Heavy rains are commonly responsible for weatherdisasters in China. Associated with the summer mon-soon, three primary rain zones of China in the warmseason are located in southern China, Jiang-Huai area, and northern China, respectively(Tao, 1980). Innorthern China, heavy rains are concentrated in July and August and are mostly caused by strong local con-vective storms(Compilers of Heavy Rainfall in NorthChina, 1992). Due to complex topography and highpopulation density, heavy rainfall events in this area, especially flash-flood-producing rainfall, can cause sig-nificant social and economic losses, and endangerlives.

During 21-22 July 2012, a disastrous rainfall hitBeijing(often referred to as the "7. 21" event). Beingthe heaviest rainfall ever occurred in the last 61 years, it caused widespread havoc in the capital, killing atleast 79 people and endangering more than 1. 9 mil-lion residents. During 0200-2200 UTC 21 July, morethan 150-mm rain fell across Beijing. The heaviest pre-cipitation reached 460 mm in Fangshan area, causingcatastrophic flood and l and slide. This extreme eventhas caused wide concerns in meteorological commu-nity in China. Several studies(e. g., Fang et al., 2012; Sun Jisong et al., 2012)investigated the synoptic and mesoscale conditions and the mechanism of the rain-storm system. Chen et al. (2012) and Sun Jun et al. (2012)discussed the causes of the extreme event, while Yu(2012)analyzed the favorable conditions for thisrecord-breaking rainfall, with the aid of radar obser-vation. Additionally, a number of papers investigateddetailed features such as the sources of water vapor, frontogenesis, upper-level jet stream, and so on(Chen et al., 2013; Li et al., 2013; Liao et al., 2013; Quan et al., 2013; Sun et al., 2013). These studies focusedon the precipitation dynamics and conditions of thisextreme event using various observations, trying togain a better underst and ing of its mechanisms. How-ever, discussions on model forecasting performance ofthis event are rare. In particular, quantitative pre-cipitation forecasts(QPFs)based on high resolutionmesoscale models are critical for operational forecast-ers.

Zhao et al. (2013)showed that the "7. 21" eventoccurred under a typical circulation pattern conduciveto heavy rainfall in northern China. The opera-tional models predicted the heavy rainfall amount withskill of various degrees, however, the forecast rainfallstart and end times were noticeably delayed(Tao and Zheng, 2013). Though some of the operational modelsreproduced rainfall accumulation, they did not capturethe correct mechanism in the rainfall process(Zhang et al., 2013). Therefore, analysis of the model capabil-ity to predict such heavy rainfall is necessary, in orderto identify the causes of forecast errors and improveQPF.

In recent years, high performance computinghas rapidly advanced the development of numericalweather prediction models on flne scales, with im-proved physics and data assimilation systems. Veri-flcation is needed to underst and the advantages and weaknesses of such model forecasts. However, tradi-tional veriflcation statistics tend to penalize a mis-located precipitation, especially for high resolutionmesoscale models, possibly resulting in poor indica-tions of QPF quality using such statistics as threatscore(TS) and equitable threat score(ETS)(Schaffer, 1990), and BIAS score. As an alternative, Ebert and McBride(2000)developed an objective-oriented ver-iflcation method within the framework of contiguousrain areas(CRAs). Through this method, the totalrainfall error of model forecasts can be decomposedinto the components of location, rain volume and pat-tern, providing quantitative analysis of precipitationerrors.

In this work, ETS and BIAS scores are used toprovide overall performance of QPF of the mesoscalemodel forecasts for the "7. 21" event, while the CRAmethod is used to better underst and the forecast er-rors of QPF and model predictability of heavy rainfall. Accordingly, Section 2 describes the data sources(in-cluding the BJ-RUCv2. 0 model output) and briefl in-troduces the CRA method. In Section 3, the mesoscalemodel forecast results are investigated and discussed. Section 4 presents the veriflcation results of precip-itation forecasts based on both traditional methods and the CRA method. Section 5 discusses the possi-ble causes of model errors. Finally, the summary and conclusions are given in Section 6. 2. Data and method2.1 Data sources

This work is based on various observations and analysis data, which are listed as follows.

1)NCEP(National Centers for EnvironmentalPrediction)GFS(The Global Forecast System)6-hanalysis with 0. 5° × 0. 5° resolution;

2)Hourly merged precipitation product with 0. 1°× 0. 1° resolution based on AWS(automatic weatherstation)observations in China and CMORPH(Cli-mate Prediction Center MORPHing technique)satel-lite data;

3)FY-2E satellite infrared images;

4)Beijing Meteorological Bureau BJ-RUCv2. 0model(Beijing-Rapid Update Cycling data assimila-tion and forecast system)hourly output data with 3-km resolution(Fan et al., 2013).

The CMORPH merged precipitation data havebeen developed through the two-step merging algo-rithm of probability density function and optimal in-terpolation, which effectively takes the advantagesof the AWS observations and satellite product ofCMORPH(Shen et al., 2010). Therefore, the mergedprecipitation product can capture more reasonableprecipitation amount and spatial distribution, as wellas the mesoscale features of precipitation distribution(Shen et al., 2013). Figure 1 shows the distributionof CMORPH merged precipitation from 0200 to 2200UTC 21 July 2012. Compared with the observed pre-cipitation released by the Beijing Meteorological Bu-reau(Sun Jisong et al., 2012), the CMORPH mergeddata successfully capture the precipitation features and agree well with surface observation. Despite ofa weaker rainfall extreme intensity(389 mm)com-pared to the observed(460 mm), the location of pre-cipitation extreme is accurate. On the other h and , thestudy area has intensive AWS network with a completequality control system. In addition, taking into acco-unt the good continuity of satellite retrieved data, theCMORPH merged data have relatively high quality and are therefore used in this work as the observedprecipitation for veriflcation.

Fig. 1. Distribution of CMORPH merged precipitation(mm)from 0200 to 2200 UTC 21 July 2012

Figure options

The BJ-RUCv2. 0 is a 3-h update and cyclingdata assimilation and forecast system based on theWRF(Weather Research and Forecasting)model and WRFDA(WRF data assimilation)system(Skamarock et al., 2005). The system contains two domains, with9- and 3-km grid spacings respectively. The 9-km fore-casts are initialized with the cold start at 0000 and 1200 UTC with hourly output. The 3-km forecastsare generated with the cold start initialization at 0000UTC each day and then run with the diabatic initial-ization(hot start)every 3 h from 0300 UTC, usingthe 0-3-h forecasts as the background and the 9-kmforecasts as the boundary conditions. Conventionalsoundings, surface observations, aircraft reports, winddata, the AWS data, and the GPSPW(GPS precip-itable water)are assimilated into the system. Basedon the BJ-RUC system(Fan et al., 2008; Chen et al., 2010; Lei et al., 2012), the BJ-RUCv2. 0 system alsoassimilates the radar reflectivity and radial velocityof the six Doppler radars in the Beijing-Tianjin-Hebeiregions with §l-h assimilation window through WRF-3DVar(3-dimensional variational system). The effectof assimilating radar reflectivity in the BJ-RUCv2. 0system has been assessed(Fan et al., 2013). In thisstudy, the 3-h analysis data and the 0-24-h forecasts(hourly output)on a 3-km grid of the BJ-RUCv2. 0system are analyzed. 2.2 The CRA method

Traditional veriflcation statistics can only provideguidance for the overall performance. For individualevent, it is diffcult to determine whether the QPF er-ror is brought by the displacement of precipitation orrainfall intensity. This work uses the CRA method(Ebert and McBride, 2000)to verify the QPF of theBJ-RUCv2. 0 system, with total error decomposed intothe components due to horizontal displacement of thesystem, error in the mean rain intensity and patternerrors. The CMORPH merged precipitation productis used as the veriflcation data.

The CRA method is carried out following thesteps below.

1)Map the forecast and observed flelds onto acommon grid(in this work, the BJ-RUCv2. 0 forecastfleld is mapped onto the 0. 1° × 0. 1° grid). Determinethe boundary of CRA by setting precipitation thresh-old.

2)Shift the forecast fleld incrementally over theobserved fleld until a "best flt" criterion is optimized. The criterion chosen in this work to determine the bestflt is the minimization of TSE(total squared error).

3)Decompose forecast error.

Accordingly, the total mean squared error(MSE)can be decomposed into three components:

MSE_total = MSE_displacement + MSE_volume + MSE_pattern.

The total MSE corresponds to the MSE of theoriginal forecast, MSE_total =1/N(f_i - o_i)², wheref_i and o_i are the forecast and observed rainfall at gridpoint i, and N is the number of gridpoints in the veriflcation domain. After shifting the forecast fleld, theMSE is recalculated as MSE_shift =1/N(f_i' - o_i)², where f_i' is the shifted forecast at grid point i.

The total error due to displacement, volume, and pattern can be then calculated as

MSEdisplacement = MSE_total - MSE_shift;

MSE_volume =( - )²;

MSE_pattern = MSE_shift - MSE_volume:

The overbar denotes the mean value over the domain.

The CRA method can quantitatively determinethe attribution of different error components and pro-vide a clearer and more comprehensive insight of theperformance of QPF. 3. Precipitation features of mesoscale forecasts

As the synoptic circulation provided favorable en-vironment to the "7. 21" extreme rainfall, high precip-itation effciency, strong ascending motion, long dura-tion, and extreme abundant water vapor brought to-gether that led to the heavy rainfall(Sun Jun et al., 2012). Although under the typical circulation con-dition various operational models predicted the rainprocess in advance(e. g., T639 in China, NCEP/GFS, regional spectral model in Japan), signiflcant errorsexist. The GFS 24-h forecast signiflcantly underes-timated the precipitation, with the maximum totalprecipitation intensity lower than 200 mm per 24 h. The 24-h forecast released by the National Meteoro-logical Center of China Meteorological Administrationat 2100 UTC 20 July 2012 indicated heavy rainfall inBeijing area with a blue rainstorm warning(over 50mm per 24 h). However, compared with the obser-vation, the forecast underestimated the precipitationwith a southwest displacement error. This work fo-cuses on the real time forecasts produced by the high-resolution BJ-RUCv2. 0 system of the Beijing Meteo-rological Bureau, and evaluates its performance and investigates possible causes of the forecast errors.

The precipitation features of the "7. 21" event areflrstly reviewed. As shown by the hourly CMORPHmerged precipitation distribution(Figs. 2a1-a4), theprecipitation area is quite concentrated with promi-nent mesoscale feature. The rain starts to intensify at0600 UTC and gradually increases to reach extreme at1200 UTC 21 July, presenting a southwest-northeast(SW-NE)oriented mesoscale, quasi-linear rain belt(Fig. 2a3). Accordingly, the rainstorm system movesfrom southwest to northeast, gradually evolves into aconvective line, bringing sustained downpour over Bei-jing. After 1200 UTC, the system heads eastward and moves out of Beijing(Fig. 2a4). The correspondingsatellite infrared images(flgure omitted)also indicatethe presence of mesoscale convective complex(MCC). The time series of FY-2E infrared images show thatthe rain clusters have been developing and maintain-ing. By 1200 UTC, the black body temperature hasdecreased to lower than -70℃, showing very high cloudtop and deep convection. The water vapor channelof satellite images also shows strong SW-NE orientedwater vapor transport to the rainstorm center. Theembedded MCSs move in a direction consistent withthe development of the surface rain belt.

Fig. 2. Hourly precipitation(mm h-1)at 0600, 0900, 1200, and 1500 UTC 21 July 2012 from(a1-a4)CMORPHmerged precipitation and forecast precipitation initiated at(b1-b4)2100 UTC 20 July, (c1-c4)0000 UTC, (d1-d4)0300UTC, (e2-e4)0600 UTC, (f3, f4)0900 UTC, and (g4)1200 UTC 21 July 2012. Valid at(b1-d1)0600 UTC, (b2-e2)0900UTC, (b3-f3)1200 UTC, and (b4-g4)1500 UTC 21 July.

Figure options

The forecasts and CMORPH merged precipita-tion are regionally averaged(shown by the red box inFig. 2a1). The time series(Fig. 3)suggest that theforecasts initialized at 2100 UTC 20 July and 0000UTC 21 July signiflcantly under-predict the observedprecipitation with time delays of 3-4 h, as forecastprecipitation peaks at 1500 UTC, compared to the ob-served peak at around 1200 UTC. The underestima-tion of regional averaged precipitation is partly due tolocation displacement(Figs. 2a2 and 2a3), as the fore-cast rain area falls to the southwest of the observation and thus most rainfall falls out of Beijing area. Precip-itation intensity and time series of short-term forecasts(initialized at 0600 and 0900 UTC 21 July)are in abetter agreement with the observation. It may resultfrom a better initial fleld as the BJ-RUCv2. 0 systeminitializes with the cold start at 0000 UTC and as-similates observation data every 3 hours. Therefore, the forecasts initialized after 0600 UTC have alreadyassimilated moisture and hydrometeor information inthe initial fleld, resulting in better forecasts. On theother h and , the cold front passes Beijing at about 1600UTC(Zhang et al., 2013). Before frontal passage, theprecipitation is generated by convective cells along thequasi-linear convective system, while precipitation af-ter 1600 UTC mainly results from frontal rainfall. Asindicated by the time series of CMORPH merged pre-cipitation, most of the "7. 21" extreme rainfall occursin the warm sector ahead of the cold front. It is evidentthat model forecasts underestimate the warm sectorprecipitation, but predict the front rainfall quite well.

Fig. 3. Time series of area averaged CMORPH mergedprecipitation(mm) and forecast precipitation(mm)at different initial times from BJ-RUCv2. 0 output.

Figure options

In particular, the forecast initialized at 0300 UTC21 July is prominently different. The forecast rain-fall is rather large at the beginning 3 hours, yet theextreme precipitation is poorly predicted with muchlower intensity and signiflcant timing errors. Furtherinvestigation of the precipitation distribution(Figs. 2d1-d4)shows that data assimilation has a positiveimpact on the forecast in the initial hours. However, the impact decreases with the forecast hours, leadingto the poor forecasts of extreme rainfall. 4. Objective veriflcation of QPFs4.1 Statistical scores

Statistical veriflcations are performed for QPFsfrom the BJ-RUCv2. 0 system. Veriflcation regioncovers 38°-42°N, 113°-119°E, with 41 × 61 grid-points. The ETSs(Figs. 4a and 4b)verifled againstCMORPH merged precipitation are generally higherfor short-term forecasts, consistent with the subjec-tive evaluation(Figs. 2 and 3). There are large dis-crepancies in ETSs at the 5 mm h^-1 threshold for dif-ferent initial times before 1600 UTC 21 July, whereshort-term forecasts are noticeably superior. After1600 UTC, the scores are comparable when precipi-tation is mainly caused by frontal rainfall and the raingradually decreases(Zhang et al., 2013). In the ex-treme precipitation period(1100-1400 UTC 21 July), ETSs at 5 mm h^-1 slightly decrease while those at 20mm h^-1 are relatively high. In addition, it is evidentthat the scores for short-term forecasts at 20 mm h^-1reach peak and then decrease quickly at 3 h. Such de-layed peak of ETSs may be caused by the adjustmentof the model to the 3DVAR analysis. With moistureobservations assimilated into the initial fleld, precip-itation forecasts are better with high ETSs at shortlead times, and the scores quickly fall with the fore-cast lead time. Particularly, the ETSs for forecastsinitialized at 0300 UTC remain very low at 20 mmh^-1, in agreement with the earlier discussions.

Fig. 4. Equitable threat scores(ETS; a, b) and BIAS scores(c, d; observed frequency is indicated by the right ordinate)for different initial times.

Figure options

On the other h and , the BIASs at 5 mm h^-1(Fig. 4c)indicate forecasts of all initial times over-predictthe rainfall to various degrees. In the extreme pre-cipitation period(1100-1400 UTC 21 July), BIASs ofall forecasts exceed 1. 0. In comparison, the observedfrequency decreases in this period, as the precipitationdistribution(Fig. 2a3)shows a relatively concentratedmesoscale rain belt. The forecasts demonstrate betterBIASs for heavy rain(> 20 mm h^-1)during 1100-1400 UTC, with BIASs of short-term forecasts closeto 1. 0(Fig. 4d). The BIASs of forecast initialized at2100 UTC 20 July are good, while ETSs remain low. Therefore, the forecast number of precipitation gridsexceeding the given threshold is comparable to the ob-servation, while the precipitation location and patternare inaccurate(Figs. 2b3 and 2b4). 4.2 The CRA veriflcation

To compensate for the weaknesses of traditionalstatistics, the CRA veriflcation is carried out againstCMORPH merged precipitation. The forecast rainfallhas been interpolated onto the 0. 1° × 1° grid as intro-duced in Section 2. In general, the model forecasts ofdifferent initial times are poorer in the extreme pre-cipitation period(Fig. 5). As for the same validationtime, the shorter the lead time, the smaller the rootmean square error(RMSE). The model's ability to pre-dict heavy rainfall is eminently poorer with increasingthresholds, as indicated by the much larger RMSE at20 mm h^-1 than that at 5 mm h^-1. Moreover, theerror decomposition suggests that the majority of to-tal RMSE at 5 mm h^-1 is due to displacement and pattern errors, while the error in forecast mean inten-sity is quite small. For results at 20 mm h^-1, the totalRMSE mainly results from the volume error. Forecastsfrom various initial times under-predict the intensityof heavy rainfall with large precipitation location and shape errors(Fig. 2). For precipitation higher than 5mm h^-1, the model predicts a much larger rain area, misrepresenting the observed feature of a linear rainbelt and the detailed rain structure.

Fig. 5. Stacked column charts of the CRA error decomposition. Each column from left to right at each valid timerepresents forecasts initiated at 2100 UTC 20 July; 0000, 0300, 0600, 0900, and 1200 UTC 21 July 2012; blue, red, and green are for errors of displacement, volume, and pattern, respectively.

Figure options

In addition to the veriflcation of hourly forecasts, the observed and forecast 21-h accumulated precipi-tation is verifled(Fig. 6). Subjective evaluation sug-gests that the model succeeds in producing a region ofextreme rainfall, with the forecast initialized at 0000UTC 21 July slightly closer to the observations. Theforecasts are evidently superior to the GFS forecast, which predicts extreme precipitation lower than 200mm per 21 h. However, the forecast rainfall locationfalls to the southwest of the observed fleld. The corre-sponding CRA veriflcation shows that the total RMSEat 20 mm h^-1 is about 60 mm, suggesting rather goodresults compared with the hourly forecasts. Error de-composition also shows that the majority of the errors, over 50%, is attributable to pattern error, with dis-placement error accounts for 30%-40% and the meanintensity error nearly negligible.

Fig. 6. Distribution of accumulated 21-h precipitation from 0000-2100 UTC 21 July 2012 from the(a)CMORPHmerged accumulated precipitation and model forecasts initiated at(b)2100 UTC 20 July and (c)0000 UTC 21 July2012.

Figure options

In conclusion, despite a reasonably good predic-tion of the accumulated precipitation, more detailedexaminations of hourly forecasts show that the realnature of the heavy rain has not been well capturedby the model, in other words, the precipitation mech-anism is inaccurately represented in the model. 5. Diagnosis of the rainfall weather conditions

The above discussion suggests that the model failsto accurately predict heavy rainfall with a large dis-placement error. To better underst and the precipita-tion mechanism and possible causes of the forecast er-rors, the synoptic circulation and physical parametersin mesoscale forecasts are diagnosed. 5.1 Synoptic circulation

GFS analyses of 500-hPa circulation show thatthe eastward moving trough encounters the southwestflow around subtropical high, bringing cold air to in-tersect with the warm moist flow. Beijing is mainlycontrolled by the upper southerly winds. For all forecast circulations at different initial times, the troughsystem in the initial flelds quickly disappeared as themodel integrates for 2-3 h. At the next analysis time, the observations containing the trough information areassimilated into the initial fleld; however, the troughis absent once again after model adjustment. In otherwords, the positive effect of data assimilation onlyshows at the initial times near the convective devel-opment and quickly decreases with the model integra-tion. The analysis flelds consist well with the GFSbackground circulations, where the short wave systemsare absent due to its coarse resolution. Comparing theforecast circulations at different initial times(Fig. 7)with the analysis fleld(Fig. 7f)valid at 1200 UTC21 July, each forecast succeeds in predicting the over-all circulation of the 500-hPa trough and north-southtemperature contrast, as well as the lower 850-hPawinds. However, the short wave trough located overBeijing in the analysis fleld(shown by the black box)isevidently missed in all forecasts. Also, the vortex sys-tems develop slightly to the south in the simulations and affect the precipitation location.

Fig. 7. BJ-RUCv2. 0 forecasts of 500-hPa temperature(shade; ℃), geopotential height(solid line; gpm), and 850-hPawind(vectors; m s-1)at 1200 UTC 21 July 2012 for initial times(a)2100 UTC 20, (b)0000 UTC 21, (c)0300 UTC 21, (d)0600 UTC 21, and (e)0900 UTC 21 July 2012; and (f)analysis field.

Figure options

The evolution of 850-hPa moisture flux indicates aprominent intensiflcation of water vapor transport and its convergence over Beijing from 0600 to 1200 UTC 21July(flgure omitted). The southeast flow and south-west monsoon stream transport water vapor to Bei-jing continuously, with remarkable water vapor con-vergence into the rainstorm system. Model forecastspredict well the moisture flux with general intensity and distribution similar to the analyses, representingthe southwest and southeast vapor channels. However, the extreme center of moisture flux in the simulationslocates a little to the southwest, consistent with thedisplacement error of the circulation. Accordingly, thedistribution of atmospheric precipitable water(PW)shows a favorable moisture condition, as PW exceeds75 mm by 0600 UTC 21 July(Fig. 8h), also shownin the GFS analysis fleld(Fig. 8d). The forecast PW(Figs. 8a, 8b, 8e-g)is slightly lower than the analyses(Figs. 8c and 8h). Also, compared with the long and narrow vapor channel extending north on the analy-sis flelds, the high value of PW in the simulations isrelatively dispersed with the southward displacement(Figs. 8a, 8b, 8e-g).

Fig. 8. BJ-RUCv2. 0 forecasts of precipitatable water(shade; mm)for(a, b, e-g)different initial times and (c, h)analysis fields valid at(a-c)0300 UTC and (e-h)0600 UTC, and (d)GFS analysis field valid at 0600 UTC 21 July 2012.

Figure options

On the other h and , the convergence area of 850-hPa moisture flux corresponds well with the precipita-tion location. To explore the causes for poor forecastsof rainfall location, the moisture flux convergence isthen diagnosed. Analyses show that water vapor di-vergence turns to convergence over Beijing, with in-creasing convergence intensity and range. By 0900UTC 21 July(Fig. 9f), a wide range of strong con-vergence covers Beijing area, as indicated in the GFSanalysis(flgure omitted)as well. Meanwhile, the 3-km analysis of the BJ-RUCv2. 0 features a southwest-northeast oriented convergence belt, consistent wellwith the surface rain belt. Short-term forecasts showbetter performance(Figs. 9d and 9e), while the fore-casts initialized before 0600 UTC present a dislocated and dispersed convergence area. In particular, the0300 UTC run predicts a strong moisture divergenceover the rainstorm center, which may lead to poor fore-casts with weaker rainfall intensity. Meanwhile, thepositive results of forecasts initialized after 0600 UTCexhibit the beneflcial effect of the assimilation of morerecent observations.

Fig. 9. BJ-RUCv2. 0 forecasts of 850-hPa moisture flux divergence(shade; 10-5 g s kg-1 m-1) and its correspondingmoisture flux(vectors)at 1200 UTC 21 July 2012 for(a-e)different initial times and (f)analysis field.

Figure options

Analysis soundings(flgure omitted)calculatedfrom the BJ-RUCv2. 0 analyses show that before therainfall starts, dry intrusion occurs at mid levels, accompanied by the low-level advection of warm and moist air. By 0300 UTC 21 July, low-level moisturehas increased, leading to a rather low lifting condensa-tion level(LCL) and a surge in convective available po-tential energy(CAPE). Low-level moisture layer overBeijing deepens to almost 10-km thick at 0600 UTC. High value of equivalent potential temperature(θe)also extends to lower atmosphere, indicating a favor-able condition of moist instability over Beijing. As forthe forecasts, the simulated unstable conditions agreewell with the analyses, except for the 0300 UTC run, which shows a small gradient of θe.

In particular, most of the extreme rainfall occursin the warm sector ahead of the cold front. There-fore, the movement of the cold front has a rather weakindication of the precipitation. Previous study intro-duces the concept of "wet baroclinic zone"(usuallydeflned by surface θe), which matches well with theprecipitation area, especially for the warm sector rain-fall(Wang, 2013). Before the rain starts, there exists asouthwest-northeast oriented zone of increasing largeθe gradient, with intensity peaking at 0600 UTC(Fig. 10). The BJ-RUCv2. 0 analyses and the AWS obser-vations(Wang, 2013)are highly consistent, indicatingrather good quality of the mesoscale model analysisagainst the GFS analysis(Fig. 10d). Meanwhile, theCAPE reaches 2000 J kg^-1 at this time, suggestinghigh convective instability. In contrast, the forecastsfail to predict the baroclinic zone. A strong baro-clinic zone extending northward is noticeably evidenton the analysis flelds(Figs. 10c and 10h); however, it is missed by all forecasts. At the same time, theforecast CAPE is much lower than the analyses. Asa result, the favorable instability conditions for heavyrainfall are not fully represented in the model simula-tions. On the other h and , the large θe gradient regionpresented at the 0300 UTC analysis(Fig. 10c)is ap-parently absent in the 3-h forecast(Fig. 10g), indicat-ing again the rapid decrease of the positive effect ofdata assimilation.

Fig. 10. BJ-RUCv2. 0 forecasts of surface equivalent temperature(shade; K)for(a, b, e-g)different initial times and (c, h)analysis fields valid at(a-c)0300 UTC and (e-h)0600 UTC, and (d)GFS analysis field valid at 0600 UTC 21July 2012.

Figure options

To sum up, the forecasts of different initial timesdepict the general synoptic scale circulations well, butwith southward location errors and the absence of ashort wave trough. In terms of physical parameters, moisture peak values are well predicted, reflecting theextremely high water vapor content. However, due tothe errors in the system displacement and the mois-ture convergence, the water vapor has not been cor-rectly concentrated over Beijing area, resulting in themodel's poor performance of precipitation location. Inaddition, despite the positive effect of data assimila-tion on the initial flelds, especially for the times ap-proaching convective development, the beneflcial in-fluence decreases with the forecast lead time. Theshort-wave trough found in the initial analyses gen-erally disappears after 3-h forecasts, affecting subse-quent forecast accuracy. 6. Summary and conclusions

In this work, using a set of observations and modeloutput, the Beijing "7. 21" heavy rainfall case is ana-lyzed, with emphasis on the analysis and veriflcation of3-km BJ-RUCv2. 0 hourly model output. In additionto traditional precipitation veriflcation statistics, theCRA method that is able to separate displacement, volume, and pattern errors is also employed to quanti-tatively verify the QPFs. In addition, possible causesof inaccurate model forecasts are discussed. The pri-mary flndings are as follows:

(1)The heavy rainfall occurred under favorablesynoptic conditions with exceptionally abundant wa-ter vapor. The precipitation gradually intensifled at0600 UTC and reached its peak at about 1200 UTC21 July 2012. The rainfall area was dominated by asouthwest-northeast oriented mesoscale rain belt.

(2)The BJ-RUCv2. 0 forecasts under-predictedthe observed precipitation with a southwest displacement error and an incorrect representation of thequasi-linear feature of the system. However, short-term simulations obtained higher scores due to theassimilation of moisture and hydrometeor informa-tion into the initial conditions. Speciflcally, the modeltends to over-predict light rain(> 5 mm h^-1)whileunderestimate heavy rainfall(> 20 mm h^-1). Furtherquantitative error decomposition shows that majorityof the total error resulted from the displacement and pattern errors, while the mean intensity error was themajor contribution of the error for heavy rainfall(>20 mm h^-1).

(3)The BJ-RUCv2. 0 model forecasts are superiorto the GFS forecasts, and the high resolution(3 km)analyses have good quality and are comparable withthe observations and the GFS analyses. However, the BJ-RUCv2. 0 remains to be further improved foraccurately simulating and predicting mesoscale pro-cesses. The predicted background circulation is quitesimilar to the observation, although with a southwarddisplacement and the absence of a short wave trough. Despite the successful simulation of high water vapor, the poor predictions of moisture convergence and in-stability condition suggest that favorable factors forthe heavy rainfall are not fully represented, leadingto quite large forecast error. In addition, data assim-ilation has a positive in°uence on the initial fleld, inparticular near convection times; however, the posi-tive impact decreases with the model integration. Thissuggests that the initial analysis flelds may not be ina good balance and further improvement to the dataassimilation system, especially in the effective use ofmesoscale data, is needed.

The Beijing heavy rainfall event, as is the casewith many similar cases, involves interactions ofweather systems and features at several scales, withcomplicated processes. Accurate precipitation fore-casting is even more challenging due to the quanti-tative nature and many sources of uncertainties. Toimprove the accuracy of heavy rainfall prediction, abetter and deeper underst and ing of the precipitationmechanism is essential. Also, further researches onthe data assimilation to optimize initial fleld and re-duce model error are also of practical necessity. Thiswork presents a preliminary analysis on the mesoscalemodel performance in predicting the Beijing "7. 21"extreme rainfall event. However, the results of theBJ-RUCv2. 0 system are limited to the speciflc case and more general conclusions on the predictability ofheavy rainfall will require more studies.

Acknowledgments. The observation and model data used in this study were provided bythe Beijing Institute of Urban Meteorology. TheCMORPH merged precipitation data were from theNational Meteorological Information Center of China. Drs. Chen Min, Fan Shuiyong, Gao Hua, and othermembers of the BJ-RUC research group are thankedfor their constructive suggestions and help. ProfessorTao Zuyu and other reviewers have made constructivecomments that greatly improved the quality of thearticle.