily extremesfrom 300–400 mm day−1 that do not appear to agree withthe AWAP dataset. The four runs with the most extremeprecipitation values are N2, N4, N12, and N7 with allbelonging to the KF-YSU simulations except N12, whichis a BMJ-YSU simulation. The simulations with the lowestprecipitation extremes are N28, N26, N34 and N10, with thefirst three runs belonging to the BMJ-MYJ simulations. Wealso find similar behaviour amongst the schemes in the otherevents from Ji et al. (2014). The daily drizzle distributionsfrom the model agree with AWAP in all cases (not shown).

3.1.5 Maximum daily totals

We finish the analysis of the daily precipitation extremes byexamining the maximum precipitation over land as a func-tion of simulation day. The time series of maximum dailytotals over land are shown in Fig. 9c,d. Clearly, the modelis 1 day late but all simulations estimate the peak maxi-mum and shape of the AWAP observations well. Before themain event good agreement is seen, however, after the event,there appears to be a general overestimation of the maxi-mum daily precipitation by the models. At storm peak, theboundary layer scheme, rather than the cumulus scheme,produces the largest differences between the models withthe YSU scheme showing higher daily totals than MYJ. Wenote that total rainfall over the land in the inner domain was

Extreme precipitation in WRF during the Newcastle East Coast Low

a g m

b

c

d

e

f

h

i

j

k

l

n

o

p

q

r

Fig. 8 Comparison of daily precipitation from AWAP vs. the mostextreme simulations. The rows show daily totals starting from9.00 a.m. 05 June 2007. For example, panels (a), (g) and (m) showthe daily totals from 9.00 a.m. 5 June to 9.00 a.m. 6 June EST. Panels

(a–f) show the land only AWAP daily totals with the ocean shaded, andfor the simulations, (g–l) the maximal daily precipitation extreme runN2, and (M)-(R) the minimal daily precipitation extreme run N28

shown in Fig. 3a of Evans et al. (2012) and when ensem-ble averaged, this shows similar results to maximum daily

totals, that is, KF-YSU and BMJ-MYJ have the highest andsmallest precipitation amounts at storm peak, respectively.

J. B. Gilmore et al.

Fig. 9 Daily precipitationdistribution and maximum dailytotals. AWAP is compared to allmodel runs (left column) andaverages for each combinationof cumulus and boundary layerscheme (right column). Panels(a)–(b) show the dailyprecipitation distributions, andversus simulation day: (c)–(d)Maximum daily totals. Theblack dots are the mean of theAWAP Monte Carlo and the graylines show the AWAP 3σ errors.The run colours and ensembleaverages are the same as Fig. 5

0 100 200 300 4001

10

100

1000

104

105

Accumulation mm day

Cou

nts

0 100 200 300 4001

10

100

1000

104

105

Accumulation mm day

Cou

nts

0 1 2 3 4 5 6 7 8 9 10 110

100

200

300

400

Days since 9am EST 2 June 2007

Max

DailyPrecipita

tion

mm

0 1 2 3 4 5 6 7 8 9 10 110

100

200

300

400

Days since 9am EST 2 June 2007

Max

DailyPrecipita

tion

mm

KF-YSU

BMJ-YSUKF-MYJBMJ-MYJ

a

c d

b

3.2 Hourly evaluation

Evaluating precipitation in WRF on an hourly basis pro-vides greater temporal sensitivity than the daily totals dis-cussed previously. The hourly evaluation consists of twoparts, a comparison of (1) the hourly distribution of pre-cipitation and (2) the total rain falling around Newcastle.For the precipitation distributions, we undertake a grid pointand an interpolated precipitation comparison. This allowsus to demonstrate a robust comparison for hourly totals. Forthe total hourly precipitation falling around Newcastle, weuse only the interpolated product since its totals are wellconstrained by AWAP.

3.2.1 Hourly precipitation distribution

The hourly precipitation distribution over land is shownin Fig. 10 for both the grid point and interpolation com-parisons. As explained in the data section, the threecomparisons used are (1) station locations only (Fig. 10a,b),(2) including all model points within 0.25◦ of the stationlocations (Fig. 10c,d) and (3) the interpolated hourly precip-itation including all model points in the interpolation landmask (Fig. 10e,f). Note that counts can be less than onebecause of the ensemble normalization.

For the grid point comparisons, the drizzle and interme-diate part of the distributions are well modelled by WRF.However, it is evident in the ensemble averages (Fig. 10b,d)that the hourly data favours the KF cumulus scheme curvesfor the highest precipitation totals P � 20 mm hr−1. This

indicates that the BMJ cumulus scheme does not producehigh enough hourly precipitation totals in this event. To testthe robustness of this relationship, we also cut the domainin two at Newcastle’s latitude and find the same preferencefor the larger precipitation extremes from the KF cumulusscheme.

For the interpolated product, reasonable agreement isseen across the models in Fig. 10e, particularly for theintermediate hourly totals from P ∼ 20 to 30 mm hr−1.There is an over estimation of the total number of countsfor P < 20 mm hr−1 for all physics choices. Some mod-els also have substantially greater hourly extremes than theinterpolated hourly station data, with these models beingruns N4 through to N8, which all belong to the KF-YSUsimulations. For the cumulus and boundary layer averagesin Fig. 10f, the effects of the physics do become evident.Figure 10b shows that the KF cumulus scheme precipita-tion distribution for P � 30 mm hr−1 provides a betterrepresentation of the observations. This result is in clearagreement with the grid point analysis above and providesadditional evidence that the KF cumulus scheme producesextremes that are more realistic. It is also clear that the KFscheme generates stronger hourly extremes than the BMJscheme. For the hourly totals, this results in a clear separa-tion of the cumulus scheme precipitation distributions thatwas not present in the daily data (compare Figs. 9b and10f). The boundary layer also plays a role here, with theYSU scheme having larger extremes than the MYJ scheme;however, the effect is less important relative to the cumulusscheme. Since the BMJ and KF distributions do not overlap

Extreme precipitation in WRF during the Newcastle East Coast Low

0 20 40 60 80 100

10 2

10 1

100101102103104105106

Accumulation mm hour

Cou

nts

Station locations only

0 20 40 60 80 100

10 2

10 1

100101102103104105106

Accumulation mm hour

Cou

nts

Station locations only

0 20 40 60 80 100

10 2

10 1

100101102103104105106

Accumulation mm hour

Cou

nts

Including 0.25o around stations

0 20 40 60 80 100

10 2

10 1

100101102103104105106

Accumulation mm hourCou

nts

Including 0.25o around stations

0 20 40 60 80 100

10 2

10 1

100101102103104105106

Accumulation mm hour

Cou

nts

Interpolation

0 20 40 60 80 100

10 2

10 1

100101102103104105106

Accumulation mm hour

Cou

nts

Interpolation

KF-YSUBMJ-YSU

KF-MYJ

BMJ-MYJ

a b

c d

e f

Fig. 10 Model hourly precipitation distributions compared to thehourly precipitation station data and interpolation. Top row: Stationonly model distributions for (a) hourly precipitation distribution for allruns and (b) the cumulus scheme boundary layer averages. Middle row:Including all model grid points within 0.25◦ of the stations for (c) allruns and (d) the cumulus scheme boundary layer averages. Last row:

Interpolated station data verses all model grid points with the interpo-lation land mask for (e) all runs and (f) the cumulus scheme boundarylayer averages. The run colours and ensemble averages are the same asFig. 5. The hourly data in each row is the black curve. Model countscan go below 1 because of the ensemble and count averages used

for hourly totals, it is possible to discriminate betweenthe BMJ and KF cumulus schemes using hourly precipi-tation distributions. We also find the cumulus and bound-ary layer schemes in the remaining events from Ji et al.(2014) display similar behaviour for hourly precipitationdistributions.

3.2.2 Hourly precipitation totals

We now present the hourly precipitation time series start-ing at 9.00 a.m. EST 02 June 2007. Figure 11a,b showstotal precipitation falling around Newcastle in a numberof peaks, with the main event from 5.4 to 7.5 days (black

line). A number of smaller events are also visible, with mostoccurring before the main event at 4.2–5.4 days.

Inspection of the WRF produced precipitation showsthere is a delay of approximately 12–24 h relative to theactual event. Despite this, the magnitude of the WRF precip-itation is generally similar (Fig. 11a), although most modelsoverestimate the amount of precipitation occurring at eventpeak. For the physics averages (Fig. 11b), the KF cumu-lus scheme overestimates peak precipitation by ∼ 30 %on average, whereas the BMJ cumulus scheme is similarto the observed peak. Figure 11b also shows the bound-ary layer influence, with the YSU scheme promoting moreprecipitation than MYJ.

J. B. Gilmore et al.

0 1 2 3 4 5 6 7 8 9 10 11 12

0

200000

400000

600000

800000

1 106

Days since 9am EST 2 June 2007

TotalHourlyPrecipitatio

nmm

0 1 2 3 4 5 6 7 8 9 10 11 12

0

200000

400000

600000

800000

1 106

Days since 9am EST 2 June 2007

TotalHourlyPrecipitatio

nmm KF-YSU

BMJ-YSUKF-MYJBMJ-MYJ

a b

Fig. 11 Total precipitation in the interpolation land mask compared with the WRF runs for each simulation hour. Panel (a) shows all runs andpanel (b) the cumulus scheme boundary layer averages. The run colours and ensemble averages are the same as Fig. 5. The hourly interpolationdata is the black curve

3.3 Storm time delay

Both the daily and hourly precipitation analysis, (Figs. 6and 10) showed that the WRF simulations contained a com-mon delay of 12–24 h relative to the actual precipitationobservations from the Newy storm. The delay originates inthe outer domain because the inner domain uses one-waynesting and has a relatively small size that cannot apprecia-bly influence the synoptic scale processes that formed thisevent. Evidence for this is provided in Fig. 12. This dis-plays a comparison of mean sea level pressure snapshotsbetween ERA Interim and the outer domain simulationsfrom the minimal (N28) and maximal (N2) precipitationextreme simulations as identified in the inner domain. Thisshows that the delay does originate in the outer domain asexpected. Other runs have similar behaviour. Detailed studyof the exact mechanisms of how different physics schemesinfluence the timing of the Newcastle storm, and its posi-tioning relative to Eastern Australia through simulations inthe outer domain is a subject for future work.

3.4 Grid scale analysis

Analysis of the WRF simulations showed that the cumulusand boundary layer schemes caused systematic changes inthe hourly precipitation distribution. These differences werehowever produced by precipitation from the microphysicsscheme rather than precipitation from the cumulus schemes(whose maximum hourly precipitation was about 10 timesless than the microphysical scheme). This implies that thecumulus and boundary layer schemes are modifying theatmosphere to allow greater grid scale microphysical precip-itation. We examined grid scale behaviour in the simulationsand found that extremes in the distributions of the verticalmoisture flux, given by qw at model grid points where q isspecific humidity and w vertical velocity, was a robust pre-dictor of hourly extremes. Figure 13 shows the land-baseddistribution for water vapour, vertical velocity and verti-cal moisture flux. Although each panel seems to be a good

predictor in this case (compare to Fig. 10f), when wesplit the domain (such as land/water, north/south) only thevertical moisture flux remains a good predictor.

Further analysis shows that the boundary layer schemecontrols the amount of water vapour in the inner domain,with YSU producing about 5 % more specific humidityon average than MYJ (Fig. 14a). In contrast, when regionswith rain only are considered (i.e. excluding no rain andlight drizzle) the cumulus scheme appears to control dif-ferences in column-average specific and relative humidity(Fig. 14b,c). Here, KF has a column-average relative humid-ity 6 % lower than BMJ. This appears to indicate thatthe KF scheme is causing vapour to rain out more vigor-ously, thereby causing stronger grid-scale circulations viathe release of latent heat, but also competing with thestronger circulations by reducing water vapour in the rain-ing regions (Wing and Emanuel 2014). If a higher thresholdis used for raining regions, i.e. focussing on extreme instan-taneous rain rates only, we start to see the microphysicsinfluence the atmospheric conditions. This effect is alreadypresent in Fig. 14c where the WDM 5 scheme can be seen toproduce lower column-average relative humidity than WSM3 and WSM 5 for a given cumulus-boundary layer com-bination (e.g. compare N1 to N6 and N7 to N9). Detailedanalysis of the specific physics influencing water vapourfrom the boundary/surface layers, cumulus schemes andmicrophysical parameterizations are left for future work.

4 Discussion

We now discuss how the physics schemes influenced pre-cipitation. For the entire storm period, the BMJ-MYJ com-bination produced the least biased precipitation totals. Forthe daily maximum extremes, all models agreed well withobservations. However, for the hourly extremes, the dif-ferent cumulus schemes had distinct behaviours: WRFwith the BMJ cumulus scheme clearly underestimatedextreme hourly totals, whereas WRF with the KF cumulus

Extreme precipitation in WRF during the Newcastle East Coast Low

Fig. 12 Evidence that the innerdomain time delay originates inthe outer domain of thesimulations. The left, middle andright columns show mean sealevel pressure for ERA Interim,maximal run N2, and minimalrun N28, respectively. These areshown from 5 to 10 June at 00UTC. The runs N2 and N28 aredepicted here using the outerdomain simulations only.Examination of the 8 to 10 Junepanels shows the low pressuresystem off the east coast ofAustralia in the WRFsimulations are systematicallylate compared to ERA Interim

scheme had much higher extreme hourly accumulations(Fig. 10b,d,f) in better agreement with the precipitation data.For the hourly precipitation totals around Newcastle, theoverall peak was reproduced, but many of the finer fea-tures were not seen due to the storm being one day late onaverage.

The difference in results between the daily and hourlytime scale at storm peak provides evidence that the rel-ative impact of cumulus and boundary layer schemes istime-scale dependent. Hourly precipitation extremes areinfluenced most by the cumulus scheme; however, highhourly totals tend to be compensated by lower totals during

J. B. Gilmore et al.

b

a

c

Fig. 13 Distribution of a water vapour mixing ratio, b vertical veloc-ity and c vertical moisture flux for ensemble averages of the fourcumulus and boundary layer scheme combinations. These distributionsare constructed from atmosphere over land in the simulations. Onlythe patterns seen in (c) are robust to changes in the area considered.Colours are the ensemble averages from Fig. 5

the same day. This results in a relatively small differentia-tion at the daily time scale, as evidenced by the reorderingof the extremes in the daily distribution compared to thehourly distribution (Fig. 9b versus Fig. 10). Maintaininghigh precipitation rates throughout the day is more stronglyinfluenced by the boundary layer scheme. This follows otherstudies, such as Kendon et al. (2012) who found that bias inhourly precipitation was most associated with deficienciesin the cumulus scheme.

The radiation and the microphysics schemes did not playa substantial role in changing the precipitation distribu-tions when ensemble averages were considered on dailyand hourly timescales, although differences were found inatmospheric conditions in raining areas for different micro-physics schemes. This result for the microphysics schemewas interesting, since it is generating the majority of theprecipitation in the simulations, and past studies indicatedprecipitation can be sensitive to the microphysics choice

a

b

c

l

l

l

l

l

l

Fig. 14 Atmospheric conditions over a the inner domain (both rain-ing and non-raining) for column water vapour, and within rain formingregions for b column water vapour and c column average relativehumidity. The different runs are coloured by their cumulus-boundarylayer group from Fig. 5

(Fiori et al. 2014). One possible reason for this lack ofinfluence on the ensemble average is that the microphysicsschemes used (WSM 3, WSM 5 and WDM 5) are all basedon the underlying microphysics of Hong et al. (2004) andLim and Hong (2010). It is also possible that large-scale for-cing dominated the convergence and rainout of watervapour on daily and hourly scales in this synoptic scaleevent.

When compared with the storm dynamics, BMJ-YSU hasthe shortest time delay compared to the observations, buthas a deficit in precipitation extremes for hourly totals. Forthe KF simulations, although they produce the larger hourlytotals seen in the observations, they are ∼ 24 h late and pro-duce too much precipitation over land. Gallus and Bresch(2006) also found KF produced higher peak precipitationrates compared to BMJ. We note the presence of the delay instorm initiation in the simulations is the major contributor tothe low skill scores seen in Evans et al. (2012) for this event.

Extreme precipitation in WRF during the Newcastle East Coast Low

5 Conclusion

We investigated precipitation extremes produced in theNewy event of 2007 in WRF. This extreme extra-tropicallow-pressure system provided an excellent test case forexamining long-range extreme rainfall simulations. Thelong lead-time probes WRF’s ability to simulate extremeevents in circumstances that are more typical of regionalclimate studies. Our simulations used 36 different physicscombinations in the WRF model. We examined storm totals,daily and hourly precipitation, and spatial measures of theextreme precipitation. We showed that extreme precipita-tion could be well modelled in WRF at daily and longertimescales using a long lead-time. We showed that allfour combinations of cumulus and boundary layer schemesmodelled the observed daily extreme precipitation well,and when ensemble averaged (Fig. 7b), all four combina-tions fell within the 3σ range of the observed distributionsof daily precipitation extremes from AWAP. Although therange of the 36 runs themselves was large, systematicdifferences were observed in the ensemble averagesfor the different cumulus and boundary layer combina-tions indicating these combinations could potentially bediscriminated.

Even though the simulations were designed with regionalclimate in mind, we have shown that higher frequency sam-pling of precipitation extremes, such as hourly totals, canreveal differences between cumulus and boundary layerschemes, whereas daily precipitation totals could not. Thismeans the parameterization of cumulus and boundary layerschemes could potentially be improved by examining hourlyprecipitation totals. Moving to hourly totals for model com-parison would be fruitful, however full spatial coverageof the area of interest would be desirable. This could beachieved by combining gauge measurements with radarmeasurements for example (Haberlandt 2007; Verworn andHaberlandt 2011).

The results presented here have direct implications forregional climate modelling of precipitation. In regional sim-ulations, one usually chooses a set of physics options andthen simulates with these over long times, typically fordecadal long simulations. If one is interested in extremeprecipitation, this approach could lead to large systematicerrors in the resulting precipitation extremes for shorteraccumulation periods. We showed that the BMJ-MYJ com-bination produced the best match with the observed pre-cipitation total for the entire storm period; however, itunderestimated hourly precipitation extremes substantially.On the other hand, the KF cumulus scheme did producerealistic hourly precipitation extremes but overestimated thetotal precipitation from the event. Therefore, using a physicscombination that provides a good representation of the meanprecipitation for a region may not produce realistic extreme

precipitation for shorter accumulation periods. This couldpotentially introduce large biases in an hourly precipitationextreme analysis.

While the microphysics parametrization is responsiblefor the vast majority of the extreme precipitation in theseWRF simulations of the Newy event, we show here thatit is the cumulus (and to a lesser extent the PBL) schemethat determines the intensity of this precipitation. That is,the large scale convergence which is generating precipita-tion through the microphysics parametrization is stronglyinfluenced by changes in the vertical atmospheric profileintroduced by the cumulus scheme. Thus, the interaction ofthese grid point schemes with the larger-scale atmosphericmotion is key to producing precipitation extremes for Newyin this WRF configuration.

Based on this work, we provide some recommendationsfor regional climate modelling simulations where precipita-tion extremes are of interest. Simulation, or at least testingwith multiple physics combinations, is advisable, and suchan ensemble has previously been found to provide a largerspread than a perturbed initial condition ensemble (Gallusand Segal 2001; Tapiador et al. 2012). This would allowan objective assessment of precipitation extremes in a givenregional climate model experiment. This work shows thatif hourly precipitation extremes are relevant to the intendedregional climate application, then using select test cases forcomparing various scheme configurations against observedhourly precipitation could provide important informationabout the potential biases introduced by scheme selection inlonger regional climate simulations.

Acknowledgments This work is made possible by funding fromthe NSW Environmental Trust for the ESCCI-ECL project, the NSWOffice of Environment and Heritage backed NSW/ACT RegionalClimate Modelling Project (NARCliM) and the Australian ResearchCouncil as part of the Discovery Project DP0772665, Super ScienceFellowship FS100100054 and Future Fellowship FT110100576. Thiswork was supported by an award under the Merit Allocation Schemeon the NCI National Facility at the Australian National University.

References

Adler RF, Huffman GJ, Bolvin DT, Curtis S, Nelkin EJ (2000) Tropi-cal rainfall distributions determined using TRMM combined withother satellite and rain gauge information. J Appl Meteor 39:2007–2023

Argueso D, Hidalgo-Munoz JM, Gamiz-Fortis SR, Esteban-Parra MJ,Dudhia J, Castro-Diez Y (2011) Evaluation of WRF parameteri-zations for climate studies over Southern Spain using a multi-stepregionalization. J Climate 24:5633–5651

Betts AK (1986) A new convective adjustment scheme. Part I: obser-vational and theoretical basis. J Atmos Sci 121:255–270

Betts AK, Miller MJ (1986) A new convective adjustment scheme.Part II: single column tests using GATE wave, BOMEX, and arcticair-mass data sets. Quart J Roy Meteor Soc 121:693–709

J. B. Gilmore et al.

Bukovsky MS, Karoly DJ (2009) The sensitivity of WRF precipitationin nested regional climate simulations. J Appl Metr Clim 48:2152–2159

Collins WD, Rash PJ, Boville BA, Hack JJ, McCaa JR, WilliamsonDL, Kiehl JT, Briegleb B (2004) Description of the NCARcommunity atmosphere model (CAM 3.0). NCAR Tech. NoteNCAR/TN-464+STR

Dee DP, Uppala SM, Simmons AJ, Berrisford P, Poli P, KobayashiS, Andrae U, Balmaseda MA, Balsamo G, Bauer P, BechtoldP, Beljaars ACM, van de Berg L, Bidlot J, Bormann N, DelsolC, Dragani R, Fuentes M, Geer AJ, Haimberger L, HealySB, Hersbach H, Holm EV, Isaksen L, Kallberg P, KohlerM, Matricardi M, McNally AP, Monge-Sanz BM, MorcretteJJ, Park BK, Peubey C, de Rosnay P, Tavolato C, ThepautJN, Vitart F (2011) The ERA-Interim reanalysis: configurationand performance of the data assimilation system. Quart J RoyMeteor Soc 137:553–597. http://www.ecmwf.int/research/era/do/get/era-interim

Dudhia J (1989) Numerical study of convection observed during thewinter monsoon experiment using a mesoscale two-dimensionalmodel. J Atmos Sci 46:3077–3107

Evans JP, Ekstrom M, Ji F (2012) Evaluating the performance of aWRF physics ensemble over South-East Australia. Clim Dynam39:1241–1258

Evans JP, Ji F, Lee C, Smith P, Argueso D, Fita L (2014) A regionalclimate modelling projection ensemble experiment – NARCliM.Geosci Model Dev 7:621–629

Fiori E, Comellas A, Molini L, Rebora N, Siccardi F, Gochis DJ,Tanelli S, Parodi A (2014) Analysis and hindcast simulations of anextreme rainfall event in the Mediterranean area: the Genoa 2011case. Atmos Res 138:13–29

Fita L, Fernandez J, Garcıa-Dıez M (2010) CLWRF: WRF modi-fications for regional climate simulation under future scenarios.Preprints, 11th WRF Users Event, Boulder, CO, NCAR, P-26

Flaounas E, Bastin S, Janicot S (2011) Regional climate modellingof the 2006 West African monsoon: sensitivity to convectionand planetary boundary layer parameterisation using WRF. ClimDynam 36:1083–1105

Gallus WA Jr, Bresch JF (2006) Comparison of impacts of WRFdynamic core, physics package, and initial conditions on warmseason rainfall forecasts. Mon Wea Rev 134:2632–2641

Gallus WA Jr, Segal M (2001) Impact of improved initialization ofmesoscale features on convective system rainfall in 10-km Etasimulations. Wea Forecasting 16:680–696

Haberlandt U (2007) Geostatistical interpolation of hourly precipita-tion from rain gauges and radar for a large-scale extreme rainfallevent. J Hydro 332:144–157

Hong SY, Dudhia J, Chen SH (2004) A revised approach to ice micro-physical processes for the bulk parameterization of clouds andprecipitation. Mon Wea Rev 132:103–120

Hong SY, Noh Y, Dudhia J (2006) A new vertical diffusion packagewith an explicit treatment of entrainment processes. Mon Wea Rev134:2318–2341

Jacob D, Petersen J, Eggert B, Alias A, Christensen OB, BouwerLM, Braun A, Colette A, Dqu M, Georgievski G, GeorgopoulouE, Gobiet A, Menut L, Nikulin G, Haensler A, Hempelmann N,Jones C, Keuler K, Kovats S, Krner N, Kotlarski S, Kriegsmann A,Martin E, Meijgaard Evan, Moseley C, Pfeifer S, Preuschmann S,Radermacher C, Radtke K, Rechid D, Rounsevell M, SamuelssonP, Somot S, Soussana J-F, Teichmann C, Valentini R, Vautard R,Weber B, Yiou P (2014) EURO-CORDEX: new high-resolutionclimate change projections for European impact research. RegEnviron Change 14:563–9678

Janjic ZI (1994) The step-mountain eta coordinate model: Furtherdevelopments of the convection, viscous sublayer and turbulenceclosure schemes. Mon Wea Rev 122:927–945

Jankov I, Gallus WA, Segal M, Shaw B, Koch SE (2005) The impact ofdifferent WRF model physical parameterizations and their interac-tions on warm season MCS rainfall. Wea Forecasting 20:1048–1060

Ji F, Ekstrom M, Evans JP, Teng J (2014) Evaluating rainfall pat-terns using physics scheme ensembles from a regional atmosphericmodel. Theor Appl Climatol 115:297–304

Ji F, Evans JP, Argueso D, Fita L, Di Luca A (2015) Using large-scalediagnostic quantities to investigate change in east coast lows. ClimDyn. doi:10.1007/s00382-015-2481-9

Jones D, Wang W, Fawcett R (2009) High-quality spatial climate data-sets for Australia. Aust Meteor Mag 58:233–248

Kain JS (2004) The Kain-Fritsch convective parameterization: anupdate. J Appl Meteor 43:170–181

Kain JS, Fritsch JM (1990) A one-dimensional entraining/detrainingplume model and its application in convective parameterization. JAtmos Sci 47:2784–2802

Karl TR, Knight RW (1998) Secular trends of precipitation amount,frequency, and intensity in the United States. Bull Amer MeteorSoc 79:231–242

Kendon EJ, Roberts NM, Senior CA, Roberts MJ (2012) Realism ofrainfall in a very high-resolution regional climate model. J Climate25:5791–5806

King AD, Alexander LV, Donat MG (2012) The efficacy of using grid-ded data to examine extreme rainfall characteristics: a case studyfor Australia. Int J Climatol. doi:10.1002/joc.3588

Lavers DA, Villarini G (2013) Were global numerical weather predic-tion systems capable of forecasting the extreme Colorado rainfallof 9-16 September 2013? GRL 40:6405–6410

Liang XZ, Xu M, Yuan X, Ling T, Choi HI, Zhang F, Chen L, Liu S,Su S, Qiao F, Wang JXL, Kunkel KE, Gao W, Joseph E, Morris V,Yu TW, Dudhia J, Michalakes J (2012) Regional climate weatherresearch and forecasting model. Bull Amer Meteor Soc 93:1363–1387

Lim KSS, Hong SY (2010) Development of an effective double-moment cloud large-scale condensation scheme with prognosticcloud condensation nuclei (CCN) for weather and climate models.Mon Wea Rev 138:1587–1612

Lonfat M, Marks FD, Chen SS (2004) Precipitation distribution intropical cyclones using the tropical rainfall measuring mission(TRMM) microwave imager: a global perspective. Mon Wea Rev132:1645–1660

Lowrey MR, Yang ZL (2008) Assessing the capability of a regional-scale weather model to simulate extreme precipitation patterns andflooding in central Texas. Wea Forecasting 23:1102–1126

Mills GA, Webb R, Davidson N, Kepert J, Seed A, Abbs D (2010) ThePasha Bulker east coast low of 8 June 2007. CAWCR technicalreport no 23 Centre for Australian weather and climate research,Melbourne, Australia

Mlawer EJ, Taubman SJ, Brown PD, Iacono MJ, Clough SA (1997)Radiative transfer for inhomogeneous atmosphere: RRTM, a val-idated correlated-k model for the long-wave. J Geophys Res102:16663–16682

Paulson CA (1970) The mathematical representation of wind speedand temperature profiles in the unstable atmospheric surface layer.J Appl Meteor 9:857–861

Schumacher RS, Clark AJ, Xue M, Kong F (2013) Factors influenc-ing the development and maintenance of nocturnal heavy-rain-producing convective systems in a storm-scale ensemble. MonWea Rev 141:2778–2801

Skamarock WC, Klemp JB, Dudhia J, Gill DO, Barker DM, DudaMG, Huang XY, Wang W, Powers JG (2008) A descriptionof the advanced research WRF version 3. NCAR Tech. NoteNCAR/TN-475+STR

Speer M, Wiles P, Pepler A (2009) Low pressure systems off the NewSouth Wales coast and associated hazardous weather: establish-ment of a database. Aust Meteorol Oceanogr J 58:29–39

Extreme precipitation in WRF during the Newcastle East Coast Low

Stephens GL, L’Ecuyer T, Forbes R, Gettlemen A, Golaz JC, Bodas-Salcedo A, Suzuki K, Gabriel P, Haynes J (2010) Dreary state ofprecipitation in global models. J Geophys Res 115:D24211

Sunyer MA, Madsen H, Ang PH (2012) A comparison of differ-ent regional climate models and statistical downscaling methodsfor extreme rainfall estimation under climate change. Atmos Res103:119–128

Tapiador FJ, Tao WK, Shi JJ, Angelis CF, Martinez MA, Marcos C,Rodriguez A, Hou A (2012) A comparison of perturbed initial con-ditions and multiphysics ensembles in a severe weather episode inSpain. J Appl Meteor Climatol 51:489–504

Verdon-Kidd DC, Kiem AS, Willgoose G, Haines P (2010) East CoastLows and the Newcastle/Central Coast Pasha Bulker storm. Reportfor the National Climate Change Adaptation Research Facility,Gold Coast, Australia

Verworn A, Haberlandt U (2011) Spatial interpolation of hourlyrainfall - effect of additional information, variogram infer-ence and storm properties. Hydrol Earth Syst Sci 15:569–584

Wing AA, Emanuel KA (2014) Physical mechanisms controllingself-aggregation of convection in idealized numerical modelingsimulations. J Adv Model Earth Sy 6:59–74