Numerical Simulations of Heavy Rainfall over Central Korea on 21 September 2010 Using the WRF Model

The Weather Research and Forecasting (WRF) (Skamarock et al., 2008) model is a mesoscale model designed for both operational applications and atmospheric research. In Korea, the WRF model has been widely used in operational institutions for weather forecasting and academic research regarding high-impact weather simulations. The Korea Meteorological Administration (KMA) has operated the Korea WRF (KWRF) since 2005 (Cho et al., 2005). (Kwun and You, 2009) compared the KWRF and the KMA Regional Data Assimilation and Prediction System (RDAPS) (Lee et al., 2002), the previous operational model, for simulated sea wind features and typhoon tracks. Although both models simulated good results in comparison to the best-track information from the Regional Specialized Meteorological Center (RSMC), the KWRF performed better in predicting typhoon tracks. The Weather Wing of the Korea Air Force (KAF) has operated the weather prediction model system based on the WRF model (i.e., KAF-WRF) since 2007. The initial and boundary conditions of the KAF-WRF are derived from National Centers for Environmental Prediction (NCEP) Global Forecasting System (GFS) (Moorthi et al., 2001) data and the KAF Global and Regional Integrated Modeling system (GRIMs) (Hong et al., 2013), respectively. The KAF-WRF configuration consists of one-way interactive triple-nested domains (18, 6, and 2 km), and each domain covers the East Asian region, the Korean Peninsula, and South Korea. The results of the KAF-WRF are used to support military operations that require weather information (Byun et al., 2011).

Several studies have examined the effects of various factors——grid size, topography, bogus vortex, data assimilation, and physics parameterization——in the WRF model configured in East Asia and centered over Korea. (Cho and Lee, 2006) investigated the effects of grid size on the simulated rainfall distribution, and demonstrated that the WRF model with a horizontal grid size less than or close to 3.3 km should be used for the proper representation of heavy rainfall events associated with multiple convection bands. (Jung et al., 2012) examined the effect of topography on heavy snowfall over Yeongdong Province, the mountainous region over the eastern part of South Korea, and showed the importance of orographic forcing with a high-resolution mountain dataset in the form of dynamical circulations associated with the snowfall. There have been several studies related to the numerical prediction of typhoons in areas of data assimilation (e.g., Kwon et al., 2010; Lee and Choi, 2010). (Lee et al., 2010) demonstrated the performance of the Joint Center for High-impact Weather and Climate Research (JHWC) real-time forecast system, which is based on the WRF model. They showed that the initial data from the 3-Dimensional Variational (3DVAR) data assimilation using the Global Telecommunication System (GTS), Automatic Weather System (AWS), wind profiler, and radar observation data, exhibit improved forecasts in summer rainfall over Korea.

Several studies have investigated the effects of physical parameterizations on the prediction of precipitation. (Lee and Park, 2002) examined the performance of various convective parameterization schemes (CPSs) in the simulation of heavy rainfall over Korea, and found that the dependency on the CPS varies from one case to another. (Hong and Lim, 2006) showed that, for a high-resolution grid of 5 km, the amount of rainfall and its temporal peak intensity increase as the number of types of hydrometeors used in the microphysics (MPS) schemes increases. (Shin and Hong, 2009) showed that the CPS and Planetary Boundary Layer (PBL) processes significantly change the location of accumulated precipitation, whereas the MPS processes influence the rainfall intensity. These Numerical weather Prediction (NWP) studies examined the effects of a specific parameterization scheme (e.g., one of MPS, CPS, PBL, or Land Surface Model) that is commonly configured using control and sensitivity experiments and that is different to the scheme in the control experiment. It is possible to validate the effects of the target physics parameterization scheme using this configuration. However, this kind of configuration is limited in determining the general role of each physical process because of the complicated interaction between the different parts of the physical processes.

For the above reasons, experiments that examine the role of physical processes using various combinations of physical parameterization schemes have been conducted. (Jankov et al., 2007) examined the performance of an NWP model over central California with combinations of two PBL schemes, four MPS schemes, and two initialization methods. They concluded that there is no configuration that confirms the best performance at all times. (Evans et al., 2012) investigated the robustness of physical processes for four storm cases over Southeast Australia using 36 simulation results that employ various combinations of physical parameterization schemes including two PBL schemes, two CPS schemes, three MPS schemes, and three radiation schemes, and confirmed that a specific physics package of the PBL and CPS processes reveals more robust performance than other combinations. A similar conclusion was achieved by (Nasrollahi et al., 2012) in simulating Hurricane Rita (2005) over the Gulf Coast. These results confirm that the configurations for the best performance vary depending on the region and target phenomena. It is also noted that such kinds of experiments using various combinations of physical parameterization schemes have not been performed for high impact weathers over the Korean Peninsula.

The objective of this study is to examine the ability of the WRF model to reproduce the heavy rainfall that occurred over Gyeonggi Province on 21 September 2010, with a local maximum value of 259 mm in Seoul. The operational forecast by the KMA incorrectly predicted the maximum rainfall amount as approximately 10 to 40 mm d^-1 over the heavy rainfall region. The influence of the PBL and MPS parameterizations associated with the precipitation amounts and the intensity are investigated using the simulated results from the combined experiments of different PBL and MPS parameterization schemes. It is important to note that the main objective of our study is not to judge superiority of a specific combination of two physical processes over others. Rather, our focus is placed on the overall impacts of each PBL and MPS process on the heavy rainfall by examining the standard deviations of averaged amounts of precipitation and its local maxima. Also, the characteristics of a specific scheme over another are not presented since the physical reasoning of a particular scheme for the differences in simulation results has already been given in previous studies based on sensitivity experiments (e.g., Bright and Mullen, 2002; Li and Pu, 2008; Shin and Hong, 2009). Section 2 gives an overview of the selected case by analyzing observations, and section 3 describes the experimental setup. The results of the model simulation are discussed in section 4, and concluding remarks are provided in the final section.

On 21 September 2010, heavy rainfall was observed over Gyeonggi Province, including Seoul, the capital city of South Korea. The local maximum of daily accumulated precipitation was 259 mm in Seoul. Figure 1 shows the observed features of the heavy rainfall. The 36-hr accumulated precipitation regions are extended in the east-west direction over the central region of the Korean Peninsula (Fig. 1a). There were peaks in each of the regions (Fig. 1b): 33.7 mm h^-1 at 0700 UTC 21 September in Seoul, and 9.0 mm h^-1 at 1100 UTC 21 September over the central region of the Korean Peninsula.

Figure 1.  (a) 36-h accumulated precipitation (mm) from 1200 UTC 20 to 0000 UTC 22 September 2010 as observed by Korea Meteorological Administration (KMA) Automatic Weather System (AWS) stations; (b) hourly time series of the observation from the KMA AWS in the Seoul region [solid box in (a)] and the central region of the Korean Peninsula [dashed box in (a)].

Figure 2.  (a) Sea level pressure (hPa) and 10-m wind vectors (full barb denotes 10 m s^-1); (b) 850-hPa geopotential height (GPH) (m, solid line) and relative humidity over 80% (shaded); (c) 500-hPa GPH (m, solid line) and temperature (^°C, dashed line); and (d) 300-hPa GPH (m, solid line), temperature (^°C, dashed line), wind vectors, and isotach over 50 m s^-1 (shaded) from the National Centers for Environmental Prediction (NCEP) Final Analysis (FNL) data at 2100 UTC 20 September 2010. The 36-h accumulated precipitation (mm) during the whole event from the Tropical Rainfall Measuring Mission (TRMM) Multi-satellite Precipitation Analysis (TMPA) data is shown in (a).

Figure 2 shows the basic weather charts obtained from the NCEP Final Analysis (FNL) dataset with a resolution of 1^{° × 1°}. At 0000 UTC 21 September, a Siberian high in Mongolia expanded to the northern part of the Korean Peninsula, and a northwestern Pacific high pressure system was located over the southern part of the East China Sea (Fig. 2a). Convergence of the two air masses provided a favorable synoptic environment for the development of a trough in the central region of the peninsula. This trough appeared only in the layer below 700 hPa. Due to the dissipation of typhoon Fanapi (2010), warm and moist air was advected to southern China (Fig. 2b). Associated with heavy rainfall over Korea, this air mass was transported to the Korean Peninsula by southwesterly flows along the boundary of the northwestern Pacific high. On the 500-hPa and 300-hPa isobaric surfaces (Figs. 2c and d), the iso-geopotential line was packed in the north of the Korean Peninsula, and a strong westerly wind developed. A jet streak was located in Manchuria in the northern part of the Korean Peninsula. Figure 2a also shows the 36-h accumulated precipitation from the Tropical Rainfall Measuring Mission (TRMM) Multi-satellite Prediction Analysis (TMPA) with a resolution of 0.25^{°× 0.25°} (Huffman et al., 2007). Significant rainfall appeared in the southwestern boundary of the model domain. Intense rainfall was observed over the central part of the Korean Peninsula. The maximum amount of precipitation over Gyeonggi Province was 239 mm, which was less than that from the AWS data by about 20 mm.

Figure 3.  Surface weather charts and enhanced IR imagery at (a) 2100 UTC 20, (b) 0000 UTC 21 (left panel), 2333 UTC 20 (right panel), (c) 0300 UTC 21, and (d) 0600 UTC 21 (left panel), 0533 UTC 21 (right panel) September 2010.

The evolution and movement of the convective systems are shown in Fig. 3. Around 2100 UTC 20 September, a convective system developed in the eastern part of Shandong peninsula (Fig. 3a). This convective cell moved eastward along the trough (Fig. 2b) and developed over the Yellow Sea. Three hours later, the convective cell was located near the west coast of Gyeonggi Province in the Korean Peninsula (Fig. 3b). The convective system accompanied the meso low system. After 0300 UTC, the convective cell stagnated over Gyeonggi Province (Figs. 3c and d). Moreover, the surface pressure pattern did not change much during that period. At 0600 UTC, the convective cells developed at the location of line formation (Fig. 3d). Satellite infrared images indicated that cloud top heights in the convective system were as high as 13 km. The formation and evolution of the mesoscale convective system were analyzed by (Jung and Lee, 2013).

The numerical experiments in this study were executed using version 3.3 of the WRF model. The model consists of one-way interactive triple-nested domains with a Lambert conformal map projection. A 3-km domain centered on Gyeonggi Province (Domain 3,160× 160) is nested in a 9-km domain (Domain 2, 121×121), which in turn is nested in a 27-km domain (Domain 1, 142× 142) (see Fig. 4a). The entire grid system has 31 vertical layers with a terrain following coordinate, and the model top is located at 50 hPa. The experiments were performed for 36 hours starting at 1200 UTC 20 September 2010. The initial and 6-hourly lateral boundary conditions were forced by the NCEP FNL data.

Figure 4.  (a) Model domain for the 27-km grid experiment (Domain 1), 9-km grid experiment (Domain 2), and 3-km grid experiment (Domain 3) with the terrain heights shaded every 100 m. (b) Geographic map of the Korean Peninsula with the terrain heights at a 9-km resolution. Contour intervals are 200 m, and values greater than 600 m are shaded.

The control (CTL) experiment and 25 subsequent sensitivity experiments were executed in order to examine the predictability of the heavy rainfall and to demonstrate the roles of the PBL and MPS parameterizations. The setup of the CTL experiment was the same as the KAF-WRF model (Byun et al., 2011), including the WRF Single Moment 6-class (WSM6) MPS scheme (Hong and Lim, 2006), the Kain-Fritsch cumulus parameterization (Kain, 2004), a new version of the rapid radiative transfer model (RRTMG) longwave radiation scheme (Iacono et al., 2008), the Goddard shortwave radiation scheme (Tao et al., 2003), the NCEP-Oregon State University-US Air Force-National Weather Service Office of Hydrologic Development (NOAH) land surface model (Chen and Dudhia, 2001), and the Yonsei University (YSU) PBL scheme (Hong et al., 2006) with the newly implemented treatment of the stable boundary layer (Hong, 2010). It should be noted that the cumulus parameterization was not used in the 3-km grid domain because of the assumption that the 3-km grid is sufficient to explicitly resolve the convective rainfall.

All of the sensitivity experiments used the same physical algorithms, except for the different PBL and MPS schemes. The 25 experiments employed five different PBL and five different MPS schemes. The PBL schemes were: YSU, Mellor-Yamada-Janjic (MYJ) (Janji\'c, 1994), Quasi-Normal Scale Elimination (QNSE) (Sukoriansky et al., 2005), Bougeault and Lacarrere (BouLac) (Bougeault and Lacarr\`ere, 1989), and University of Washington (UW) (Bretherton and Park, 2009). The MPS schemes were: WSM6, Goddard (Tao et al., 1989), Thompson (Thompson et al., 2008), Milbrandt 2-moments (Milbrandt and Yau, 2005), and Morrison 2-moments (Morrison et al., 2009). The five different bulk MPS schemes used in this study can be categorized into single-moment schemes and double-moment schemes. The single-moment schemes——WSM6 and Goddard——predict the mixing ratio of the hydrometeors. Meanwhile, the double-moment schemes——Thompson, Milbrandt 2-moments, and Morrison 2-moments——predict not only the mixing ratio, but also the number concentration of the hydrometeors. In addition, the PBL schemes can also be divided into nonlocal first-order closure schemes (i.e., YSU) and turbulent kinetic energy (TKE) closure schemes——MYJ, QNSE, BouLac, and UW. The vertical diffusion effect in the TKE scheme can be determined using the local diffusivity of the variable. Meanwhile, both the local diffusivity and the non-local mixing by large convective eddies are required in the nonlocal scheme.

The KMA operates a dense surface AWS observational network in South Korea. Observations began in 1988 with only 15 stations, and by 2013 there were 670 stations. The raw data are collected at 1-min intervals, and the dataset used in this study was archived at 1-h intervals. To evaluate the horizontal distribution of the amount of precipitation, the horizontally inhomogeneous point values of the AWS data were interpolated onto a 0.1^° grid resolution using the Cressman's objective analysis scheme (Cressman, 1959). The values were averaged with weights proportional to the inverse of the squared distance between the center of the grid box and the stations within two radii. One may argue that the skill scores of the simulated precipitation need to be computed on the station grid. We confirmed that the density of the AWS stations is quite homogeneous, with an average distance between them of 13 km, but especially dense over the heavy rainfall region, which did not affect the computed skill scores over the interpolated grid at 0.1^°.

For quantitative verification of each precipitation sensitivity experiment, the bias, root-mean-square-error (RMSE) and pattern correlation coefficient (PC) between the KMA AWS data and each sensitivity experiment were calculated at each AWS grid point. For a robust evaluation of intensity, the experimental data were interpolated into a 0.1^° grid resolution by using a 4-point weighting average, with the assumption that AWS data are the best estimates. The bias is defined as

\begin{eqnarray} {\rm Bias}=\dfrac{1}{N}\sum_{n=1}^N(a_{{\rm A},n}-a_{{\rm O},n}) ,\\[-1mm]\nonumber \end{eqnarray}

where the summations are performed over all grid points for the analysis zone, N is the total number of grid points included in the region, and the superscripts A and O refer to analysis and observations, respectively. The RMSE was also calculated as

\begin{equation} {\rm RMSE}=\left[\dfrac{1}{N}\sum_{n=1}^N(a_{{\rm A},n}-a_{{\rm O},n})^2\right]^{1/2} . \end{equation} The normalized RMSE (NRMSE) is obtained by dividing the RMSE into the observed mean value. The similarity of the horizontal distribution can be objectively measured by the PC, which is expressed as \begin{equation} {\rm PC}=\dfrac{\sum_{n=1}^N(a_{{\rm O},n}-\bar{a}_{\rm O})(a_{{\rm A},n}-\bar{a}_{\rm A})} {\left[\sum_{n=1}^N(a_{{\rm O},n}-\bar{a}_{\rm O})^2\sum_{n=1}^N(a_{{\rm A},n}-\bar{a}_{\rm A})^2\right]^{1/2}} , \end{equation} where the bars denote spatial averaging within a region.

The CTL experiment simulated the large-scale patterns well, including the Siberian and North Pacific highs and a low pressure system centered over the Sea of Okhotsk (cf., Figs. 2a and 5a). Compared with the TMPA data in Fig. 2a, the simulated precipitation from the CTL experiment was similar to what was observed in terms of its distribution, but underestimated the local maxima. Over the central Korean Peninsula, the 36-h accumulated precipitation from the CTL run was only 50 mm, which was much less than the 200 mm from the AWS observations. Hereafter, discussion is focused on the results of the 3-km experiments, since the results of the coarse resolution domains were inadequate to resolve mesoscale features. The CTL experiment properly captured the horizontal distribution of the accumulated rainfall. However, the maximum rainfall amount was approximately 100 mm less than that from observations (cf., Figs. 1a and 5a). The maximum rainfall location was displaced eastward by approximately 30 km compared to the observations. The patterns of accumulated precipitation were separated into Seoul and northern Gyeonggi Province, which is north of Seoul.

Figure 5b shows the time series of the hourly precipitation over the central region of the Korean Peninsula (dashed box in Fig. 5a) and the maximum precipitation region (solid box in Fig. 5a). In the precipitation time series data that were averaged over the central Korean Peninsula region, the rainfall peak appeared 5 hours earlier than the observed peak. The simulated peak of rainfall (9 mm h^-1) was similar to the observation in terms of amount. For the maximum precipitation region, two peaks were produced, unlike the one peak in the observations. The first and second peaks appeared at 0200 UTC and 0500 UTC 21 September, respectively, with the second peak occurring 2 hours earlier than that of the observations. The spurious precipitation was simulated along the Taebaek Mountains, which are located on the eastern flank of the Korean Peninsula, running from north to south.

Table 1 shows the bias, RMSE and PC scores for the results from all of the experiments. The objective skill scores for the accumulated precipitation were calculated against the KMA AWS observation data. A more traditional measure of precipitation, such as the threat score, could be used. It is, however, hard to demonstrate the performance of all 25 experiments with the threat score, but the overall evaluation of skills with the scores in Table 1 was found to comply with the skill based on the threat score (not shown). Interestingly, particular combinations of the PBL and MPS processes showed better performance than other combinations. The best combination in terms of the RMSE and PC scores was observed in the MYJ-Milbrandt experiment. The RMSE and PC scores of the MYJ-Milbrandt experiment were 32.33 and 0.75, respectively. The second and third best scores appeared in the QNSE-Goddard and YSU-Morrison combinations, respectively. The three best combinations did not share the same PBL or MPS scheme, although the YSU PBL showed overall superior performance over other PBL schemes for each MPS scheme.

Figure 5.  36-h accumulated precipitation (mm) from 1200 UTC 20 to 0000 UTC 22 September 2010 from (a) Domain 1, (b) Domain 3, and (c) hourly precipitation time series from the CTL experiment in the maximum precipitation region [solid box in (b)] and central region of the Korean Peninsula [dashed box in (b)], and the AWS observations (gray line). Sea level pressure (hPa) and 10-m wind vectors (full barb denotes 10 m s^-1) at 2100 UTC 20 September 2010 are shown in (a).

To verify the variation of the physical schemes in the skill of precipitation prediction, standard deviations of each skill score were calculated. If a scheme had relatively low standard deviation, this meant it had stable performance when combined with other physics schemes. The average value of the experiments with the YSU PBL scheme for PBL processes showed the lowest RMSE and second lowest standard deviation. The MPS process that demonstrated relatively good performance when combined with each of the PBL schemes was the Goddard scheme. The experiment with the Goddard MPS scheme combined with the QNSE or UW PBL scheme showed the best RMSE score among the five experiments that employed the same PBL scheme.

Figure 6 presents the 36-h accumulated precipitation and its patterns using various combinations of the PBL and MPS schemes. Each row represents the results for a specific MPS scheme and the different columns show the results with the different PBL schemes. Some of the combinations (e.g., MYJ-Milbrandt and QNSE-Goddard) captured well the distribution of the precipitation band and the maximum precipitation. The results with the Goddard scheme showed a higher local maximum than those from other experiments for the same PBL scheme, except the BouLac-Goddard combinations. The simulations employing the Thompson scheme exhibited a smaller spatial extent of precipitation. The experiments with the YSU PBL scheme, except the YSU-Goddard experiment, showed the isolated precipitation over Baengyeong Island (37.95^°N, 124.67^°E; see Fig. 4b), and along the Taebaeck Mountains, as greater than 90%. The experiments employing the UW scheme tended to exhibit a single precipitation band with an east-west direction (see Fig. 4b).

Hereafter, the term "MPS ensemble" denotes an average of the results using the various MPS schemes for a specific PBL scheme, and the term "PBL ensemble" is the average of the results using various PBL schemes for a specific MPS scheme. By comparing the MPS and PBL ensembles and all of the ensemble members, the effects of the physical processes can be confirmed.

In order to determine the distribution of the atmospheric state for each ensemble member, time series of the standard deviations for various variables were calculated. The standard deviations for the sea level pressure and rainfall over Domain 1 showed similar accuracy between the PBL and MPS ensembles (not shown). However, the standard deviations of the 3-dimensional variables appeared differently between the two ensembles (Fig. 7). At 850 hPa, the standard deviation of the PBL ensemble members had a larger value than that of the MPS ensemble members. The differences between the standard deviations of the MPS ensemble and the PBL ensemble were particularly great with regard to the geopotential height and wind vector. These differences increased over the forecast time. With an increase in altitude, the standard deviation of the specific humidity gradually decreased. This finding indicates that the atmospheric states of the PBL ensemble members were influenced more than those of the MPS ensemble members, especially at lower altitudes.

Figure 8 presents the hourly precipitation proportion averaged over the left half of the dashed box marked in Fig. 1a (Figs. 8a and b), and the right half of the same box (Figs. 8c and d). The area of the left half represents the rain band inflow, and the right half the dissipating area affected by the orography of the Taebaek Mountains (see Fig. 4b). Figure 8 indicates that the precipitation rates varied significantly with different combinations of model physics. The time series of the precipitation proportion from the MPS ensemble members showed similar patterns, except for the case of the YSU scheme——shape, hourly proportion, start, peak, and ending time (Fig. 8a). The standard deviation ranged below 0.5, except for the YSU result (0.636). When the results from the 25 experiments were sorted by the employed MPS processes, the time series showed higher variability in general. Figure 8b reveals that the standard deviations of the Thompson, Milbrandt and Morrison schemes were greater than 0.5. This shows that the variability was relatively large when the double-moment MPS schemes (i.e., Thompson, Milbrandt, and Morrison) were employed. In the dissipating region, the range of standard deviation for the PBL ensemble (0.430-0.633 mm) was similar to that for the MPS ensemble (0.360-0.694 mm). The experiments with relatively higher standard deviations over the inflow area (i.e., MYJ in the MPS ensemble results; Thompson, Milbrandt, and Morrison in the PBL ensemble results) showed reduced values over the mountains compared to those over the inflow area, except for the YSU experiment in the MPS ensemble results. Meanwhile, the experiments with lower standard deviations over the plain area (i.e., QNSE, BouLac, and UW in the MPS ensemble results; WSM6 and Goddard in the PBL ensemble results) showed increased values over the mountains. Overall, the PBL effect on the temporal distribution of the precipitation proportion was more pronounced over the precipitation inflow area than the dissipating region.

The occurrence time of maximum precipitation is shown in the bottom-right corner of each panel in Fig. 9. It is shown that the maximum precipitation occurrence was influenced more by the PBL processes than by the MPS processes.

Figure 6.  36-h accumulated precipitation (mm) from 1200 UTC 20 to 0000 UTC 22 September 2010 for each sensitivity experiment. Contours refer to the 90th percentile of precipitation over 1 mm. The 90th-percentile precipitation value for each experiment is shown in mm in the bottom-right corner of each panel.

Figure 7.  Time evolution of the standard deviation for each MPS (black lines) and PBL (gray lines) ensemble member: (a) geopotential height (m), (b) temperature (K), (c) specific humidity (g kg^-1), and (d) vector wind (m s^-1) for the 850-, 700-, 500-, and 250-hPa pressure levels.

Figure 8.  Time series of the simulated precipitation proportion averaged over the (a, b) left half of the dashed box marked in Fig. 1a (37^°-38^°N, 126^°-127.5^°E) and (c, d) right half (37^°-38^°N, 127.5^°-129^°E), and the AWS observations (gray dotted line). The 25 experiments are sorted by the selected physical processes: (a, c) PBL; (b, d) MPS.

Figure 9.  Vertical distributions of water species, averaged for 2 h at the valid time, which is the maximum precipitation over the heavy rainfall region marked as a dashed box in Fig. 1a. The valid time for the maximum of 1-h accumulated precipitation is shown in the bottom-right corner of each panel.

Figure 10.  Areal extent of the simulated 36-h accumulated precipitation over the area marked as a dashed box in Fig. 1a for the PBL ensemble members with the (a) WSM6, (b) Goddard, (c) Thompson, (d) Milbrandt 2-moments, and (e) Morrison 2-moments MPS schemes.

The maximum values of the 36-h accumulated precipitation in each experiment are shown in Table 2. Most experiments simulated the amount of precipitation to be lower than observed (278.60 mm). The simulation with the UW-Thompson combination revealed the smallest value (112.91 mm), whereas the QNSE-Goddard combination produced the largest value (374.46 mm). The values of the PBL ensemble ranged from 129.44 mm for the Thompson scheme to 214.40 mm for the Goddard scheme, and the results from the MPS ensemble had a similar range, from 144.81 for the UW to 209.15 for the QNSE scheme. The standard deviation of each ensemble demonstrated the greatest value in ensemble members including the QNSE-Goddard experiment: 86.99 mm for the PBL ensemble and 98.49 mm for the MPS ensemble. The PBL ensemble for the Thompson MPS scheme had the lowest standard deviation (17.19 mm) and the second lowest standard deviation appeared in the MPS ensemble for the UW PBL (28.04 mm).

Unlike the maximum precipitation, which was sensitive to both the PBL and the MPS processes, the area-averaged precipitation was influenced more by the MPS than by the PBL processes (Table 3). Compared to the AWS observation (88.70 mm), all the experiments underestimated the amount of precipitation. The experimental results of the PBL ensemble ranged from 47.51 mm for the Thompson scheme to 65.64 mm for the Goddard scheme. This range was greater than that of the MPS ensemble, which ranged from 52.84 mm for the UW scheme to 63.60 mm for the MYJ scheme. The standard deviations of the PBL and MPS ensembles differed significantly. The standard deviation of the PBL ensemble for the Thomson scheme was only 1.93 mm. The largest standard deviation value of the PBL ensemble (6.11 mm) was relatively small, as compared to the MPS ensemble, ranging from 6.03 to 9.54 mm. Each member of the PBL ensemble simulated similar total rainfall amounts. It is thus concluded that the maximum rainfall amount should be equally sensitive to the MPS and PBL processes, while the average amount is more sensitive to the MPS than the PBL processes.

The areal extent of precipitation was calculated as the product of the number of grid points that exceeded a specified rainfall threshold and grid size. Figure 10 indicates the areal extent computed at three thresholds (40, 80, and 120 mm). All of the experiments underestimated the areal extent for the three thresholds. The MYJ-Goddard combination showed the least error in the 40-mm threshold. In the 80- and 120-mm thresholds, the MYJ-Milbrandt combination showed the least error. The precipitation area was similarly simulated when the same MPS was employed. The Goddard scheme showed a better average performance in the 40-mm threshold, and the Milbrandt scheme showed better performance in the 80- and 120-mm thresholds. Overall, the Thomson scheme underestimated the area for all of the thresholds.

The reason for these characteristics——average precipitation, and precipitation areal extent——which depend on the MPS scheme employed, comes from the difference in the simulated water species in each experiment. In order to confirm the simulated water species characteristics, the vertical profiles of the area's averaged condensates were obtained from each experiment and the results are plotted in Fig. 9. The experiments that employed the same MPS scheme produced similar profiles for water species, especially for the ice phases such as ice, snow and graupel. The Thompson experiments, which simulated weak precipitation rate, showed a smaller maximum rainwater mixing ratio (average: 0.25 g kg^-1), and the Goddard experiments, which simulated a relatively strong precipitation rate, showed larger values (average: 0.46 g kg^-1) than the other MPS experiment results (between 0.32 and 0.39 g kg^-1). The difference in the rainwater mixing ratio between the sensitivity experiments, the reason for the difference in the rainfall amount, was influenced by the ice phase mixing ratio. There were few ice and graupel particles above the 4 km height in the Thompson scheme results; however, the Goddard and Milbrandt schemes, which simulated a large amount of precipitation, both showed higher maximum mixing ratios of graupel (between 0.50 and 0.87 g kg^-1) at a height of 6 km. The Goddard scheme showed a maximum snow mixing ratio at 9 km, and the Milbrandt scheme a maximum ice mixing ratio at 12 km.

In this study, we examined the capability of the WRF model to simulate the heavy rainfall event over Gyeonggi Province on 21 September 2010. The local maximum of observed precipitation was approximately 259 mm d^-1. A triple-nested WRF model with a highest resolution of 3 km was forced by 6-hourly NCEP FNL data for 36 hours from 1200 UTC 20 to 0000 UTC 22 September 2010. A total of 25 experiments with five different PBL schemes——YSU, MYJ, QNSE, BouLac, and UW——and five different MPS schemes——WSM6, Goddard, Thompson, Milbrandt 2-moments, and Morrison 2-moments——were conducted in order to investigate the role of the PBL and MPS processes in simulating the heavy rainfall.

The WRF model simulated the spatial distribution of the 36-h accumulated precipitation reasonably well, showing it extending from the west coast of the Korean Peninsula eastward, although its maximum was underestimated. Sensitivity experiments revealed that a particular combination of the PBL and MPS processes resulted in relatively good performance: the MYJ-Milbrandt, QNSE-Goddard, and YSU-Morrison schemes for the PBL and MPS processes, respectively. This conclusion reflects previous findings that demonstrated good skill in reproducing high-impact weathers by a particular combination of physics packages, with no specific scheme outperforming in all aspects, although the YSU scheme for the PBL and Goddard scheme for the MPS processes revealed relatively good skill in the forecast of the selected heavy rainfall.

A major focus of this study was the relative importance of the PBL and MPS processes in simulating the distribution of accumulated precipitation, its local maximum, and temporal evolution of the precipitation over the heavy rainfall region. These issues were addressed by examination of the ensemble spread of skill scores from the experiments with particular PBL or MPS schemes. It was found that the PBL process affected the onset of the rainfall in the heavy rainfall region, and its dissipation. Also, the areal distribution of the precipitation over the heavy precipitation region was influenced more by the PBL processes than by the MPS processes. This relative importance of the PBL processes was found to be due to the modulated atmospheric structure in the lower troposphere. (Shin and Hong, 2011) revealed that the stability at lower levels is significantly affected by PBL processes, which may affect the temporal evolution of simulated precipitation and its areal distribution. Meanwhile, the amount of precipitation was influenced more by the MPS than by the PBL processes. Whereas the maximum precipitation was sensitive to both the PBL and MPS processes, the amount of area-averaged precipitation was influenced more by the MPS than by the PBL processes.

It is important to note that the conclusion of this study is limited, mainly due to the fact that we have examined a single case study only. The relative importance of PBL and MPS processes may depend upon the characteristics of the selected heavy rainfall case. The selected convective system was thermodynamically organized by local instability, and further studies with different mechanisms, such as synoptically organized systems, is needed. Despite this uncertainty, our findings contribute to identifying the key algorithm for improving the forecast skill of heavy rainfall over East Asia.

Acknowledgements. This work was carried out through an R&D project on the development of global numerical weather prediction systems at the Korea Institute of Atmospheric Prediction Systems (KIAPS) and under Grant CATER 2012-3035 funded by the Korea Meteorological Administration (KMA).