HRVATSKI METEOROLO©KI »ASOPIS CROATIAN … · (odgovara spektru terena) te pokazuje malu sezonsku varijabilnost. Spektralna dekompozicija mjerenih i modeliranih podataka u frekvencijskoj

Hrvatsko meteoroloπko druπtvoCroatian Meteorological Society

HRVATSKI METEOROLO©KI »ASOPISCROATIAN METEOROLOGICAL JOURNAL

50 Hrv. meteor. Ëasopis Vol. 50 p. 1-144 ZAGREB 2015

»asopis je izdan povodom 20. godišnjice ALADIN projekta u Hrvatskoj

HMC-50.qxp_HMC-50 10/03/16 09:49 Page 1

Hrvatski meteoroloπki Ëasopis � Croatian Meteorological Journal, 50, 2015., 105∑120

Original scientific paperIzvorni znanstveni rad

VERIFICATION OF THE OPERATIONAL 10 M WIND FORECAST OBTAINED WITH THE ALADIN MESOSCALE

NUMERICAL WEATHER PREDICTION MODEL

Verifikacija operativne prognoze vjetra na 10 m visine dobivena ALADIN mezoskalnim numeričkim modelom prognoze vremena

MARIO HRASTINSKI, KRISTIAN HORVATH, IRIS ODAK PLENKOVIĆ,STJEPAN IVATEK-ŠAHDAN, ALICA BAJIĆ

Meteorological and Hydrological Service, Grič 3, 10000 Zagreb, CroatiaDržavni hidrometeorološki zavod, Grič 3, 10000 Zagreb, Hrvatska

[email protected]

Recieved 22 May 2015, in final form 17 July 2015

Abstract: This paper presents the results of the verification of operational 10 m wind forecastobtained with the ALADIN mesoscale numerical weather prediction model. In the period2010-2012 ALADIN/ALARO 8 km forecasts were initialized daily at 00 UTC and driven withthe ARPEGE global model forecasts through the 72-hourly forecasting range. Obtained fore-casts were further refined to 2 km grid spacing, using the simplified and cost-effective dynami-cal adaptation method (ALADIN/DADA 2 km forecasts). Since the primary objective of thisstudy is to assess the efficiency of wind forecast in regions of complex terrain as well as highwind energy potential, eight stations from different wind climate regions of the eastern Adriat-ic coast were selected to perform the verification procedure. Based on variety of statistical andspectral scores, it is suggested that the wind forecast generally improves with the increase ofhorizontal resolution. At bora dominated stations, the multiplicative mean systematic error isreduced by more than 50%. The largest portion of root-mean square errors can be attributedto dispersion or phase errors at majority of stations and their contribution increases with mod-el horizontal resolution. Spectral analysis in the wavenumber domain suggests that the slope ofkinetic energy spectra of both models decreases from k-3 in the upper troposphere towards ~ k-5/3 near the surface (corresponding to orography spectra) and shows minor seasonal variabil-ity. Spectral decomposition of measured and modeled data in the frequency domain indicatesa significant improvement in simulating the primary and secondary maximum of spectral pow-er (related to synoptic and diurnal motions) by using the ALADIN/DADA 2 km model, espe-cially for the cross-mountain wind component mostly related to strong and gusty bora flows.Finally, the common feature of both models is a significant underestimation of motions atscales below semi-diurnal, which is a result of their absence in initial conditions and of limitedmodel ability to represent small-scale processes.

Key words: wind forecast, complex terrain, ALADIN, statistical verification, spectral verification

Sažetak: U radu su dani rezultati verifikacije operativne prognoze vjetra na 10 m visine dobive-ni ALADIN mezoskalnim numeričkim modelom. U razdoblju 2010.-2012. ALADIN/ALARO8 km prognoze svakodnevno su inicijalizirane u 00 UTC, a pokrenute su prognozama global-nog modela ARPEGE kroz 72-satno prognostičko razdoblje. Dobivene prognoze dalje su pro-finjene na 2 km mrežu točaka korištenjem pojednostavljene i računski efikasne metode dina-mičke adaptacije (ALADIN/DADA 2 km prognoze). Budući da je glavna svrha ovog istraži-vanja procjena uspješnosti prognoze vjetra u područjima složenog terena, ali i visokog potenci-jala raspoložive energije vjetra, za potrebe provedbe verifikacije odabrano je osam postaja sistočne obale Jadrana koje predstavljaju područja različite klime vjetra. Analiza više statisti-čkih i spektralnih mjera točnosti sugerira da se prognoza vjetra općenito popravlja s poveća-njem horizontalne rezolucije. Na postajama izloženim buri srednja multiplikativna sistematska

HMC-50.qxp_HMC-50 10/03/16 09:50 Page 105

1. INTRODUCTION

A common method used for the wind fore-casting in the Atmospheric Boundary Layer(ABL) in complex terrain is dynamical down-scaling with the use of mesoscale NWP mod-els. Typical horizontal grid spacing of thesemodels is below 10 km, with the option of fur-ther refinement depending on the spatial andtemporal scales of observed phenomena in thetarget area. Generally, downscaling intro-duces new scales, both temporal and spatial,which results in better representation ofmesoscale wind systems due to common ac-tion of model physics and dynamics. Predica-ble mesoscale wind systems are most often aresult of surface and terrain inhomogeneities(e.g. sea-land breeze or valley winds) or inter-action between the terrain and large-scaleflow (e.g. downslope windstorms, gravitywaves and gap flows) (Pielke, 2002).

Due to its strength and frequency, bora (e.g.Smith 1985; Grisogono and Belušić 2009;Belušić et al. 2013) and sirocco (Jurčec et al.1996) are particularly important for forecast-ing wind conditions in Croatia. As they usuallyappear within the mesoscale cyclonic systemsover the Mediterranean (Horvath et al. 2008),the desired property of mesoscale modelsused for the refinement of wind forecastshould be the ability to simulate nonlinearflows, as well as flows over mountains for dif-ferent types of background flow. As regards tomany potential applications of high-resolutionwind forecasting, such as in transport, energy,tourism, agriculture etc., forecast errors mayprofoundly affect all of those applications, butperhaps the most sensitive is wind energy sec-

tor since the wind power error is proportionalto the cube of wind speed error. Therefore, weaim to quantify and distinguish among differ-ent sources of model errors and to comparethe results of different verification techniquesin order to help in further development ofhigh resolution mesoscale NWP modelling.

Verification of mesoscale flows is a challeng-ing task for which adequate approaches stillneed to be revised and defined. Traditionallyused statistical scores (e.g. multiplicativemean systematic error (MBIAS; the ratio ofmean modeled and observed wind speed),root-mean-square error (RMSE) and meanabsolute error (MAE)) seem to be insufficientfor that purpose since small spatial and tem-poral errors of generally well simulated phe-nomena can profoundly change the verifica-tion results (Mass et al. 2002; Rife et al. 2004).Therefore, we also utilized spectral analysis inwavenumber and frequency domains as a sup-plementary verification method to provide ascale-dependent measure of model perform-ance. For example, the model ability to recon-struct theoretically expected shape of kineticenergy spectra can serve as a prime tool forqualitative model evaluation and for estima-tion of model effective resolution.

Furthermore, spectral decomposition in thefrequency domain allows quantification ofpower distribution among different temporalscales and thus provides an information aboutexposure of particular station to diurnal andsub-diurnal forcing, as well as on model abilityto simulate observations. One of major poten-tial benefits of high resolution mesoscale mod-

106 Hrvatski meteoroloπki Ëasopis � Croatian Meteorological Journal, 50, 2015

pogreška smanjuje se za više od 50%. Najveći doprinos korjenu srednje kvadratne pogreškemože se, na većini postaja, pripisati faznoj pogrešci čiji udio raste s povećanjem horizontalnerezolucije modela. Spektralna analiza u domeni valnog broja pokazuje da se nagib spektara ki-netičke energije obaju modela smanjuje od k-3 u gornjoj troposferi prema ~ k-5/3 blizu površine(odgovara spektru terena) te pokazuje malu sezonsku varijabilnost. Spektralna dekompozicijamjerenih i modeliranih podataka u frekvencijskoj domeni ukazuje na značajno poboljšanje prisimuliranju primarnog i sekundarnog maksimuma spektralne snage (povezani sa sinoptičkim idnevnim gibanjima) korištenjem ALADIN/DADA 2 km modela, posebno za komponentuvjetra okomitu na Dinaride koja je uglavnom vezana uz buru. Konačno, zajednička karakteri-stika obaju modela je značajno podcjenjivanje spektralne snage na skalama manjim od polud-nevne, što je posljedica neprisutnosti tih gibanja u početnim uvjetima, odnosno smanjenjihmogućnosti modela da reprezentira procese najmanjih prostornih i vremenskih skala.

Ključne riječi: prognoza vjetra, kompleksan teren, ALADIN, statistička verifikacija, spektral-na verifikacija

HMC-50.qxp_HMC-50 10/03/16 09:50 Page 106

elling is to capture diurnal forcing by topogra-phy or surface inhomogeneity (Rife et al.2004). Hence, at stations with significant diur-nal component there exists a potential for im-proving the results using higher-resolution nu-merical models. On the other hand, scales be-low diurnal and in particular below semi-diur-nal still represent a challenge for operationalforecasting. Those scales include circulationsthat may result from far upstream landscapeforcing or from nonlinear interactions. Due tosparse spatial character of wind measuringnetwork (Bajić 2011.), these motions are notpresented in the model initial conditions.Therefore, if not predicted through sub-diur-nal forcing, they cannot be deterministicallypredicted regardless of the grid spacing andmodel physics.

The major objective of this study is to assessthe ability of dynamical adaptation to forecastrelevant wind conditions in the complex ter-rain of the eastern Adriatic coast. Other ob-jectives include quantification of differentsources of model errors and their changes inmagnitude with decrease of grid spacing, aswell as to perform more physical insight intothe model ability to reproduce near surfacewinds by using the spectral decomposition.The paper is organized as follows. The modelsetup and verification strategy are describedin section 2. The results of statistical and spec-tral verification are given in section 3. Conclu-sions are given in section 4.

2. DATA AND METHODS

2.1. Model settings

Meteorological and Hydrological Service(DHMZ) of Croatia uses Aire Limitée Adap-tation Dynamique développement InterNa-tional (ALADIN; Bubnova et al. 1995) limit-ed area model for the operational weatherforecast. ALADIN is a primitive equationmodel built on the basis of the global modelsIFS/ARPEGE (IFS - Integrated Forecast Sys-tem, ARPEGE - Action de Recherche PetiteEchelle Grande Echelle) and it uses spectraltechnique for the horizontal representation ofprognostic fields. Set of primitive equations issolved for wind speed components, tempera-ture, specific humidity and surface pressure,using the two-time-level semi-implicit semi-Lagrangian integration scheme. In the vertical

direction ALADIN model uses hybrid pres-sure-type � coordinate (Simmons and Bur-ridge 1981.) on 37 model levels with the finitedifferences method. Prognostic fields in spec-tral space are obtained using the doubleFourier transform, while the biperiodicity issatisfied by applying the elliptic truncation, i.e.including the extension zone (Machenhauerand Haugen 1987). The vertical transfer ofheat, momentum and moisture are describedby turbulence parameterization based onLouis scheme (1982) with modified dependen-cy on stability (Redelsperger et al. 2001).Physics package includes shallow convectionparameterization (Geleyn 1987). Stratiformand convective processes are considered indi-vidually by using the Kessler type of parame-terization (Kessler 1969) for resolved precipi-tation and modified Kuo scheme for deep con-vection (Geleyn et al. 1982). Radiation is pa-rameterized according to Geleyn andHollingsworth (1979) and Ritter and Geleyn(1992). The vertical transport of moisture andheat in the soil are parameterized in two lay-ers following the Giard and Bazile (2000).

The operational model version at DHMZ isrun in a hydrostatic mode with 37 vertical lev-els (the lowest level is approximately at 17 mabove ground level (AGL)) and 8 km horizon-tal grid spacing. The model integration do-main is shown on Figure 1. During the verifi-cation period initial and lateral boundary con-ditions (LBC) were obtained from analysisand forecasts of the ARPEGE global model.In the case of analysis the digital filter initial-ization (DFI; Lynch and Huang 1994) wasused, while the frequency of LBCs was 3hours. Model was initialized daily at 00 UTCand integrated over forecasting period of 72hours with time-step of approximately 5.45minutes (11 time-steps in 1 hour); but the out-put data were archived every 3 hours. Afterthe start of the model, mesoscale energy startsto accumulate and it takes few hours for thestabilization of the results. Based on the analy-sis of yearly averaged lead-time dependent ki-netic energy spectra the spin-up, i.e. stabiliza-tion period of 9 hours was allowed. This ver-sion of the ALADIN model setup is referredto as ALADIN/ALARO 8 km (AL8).

After integration, the prognostic fields of AL8model were refined to 2 km grid spacing over

107M. Hrastinski et al.: Verification of the operational 10 m wind forecast obtained with the ALADIN mesoscale numerical weather prediction model

HMC-50.qxp_HMC-50 10/03/16 09:50 Page 107

a sub-domain (Fig. 1), using the dynamicaladaptation method (Žagar and Rakovec 1999;Ivatek-Šahdan and Tudor 2004.). Dynamicaladaptation is simplified and cost-effectivemethod used for dynamical downscaling ofnear-surface winds. The main simplificationsinclude reduced number of vertical levels (15;opposed to 37 of AL8 model) above 1000 me-ters, as well as exclusion of moist processesand effects of radiation. A total 8 of these 15vertical levels are placed within lowest 1000meters and have the same spacing as in AL8model. The dynamical adaptation proceduretakes the prognostic fields from each AL8 out-put file and interpolates them to the 2 km res-olution. Those files are then used as initial andlateral boundary conditions. As orography at2 km resolution contains more details, thequasi-stationary balance is disrupted. The

model fields adapt to higher resolution terrainby running the numerical model from this newinitial state until the quasi-stationary state isachieved (in operative; 30 steps of 60 s each).As discussed by Žagar and Rakovec (1999),dynamical adaptation will be more successfulwhen pressure gradients are stronger because:i) dynamical forcing is dominant mechanism,i.e. the large scale wind is strong enough to at-tenuate local thermal or convection-inducedcirculations and ii) adaptation period is signifi-cantly shorter. On the other hand, adaptationmight fail in case of: i) valley inversions, ii)thermal circulations that are not resolved bythe driver model, iii) quick passage of coldfronts and other errors in the input data. Thisversion of the ALADIN model setup is re-ferred to as ALADIN/DADA 2 km (DA2).More details on weather forecasting at MHS,


Figure 1. Integration domain of the ALADIN/HR model at 8 km horizontal grid spacing (AL8; outer domain)and the domain of dynamical adaptation at 2 km horizontal grid spacing (DA2; inner domain) with thecorresponding orography and stations selected for the verification of model outputs.

Slika 1. Integracijska domena ALADIN/HR modela na horizontalnoj razlučivosti od 8 km (AL8; vanjskadomena) i domena dinamičke adaptacije na 2 km horizontalnoj razlučivosti (DA2; unutarnja domena) spripadnom orografijom i mjernim postajama odabranim za verifikaciju modelskih izlaza.

HMC-50.qxp_HMC-50 10/03/16 09:50 Page 108

as well as on both AL8 and DA2 model setupscan be found in Tudor et al. (2013).

2.2. Observations and verification procedure

The verification was performed for AL8 andDA2 datasets in period 2010-2012, using themeasured data from eight meteorological sta-tions spread around eastern Adriatic coast.These stations represent different wind cli-mate regions and by name they are: Jasenice,Most Krk, Most Pag, Makarska, Novalja, SplitMarjan, Šibenik and Zadar. Jasenice, MostKrk and Most Pag stations are representativefor bora dominated regions of northern partof Eastern Adriatic coast. Makarska and SplitMarjan stations were selected for model veri-fication of wind climate at central part of east-ern Adriatic, where bora and sirocco are twodominant types of flow. Other stations (No-valja, Šibenik and Zadar) represent regionswith relatively weaker winds and with signifi-cant portion of locally developed and thermal-ly driven flows. The criterion for selection ofstations was the record completeness (lessthan 10% of missing data), which is especiallyimportant for spectral analysis. Gaps in timeseries were filled by means of linear temporalinterpolation. As we wanted to avoid spatialinterpolation of model outputs, the closestland-placed point among four neighboringmodel grid points was selected as representa-tive one. Wind components at 10 m AGL areobtained by vertical interpolation between thesurface and the lowest model level, using thesimilarity theory (Geleyn 1988). Instrumental,calibration and representativeness errors werenot taken into account during the verificationprocess. This should be kept in mind duringthe evaluation of model performance, sincesome studies adduce the value of 1.15 ms-1 asrepresentativeness error of near surface windspeed under well-mixed ABL in the complexterrain (Rife et al 2004).

Before the start of the verification procedureAL8 and DA 2 model data were reduced to 21hours forecasting period, bearing in mind thespin-up period of 9 hours for both models. Sta-tistical verification included calculation ofMBIAS, RMSE and MAE, according to stan-dard equations (e.g. Wilks 2006). As RMSE isaffected by uncertainties both in time and space,we have performed the decomposition on its in-tegral components (Murphy 1988; Horvath et

al. 2012): bias of the mean (BM), bias of thestandard deviation (BSD) and dispersion orphase error (PHE). Obtained values of mo-ment-based scores were averaged over monthlyperiods to show their seasonal variability.

The kinetic energy (KE) spectra were calcu-lated directly from spectral fields of AL8 andDA2 model files at levels in lower, middle andupper troposphere, using the ECTOplasmsoftware (http://www.cnrm.meteo.fr/gmapdoc/meshtml/ecto3.0). Spectra were calculatedevery 3 hours (interval of availability for bothAL8 and DA2 data) for the entire year of2012. Here presented spectra are yearly aver-aged values corresponding to the 12-houryforecast. It should be noted that KE spectra ofAL8 and DA2 models were not calculated ondomains of the same size (Fig. 1), hence wehave a shift in energy at particular wavenum-bers. Since the method of calculation usedhere does not affect the slope of KE spectra atshortest wavelengths, the energy shift men-tioned above does not affect the performedanalysis and conclusions related to the effec-tive model resolution.

Because majority of stations are located nearthe seashore or in very complex terrain and arecharacterized by motions of different temporalscales, the success of the analyzed models wasverified using the spectral decomposition in thefrequency domain. As typical diurnal rotationof winds in the Adriatic (Telišman Prtenjakand Grisogono 2007) partially hides the diurnalspectral peak if the decomposition is performedusing the wind speed values, we have per-formed the decomposition of both horizontalwind components. The coordinate system is ro-tated such that u and v components correspondto the cross-mountain and along-mountain di-rections. Spectral decomposition of detrendedtime series was performed using the Welch pe-riodogram-based method (Welch 1967) on seg-ments that overlap by 50 %. The length of spec-tral window (L=256) was adjusted to optimallyemphasize the diurnal and semi-diurnal spec-tral peaks. Measured and modelled data bothcorresponding to 10-min average values weresampled at regular 3-hourly intervals.

Quantification of these short-scale motions isimportant because the ability of their simula-tion is one of potential benefits of high-resolu-


HMC-50.qxp_HMC-50 10/03/16 09:50 Page 109

tion mesoscale models compared to coarsermesoscale or global models. To compare re-sults with Žagar et al. (2006) and Rife et al.(2004), although for somewhat different typesof stations, we have defined: diurnal range(DIU) as periods between 22 and 26 hours,sub-diurnal range (SUB) which comprises pe-riods between 6 and 22 hours and synoptic orlarger than diurnal range (LTD) as one withperiods longer than 26 hours and shorter than7 days. The area under the power spectraldensity (PSD) curve by spectral ranges was

calculated by means of numerical integration,using the trapezoidal rule. It should be notedthat PSD analysis performed here contains theeffect of aliasing, since 10-min values sampledevery 3 hours will be necessarily contaminatedby oscilations with periods shorter than 6hours (here corresponding to Nyquist fre-quency) and thus aliased distorting the shapeof analyzed frequency spectrum. While alias-ing will affect all scales, testing of this effect onmeasured data suggested that the effect onscales larger than diurnal is rather small, while


Figure 2. Wind roses from measurements at 10 m AGL for selected stations during period 2010-2012 (cf. Fig.1 for exact location of stations). The percentage of missing data amounts 5.9% for Most Krk station, 2.2% forMost Pag station, 5.3% for Split Marjan station and 1.1% for Šibenik station.

Slika 2. Ruže vjetra, dobivene iz mjerenja, na 10 m iznad tla za odabrane postaje u razdoblju 2010.-2012. (vidiSl. 1 za točne lokacije postaja). Postotak nedostajućih podataka iznosi 5.9% za postaju Most Krk, 2.2% zapostaju Most Pag, 5.3% za postaju Split Marjan i 1.1% za postaju Šibenik.

HMC-50.qxp_HMC-50 10/03/16 09:50 Page 110

significant effects may be found on sub-diur-nal scales, especially near the periods corre-sponding to the Nyquist frequency. The effectof aliasing in the sub-diurnal range discussedabove is estimated from measured data, andresults in the energy of sub-diurnal motionsbeing overestimated for around 40% on aver-age for coastal stations. While we cannot cir-cumvent this effect since the model data arestored every 3 hours, it may be noted that wechose to perform PSD analysis on 10-minmeasured data which are also sampled every 3hours and are thus aliased in a similar manner.Therefore, for the purpose of verification per-formed in this paper, it is likely that the alias-ing effect cancels out, at least to an extent.

3. RESULTS

3.1. Moment-based verification

In order to assess the reliability of wind fore-cast obtained with the ALADIN mesoscaleNWP model for the Adriatic region, we have


performed the statistical verification of AL8and DA2 model outputs at eight stations fromdifferent wind climate regions. As verificationscores we utilized MBIAS and RMSE, whichwas further decomposed on its integral compo-nents to quantify among different sources ofspatial and temporal errors (cf. Section 2.2.).

As representative locations for presentationof verification results we have chosen MostKrk, Most Pag, Split Marjan and Šibenik sta-tions exposed to bora, sirocco and other weak-er flows on northern, middle and southernpart of Eastern Adriatic (Figure 2). Obtainedvalues of moment-based scores were monthlyaveraged over all forecasting lead-times andshown in form of an annual cycle (Figure 3;MBIAS and Figure 4; RMSE), while averagedvalues for the entire period of 2010-2012 areshown separately in Table 1. By applying thedynamical adaptation (DA2 model) the 10 mwind speed forecast improved at all stations,and in particular at bora dominated stations

Figure 3. Mean monthly multiplicative systematic error (MBIAS; MBIAS=1 stands for the perfect agreementof measured and modeled data) of 10 m wind speed for AL8 and DA2 models at selected stations during period2010-2012.

Slika 3. Srednja mjesečna multiplikativna sistematska pogreška (MBIAS; MBIAS=1 označava savršenopodudaranje mjerenih i modeliranih podataka) brzine vjetra na 10 m visine za AL8 i DA2 modele naodabranim postajama u razdoblju 2010.-2012.

HMC-50.qxp_HMC-50 10/03/16 09:50 Page 111

Most Krk and Most Pag (Figure 3.). CoarserAL8 model underestimated the average windspeed values there by more than 20%, whilethe DA2 model reduced the systematic errorto under 10%. However, the wind speed val-ues obtained by DA2 model overestimate themeasured data at bora dominated stations.The largest deviations from ideal MBIAS val-ue were observed during late spring and sum-mer months at Most Krk and Šibenik stations,as well as during late autumn and wintermonths at Most Pag and Split Marjan stations.As seen from the relative frequency his-tograms (not shown here), the first mentionedcan be explained by the relatively poor per-formance in simulating near calm conditions(MBIAS is the most sensitive here) by bothAL8 and DA2 models. On the other hand, be-sides winter calms, the second mentioned in-cludes a problem with underestimation of thefrequency of stronger winds (V > 10 ms-1; SplitMarjan station).

RMSE values are larger at bora dominatedstations of northern and central part of theeastern Adriatic (Most Krk, Most Pag) than atother two analyzed stations (Split Marjan,Šibenik) (Figure 4). By applying the dynami-cal adaptation method, those values were re-duced up to mostly 20% when compared tothe AL8 model. The largest values of RMSEwere observed during winter months, whenthe wind speed values are in average higher atall stations. This is somewhat expected asRMSE is generally proportional to the aver-age wind speed. The largest portion of RMSEerrors can be attributed to PHE, exceptmaybe for AL8 model at bora dominated sta-tions where BM has similar or even greatercontribution (Figure 4; winter months). Withthe increase of model resolution BM and BSDgenerally decrease, while the PHE increases.The exception is Šibenik station where thegeneral rule holds only for the BM.


Figure 4. Decomposition of the root-mean-square error (RMSE) of 10 m wind speed for AL8 and DA2 modelsat selected stations in period 2010-2012. BM stands for the bias of the mean, BSD for the bias of the standarddeviation and PHE for dispersion or phase error.

Slika 4. Rastav korijena srednje kvadratne pogreške (RMSE) brzine vjetra na 10 m za AL8 i DA2 modele naodabranim postajama u razdoblju 2010.-2012. BM označava pristranost srednjaka, BSD pristranost standardnedevijacije, a PHE disperziju ili faznu pogrešku.

HMC-50.qxp_HMC-50 10/03/16 09:50 Page 112

3.2. Analysis of wind spectra in the wavenumber domain

By using multiple datasets, observational ana-lyses of KE spectra have shown that inherent-ly 2D large-scale motions (several thousand toseveral hundred of kilometers) in the free tro-posphere and lower stratosphere conform tothe k-3 power law, while at smaller, i.e. inertialturbulence scales this dependence relaxes to k-

5/3 (Nastrom and Gage 1985; Lindborg 1999).The slope of KE spectra is remarkably univer-sal as it has shown very little dependence onlatitude, season or altitude (Nastrom andGage 1985; Skamarock 2004). Between theabove mentioned scales there are mesoscalespectra which are not understood well, asmesoscale motions are predominantly 2D butthe KE spectrum is closer to k-5/3 (like 3D tur-bulence). Less steep dependence of KE spec-tra on wavenumber at mesoscale and shorterscales suggests that motions of these scales areenergetic and that error growth may be fasterthan at synoptic scales, as time scale ofmesoscale phenomena is much shorter thantypical forecasting range of mesoscale NWPmodels, so these existing errors have enoughtime to propagate upscale. Thus, the timelydetection and elimination of errors on smallscales potentially has a great importance forthe successful operation of mesoscale NWPmodels. Without going further into the theoryof KE spectra, we would like to state that KEspectra are an appealing tool for the analysisof properties of mesoscale models on variousspatial scales and qualitative assessment ofmodel performance.

On Figure 5 are shown the KE spectra forboth AL8 and DA2 models at levels in upper,middle and lower troposphere. Although thevertical levels for the two considered modelsare numerated differently (legend), the curvesof similar shades correspond to approximatelythe same altitude (according to the captions).As it can be seen, the dynamical adaptationhas created a significant portion of mesoscaleenergy at scales below 100 km. Furthermore,the energy of meso-scale motions of DA2model near the surface is much higher than inupper troposphere, which indicates the impor-tance of small scale surface processes. At high-er levels and for larger spatial scales, AL8 andDA2 models show similar behavior and con-form to the k-3 law. Unlike expectations, thereis no gradual transition towards the less steepbehavior at mesoscale. This suggests thatthere is not enough mesoscale energy createdin the upper troposphere in the ALADINmodel, which was already reported by Horvath et al. (2011) and to some extent byŽagar et al. (2006).

In the mid-troposphere the slope of the kinet-ic energy spectra slightly flattens, while nearthe surface it flattens even more and resem-bles the spectra of orography. This feature ofKE spectra suggests that surface winds in AL8and DA2 models are primarily adapted to theorography. On the other hand, variation of theslope of KE spectra with height is most likelythe result of presence of orographycally in-duced and vertically propagating gravitywaves over the model domain. Similar behav-ior of KE spectra is reported by Horvath et al.


Table 1. The multiplicative mean systematic error (MBIAS), root-mean square error (RMSE) and measured10 m wind speed during period 2010-2012 for selected stations (cf. Fig. 1). The MBIAS=1 denotes unbiased set,while MBIAS > 1 and MBIAS < 1 denote model overestimation and model underestimation of measured data.

Tablica 1. Srednja multiplikativna sistematska pogreška (MBIAS), korijen srednje kvadratne pogreške(RMSE) i izmjerena 10 m brzina vjetra tijekom razdoblja 2010.-2012. za odabrane postaje (vidi Sl. 1). JediničniMBIAS (MBIAS=1) označava podatke bez sistematske pogreške, dok MBIAS > 1 i MBIAS < 1 označavajumodelsko precjenjivanje, odnosno podcjenjivanje izmjerenih vrijednosti.

STATION Most Krk Most Pag Split Marjan Šibenik

MODEL AL8 DA2 AL8 DA2 AL8 DA2 AL8 DA2

MBIAS 0.72 1.09 0.78 1.08 0.87 0.91 1.21 1.12

RMSE (ms-1) 3.51 2.88 3.16 2.92 2.07 2.03 2.12 1.84

Vmeas (ms-1) 3.50 4.59 3.99 2.93

HMC-50.qxp_HMC-50 10/03/16 09:50 Page 113

(2011) for the dynamical downscaling ofERA-40 reanalysis with the ALADINmesoscale NWP model. The KE spectra mayalso be used for the estimation of model effec-tive resolution by studying the deviation ofmodeled spectra from the expected values atspectral tails, i.e. short scales (Skamarock etal. 2004; Horvath et al. 2011). By taking k-3 asa reference slope in the upper troposphereand the orography spectra in the lower tropo-sphere, we have estimated that effective reso-lutions of AL8 and DA2 models are approxi-mately 5-6Dx of the corresponding grid spac-ing. It may be also inferred that the effectivemodel resolution somewhat varies with height.The seasonal variability of KE spectra forAL8 and DA2 models was investigated as well(not shown here) and we have found thatslope of the spectra remains unchanged, butthe amount of energy decreases from winterto summer and opposite.


3.3. Spectral verification in the frequency domain

Here are presented the results of spectral veri-fication in the frequency domain performedfor Most Krk, Most Pag, Split Marjan andŠibenik stations. On Figs. 6. and 7. the PSD ofcross-mountain and along-mountain measuredand modeled wind components for Most Krkand Most Pag stations are shown. The largestportion of measured power at both stations isassociated with LTD motions, which are moreenergetic for the cross-mountain component,related to the strong and gusty bora wind.These motions are reasonably well simulatedwith the DA2 model, unlike the along-moun-tain LTD motions which are severely underes-timated at Most Krk station and slightly over-estimated at Most Pag station. However, thePSD of both components in LTD range arebetter simulated than in the AL8 model. Thesecondary spectral maximum appears at diur-

Figure 5. Kinetic energy spectra for AL8 and DA2 models at levels in upper (~ 400 hPa; blue shades), middle(~ 600 hPa; red shades) and lower troposphere (~ 1000 hPa; green shades). Normalized orography spectra ofAL8 and DA2 integration domains are added too (black-grey shades).

Slika 5. Spektri kinetičke energije AL8 i DA2 modela na nivoima u višoj (~ 400 hPa; plave nijanse), srednjoj(~ 600 hPa; crvene nijanse) i nižoj troposferi (~ 1000 hPa; zelene nijanse). Normalizirani spektri terena AL8 iDA2 modela također su dodani (crno-sive nijanse).

HMC-50.qxp_HMC-50 10/03/16 09:50 Page 114

nal scales and it is characterized by very wellsimulated total amount of power in the DA2forecast, although the larger amplitude thanthe one from measured spectra should affectthe share of diurnal power in the total modeledpower. Anyway, the DA2 model improvesboth the amplitude and amount of power inDIU spectral range over AL8 model for eachof the bora dominated stations. This feature,along with the improved amount of modeledpower in LTD spectral range presents themain added value of dynamical adaptation atMost Krk and Most Pag stations. One of themajor characteristics of modeled spectra is adeficiency of power in sub-diurnal spectralrange, and in particular for the scales shorterthan semi-diurnal. Due to similarities with theother two stations this feature of the ALADINmodel will be discussed in detail later.

Results of the spectral decomposition for SplitMarjan and Šibenik stations are presented onFigures 8 and 9. As it can be seen, the improve-ments of DA2 model over AL8 are somewhat

less than for the bora dominated stations, andare primarily observed for LTD and SUBranges of cross-mountain component at SplitMarjan station and LTD range of along-moun-tain component at Šibenik station. The amountof diurnal power and amplitude of the spectraldiurnal peak are similar for both models andwell compared with measurements. Unlike thebora dominated stations, the amount of powerin LTD spectral range is similar for cross-mountain and along-mountain components.On the other hand, the amount of power with-in sub-diurnal range of cross-mountain mo-tions is larger than at bora dominated stations,while the opposite stands for along-mountainmotions. Contrary to the bora dominated sta-tions, Split Marjan and Šibenik have pro-nounced semi-diurnal circulations which arepresumably related to the land-sea breeze cir-culation. This motions are reproduced very ac-curately by both models and somewhat betterfor the cross-mountain component. The samewas reported by Horvath et al. (2011).


Figure 6. Power spectral density (PSD) of 10 m cross-mountain (left) and along-mountain (right) windcomponent of measured and modeled AL8 and DA2 data at Most Krk station in period 2010-2012.

Slika 6. Spektar snage (PSD) poprečne (lijevo) i uzdužne (desno) komponente vjetra na 10 m visine za mjerenei modelirane AL8 i DA2 podatke na postaji Most Krk u razdoblju 2010.-2012.

Figure 7. Power spectral density (PSD) of 10 m cross-mountain (left) and along-mountain (right) windcomponent of measured and modeled AL8 and DA2 data at Most Pag station in period 2010-2012.

Slika 7. Spektar snage (PSD) poprečne (lijevo) i uzdužne (desno) komponente vjetra na 10 m visine za mjerenei modelirane AL8 i DA2 podatke na postaji Most Pag u razdoblju 2010.-2012.

HMC-50.qxp_HMC-50 10/03/16 09:50 Page 115

Finally, the common characteristic of bothAL8 and DA2 models is a severe underesti-mation of power at scales below semi-diurnal.Similar was already reported by Žagar et al.(2006) and was present in the results regard-less on the domain size and positioning, aswell as the used nesting strategy; direct nestingor by using the intermediate domain. This in-dicates the limited ability of currently usedALADIN model versions in simulating theproper amount of mesoscale energy at thevery shortest scales, since those are not re-solved by the model. As we have allowed verylarge spin-up period of 9-hours, it is likely thatthe reason for the deficiency of power at thesescales can be related numerical diffusion inthe ALADIN model.

In order to quantify the results of spectral de-composition in the frequency domain, we havecalculated the amount of power in LTD, DIUand SUB spectral ranges for all eight stationsincluded in the analysis. On Figure 10 we cansee the spectral power distribution of cross-

mountain wind component (measured ormodeled) in different spectral ranges normal-ized by total power (measured or modeled; x-axis) and by the observed power in the samespectral range (y-axis) at the same station. Ifwe firstly focus on measurements and the vari-ability along x-axis, it can be noted that eightstations are classified into two groups of fourstations. Due to similar spectral characteristicsMakarska station is classified into group withpreviously mentioned bora dominated (BD)stations, although it has somewhat more ener-gy in the mesoscale range. The other group ofstations comprises of Split Marjan station withsimilar portion of bora and sirocco flows, andthree stations with significant portion of local-ly developed and thermally driven (LDTD)flows. The distribution of spectral poweramong spectral ranges of LDTD stations issimilar to the one for coastal stations in Žagaret al. (2006), i.e. 50-55% of power is containedin LTD spectral range, with significant portionof 15-20% and 25-30% in DIU and SUBranges. On the other hand, the majority of


Figure 8. Power spectral density (PSD) of 10 m cross-mountain (left) and along-mountain (right) windcomponent of measured and modeled AL8 and DA2 data at Split Marjan station in period 2010-2012.

Slika 8. Spektar snage (PSD) poprečne (lijevo) i uzdužne (desno) komponente vjetra na 10 m visine za mjerenei modelirane AL8 i DA2 podatke na postaji Split Marjan u razdoblju 2010.-2012.

Figure 9. Power spectral density (PSD) of 10 m cross-mountain (left) and along-mountain (right) windcomponent of measured and modeled AL8 and DA2 data at Šibenik station in period 2010-2012.

Slika 9. Spektar snage (PSD) poprečne (lijevo) i uzdužne (desno) komponente vjetra na 10 m visine za mjerenei modelirane AL8 i DA2 podatke na postaji Šibenik u razdoblju 2010.-2012.

HMC-50.qxp_HMC-50 10/03/16 09:50 Page 116

spectral power for BD stations is distributedin the LTD spectral range (65-80%). SUBspectral range of BD stations contains 15-25%of spectral power, while the amount DIUpower is almost negligible (except for SplitMarjan station). It shall be noted that thespectral power for SUB spectral range is over-estimated, as discussed in Section. 2.2.

In contrast to the measurements, modellingresults do not display such a difference be-tween various wind climate regions. Generalcharacteristic of AL8 and DA2 models forboth group of stations is that they significantlyunderestimate the share of SUB motions intotal, as well as the ratio of modeled andmeasured power in the same range. At BDstations both models overestimate the share ofDIU power, although the results are signifi-cantly improved with the dynamical adapta-tion. On the other hand, there is no much dif-ference for the LDTD group of stations. Final-ly, the improvements in LTD range are onceagain related only to the BD stations, both forthe share of LTD power in total and for theratio of modeled and observed power in thesame range. The results for along-mountaincomponent are relatively similar, although thegrouping is not pronounced that much be-cause few BD stations have spectral powerdistribution closer to LDTD stations.

4. CONCLUSIONS

Here are presented the verification results ofoperational 10 m wind speed obtained withALADIN/ALARO 8 km (AL8) and ALADIN/DADA 2 km (DA2) models in pe-riod 2010-2012. The results of statistical verifi-cation suggest that using the dynamical adap-tation for surface mean wind speed predictionin complex terrain of eastern Adriatic coastreduces the multiplicative mean systematic er-ror (MBIAS) at bora dominated (BD) sta-tions by more than 50% compared to thecoarser AL8 model. On the other hand, theimprovement on stations with significant por-tion of locally developed and thermally driven(LDTD) flows is somewhat less, but that is ex-pected considering the dominant windregimes and simplifications included in the dy-namical adaptation method. The signal offorecast improvement is also seen in the root-mean-square error (RMSE) score, which haslargest values during winter months and re-


Figure 10. The spectral power distribution of cross-mountain wind component (measured or modeled)in different spectral ranges normalized by totalpower (measured or modeled; x-axis) and by theobserved power in the same spectral range (y-axis)at the same station

Slika 10. Razdioba spektralne snage poprečnekomponente vjetra (mjerena ili modelirana) urazličitim spektralnim rasponima normaliziranaukupnom snagom (mjerena ili modelirana; x-os) temjerenom snagom u istom spektralnom rasponu (y-os) na istoj postaji.

HMC-50.qxp_HMC-50 10/03/16 09:50 Page 117

duces from north to south. The largest portionof RMSE errors for AL8 model can be attrib-uted to dispersion or phase error (PHE), al-though the bias of the mean (BM) is quitelarge at BD stations. With the refinement ofmodel horizontal resolution the share of PHEin RMSE generally increases, while BM andbias of the standard deviation (BSD) signifi-cantly decrease and become several timessmaller compared to PHE. Thus, it appearsthat errors in creation and termination of cer-tain process become the main reason for theerror in high resolution mesoscale NWP mod-els, and suggests that major improvements tothe wind forecasting may be reached throughreducing dispersion errors, such as throughmesoscale data assimilation process.

The scale dependent evaluation of model per-formance, conducted by using the spectralanalysis in wavenumber and frequency do-mains, enabled the model assessment on widerange of scales. Calculated kinetic energy(KE) spectra compare well to the theoreticaland observational evidence in midlatitudes. Inthe upper troposphere and at scales above fewhundred kilometers, both AL8 and DA2 KEspectra conform to the k-3 power law. Frommid-troposphere towards the ground, the KEspectra of both models flatten and finally nearthe surface they resemble the spectra of orog-raphy (~ k-5/3), thus showing the influence ofthe orography on near surface flows in theALADIN model. Furthermore, the variationof the slope of the KE spectrum with height ismost likely related to the appearance and ver-tical propagation of orographycally-inducedgravity waves in the model domain. Thelargest drawback of the AL8 model KE spec-tra seems to be the deficit of energy at scalesbelow 200 km in upper troposphere, i.e. theexistence of unfavorable steepening of spec-tral slope from k-3, rather than theoreticallyexpected flattening towards k-5/3. On the oth-er hand, DA2 does not show this behaviorapart from the regular effective resolution ef-fect evident. The main benefit of dynamicaladaptations observed in the kinetic energyspectra seems to be the creation of significantportion of near surface energy at scales below50 km.

Analysis of measured and modeled powerspectra obtained by decomposition in the fre-

quency domain suggested that using the dy-namical adaptation improved the ability ofsimulating the amount of power in all spectralranges over AL8 model. However, the majorimprovements are observed for both theamount of power compared to measurementsand the share in total power for the longerthan diurnal (LTD) and diurnal (DIU) scalesof cross-mountain motions at BD stations.This is in large part related to the better simu-lation of higher wind speeds related to boraflows by the DA2 model. Major drawbacks ofboth models are related to the inability ofgrouping different wind climate regimes asseen in the measurements and the insufficientmodel performance at scales below semi-diur-nal. The later one might even influence the ap-plicability of mesoscale numerical weatherprediction (NWP) models in the assessment ofwind climatology in regions where the amountof power at the smallest scales in comparablewith the amount of power at the larger scales.This study points out to some advantages ofhigh resolution mesoscale NWP modelling us-ing the dynamical adaptation, mostly relatedto decreasing the systematic error by improv-ing the simulation of bora flows. However,there are few questions raised regarding themodel uncertainties and drawbacks of the dy-namical adaptation method. First of all, itwould be crucial to quantify on disadvantagesof reduced complexity of dynamical adapta-tion compared to the full-physics based mod-els of the same horizontal grid spacing. A po-tential study should include evaluation overlonger periods as the one presented here, aswell as case studies, which should be involvedas they bring more information on dynamicsand extreme events. Furthermore, it would bebeneficial to make an effort on advanced un-derstanding of model errors through relatingvarious aspects of physical and spectral verifi-cation measures. Finally, the analysis andforecasting of bora gustiness and turbulenceremains to be one of major research chal-lenges, both for the field of meteorology andapplications to wind energy sector.

ACKNOWLEDGEMENTS

This study was supported by the grantIPA2007/HR/16IPO/001-040507 through theWILL4WIND project (www.will4wind.hr).


HMC-50.qxp_HMC-50 10/03/16 09:50 Page 118

REFERENCES

Bajić, A., 2011: Prostorna raspodjela očeki-vanih maksimalnih brzina vjetra nasloženom terenu Hrvatske kao podloga zaocjenu opterećenja vjetrom. Doktorski rad.Geofizički odsjek PMF-a, 110 str.

Belušić, D., M. Hrastinski, Ž. Večenaj, and B.Grisogono, 2013: Wind Regimes Associat-ed with a Mountain Gap at the Northeast-ern Adriatic Coast. J. Appl. Meteor. Clima-tol., 52, 2089-2105., doi:http://dx.doi.org/10.1175/JAMC-D-12-0306.1

Bubnova, R., G. Hello, P. Benard, and J.-F.Geleyn, 1995: Integration of fully elasticequations cast in the hydrostatic pressureterrainfollowing coordinate in the frame-work of ARPEGE/ALADIN NWP system.Mon. Wea. Rev., 123, 515-535.

Geleyn, J.-F., and A. Hollingsworth, 1979: Aneconomical analytical method for computa-tion of the interaction between scatteringand line absorption of radiation. Contrib.Atmos. Phys., 52, 1-16.

Geleyn, J.-F., C. Girard, and J.-F. Louis, 1982:A simple parametrization of moist convec-tion for large-scale atmospheric models.Beitr. Phys. Atmos., 55, 325-334.

Geleyn, J.-F., 1987: Use of a modifiedRichardson number for parameterizing theeffect of shallow convection. In: Matsuno Z.(ed)., Short and medium range numericalweather prediction, special volume of J. Me-teor. Soc. Japan, 141-149.

Geleyn J.-F., 1988: Interpolation of wind, tem-perature and humidity values from modellevels to the height of measurement. Tellus,40A, 347-351.

Giard, D. and E. Basile, 2000: Implementationof a New Assimilation Scheme for Soil andSurface Variables in a Global NWP Model.Mon. Wea. Rev., 128, 997-1015.

Grisogono, B., and D. Belušić, 2009: A reviewof recent advances in understanding themeso- and microscale properties of the se-vere Bora wind. Tellus, 61A, 1-16.

Horvath K., Y-L. Lin, B. Ivancan-Picek, 2008:Classification of Cyclone Tracks overApennines and the Adriatic Sea. Mon.Wea. Rev., 136, 2210-2227.

Horvath, K., A. Bajić, and S. Ivatek-Šahdan,2011: Dynamical downscaling of windspeed in complex terrain prone to bora-type flows, J. Appl. Meteorol. Climatol., 50,1676–1691, doi:10.1175/2011JAMC2638.1.

Horvath, K., D. Koracin, R. Vellore, J. Jiang,and R. Belu, 2012: Sub-kilometer dynami-cal downscaling of near-surface winds incomplex terrain using WRF and MM5mesoscale models, J. Geophys. Res., 117,D11111, doi:10.1029/2012JD017432.

Ivatek-Šahdan, S., and M. Tudor, 2004: Use ofhigh-resolution dynamical adaptation in op-erational suite and research impact studies.Meteor. Z., 13, 99-108.

Jurčec, V., B. Ivančan-Picek, V. Tutiš, and V.Vukičević, 1996: Severe Adriatic jugo wind.Meteor. Z., 5, 67-75.

Lindborg, E., 1999: Can the atmospheric ki-netic energy spectrum be explained by two-dimensional turbulence? J. Fluid Mech.,388, 259-288.

Louis J.-F., Tiedke M. and Geleyn J.-F., 1982:A short history of PBL parameterization atECMWF. In: Proceedings from ECMWFworkshop on planetary boundary layer pa-rameterization, 25-27 November 1981, pp59-79

Lynch, P. and X.-Y. Huang, 1994: Diabatic ini-tialization using recursive filters. Tellus,46A, 583-597.

Kessler, E., 1969: On the Distribution andContinuity of Water Substance in Atmos-pheric Circulations. Meteor. Monogr., No.10, Amer. Meteor. Soc., 84 pp.

Machenhauer, B., J. E. Haugen, 1987: Test ofa spectral limited area shallow water modelwith time-dependent lateral boundary con-ditions and combined normal mode/semi-lagrangian time integration schemes. Tech-niques for Horizontal Discretization in Nu-merical Weather Prediction Models, 2-4 No-vember 1987, ECMWF, 361-377.

Mass, C. F., D. Ovens, K. Westrick, and B. A.Colle, 2002: Does increasing horizontal res-olution produce more skillful forecasts?Bull. Amer. Meteor. Soc., 83, 407-430.


HMC-50.qxp_HMC-50 10/03/16 09:50 Page 119

Murphy, A. H., 1988: Skill scores based on themean square error and their relationshipsto the correlation coefficient, Mon. Wea.Rev., 116, 2417-2424.

Nastrom, G. D., and K. S. Gage, 1985: A cli-matology of atmospheric wavenumberspectra of wind and temperature observedby commercial aircraft. J. Atmos. Sci., 42,950-960.

Pielke, R. A., 2002: Mesoscale meteorologicalmodeling. 2nd Edition, Academic Press, SanDiego, CA, 676 pp.

Redelsperger, J. L., Mahé, F., Carlotti, P.,2001: A simple and general subgrid modelsuitable both for surface layer and free-stream turbulence. Bound.-Layer Meteor.101, pp.375-408.

Rife, D. L., C. A. Davis, and Y. Liu, 2004: Pre-dictability of low-level winds by mesoscalemeteorological models. Mon. Wea. Rev.,132, 2553-2569.

Ritter, B., and J. F. Geleyn, 1992: A compre-hensive radiation scheme for numericalweather prediction models with potentialapplications in climate simulations. Mon.Wea. Rev., 120, 303-325.

Simmons, A. J., Burridge, D. M., 1981: An En-ergy and Angular-Momentum ConservingVertical Finite Difference Scheme and Hy-brid Vertical Coordinates. Mon. Wea. Rev.,109, 758-766.

Skamarock, W. C., 2004: Evaluatingmesoscale NWP models using kinetic ener-gy spectra. Mon. Wea. Rev., 132, 3019-3032.

Smith, R. B., 1985: On severe downslopewinds. J. Atmos. Sci., 42, 2597-2603.

Telišman Prtenjak, M., and B. Grisogono,2007: Sea/land breeze climatological char-acteristics along the northern CroatianAdriatic coast. Theor. Appl. Climatol., 90,201-215.

Tudor, M., Ivatek-Šahdan, S., Stanešić, A.,Horvath, K., and Bajić, A., 2013. Forecast-ing weather in Croatia using ALADIN nu-merical weather prediction model. In: Cli-mate Change and Regional/Local Respons-es / Zhang, Yuanzhi ; Ray, Pallav (eds.). Ri-jeka: InTech 59-88.

Welch, P. D., 1967: The use of fast Fouriertransform for the estimation of power spec-tra: A method based on time averagingover short, modified periodograms. IEEETrans. Audio Electroacoust., AU-15 (6), 70-73.

Wilks, D. S., 2006: Statistical Methods in theAtmospheric Sciences. 2nd ed. InternationalGeophysics Series, 59, Academic Press, 627pp.

Žagar, M., and J. Rakovec, 1999: Small-scalesurface wind prediction using dynamicadaptation. Tellus, 51A, 489-504.

Žagar, N., M. Žagar, J. Cedilnik, G. Gregorič,and J. Rakovec, 2006: Validation ofmesoscale low-level winds obtained by dy-namical downscaling of ERA-40 over com-plex terrain. Tellus, 58A, 445-455.


HMC-50.qxp_HMC-50 10/03/16 09:50 Page 120

Documents

HRVATSKI METEOROLO©KI »ASOPIS CROATIAN … · (odgovara spektru terena) te pokazuje malu sezonsku varijabilnost. Spektralna dekompozicija mjerenih i modeliranih podataka u frekvencijskoj