Parameterization-Induced Error Characteristics of MM5 and WRF Operated in Climate Mode Over the Alpine Region: an Ensemble-Based Analysis

15 JUNE 2011 A W A N E T A L . 3107

Parameterization-Induced Error Characteristics of MM5 and WRF Operated in Climate Mode over the Alpine Region: An Ensemble-Based Analysis

NAUMAN K. AWAN,H.TRUHETZ, AND A. GOBIET Wegener Center for Climate and Global Change, Institute for Geophysics, Astrophysics, and Meteorology, University of Graz, Graz, Austria

(Manuscript received 11 February 2010, in ﬁnal form 14 February 2011)

ABSTRACT

This study investigates the role of physical parameterization in regional climate model simulations. The authors also present a comprehensive assessment of errors arising from use of physical parameterization schemes, and their consequent impact on model performance in a region of complex topography. An error range

related to the choice of physical parameterization is provided for 2-m air temperature T2m and precipitation. Two state-of-the-art nonhydrostatic mesoscale regional climate models, the ﬁfth-generation Pennsylvania State University–National Center for Atmospheric Research Mesoscale Model (MM5) and the Weather Research and Forecasting (WRF) model, are used to dynamically downscale the 40-yr ECMWF Re-Analysis

(ERA-40) to a spatial resolution of 10 km 3 10 km in the European alpine region. Simulated T2m and precipitation are compared with gridded observational datasets. The model performance on regional and subregional scales is evaluated on daily, monthly, seasonal, and annual time scales. The results are based on a mixed physics ensemble of twenty-nine 1-yr-long hindcast simulations generated by choosing different model conﬁgurations. These results indicate that performance of both models is sensitive to the choice of physical parameterization and WRF is more sensitive than MM5. This sensitivity is higher during summer than during winter. The cumulus and microphysics scheme have the dominant effect on model performance during summer while boundary layer and radiation schemes affect the results during all seasons.

It is found that annual mean error in the alpine region for T2m and precipitation lies between 22.758 and 21.088C(21.128 and 1.338C) and 0.27 and 0.80 mm day21 (0.13 and 1.51 mm day21) for MM5 (WRF), respectively. The authors also found that the error range during winter for subregional scales is higher than the regional-scale error range, while during the other seasons the subregional error ranges are higher only in case

of precipitation but not for T2m. These results suggest that signiﬁcant reductions of errors can be achieved by choosing a suitable model conﬁguration for the region of interest. The other outcome of these experiments is a suitable setup, which can be used to conduct further long-term high-resolution climate simulations.

1. Introduction several assumptions and approximations to simplify unresolved processes. These parameterizations typically in- Atmospheric processes occur on various spatial and clude moist convection, atmospheric turbulence, radiative temporal scales ranging from 1022 to 108 mand1021 to transfer, microphysics, soil, and vegetation interaction. 108 s (Orlanski 1975). Global and regional climate During the past few years a wide range of parameteriza- models (GCMs and RCMs) cannot resolve these phys- tion schemes have been implemented in climate models, ical processes on all scales per construction. However, which opens a broad range of choice in model configu- unresolved processes are considered in these models with ration and provides an opportunity to identify deficiencies the aid of physical parameterization schemes that apply in these schemes by comparative evaluation. In published literature, one can find an extensive list of different parameterization schemes depicting the same physical pro- Corresponding author address: Nauman K. Awan, Regional and cess. The fact that each of these schemes is based on many Local Climate Modeling and Analysis Research Group, Wegener Center for Climate and Global Change Leechgasse 25, A-8010, assumptions, and these assumptions may fail or give an Graz, Austria. inadequate response to certain synoptic forcing, limits their E-mail: [email protected] application and acts as a source of errors in the models.

DOI: 10.1175/2011JCLI3674.1

Ó 2011 American Meteorological Society Unauthenticated | Downloaded 10/11/21 03:12 AM UTC 3108 JOURNAL OF CLIMATE VOLUME 24

This study focuses on the performance of physical pa- RCMs but can also help to identify different sources of rameterization schemes and on the errors associated with errors in the model results. Jacob et al. (2007) found that their application in RCMs. We use two limited area models depending on different models and model configurations in this study: the Pennsylvania State University–National the mean bias (in comparison with Climatic Research Center for Atmospheric Research (NCAR) Mesoscale Unit dataset) in temperature (precipitation) for the alpine Model (MM5) (Dudhia 1993) and the National Centers region during the months of December–February (DJF) for Environmental Prediction –NCAR Weather Research and June–August (JJA) lies between 20.348 and 2.018C and Forecasting (WRF) model (Skamarock et al. 2007). (20.62 and 0.41 mm day21) and 20.908 and 1.958C Both models are primarily used in numerical weather (21.15 and 0.53 mm day21). In another study, Suklitsch prediction (NWP) but are applied with some modifica- et al. (2008) (based on eight 1-yr climate simulations tions for long-term simulations in climate mode as well. with 10 km 3 10 km grid spacing) showed that depending The applicability of MM5 as an RCM has been frequently on model setup the European RCM Consortium for demonstrated, for example, in Zhu and Liang (2007) and Small-Scale Modeling (COSMO) Model in Climate Mode Fernańdez et al. (2007), while the application of WRF as (CCLM) (Boehm et al. 2006) has annual mean bias lying an RCM is still being explored, for example, in Done between 20.988 and 21.448C for temperature and 20.14 et al. (2004, 2005). and 20.42 mm day21 for precipitation in the entire Eu- Physical parameterizations as an error source in model ropean alpine region. In a more recent study, Suklitsch results have been a topic of numerous studies carried out et al. (2010) found that the error range in the alpine region with MM5 and WRF mostly focusing on short-term defined as the 2.5th and 97.5th percentiles of error lie weather events. Studies from Pan et al. (1996), Ferretti between 23.28 and 2.08C for temperature and between et al. (2000), Kotroni and Lagouvardos (2001), Jankov 22.0 and 3.1 mm day21 for precipitation. et al. (2005), and Otkin and Greenwald (2008) give a We would like to point out that besides physical pa- good overview on the different physical parameteriza- rameterization schemes and the unconfined empirical tions available in these models and their response to parameters within these schemes, there are other sour- different synoptic conditions. Although these studies give ces of errors in the numerical models. The dependence of an overview of what can be expected in terms of model numerical models on different numerical solvers, initial performance, it is inconclusive whether the underlying and boundary conditions, domain sizes and position, hor- assumptions within the parameterizations hold or fail in izontal and vertical resolution, soil and vegetation char- longer simulations and the error ranges associated with acteristics, along with nudging and assimilation techniques parameterizations has not been quantified yet. The in- account for these errors in the results. Although it would vestigation of these issues requires relatively longer sim- be interesting to explore the effect of all these factors, the ulation period on the one hand and an ensemble consisting main focus of our study is the role of physical parame- of simulations with different models (multimodel ensem- terizations in RCM simulations. Therefore, such inves- ble) and model configurations (mixed physics ensemble) tigations are out of the scope of this study. In this study we on the other hand. have tried to identify the extent and impact of errors The application of ensembles in NWP has helped re- resulting from different physical parameterizations by searchers to improve the weather forecasts and they have using an ensemble approach. been widely used in NWP research (Murphy 1993; Molteni The general layout of this paper is as follows. The et al. 1996), but their application in climate/climate change region of interest and the boundary conditions are de- research was limited because of their burden on compu- scribed in section 2. In section 3, differences in the two tational resources. During recent years some promising RCMs are briefly illustrated along with the observa- projects like the Prediction of Regional Scenarios and tional datasets that are used for model evaluation. The Uncertainties for Defining European Climate Change experiment design and evaluation methodologies are Risks and Effects (PRUDENCE) (Christensen et al. 2007), explained in section 4. The results are presented in North American Regional Climate Change Assessment section 5, and the discussion of results in section 6, while Program (NARCCAP; http://www.narccap.ucar.edu/), conclusions can be found in section 7. and ENSEMBLES (Hewitt and Griggs 2004) have suc- cessfully demonstrated the application of model ensembles in climate research. The results of the PRUDENCE 2. Model domains and boundary conditions project are based on a multimodel ensemble of 10 models operated on 50 km 3 50 km grids over Europe. Jacob The topography has strong influence on the climate of et al. (2007) and De´que´ et al. (2007) showed that en- a region. The substantial orographic features (;25% of sembles cannot only help to improve the reliability of earth’s total dry land area) significantly influence the

Unauthenticated | Downloaded 10/11/21 03:12 AM UTC 15 JUNE 2011 A W A N E T A L . 3109

parameterizations, and this setup is identical for all sensitivity experiments. Datasets provided by the United States Geological Survey (USGS) are used to derive topography, land use, and land–water masks information. The static fields from USGS 29329 and 3093309 are used for D1 and D2, respectively. The USGS 24-category land classification data are used to represent dominant vegetation types. The ERA-40 dataset (Uppala et al. 2005) with 6-hourly temporal and 1.125831.1258 grid spacing is used to provide initial and boundary conditions for D1 in both models. All the required parameters including the initial conditions for soil moisture are taken from the reanalysis dataset. As a consequence of nesting setup the update frequency of lateral boundary conditions (LBC) for D2 is much higher, FIG. 1. The model domain and topography of the fine (30 km) D1 that is, 180 s, equal to the time step of D2. The specified and high-resolution (10 km) D2 simulations. The highest point in LBCs are used for all the experiments (Grell et al. 1995; the D2 topography is 2945.82 m; Mont Blanc, the highest peak of Skamarock et al. 2007). The relaxation zone for MM5 the European Alps, is 4808 m. (WRF) is set to 7 (5) grid points. The sea surface tem- peratures are updated every 6 h in both models, while the regional and global climate by changing the dynamics of vegetation fraction is updated every month for MM5 and the atmospheric circulation (Kapos et al. 2000). Our study every 6 h for WRF. The temporal resolution for output region is the European alpine region, which extends in the of fine- (high) resolution domain is 6 (1) h. The evaluated form of an arc of 800 km, with a mean width and an av- ensemble members have a common evaluation period, erage ridge height of approximately 200 and 2.5 km, re- from 0000 UTC 1 January 1999 to 0000 UTC 1 January spectively. The Alps deflect the atmospheric flow, both 2000, which is simulated with a spinup of three months. horizontally and vertically, and introduce elevated sources and sinks of sensible and latent heat. They also induce 3. Models and observations waves that propagate in the free atmosphere. The occur- a. Regional climate models rence of several distinctive orographically related atmospheric flow phenomena (e.g., the wind systems of the The Advanced Research WRF (ARW-WRF; v2.2.1) Boise, Bora, Fo¨ hn, and Mistral; alpine lee cyclogenesis and MM5 (v3.7.4) are used in this study. The main dif- events and orographic precipitation enhancement) has ference between the two models is the dynamical core. attracted many researchers. In the case of the ARW-WRF the solution of governing The study area is shown in Fig. 1. The two-step nesting equations is approximated by using third-order Runge– approach is used to dynamically downscale the 40-yr Kutta split-explicit time integration, while MM5 uses a European Centre for Medium-Range Weather Forecasts first-order (time filtered) Leapfrog time integration. A few (ECMWF) Re-Analysis (ERA-40) dataset. The parent other important differences are as follows: MM5 (ARW- model domain 1 (D1) is a fine-resolution domain (30-km WRF) uses Arakawa B grid (Arakawa C grid); a different grid spacing) covering most parts of Europe, the Medi- vertical coordinate system, that is, terrain-following height terranean Sea, and some part of the Atlantic Ocean. D1 is coordinates (terrain-following hydrostatic pressure verti- providing the synoptic features and general circulation cal coordinates); different treatment of mass, momentum, patterns to the nested high-resolution (10-km grid spac- entropy, and scalars, that is, no conservation properties ing) domain 2 (D2). D2 covers the Alps and its sur- (conserves mass, momentum, entropy, and scalars using roundings, that is, southern parts of Germany and Czech flux form prognostic equations); and different advection Republic, eastern part of France, Switzerland, northern terms, that is, second-order-centered differencing for ad- part of Italy, northern part of Croatia, Slovenia, Austria, vection (fifth-order upwind or sixth-order-centered dif- northwestern region of Bosnia and Herzegovina, western ferencing for advection). The treatment of lower- and region of Hungary and Slovakia, along with some parts of upper-boundary conditions is also different. There are the Adriatic and Ligurian Sea. The fine- (high) resolution also differences due to some new improvements that are domain has 100 (79) grid points in the south–north di- implemented in ARW-WRF and are absent in MM5 like rection and 124 (109) grid points in the west–east direction. filters for external mode, damping of vertically propa- Both domains (D1 and D2) share the same set of physical gating acoustic modes, damping of anomalously large

Unauthenticated | Downloaded 10/11/21 03:12 AM UTC 3110 JOURNAL OF CLIMATE VOLUME 24

TABLE 1. The list of evaluated parameterization schemes and their short names.

Physical parameterizations Reference Short name Convection Grell–Devenyi ensemble Grell and Devenyi (2002) GD Grell (no shallow convection) Grell et al. (1995) GR Kain–Fritsch Kain (2004) KF Betts–Miller scheme Betts and Miller (1993) BM Betts–Miller–Janjic Janjic (1994) BMJ Microphysics Eta Ferrier Rogers et al. (2001) Ferrier ARW-WRF Single Moment 6-Class Hong and Lim (2006) WSM6 Thompson Thompson et al. (2004) Thompson Reisner 1 Reisner et al. (1998) Reisner 1 Reisner 2 Adapted from Reisner et al. (1998) Reisner 2 Radiation Rapid Radiative Transfer Model Mlawer et al. (1997) RRTM Goddard shortwave Schwarzkopf and Fels (1991) Goddard Dudhia shortwave Dudhia (1989) Dudhia Geophysical Fluid Dynamics Laboratory Lacis and Hansen (1974) GFDL Boundary Layer Monin–Obukhov scheme Skamarock et al. (2007) MOS Monin–Obukhov–Janjic scheme Janjic (2002) MOJ Yonsei University Hong et al. (2006) YSU Mellor–Yamada-Janjic Janjic (2002) MYJ Eta PBL Janjic (1990); Janjic (1994) Eta MRF PBL Hong and Pan (1996) MRF Land surface model Noah LSM Chen and Dudhia (2001) Noah vertical velocities, second-order horizontal mixing, and Services (EUMETNET) (E-Obs version 1.0). E-Obs is implementation of many new physical parameterization based on observations taken from 2317 meteorological schemes. stations from 41 different countries in Europe. This dataset has mean, maximum, and minimum temperature along b. Observational datasets with the sum of precipitation on a daily basis with a grid The scarcity of observed datasets is one of the biggest spacing of 25 km. It is a relatively new dataset and con- obstacles in model evaluation, particularly at high spa- siders area means instead of gridpoint values, which makes tial resolution. The Alps on the other hand are among it a better choice for evaluating RCM results. The tem- the world’s best documented regions. The presence of perature field of E-Obs has a higher resolution as com- many meteorological stations has made it possible to pared to other available observational datasets. create high-resolution observed gridded datasets. The evaluation of simulated high-resolution precipitation is based on an observational dataset provided by 4. Experiment design and evaluation methods the Swiss Federal Institute of Technology Zu¨ rich (re- a. Experiment design ferred to as ETH dataset) (Frei and Scha¨r 1998; Frei et al. 2006). ETH climatology data make use of daily rain The available options for microphysics, radiation, con- gauge observations (over 5000 observations per day) from vection, and planetary boundary layer in MM5 and ARW- the operational networks of all alpine countries. It has WRF models offer a broad range of configurations. In a horizontal resolution of 1/68 (approximately 20-km grid this study we have conducted a series of experiments. spacing) and is derived based on techniques described in A total of 13 (16) experiments with MM5 (ARW-WRF) Frei et al. (2006). using different available configurations is carried out. We The dataset used in this study for evaluating the sim- tested out available physical parameterizations and some ulated high-resolution 2-m air temperature T2m is the miscellaneous options like change in vertical resolution, European Climate Assessment and Dataset (ECA&D) feedback, horizontal diffusion, and damping. Table 1 gives (Haylock et al. 2008). This project is initiated by Euro- an overview of the physical parameterizations tested in pean Climate Support Network (ECSN) and currently this study, while Tables 2 and 3 provide the details of the supported by the Network of European Meteorological complete ensemble.

Unauthenticated | Downloaded 10/11/21 03:12 AM UTC 15 JUNE 2011 A W A N E T A L . 3111

TABLE 2. The MM5 ensemble members with key TABLE 3. The ARW-WRF ensemble members with key parameterization settings. parameterization settings.

Expt Physical parameterization settings Expt Physical parameterization settings RE KF, Reisner 1, RRTM, Eta PBL, RE GD, Ferrier, Goddard, RRTM, MOJ, NOAH LSM, shallow convection, NOAH, MYJ, vertical levels 5 30, vertical levels 5 30, SST and feedback SST and feedback on, pressure off, pressure at model top 5 100 mb at model top 5 50 mb HD Za¨ngl z diffusion PT Pressure at model top 5 100 mb CU1 BM CU1 KF CU2 GR (no shallow convection) CU2 BMJ SS MRF PBL MP WSM6 MP Reisner 2 DA Model ﬁlter: damping on VE1 Vertical levels 5 40 SW1 Dudhia VE2 Vertical levels 5 20 SW2 GFDL FB Feedback on SS MOS, YSU L2A Reisner 2, MRF PBL VE Vertical levels 5 20 L3A Reisner 2, MRF PBL, feedback on L2A BMJ, WSM6 L3B Reisner 2, MRF PBL, feedback on, L2B KF, MOS, YSU vertical levels 5 40 L2C KF, Dudhia L3C Reisner 2, MRF PBL, feedback on, L3A KF, MOS, YSU, Dudhia vertical levels 5 20 L3B KF, MOS, YSU, Dudhia, WSM6 L3C KF, MOS, YSU, Dudhia, Thompson

At first, a complete year is simulated, which is referred to as reference simulation (RE). The physical parame- have different errors in different regions. To address this terization chosen for MM5 RE is based on the first highly important question, the RCMs are operated at very high resolved (10 km 3 10 km grid spacing) regional climate resolution (10-km grid spacing) and a rigorous evalua- simulation for the alpine region by Gobiet et al. (2006), tion in nine climatological subregions of the alpine re- which was conducted during the Austrian climate region is carried out in addition to the D2 evaluation. search project ‘‘Research for Climate Protection – Model A clustering method is applied to the ETH daily pre- Run Evaluation’’ (reclip:more) Loibl et al. (2007). The cipitation dataset to find the climatological subregions in RE configuration for ARW-WRF is mostly based on the alpine region. The clustering technique uses Ward’s Done et al. (2005) and previous experiences with MM5. method and the K-mean method (Wilks 2006) to define To quantify the impact of each parameterization the groups (subregions) on the basis of data properties. The experiments are carried out at three levels. The level-1 subregional analysis is done on nine subregions (Fig. 2). (L1) simulations have only one modification compared The model and observed datasets do not have the to the RE configuration, while level-2 (L2) and level-3 same grid spacing, therefore to compare modeled and (L3) experiments represent two and more than two observed parameters we have used a refinement and modifications, respectively. resampling technique. The key advantages of this tech- In L1 a total of eight (nine) experiments is conducted nique are as follows: i) the features that are not resolved with MM5 (WRF). The analysis of these simulations laid by the coarser dataset do not disturb the evaluation and down the basis for further simulations. We combine the no artificial interpolation structures are introduced, and configurations that show better performance in the L1 ii) the area averages are left unchanged. The refinement simulations to do additional one (three) L2 experiments. and resampling technique as well as the clustering method These simulations provide us further insight and help us is explained in more detail in Suklitsch et al. (2008). in determining the necessary parameterization changes The model performance is based on deviations of T2m to acquire better results. Therefore, the additional three and precipitation from observations. The mean (sum) of

L3 experiments are carried out with both models. The hourly instantaneous (accumulated) T2m (precipitation) application of this three-level experimental approach is validated on daily (0000–2300 UTC), monthly, seasonal results in 29 ensemble members. [winter (DJF); spring, March–May (MAM); summer, (JJA); and autumn, September–November (SON)], and b. Evaluation methodology annual time scales. Area-average values for D2 and The complex terrain of the Alps gives rise to many com- subregions are compared with the E-Obs and ETH data, plex physical processes, and from a phenomenological respectively, and the results are quantiﬁed in terms of point of view it was anticipated that model results will biases, normalized centered root-mean-square errors,

Unauthenticated | Downloaded 10/11/21 03:12 AM UTC 3112 JOURNAL OF CLIMATE VOLUME 24

FIG. 2. (top) Observed annual mean precipitation (1979–99, based on ETH). (bottom) Ten climatological subregions (number of grid points are given in parentheses) found by applying a clustering method: 1) ‘‘Alps South’’ (133); 2) ‘‘Provence’’ (163); 3) ‘‘Alps North’’ (143); 4) ‘‘Slovenia’’ (82); 5) ‘‘Not in D2 therefore, it is excluded’’; 6) ‘‘Padan Plain’’ (172); 7) ‘‘Western Prealps’’ (160); 8) ‘‘Bourgogne’’ (380); 9) ‘‘Southern Germany’’ (208); and 10) ‘‘Alps East’’ (267). normalized standard deviations, and temporal correla- Figures 3a and 3c depict the area-averaged (D2) monthly tion to determine the strengths and weaknesses of each deviations of simulated T2m from E-Obs, for all ensemble ensemble member. members. The MM5 ensemble members (Fig. 3a) are cold biased with some better results during winter, while the ARW-WRF ensemble members (Fig. 3c) feature 5. Results warm biases during winter and varying results during the rest of the year, depending on model conﬁguration. For All results presented in this section refer to the high- MM5 RE, T is underestimated to a lesser extent during resolution simulations (10-km grid) in D2. 2m winter than during the rest of the year (annual mean bias is 21.768C, Fig. 3b). The ARW-WRF RE simulation on the a. Surface temperature other hand has a warm bias during winter and a slight cold Figure 3 provides an overview of differences resulting bias during summer (annual mean bias is 0.358C, Fig. 3d). from different model conﬁgurations in mean T2m. The spread between the coldest and warmest simulation

Unauthenticated | Downloaded 10/11/21 03:12 AM UTC 15 JUNE 2011 A W A N E T A L . 3113

FIG. 3. (a) MM5 and (c) ARW-WRF monthly mean biases in T2m. (b) MM5 and (d) ARW-WRF, Taylor plots (Taylor 2001) showing temporal correlation (arc), normalized centered RMSE (distance from the reference, and normalized standard deviations (on axes) calculated on daily basis after removing the annual cycle. The legend contains colors representing each ensemble member and numbers depicting the corresponding annual mean bias. implies that parameterizations have a stronger effect simulations CU1 and CU2 (a larger spread during sum- during summer than during winter (Figs. 3a,c). mer is seen). On the other hand, ARW-WRF ensemble

The area-averaged values of daily mean T2m are also has a minimum spread of approximately 1.58Candmax- evaluated in terms of temporal correlation, normalized imum is up to approximately 48C during summer. Figures standard deviation, and normalized centered root-mean- 3b and 3d also support these findings. square errors. Initial analysis revealed that T2m results have Figure 4 depicts the seasonal biases for each ensemble a very high correlation coefficient (.0.98) (not shown). member in each subregion for T2m. Columns represent Afterward, we removed the annual cycle by subtracting single ensemble members (starting from the RE simula- the daily mean climatology (1979–99) from all the ensem- tion) and rows represent different subregions. In the MM5 ble members as well as observations before the analysis. RE simulation, the T2m is underestimated throughout the The results are shown in Figs. 3b and 3d. One can see that year in all subregions, however, the results during DJF are the value of correlation coefficient has decreased sub- better than during the other seasons. The minimum bias stantially. The correlation coefficient is now between (20.248C) is in the region ‘‘Alps East’’ (DJF) and maxi- approximately 0.45 and 0.7. The normalized centered mum bias (22.738C) in ‘‘Bourgogne’’ (JJA). The cold bias root-mean-square values are between approximately 0.88 in MM5 is persistent in all L1, L2, and L3 simulations. The and 1.58C. Although variability in T2m is overestimated only significant mitigation is achieved by changing the Eta in both the models, it can be significantly improved by PBL scheme to Medium-Range Forecast (MRF; SS, L2A, choosing a suitable configuration. The annual mean bia- and L3C). The changes in microphysics from Reisner 1 ses in the MM5 (ARW-WRF) T2m ranges from 22.758 to explicit moisture scheme to Reisner 2 (MP), and changes 21.088C(21.128 to 1.338C). In MM5, T2m is affected less in the vertical resolution of the model from 30 to 40 (VE1) by changes in model parameterization than in ARW- and 20 (VE2) sigma levels also showed some minor im- WRF. From Figs. 3a and 3c one can see that for most of provements. Overall, the MM5 L3C simulation (Reisner MM5 ensemble the spread between ensemble members 2, MRF, 20 vertical levels) gave us the best results (annual is approximately 18–1.58C with the exception of two mean bias is 21.088C) for T2m (Fig. 3b).

Unauthenticated | Downloaded 10/11/21 03:12 AM UTC 3114 JOURNAL OF CLIMATE VOLUME 24

FIG. 4. The seasonal (top to bottom: DJF, MAM, JJA, SON) subregional biases of T2m (8C) for (left) MM5 and (right) ARW-WRF ensemble members.

Unauthenticated | Downloaded 10/11/21 03:12 AM UTC 15 JUNE 2011 A W A N E T A L . 3115

TABLE 4. The 5th (P5) and 95th (P95) percentiles of error of ensemble members in D2 and subregions (SR).

DJF MAM JJA SON Annual

Parameter Model P5 P95 P5 P95 P5 P95 P5 P95 P5 P95

T2m (8C) MM5 (D2) 21.29 20.09 22.95 21.51 24.00 21.41 22.68 21.23 22.74 21.06 (SR) 22.19 0.31 22.83 21.36 23.95 20.93 22.82 20.98 22.88 20.30 ARW-WRF (D2) 0.12 1.78 21.32 1.21 22.40 1.14 21.17 1.16 21.19 1.32 (SR) 20.40 2.41 21.41 1.19 22.17 1.02 21.22 1.14 21.59 1.71 Precip (mm day21) MM5 (D2) 0.39 0.79 0.38 0.96 0.06 0.92 20.23 0.59 0.26 0.80 (SR) 21.39 1.58 20.24 1.43 20.32 1.50 20.61 1.10 20.52 1.45 ARW-WRF (D2) 20.09 1.14 0.46 2.32 0.32 2.47 20.31 0.64 0.13 1.50 (SR) 21.12 2.60 0.16 3.63 20.10 4.00 20.86 1.23 20.67 3.29

The ARW-WRF RE T2m in DJF is overestimated in [L3B (L3A) is the favored setting for MM5 (ARW- all subregions (Fig. 4). During other seasons, positive and WRF), for more on favored settings please see section 7]. negative biases occur, depending on the subregion. The The MM5 (ARW-WRF) reference simulation has a cold largest cold bias is found in Bourgogne (JJA) and amounts (warm) bias and selecting a suitable parameterization to 21.688C, and the largest warm bias is in Alps East improves the bias in D2 area mean value up to approxi-

(2.418C). Significant decrease of T2m is observed when we mately 0.48C. The other notable result from this figure is switch from Goddard shortwave (SW) scheme to Dudhia that both models have a similar spatial pattern of bias. (SW1, L2C, L3A, L3B, L3C) and from Ferrier to ARW- b. Precipitation WRF Single-Moment 6-class (WSM6) microphysics scheme (MP, L2A, L3B). For example, the annual bias is Figures 6a and 6c depict the monthly biases of all en- changed from 0.358CinREto21.208C in SW1 and 20.17 semble members. In general, precipitation is overestimated in MP (Fig. 3d). In contrast, an increase in T2m is found in both models and the spread of deviations suggests that with changes in the cumulus scheme from the Grell– the models are less sensitive to parameterization change Devenyi ensemble (GD) to Kain–Fritsch (KF) and Betts– during winter than during summer, particularly ARW- Miller–Janjic (BMJ) (CU1, CU2, L2A, L2B, L2C, L3A, WRF. WRF exhibits a wet bias during spring and summer L3B, L3C). For instance, the annual mean bias is 1.338C in most simulations because of strong orographic convec- for CU1 and 1.018C for CU2 (Fig. 3d). Other configu- tion initiation. Generally, the MM5 results show less spread ration changes show only minor effects on T2m. Overall, within the ensemble and a smaller error range than ARW- the ARW-WRF L3A simulation [KF, Dudhia, Monin– WRF. The annual mean bias for the RE simulation is Obukhov Scheme (MOS), Yonsei University (YSU)] gave 0.59 (1.32) mm day21 for MM5 (ARW-WRF). The over- us the best results for T2m (annual mean bias for L3A is all annual mean bias for precipitation lies between 0.27 and 20.018C, Fig. 3d). 0.80 (0.13 and 1.51) mm day21 for MM5 (ARW-WRF). Table 4 illustrates 5th and 95th percentiles of the sea- The annual cycle was also removed from precipitation; sonal and annual errors as found in D2 and in the nine however, as expected this did not have much effect on subregions (error range). The annual error range in the correlation coefficient. The analysis is shown in Figs. 6b D2 (subregions) for MM5 is 22.748 (22.888)to21.068 and 6d. These figures show that a suitable selection of pa- (20.308)C and for ARW-WRF it is 21.198 (21.598)to rameterizations can reduce the overestimation of day-to- 1.328 (1.718)C. The maximum (minimum) seasonal D2 day variability. This is more clearly seen in ARW-WRF and subregional error range for MM5 is during autumn results. For example, the normalized standard deviation and summer (winter and spring) while for ARW-WRF it is reduced from approximately 1.5 to approximately 1.1. is during summer (autumn) season for both cases. In addition, these results demonstrate that improvements The results presented so far are based on the area mean in correlation can also be achieved. The correlation co- values. The defined subregions are small; however, in some efficient is improved from approximately 0.7 to approxi- cases the area mean values could be misleading owing to mately 0.8. the cancellation effect of positive and negative biases in The subregional analysis (Fig. 7) reveals regional the same subregion. Therefore, it is important to take into differences in precipitation bias. Both models tend to account the spatial distribution of these biases. Figure 5 overestimate precipitation in the mountainous regions, shows the spatial distribution of annual T2m biases from particularly during spring, summer, and autumn (‘‘Western E-Obs dataset. Figures 5a–c are showing MM5’s refer- Prealps,’’ ‘‘Alps South,’’ ‘‘Alps North,’’ ‘‘Alps East’’) and ence and favored simulations along with the ensemble underestimate precipitation in the southeastern part of mean; Figs. 5d–f are showing the same for ARW-WRF the domain (‘‘Slovenia’’) during winter. In comparison

Unauthenticated | Downloaded 10/11/21 03:12 AM UTC 3116 JOURNAL OF CLIMATE VOLUME 24

FIG. 5. (a)–(c) The spatial distribution of MM5 annual mean T2m deviations from E-Obs in (a) RE, (b) favored, and (c) ensemble mean. (d)–(f) As in (a)–(c), but for the ARW-WRF. to ARW-WRF the biases are smaller in the MM5 en- during winter in ‘‘Alps North,’’ while the maximum semble. MM5 simulations also show less effect of com- ARW-WRF bias is 4.95 mm day21 during summer in the plex topography than ARW-WRF (Figs. 7 and 8). The same region. This wet-bias during summer is clearly dom- maximum bias in the MM5 ensemble is 2.11 mm day21 inant in ARW-WRF simulations. Further investigation

21 FIG. 6. As in Fig. 3, but for precipitation (mm day ).

Unauthenticated | Downloaded 10/11/21 03:12 AM UTC 15 JUNE 2011 A W A N E T A L . 3117

21 FIG. 7. As in Fig. 4, but for precipitation (mm day ).

Unauthenticated | Downloaded 10/11/21 03:12 AM UTC 3118 JOURNAL OF CLIMATE VOLUME 24

FIG. 8. As in Fig. 5, but for precipitation. of this issue revealed that the frequency of precipitation improvement during spring and summer, while the L3 is overestimated during spring and summer (Fig. 9). simulation L3B (0.27 mm day21) (Reisner 2, MRF, While most of the configurations have not been able to feedback on) show improvements throughout the year. resolve this problem, the L3A configuration has shown The ARW-WRF RE simulation (1.32 mm day21) could significant improvements in this regard. be significantly improved by changing the boundary The analysis of the L1 MM5 simulations (Fig. 6) shows layer scheme from MYJ to YSU–MOS (simulation SS, that changing the convection scheme from KF (RE, 0.93 mm day21) and by switching from the Goddard ra- 0.59 mm day21) to BM (CU1, 0.43 mm day21)andGrell diation scheme to the Dudhia scheme (simulation SW1, (GR) (CU2, 0.44 mm day21) tends to slightly improve 0.63 mm day21) (Fig. 6). Also changing the GD convec- precipitation results, however, both convection schemes tions scheme to KF showed some improvements. In level 2, 21 significantly increase the T2m bias. Therefore, they are the combinations KF, YSU–MOS (L2B, 0.34 mm day ) not regarded in the L2 and L3 simulations. The use of and KF, and Dudhia (L2C, 0.42 mm day21) performed Reisner 2 microphysics (simulation MP, 0.44 mm day21) better; in L3 the combination of L2 configurations KF, improves the winter and autumn precipitation results. Dudhia, and YSU–MOS (simulation L3A, 0.13 mm day21) Simulation L2A (0.38 mm day21) (Reisner 2 combined showed the best overall performance. However, the better with MRF planetary boundary layer scheme) shows some performing combination L3B (L3A) of MM5 (ARW-WRF)

FIG. 9. The errors in the simulated frequency of precipitation for (left) the MM5 and (right) the ARW-WRF.

Unauthenticated | Downloaded 10/11/21 03:12 AM UTC 15 JUNE 2011 A W A N E T A L . 3119 is not the best performing configuration in all subregions processes (e.g., convection), shows a higher sensitivity to and during all seasons (Fig. 7). changes in parameterization than during winter. De´que´ In Table 4 the bottom four rows show the error range et al. (2007) found that during summer the component of of the D2 and subregional precipitation. The annual uncertainty arising from RCM is the most dominant and error range in the D2 (subregions) for MM5 is 0.26 could be as large as GCM component. (20.52) to 0.80 (1.45) mm day21 and for ARW-WRF it The subregional analysis points out that within the 21 is 0.13 (20.67) to 1.50 (3.29) mm day . The maximum ensemble both models have maximum T2m biases in the (minimum) seasonal D2 and subregional error range for same region, that is, a cold bias in Bourgogne (during MM5 is during summer and winter (winter and spring), summer) and a warm bias in Alps East (during winter) while for ARW-WRF it is during summer (autumn). (Fig. 4). In case of precipitation there is an overestimation The spatial distribution of precipitation bias is shown in in the inner alpine region (Figs. 7 and 8) particularly in Fig. 8. The top panels from left to right show the MM5’s ARW-WRF. The overestimation of winter precipitation reference and favored simulations along with the ensem- in the alpine region can partly, but not entirely, be related ble mean; the bottom panels show the same for ARW- to systematic measurement biases in the observations WRF. In both models the wet bias can be reduced (D2 (see discussion in Frei et al. 2003). This wet bias might be mean bias is improved up to 0.4 and 1.2 mm day21 for considered as the model’s inadequate response to orog- MM5 and ARW-WRF, respectively); however, there is raphy, which is similar to the findings of Done et al. (2004). relatively smaller improvement in the inner alpine region. They found that ARW-WRF climate mode simulation with a grid spacing of 30 km during the winter season over the western United States showed an enhancement 6. Discussion of precipitation over the Rocky Mountains. Physical parameterization schemes are nonlinearly We also investigated the overestimation of ARW-WRF interacting with each other: The shortwave radiation com- summer season precipitation. We analyzed the frequency ing from the sun reaches the ground and is used by the and intensity of precipitation. We found that although in- surface scheme to modify soil temperature. The surface tensity was affected (not shown) the main problem was scheme also modifies sensible and latent heat fluxes, frequency that was significantly overestimated, that is, hence changing the stability of the PBL. This information convection is initiated too often resulting in a wet bias is used by the cumulus scheme to form the cumulus clouds (Fig. 9). In this context, the KF scheme trigger function and precipitation, which feed back the soil moisture. The better handles atmospheric instability than GR or BMJ. cumulus scheme also modifies the available moisture con- The use of KF (CU1, L2B, L2C, L3A, L3B, L3C) helped tent in the atmosphere, which affects the microphysics to reduce this bias significantly. The change of radiation scheme that determines the liquid and solid atmospheric scheme from Goddard to Dudhia (SW1, L3A, L3B, water content, resultantly affecting the shortwave/longwave L3C) and PBL scheme from MYJ to YSU (SS, L2B, radiation scheme. In addition, parameterization schemes L3A, L3B, L3C) also helped to reduce this error. Jankov interact with the dynamical core of the model. These com- et al. (2005) have also attributed that BMJ often over- plex relationships make the interpretation of model de- predicts areas of light precipitation and the biases as- ficiencies very challenging. sociated with the use of BMJ are much higher than with Owing to a lack of observational data, a complete di- the KF (based on eight different short-term weather agnosis is out of scope here. However, our evaluation based events during May and June on a 12-km grid). They also on gridded observational datasets for temperature and conclude that cumulus parameterizations have most effect precipitation shows a broad spectrum of error characteris- on their results, while PBL schemes show less sensitivity tics on subregional and subannual scales, thus provid- and microphysics least sensitivity. Fernańdez et al. (2007) ing some insight into the behavior of parameterization found that for MM5 excessive summer precipitation in the schemes as well as their suitability for high-resolution mountainous region was reduced by changing the radia- climate simulations. tion scheme to the Community Climate Model (CCM2). In our results the precipitation during winter season is Our results suggest that in summer cumulus, micro- less affected by changes in parameterizations than during physics schemes play a vital role, but PBL and shortwave summer, particularly in the case of ARW-WRF (Figs. 6a,c). radiations parameterizations can significantly affect the This behavior can be explained by the fact that during model performance during all seasons. For example, in summer the small-scale processes have large influence as MM5, the change in the PBL scheme to MRF (SS) af- compared to during winter because the large-scale forc- fected the T2m significantly and switching to Reisner 2 ing becomes weak. Therefore, during summer a parame- (MP) improved both T2m and precipitation results. In ter such as precipitation, which is more affected by local ARW-WRF, the change in PBL scheme to YSU–MOS

Unauthenticated | Downloaded 10/11/21 03:12 AM UTC 3120 JOURNAL OF CLIMATE VOLUME 24

(SS) improved the precipitation pattern during summer representative owing to climate variability (for details (not shown), and the use of the Dudhia shortwave scheme please see section 2.3 in Suklitsch et al. 2010). Another (SW1) instead of Goddard signiﬁcantly inﬂuences the shortcoming of this short simulation period is that slow

T2m and precipitation during all seasons. processes are not captured at this time scale. For instance, The subregional analysis is also helpful in providing a more time is required for soil moisture to reach equilib- comparison of L1 simulations with L2 and L3 simulations. rium, and model performance may vary once equilibrium For instance, from Fig. 5 one can quantify the individual as is reached. However, all simulations were initialized with well as combined effect of different physical parameteri- the same soil moisture ﬁeld, and we have not found any zations. For example, in case of ARW-WRF simulations major discrepancies in our results, for example, no large

CU1 and SW1 the mean bias for SON is 1.168 and 21.018C, biases in T2m or precipitation are found. It is also worth respectively. The L2C simulation, which is a combination mentioning that the Noah land surface model (LSM) was of these two L1 simulations, has a mean bias of 20.818C. used for all simulations, and we have not tested any other This shows that the simulations carried out in L2 do not land surface model. Therefore, the error characteristics correspond to a linear combination of L1 simulations. reported in this study might differ when another land This is true for L3 simulations as well. A similar argument surface model is applied. It should also be considered is also valid for precipitation results. The subregional that in all these simulations the initial and boundary analysis also revealed a deficiency in both models: Winter conditions are provided by a reanalysis dataset. There- precipitation is systematically underestimated in Slovenia. fore, while referring to our results, errors arising from This might indicate a misleading representation of the use of another driving dataset should be considered. For lee cyclogenesis over the Mediterranean Sea, which nota- example, for simulations driven with GCM dataset one bly affects the climate in southern and southeastern parts has to consider the GCM errors as well as the RCM re- of the alpine region Scha¨r et al. (1998). However, a de- sponse to these errors. Therefore, for such applications it tailed analysis of this effect is out of the scope of this study. must be considered that RCMs are not able to correct the Table 4 shows that in both models the winter error range errors in the LBCs. The errors in the LBCs will be in- for precipitation and temperature is higher for subregions herited into the RCM simulation and might dominate the than for D2, while during the other seasons, error ranges entire error budget. are higher for precipitation only. For instance, for temperature the error range does not increase significantly 7. Conclusions on small scales except during winter. These results are in agreement with the findings of Suklitsch et al. (2010). It is evident from our analysis that a significant part of Though obvious improvements can be achieved by a the model error arises from the application of physical suitable selection of parameterization schemes (see sec- parameterizations. However, a careful selection of pa- tion 7), it is notable that no parameterization setting has rameterizations can help to reduce this error for the re- performed ideally for both variables in all subregions and gion of interest. We have seen that an ensemble technique during all seasons. These results are also in accordance can help to achieve this goal. Although no combination with Fernańdez et al. (2007), who conducted sixteen 5-yr has performed ideally during all seasons and for all sub- simulations with MM5 on a 30-km grid over the Iberian regions the favored configurations L3B and L3A (MM5, Peninsula. It is also worth mentioning that this improve- ARW-WRF) have shown significant reductions in errors ment in results is not a consequence of pure model tuning for T2m and precipitation. Annual D2 mean biases for since we have used the standard implementations of T2m in both models are improved up to approximately physical parameterizations and did not vary single em- 0.48C. The precipitation errors are reduced up to 0.4 pirical parameters to minimize errors. We have seen that and 1.2 mm day21 for MM5 and ARW-WRF, respec- when the appropriate level of complexity of a physical tively. However, some biases remain, that is, for MM5 process is adequately represented in a physical param- (ARW-WRF) 21.418C and 0.27 mm day21 (20.018Cand eterization it gives better results. For example, we have 0.13 mm day21), along with notable standard deviations. seen that in our region low level control convective We found that both models are sensitive to changes in scheme (such as KF) is performing better than a deep- parameterizations. We cannot conclude that precipitation layer control convective scheme (BM and BMJ). is more sensitive than T2m (Figs. 3 and 6). However, we The authors would like to point some limitations of have seen larger influence on model performance during this study. As a precautionary reminder we suggest the summer than during winter. The T2m for both models reader keep in mind that our results are based on 1-yr shows a higher spread during summer than during win- simulations only and, although they include a number of ter. Precipitation also shows higher spread during summer meteorological conditions, they might not be entirely than during winter; however, for MM5 it is relatively less

Unauthenticated | Downloaded 10/11/21 03:12 AM UTC 15 JUNE 2011 A W A N E T A L . 3121 prominent (Figs. 3 and 6a,c). These findings are similar to any certain advantages and disadvantages of the ARW- De´que´ et al. (2007) and are also supported by theoretical WRF over the MM5. We have attempted to use regional understanding, that is, during summer the influence of the and subregional analysis to provide a complete overview large scale diminishes and the role of local-scale pro- on strengths and weaknesses of these models on daily, cesses increases. It should also be mentioned that T2m and monthly, seasonal, and annual basis. In general, we have precipitation error range during winter is larger in smaller seen that the ARW-WRF is more sensitive to changes in subregions than in the entire study region (D2), while parameterization in comparison with the MM5, that is, during the other seasons the subregional error ranges the error range for the ARW-WRF ensemble is broader are higher only for precipitation but not for temperature than for the MM5. Although potential benefit of the (Table 4). ARW-WRF over the MM5 in the case of precipitation Furthermore, we found that cumulus, PBL, and radi- depending on the region of interest is arguable (Fig. 8), ation parameterizations have the dominant impact on for temperature ARW-WRF has certainly performed our results, while microphysics and vertical resolution have better (Fig. 5). shown some improvements. Compared with numerous Future work should include long-term high-resolution short-term experiments mostly conducted in NWP (see climate simulations, which should be carried out to include section 1), we have found partial agreement. Jankov the effect of long-term processes. Further investigations et al. (2005) report that warm season rainfall showed the focusing on process-oriented analysis to relate the model’s highest sensitivity to cumulus parameterizations, less sen- shortcomings to important processes are also required. sitivity to PBL schemes, and the least sensitivity to microphysics schemes. Our results based on 1-yr-long simulations also show that the cumulus and microphysics schemes Acknowledgments. We thank the two anonymous re- have dominant impacts during the warm season (MAM, viewers for their valuable suggestions and comments. JJA). However, in addition radiation and PBL schemes Their help substantially improved the quality of our work. affect the model performance during summer as well as This study is supported by Austrian Science Fund (FWF), during other seasons; Done et al. (2005) report that ARW- under project ‘‘NHCM-1’’ (non hydrostatic climate mod- ¨ WRF overestimates precipitation over the Rocky Moun- eling, ID P19619), Austrian Exchange Service (OAD), and tains during the winter. We also found out that ARW-WRF Higher Education Commission (HEC) of Pakistan. We showed a considerable wet bias in the mountainous region, acknowledge the E-OBS dataset from the EU-FP6 project during winter as well as during summer (Fig. 7). However, ENSEMBLES (http://ensembles-eu.metoffice.com) and we saw improvements when we changed the cumulus, ra- the data providers in the ECA&D project (http://eca. diation, and PBL schemes. Fernańdez et al. (2007) also knmi.nl). The authors also acknowledge the Swiss Fed- report that for MM5 excessive summer precipitation over eral Institute of Technology Zu¨rich (ETH, Zu¨rich, Swit- the mountainous region was reduced by changing the ra- zerland). We are thankful to ECMWF for providing the diation scheme. All these findings highlight the vital im- ERA-40 dataset and computational resources for this portance of realistic parameterizations of subgrid scale study. We also thank the University of Graz for pro- processes; for example, the behavior of the model errors viding computational facilities vital for completion of when the PBL schemes are changed indicates a notable this research. potential to improve the orographic-induced precipitation. Therefore, for high-resolution climate simulations cumulus, PBL, and radiation schemes should be chosen care- REFERENCES fully. It should also be mentioned that we have only used Betts, A., and M. J. Miller, 1993: The Betts-Miller scheme. The the Noah LSM for all simulations. Therefore, these results Representation of Cumulus Convection in Numerical Models, do not imply that LSMs are of less importance. It should Meteor. Monogr., No. 46, Amer. Meteor. Soc., 107–121. be noted that the use of another LSM might change the Boehm,U.,M.Kucken,W.Ahrens,A.Block,D.Hauffe,K.Keuler, error characteristics reported in this study. B. Rockel, and A. Will, 2006: CLM—The Climate Version of LM: Brief description and long-term applications. COSMO Another important outcome of this work is a com- Newsletter, Vol. 6, 225–235. prehensive assessment and comparison of the MM5 and Chen, F., and J. Dudhia, 2001: Coupling an advanced land surface– ARW-WRF models in terms of model performance in hydrology model with the Penn State NCAR MM5 modeling the Alps. In general, the ARW-WRF, which is the new system. Part I: Model implementation and sensitivity. Mon. state-of-the-art nonhydrostatic atmospheric model, is Wea. Rev., 129, 569–585. Christensen, J., T. Carter, and M. Rummukainen, 2007: Evaluating based on the MM5 model, and the main difference be- the performance and utility of regional climate models: The tween the two models is the dynamical core. Previous PRUDENCE project. Climatic Change, 81, 1–6, doi:10.1007/ work based on small case studies is inconclusive about s10584-006-9211-6.

Unauthenticated | Downloaded 10/11/21 03:12 AM UTC 3122 JOURNAL OF CLIMATE VOLUME 24

De´que´, M., and Coauthors, 2007: Intercomparison of regional cli- ——, and J.-O. J. Lim, 2006: The WRF single-moment 6-class mate simulations for Europe: Assessing uncertainties in model microphysics scheme (WSM6). J. Korean Meteor. Soc., 42, projections. Climatic Change, 81, 53–70, doi:10.1007/s10584- 129–151. 006-9228-x. ——, Y. Noh, and J. Dudhia, 2006: A new vertical diffusion Done, J., L. Leung, C. Davis, and B. Kuo, 2004: Regional climate package with an explicit treatment of entrainment processes. simulations using the WRF model. Proc. Fifth WRF/14th MM5 Mon. Wea. Rev., 134, 2318–2341. User’s Workshop, Boulder, CO, National Center for Atmospheric Jacob, D., and Coauthors, 2007: Intercomparison of regional cli- Research, P8. [Available online at http://www.mmm.ucar.edu/ mate models for Europe: Model performance in present-day mm5/workshop/ws04/PosterSession/Done.James.pdf.] climate. Climatic Change, 81, 31–52, doi:10.1007/s10584-006- ——, ——, ——, and ——, 2005: Simulation of warm season 9213-4. rainfall using WRF regional climate model. Proc. Sixth WRF/ Janjic, Z. I., 1990: The step-mountain coordinate: Physical package. 15th MM5 User’s Workshop, Boulder, CO, National Center Mon. Wea. Rev., 118, 1429–1443. for Atmospheric Research, 9.2. [Available online at http:// ——, 1994: The step-mountain Eta coordinate model: Further www.mmm.ucar.edu/wrf/users/workshops/WS2005/abstracts/ developments of the convection, viscous sublayer, and turbu- Session9/2-Done.pdf.] lence closure schemes. Mon. Wea. Rev., 122, 927–945. Dudhia, J., 1989: Numerical study of convection observed dur- ——, 2002: Nonsingular implementation of the Mellor-Yamada ing the winter monsoon experiment using a mesoscale two– level 2.5 scheme in the NCEP Meso model. NCEP Office Note dimensional model. J. Atmos. Sci., 46, 3077–3107. 37, 61 pp. ——, 1993: A nonhydrostatic version of the Penn State/NCAR Jankov, I., W. A. Gallus Jr., M. Segal, and S. E. Koch, 2005: The mesoscale model: Validation tests and simulation of an At- impact of different WRF model physical parameterizations lantic cyclone and cold front. Mon. Wea. Rev., 121, 1493– and their interactions on warm season MCS rainfall. Wea. 1513. Forecasting, 20, 1048–1060. Fernańdez, J., J. P. Monta´vez, J. Saénz, J. F. Gonza´lez-Rouco, and Kain, J. S., 2004: The Kain–Fritsch convective parameterization: E. Zorita, 2007: Sensitivity of the MM5 mesoscale model An update. J. Appl. Meteor., 43, 170–181. to physical parameterizations for regional climate studies: Kapos, V., J. Rhind, M. Edwards, C. Ravilious, and M. Price, 2000: Annual cycle. J. Geophys. Res., 112, D04101, doi:10.1029/ Developing a map of the world’s mountain forests. Forests in 2005JD006649. a Sustainable Mountain Environment, Springer 521–529 pp. Ferretti, R., T. Paolucci, W. Zheng, G. Visconti, and P. Bonelli, Kotroni, V., and K. Lagouvardos, 2001: Precipitation forecast skill 2000: Analyses of the precipitation pattern on the Alpine re- of differrent convective microphysical schemes: Application for gion using different cumulus convection parameterizations. the cold season over Greece. Geophys. Res. Lett., 28, 1977–1980. J. Appl. Meteor., 39, 182–200. Lacis, A., and J. E. Hansen, 1974: A parameterization for the ab- Frei, C., and C. Scha¨r, 1998: A precipitation climatology of the Alps sorption of solar radiation in the earth’s atmosphere. J. Atmos. from high-resolution rain gauge observations. Int. J. Climatol., Sci., 31, 118–133. 18, 873–900. Loibl, W., A. Beck, M. Dorninger, H. Formayer, A. Gobiet, and ——, J. Christensen, M. Dˇ eque´, D. Jacob, R. Jones, and P. Vidale, W. Scho¨ ner, 2007: Kwiss-program reclip: More research for 2003: Daily precipitation statistics in regional climate models: climate protection—Model run evaluation. ARC Tech. Rep., Evaluation and intercomparison for the European Alps. 107 pp. J. Geophys. Res., 108, 4124, doi:10.1029/2002JD002287. Mlawer, E., S. Taubsman, P. Brown, M. Iacono, and S. Clough, ——, R. Scho¨ ll, S. Fukutome, J. Schmidli, and P. Vidale, 2006: 1997: Radiative transfer for inhomogeneous atmosphere: Future change of precipitation extremes in Europe: An in- RRTM, a validated correlated k model for the longwave. tercomparison of scenarios from regional climate models. J. Geophys. Res., 102, 16 663–16 682. J. Geophys. Res., 111, D06105, doi:10.1029/2005JD005965. Molteni, F., R. Buizza, T. Palmer, and T. Petroliagis, 1996: The Gobiet, A., H. Truhetz, and A. Riegler, 2006: A climate scenario ECMWF ensemble prediction system: Methodology and val- for the Alpine region, reclip:more project year 3, WegCenter idation. Quart. J. Roy. Meteor. Soc., 122, 73–119. progress report. Wegener Center for Climate and Global Change Murphy, J. M., 1993: What is a good forecast? An essay on the nature Tech. Rep. 23 pp. of goodness in weather forecasting. Wea. Forecasting, 8, 281–293. Grell, G. A., and D. Devenyi, 2002: A generalized approach to Orlanski, I., 1975: Rational subdivision of scales for atmospheric parameterization convection combining ensemble and data processes. Bull. Amer. Meteor. Soc., 56, 527–530. assimilation techniques. Geophys.Res.Lett.,29, 1693, doi:10.1029/ Otkin, J. A., and T. J. Greenwald, 2008: Comparison of WRF 2002GL015311. model-simulated and MODIS-derived cloud data. Mon. Wea. ——, J. Dudhia, and D. Stauffer, 1995: A description of fifth– Rev., 136, 1957–1969. generation Penn State/NCAR mesoscale model (MM5). NCAR Pan, Z., E. Takle, M. Segal, and R. Turner, 1996: Influences of Tech.Rep.NCAR/TN-398+STR,122pp. parameterization schemes on the response of rainfall to soil Haylock, M., N. Hofstra, A. K. Tank, E. Klok, P. Jones, and moisture in the central United States. Mon. Wea. Rev., 124, M. New, 2008: A European daily high-resolution gridded data 1786–1802. set of surface temperature and precipitation for 1950–2006. Reisner, J., R. J. Rasmussen, and R. T. Bruintjes, 1998: Explicit J. Geophys. Res., 113, D20119, doi:10.1029/2008JD010201. forecasting of supercooled liquid water in winter storms using Hewitt, C., and D. J. Griggs, 2004: Ensembles-based predictions of the MM5 mesoscale model. Quart. J. Roy. Meteor. Soc., 124, climate changes and their impacts. Eos, Trans. Amer. Geo- 1071–1107. phys. Union, 85, 52, doi:10.1029/2004EO520005. Rogers, E., T. Black, B. Ferrier, Y. Lin, D. Parrish, and G. D. Hong, S.-Y., and H. Pan, 1996: Nonlocal boundary layer vertical Mego, 2001: Changes to NCEP meso Eta analysis and forecast diffusion in a medium–range forecast model. Mon. Wea. Rev., system: Increase in resolution, new cloud microphysics, mod- 124, 2322–2339. ified precipitation assimilation, modified 3DVAR analysis.

Unauthenticated | Downloaded 10/11/21 03:12 AM UTC 15 JUNE 2011 A W A N E T A L . 3123

NWS Technical Procedures Bulletin, No. 479, National Weather ——, ——, H. Truhetz, N. K. Awan, H. Go¨ ttel, and D. Jacob, 2010: Service. [Available at http://www.emc.ncep.noaa.gov/mmb/ Error characteristics of high resolution regional climate mmbpll/spring2001/tpb/.] models over the Alpine area. Climate Dyn., doi:10.1007/ Scha¨r, C., T. Davies, C. Frei, H. Wanner, M. Widmann, M. Wild, s00382-010-0848-5. and H. Davies, 1998: Current Alpine climate. Views from the Taylor, K., 2001: Summarizing multiple aspects of model perfor- Alps: Regional Perspectives on Climate Change, P. Cebon mance in a single diagram. J. Geophys. Res., 106, 7183–7192. et al., Eds., The MIT Press, 21–63. Thompson, G. R., M. Rasmussen, and K. Manning, 2004: Explicit Schwarzkopf, M., and S. B. Fels, 1991: The simpliﬁed exchange forecasts of winter precipitation using an improved bulk mi- method revisited: An accurate, rapid method for computation of crophysics scheme. Part I: Description and sensitivity analysis. infrared cooling rates and ﬂuxes. J. Geophys. Res., 96, 9075–9096. Mon. Wea. Rev., 132, 519–542. Skamarock, W. C., J. B. Klemp, J. Dudhia, D. O. Gill, D. Barker, Uppala, S. M., and Coauthors, 2005: The ERA-40 Re-Analysis. W. Wang, and J. Powers, 2007: A description of the advanced Quart. J. Roy. Meteor. Soc., 131, 2961–3012. research WRF version 2. NCAR Tech. Note NCAR/TN- Wilks, D., 2006: Statistical Methods in the Atmospheric Sciences.In- 4681STR, 88 pp. ternational Geophysics Series, Vol. 91. Academic Press, 648 pp. Suklitsch, M., A. Gobiet, and C. Frei, 2008: High-resolution sen- Zhu, J., and X.-Z. Liang, 2007: Regional climate model simulations sitivity studies with the regional climate model CCLM in the of U.S. precipitation and surface temperature during 1982– Alpine region. Meteor. Z., 17, 467–476. 2002: Interannual variation. J. Climate, 20, 218–232.

Unauthenticated | Downloaded 10/11/21 03:12 AM UTC