Effect of Horizontal Resolution on a Meso‐β‐Scale Vortex Simulation in an Extreme Rainstorm on 22 May 2020 Over South China: A Contrastive Study Based on Different‐Resolution Ensembles

A transient extreme rainstorm that occurred over Guangzhou city, China, on 22 May 2020, has been investigated based on two ensemble prediction systems (EPSs), one with 9‐km (TRAMS9km‐EPS) and the other with 3‐km (TRAMS3km‐EPS) grid spacings, respectively. The results show that the better performance in the rainstorm event of TRAMS3km‐EPS than that of TRAMS9km‐EPS has been attributed to the reasonable simulation of an eastward propagating meso‐β‐scale convective vortex (MβCV). The composite evolution of the good‐performing members for the two EPSs has verified the important role of the MβCV in modulating the structure of low‐level jet (LLJ) to converge toward the Guangzhou city, facilitating the formation of the rainstorm. In comparison to the gravity wave disturbance when LLJ passed through the mountains in TRAMS3km‐EPS, the topographic effect induced the moist air to ascend deeply in TRAMS9km‐EPS, which directly led to earlier outbreaks of convection and rainstorm processes and affected the structural intensity of the upstream MβCV. Quantitative PV diagnosis has demonstrated that the dispersedly strong high‐PV systems in TRAMS3km‐EPS were caused by positive feedback effect between meso‐scale convective systems and diabatic heating, and were favorable for the entire cyclonic development of MβCV. The unsymmetrical PV tendency around the moving MβCV under the combined effect of diabatic heating and advection process in TRAMS9km‐EPS resulted in a rapid propagation of this system. However, the symmetrically developed PV system within the MβCV in TRAMS3km‐EPS allowed the enhanced cyclonic circulation to persistently affect Guangzhou city, inducing extreme rainstorms similar to the observations.

10.1029/2023EA002921 2 of 22 Generally, a transient rainstorm over the coastal area of South China has greater intensity and longer duration in comparison to other regions in China according to the historical extreme value of rainfall with different durations (Luo et al., 2016;Zheng et al., 2016).One of the most remarkable characteristics is that heavy rain or rainstorms over South China are always accompanied by organized mesoscale convective systems (MCSs) (Ding & Liu, 2001;Luo et al., 2013).In fact, directly relevant to the formation of MCSs, convection initiation over South China depends on multiscale atmospheric processes (Trier, 2003), such as low-level jets (LLJs, Du & Chen, 2019;Gebauer, 2017;Moncrieff & Liu, 1999), topographic and land surface effects (Anthes et al., 1982;Lanicci et al., 1987), frontal boundaries (Koch & Kocin, 1991;Trier et al., 1991), gust fronts related to cold pools (Schumacher & Johnson, 2005;Wilson & Schreiber, 1986), and gravity waves (Richiardone & Manfrin, 2003;Trexler & Koch, 2000).For instance, in late August 2018, an ultralong-duration, linearly shaped MCS caused an extreme rainstorm in Guangdong Province under multiple atmospheric factors.The 24-hourly rainfall intensity of this extreme rainstorm reached 1,056.7 mm, breaking the provincial historical record (Zeng et al., 2020).As suggested by Zeng et al. (2020), such a well-developed MCS was triggered under a suitable environment field of southerly flow, which was sustained for a long period by the reverse forces between the monsoon depression, local terrain, and barometric gradient, directly leading to the disastrous rainstorm event.These results indicate that the MCS activities caused by external forcing factors are an important type of event that generates extreme rainstorms over Guangdong Province.
Furthermore, mesoscale convective vortexes (MCVs) are another typical weather system inducing MCSs over South China (e.g., Huang et al., 2010;Liu et al., 2007) and were also found to directly contribute to the Guangzhou 5.22 rainstorm (Xiao et al., 2021).According to Raymond and Jiang (1990), anomalous diabatic heating in the vertical direction associated with MCSs fundamentally stimulates the formation of mesoscale cyclonic circulation in the middle layer; that is, the MCV is the product of the diabatic heating effect of smaller-scale MCS activities.This view has also been supported by most studies (e.g., Chen & Frank, 1993;Conzemius & Montgomery, 2010;Fritsch et al., 1994).In turn, the mature MCV can generate newly organized MCSs through its own dynamic structure and favorable environmental field, further strengthening sustainable development and evolution of MCVs (Fritsch et al., 1994;Trier & Davis, 2002).Nevertheless, most of the studies related to such development mechanism of MCV are qualitative research, which requires more in-depth exploration of the 10.1029/2023EA002921 3 of 22 mutual conversion between the dynamic and thermodynamic processes, as well as model verification with higher resolution data.
Given that the common convective rainstorms over South China caused by MCSs and MCVs usually occur in fragmented pieces with a smaller spatial scale (meso -β to -γ scale) than general frontal rainfall (Xia et al., 2006;Xia & Zhao, 2009), the development of numerical weather prediction (NWP) models is necessary for operational prediction and scientific research.Remarkably, compared with deterministic forecasts, ensemble prediction contains certain disturbance deviations and reflects a variety of possible changes in future weather forecasts, thus possessing prominent advantages in dealing with the uncertainty of NWP models and improving the predictability of extreme events (Du & Chen, 2018;Medina et al., 2019;Zhang, 2018aZhang, , 2018b)).In 2006, a new tropical regional atmosphere model for the South China Sea (CMA-TRAMS) was established by the China Meteorological Administration (CMA), and has been significantly developed in recent years (Xu et al., 2022;Zhong et al., 2020; see Section 2 for a detailed introduction).Based on CMA-TRAMS, a mesoscale ensemble prediction system (EPS) with a horizontal resolution of 9 km, named TRAMS-EPS, was developed by Zhang (2018a) and was proven to have more advantages in forecasting the intensity of tropical cyclones and the related heavy rainfall (Zhong et al., 2021), as well as the strong wind over the European Center for Medium-Range Weather Forecasts (ECMWF) global EPS.To address the limited coverage of the model domain, TRAMS-EPS was updated in 2020 with a new version of CMA-TRAMS covering a relatively large domain and has been preoperational since April 2020 (Shi et al., 2023).
By using the 9-km grid-spacing TRAMS-EPS (TRAMS9km-EPS), Xiao et al. (2021) found that the Guangzhou 5.22 rainstorm process was triggered by an eastward propagating MCV and its interaction with LLJs, with the GOOD members of TRAMS9km-EPS reproducing this rainstorm event (this is also discussed in Section 3).However, there are obvious deficiencies in simulating or forecasting the propagation law and dynamic structure of such mesoscale systems for TRAMS9km-EPS, leading to an underestimate of the intensity of the Guangzhou 5.22 rainstorm (Xiao et al., 2021).In fact, prediction models with increasing resolution have been shown to be beneficial for quantitative rainfall forecasting, especially for predicting transient convective processes in terms of mode, intensity, and diurnal cycle (e.g., Clark et al., 2007;Done et al., 2004;Weisman et al., 2008).Therefore, the objective of this study is to reveal the effect of the horizontal resolution on the simulation of a meso-β-scale convective vortex (MβCV) in triggering the record-breaking rainstorm in Guangzhou city by comparing a higher-resolution (i.e., 3 km) version of TRAMS-EPS (TRAMS3km-EPS) with TRAMS9km-EPS.The dynamic structure and life cycle of the MβCV from a Lagrangian perspective are compared and analyzed between the two EPSs through PV diagnoses.
This study is organized as follows.Section 2 introduces the data and methods.Section 3 describes the basic features of the extreme rainstorm event and compares the simulation differences between TRAMS9km-EPS and TRAMS3km-EPS.In Section 4, the differences in MβCV in triggering the Guangzhou 5.22 rainstorm between the two EPSs are further compared and analyzed, and the effect of horizontal resolution on the MβCV simulation is especially examined in terms of multiple aspects.Finally, a summary and discussion are given in Section 5.

Data
The hourly gridded rainfall data of the China Meteorological Precipitation Analysis (CMPA) V1.0 product used in this study are provided by the National Meteorology Information Center of China Meteorological Administration (http://data.cma.cn) as the benchmark to evaluate the model-based rainfall products.The CMPA data with a high spatial resolution (0.1° × 0.1°) are generated by integrating the observations from automatic meteorological stations in China and the satellite rainfall product from the Climate Prediction Center MORPHing technique of the National Oceanic and Atmospheric Administration (Shen et al., 2010).
For comparison with the atmospheric products of the prediction model in revealing the physical mechanism for the Guangzhou 5.22 rainstorm, hourly atmospheric circulation data are extracted from fifth generation ECMWF reanalysis data (ERA5) (Hersbach et al., 2020), with a horizontal resolution of 0.25° × 0.25°.

Model Introduction
CMA-TRAMS is currently in operation to provide weather prediction services for forecasters.Based on the Global/Regional Assimilation and Prediction System (GRAPES; Chen et al., 2008) framework, CMA-TRAMS is a nonhydrostatic regional model that adopts a semi-implicit, semi-Lagrangian scheme for integration over time and employs a horizontal grid designed on a longitude-latitude mesh with Arakawa C-grid staggering.The vertical coordinate of CMA-TRAMS is terrain-following based on Charney-Philips vertical layer skipping (Charney & Phillips, 1953) with 65 vertical layers.The physics parametrization schemes used in CMA-TRAMS include the Weather Research and Forecasting (WRF) Single-Moment 6-class (WSM6) microphysics scheme (Hong & Lim, 2006), the simplified Arakawa-Schubert (SAS) cumulus parametrization (CU) scheme (Han & Pan, 2011;Pan & Wu, 1995), and the medium-range forecast model (MRF) planetary boundary layer (PBL) scheme (Hong & Pan, 1996) (see Chen et al. (2020) for a detailed introduction of the physics parametrization schemes of CMA-TRAMS).The operational version of CMA-TRAMS (TRAMS9km, hereafter) includes 1,001 × 601 horizontal grid points with a horizontal resolution of 0.09° × 0.09°, covering a large area of 0.8°-54.8°N,70°-160°E (Domain A in Figure 2).
For the deterministic forecasts of TRAMS9km, the ECMWF analyses and forecast data with a horizontal resolution of 0.125° × 0.125° are used as the initial conditions (ICs) and lateral boundary conditions (LBCs), respectively.Observed information on cloud water and rainwater from radar and satellites is analyzed based on the cloud analysis system from the Advanced Regional Prediction System (ARPS) Data Analysis System (ADAS) and then assimilated based on the nudging technique.TRAMS9km-EPS is based on TRAMS9km and contains 30 perturbed members, whose ICs and LBCs are provided by downscaling the first 30 perturbed members of the ECMWF ensemble model (Buizza, 2014), and the model physics are unperturbed.The other sets in terms of the physics parametrization schemes and the assimilation process for TRAMS9km-EPS are the same as the deterministic forecasts of TRAMS9km.
Recently, to further improve the ability to predict small-scale convective rainfall processes, a higher-resolution version of TRAMS (TRAMS3km hereafter) and the corresponding EPS (TRAMS3km-EPS) have been developed.In comparison with TRAMS9km, TRAMS3km contains 1,131 × 917 horizontal grid points with horizontal resolutions of 0.03° × 0.03°, covering an area of 16°-43.48°N,93.04°-126.94°E(Domain B in Figure 2), with the same physics parametrization schemes, ICs, LBCs, and assimilation process as TRAMS9km.TRAMS3km-EPS is based on TRAMS3KM and contains 30 members, which are perturbed in the same way as TRAMS9km-EPS.In the present study, both EPSs start at 12:00 UTC on 20 May 2020 and are integrated for 48 hr, during which the forecast period of 12-36 hr covering the Guangzhou 5.22 rainstorm process is the focus.

Ensemble-Based Sensitivity Analysis
The method of ensemble-based sensitivity analysis (ESA) combines linear statistics with ensemble prediction of nonlinear models and estimates the sensitivity of prediction to the characteristics of early atmospheric development and evolution by applying linear statistics to the prediction trajectory of a group of nonlinear models (Hakim & Torn, 2008).It can reveal the dynamic mechanism of error growth that cannot be explained by simple linear statistics.
The forecast sensitivity (S) is expressed as follows: where J is the forecast variable, X i is the state variable at time i, cov is the covariance operator, and var is the variance operator.Equation 1 indicates that forecast sensitivity (S) is related to the correlation coefficient between the state variable (X i ) and the forecast variables (J).

Subgroup Ensemble Analysis
The subgroup ensemble analysis is conducted by grouping the members of the EPS according to different standards and then synthesizing and analyzing the environmental parameters corresponding to the members of each subgroup and comparing their differences (Lamberson et al., 2016;Wu et al., 2020).According to the prediction accuracy of the rainfall intensity and location, the respective 30 members in TRAMS9km-EPS and TRAMS3km-EPS are ranked in the present study in such a manner that three groups are selected, GOOD members (five members), POOR members (five members) and OTHERS (20 members) (this is discussed further in Section 3).For better comparison, the ensemble means of the five GOOD and five POOR representations in the two EPS models were the main focus to identify the effect of horizontal resolution on the simulation of the Guangzhou 5.22 rainstorm as well as the key physical processes involved in triggering this extreme weather event.

PV Equation
The PV equation in the isobaric coordinate system can be expressed as follows (Ertel, 1942;Hoskins, 1997;Hoskins et al., 1985;Zhang et al., 2021): where P is the Ertel PV, which is the dot product of the absolute vorticity vector for unit mass and potential temperature gradient, i.e., P = αξ a • ∇θ (in which α is the specific volume, ξ a is the three-dimensional absolute vorticity, θ is the potential temperature, and ∇ is the three-dimensional gradient operator in pressure coordinates), u and v are the zonal and meridional wind components, respectively, ω is the vertical velocity, ζ is the vertical relative vorticity, g is the gravitational acceleration,  θ is the diabatic heating rate, and F is frictional acceleration in the momentum equation.The left-hand side of Equation 2 refers to the local rate of change in PV (or the PV tendency), while the five underlined forcing terms (indicated by letters F 1 -F 5 ) on the right-hand side refer to the zonal, meridional, and vertical PV advection, and the PV generation by horizontally and vertically nonuniform diabatic heating, respectively.Note that the PV dissipation associated with frictional force [the sixth term on the right-hand side of Equation 2] is not analyzed here because its effect is relatively small in the free atmosphere for analyzing the evolution of the MCV.The quantitative diagnoses associated with PV using Equation 2 will be given in Section 5.

Case Overview and Model Verification
As introduced in Section 1, central Guangdong Province suffered an extreme rainstorm from 21 to 22 May 2020, with the most intense rainfall occurring in Guangzhou city over a 12-hr period.Thus, Figure 1 first illustrates the spatial distribution of the 12-hourly accumulated rainfall and the associated surface wind based on the CMPA and ERA-5 data from 12:00 UTC on 21 May to 00:00 UTC on 22 May to provide an overview of these unusual weather events.Note in Figure 1 that apparent convergent flows consisting of northerlies from higher latitudes in conjunction with southwesterlies from the South China Sea were present in central Guangdong Province, forming an asymmetric cyclonic circulation with scales of 200-400 km.To the south of such cyclonic flows, there was strong vertical motion distributed along the southwesterlies with magnitudes reaching −1 Pa s −1 , which resulted in extremely heavy rainfall of greater than 150 mm over the estuary area east of Guangzhou city.Generally, according to previous studies (Davis & Galarneau, 2009;Davis & Trier, 2007), this rain-producing convective mesoscale system accompanied by closed cyclonic circulations in the lower troposphere could be identified as a typical system of the meso-β-scale convective vortex (MβCV), similar to that in Zhang et al. (2023).Note also that the intensity of the extreme rainfall around Guangzhou city was nearly 10 times stronger than that over the nearest area, suggesting that the formative mechanism for such locally extreme rainfall processes in Guangzhou city may involve mesoscale convective activities and their interactions with synoptic-scale circulations.To better summarize the evolution characteristics of this regional small-scale rainfall event, a target domain (22.8°-24°N, 112.8°-114.3°E,as shown in Figure 1) with the highest rainfall intensity and distinct vertical motion is selected as a key area to further compare and analyze the simulation effects of models with different resolutions.
To first evaluate the performance of the model simulation, Figure 3 shows the 12-hr accumulated rainfall averaged from the respective 30 members of TRAMS9km-EPS and TRAMS3km-EPS for the period of 12:00 UTC 21-00:00 UTC 22 May during the Guangzhou 5.22 rainstorm event.Compared with the rainfall observations shown in Figure 1, both EPSs could reproduce the large-scale rainfall distribution over the central and eastern Guangdong Province.However, the intensities of the rainstorm over Guangzhou key area are notably underestimated with a composite rainfall amount of 75-125 mm in TRAMS3km-EPS and only 25-50 mm in TRAMS9km-EPS (Figures 3a and 3b).Although such an underestimate of rainfall amount in comparison with the actual observations was attributed to the direct ensemble average for the different members to some extent, the performance of TRAMS9km-EPS was evidently weaker than that of TRAMS3km-EPS for this event.Note also that the standard deviation of the 12-hr rainfall over the key area and its eastern region in TRAMS3km-EPS reached 40 mm, which was almost two times larger than that in TRAMS9km-EPS (Figures 5a and 5b).The high-value standard deviation with the evidence of stronger ensemble mean rainfall intensity shown in Figure 3b indicates that certain members in TRAMS3km-EPS could actually reflect the extreme nature of this event.
The performance improvement of TRAMS-EPS can be further manifested according to the temporal evolution of the related area-averaged hourly rainfall over the Guangzhou key area between the two EPSs.Note in Figures 3c  and 3d that almost all the members of the two EPSs could precisely simulate the first rainfall process during 12:00 UTC-18:00 UTC 20 May with an intensity of 4-8 mm compared with the observations.However, for the Guangzhou 5.22 rainstorm process, the comprehensive rainfall intensity and the temporal evolution in TRAMS3km-EPS were more accurate, with some members reproducing the peak magnitude of the rainstorm equal to 12 mm well (Figure 3d).The two concentrated peak rainfall periods in this EPS were also similar to the observations (Figure 3d).In contrast, the peak rainfall periods for most members in TRAMS9km-EPS were significantly earlier by approximately 20 hr (Figure 3c), accompanied by weaker rainfall intensity.This finding also explains the much weaker rainfall intensity during 12:00 UTC 21-00:00 UTC 22 May, as shown in Figure 3a.
Given that the rainfall occurrence is strongly dependent on the convergence of warm and moist air, as well as dynamic ascending motion, vertically integrated moisture flux divergence based on the ERA-5 reanalysis data and the ensemble mean model data is shown in Figure 4a to clarify the reasons for the performance differences between two EPSs.The associated pressuretime cross sections of the horizontal moisture flux and vertical velocity for the two EPSs are also shown (Figures 4b and 4c).Note in Figure 4a that the vertically integrated moisture flux convergence of TRAMS9km-EPS first occurred after 00:00 UTC on 21 May, accompanied by significant vertical motion establishment from 850 hPa to nearly 150 hPa (Figure 4b).This strong convective motion process corresponded well with the earlier occurrence of rainstorms for TRAMS9km-EPS over the Guangzhou key area (Figure 3c).The vertically integrated moisture flux convergence of TRAMS3km-EPS, however, was similar to that of the ERA 5-based fields in terms of intensity and time evolution (Figure 4a), thus performing better for the rainstorm simulation than that in TRAMS9km-EPS.Note that the strong vertical motion in TRAMS9km-EPS at approximately 00:00 UTC on 21 May occurred under a strong boundary moisture flux with a maximum magnitude of 2.4 × 10 −2 kg m −1 Pa −1 s −1 (Figure 4b).In contrast, the boundary moisture flux was relatively weak in TRAMS3km-EPS after the deep ascending motion was established at 12:00 UTC on 21 May (Figure 4c).These contrastive phenomena actually reflect the differences in the effects of topography and MβCV on boundary moisture flux between the two EPSs (as discussed further in Section 4).The above results of the ensemble mean indicate that moisture flux and dynamic effect jointly led to the performance differences between the two EPSs.Since considerable standard deviations of the members existed for both EPSs, there were significant differences between individual members, thus requiring subsequent categorization analysis.

Comparative Subgroup Analysis
Equitable threat scores (ETSs) represent the prediction technique of a rainfall forecast result that meets a certain rainfall threshold within the forecast area compared to a random forecast that meets the same rainfall threshold.ETSs are generally used to measure the effectiveness of convective ensemble forecasting.To compare the simulations between different members of the two EPSs, the ETSs of the 12-hr accumulated rainfall over Guangdong Province for the respective 30 members in TRAMS9km-EPS and TRAMS3km-EPS are shown in Figure 5. Since the actual Guangzhou 5.22 rainstorm mainly occurred during the period of 12:00 UTC 21 to 00:00 UTC 22 May, this period was still chosen as the key rainstorm period to calculate the score of the different submembers of the two EPSs for criterion unification.For the rainfall processes with different thresholds, the average ETS for TRAMS9km-EPS reached 0.055 (light rain, >10 mm), 0.087 (moderate rain, >25 mm), and 0.086 (heavy rain, >50 mm) (Figure 5a), which was much lower than that for TRAMS3km-EPS with average ETSs of 0.233, 0.177, and 0.203, respectively (Figure 5b).This finding demonstrates that the comprehensive performance of the Guangzhou 5.22 rainstorm in TRAMS3km-EPS is better than that in TRAMS9km-EPS.However, note that there were obvious differences in ETSs among different members in the two EPSs (Figures 5a and 5b).To capture the common characteristics of specific groups and then better compare the differences between the two EPSs, five GOOD members for each EPS were selected based on the five highest comprehensive ETSs of three rainfall thresholds (namely, light rain, moderate rain, and heavy rain).In this manner, five POOR members were similarly distinguished due to the lowest five ETSs, and the remaining members were classified into the OTHER group (20 members for each EPS).
As discussed in Section 3.1, the Guangzhou 5.22 rainstorm actually occurred after 12:00 UTC on 21 May; thus, Figure 6 depicts the distribution of the 12-hr accumulated rainfall from 12:00 UTC on 21 May to 00:00 UTC 10.1029/2023EA002921 9 of 22 on 22 May, as well as the 850 hPa wind superimposed by the specific humidity at 12:00 UTC on 21 May for the GOOD, OTHER, and POOR members of the two EPSs to illustrate the circulation factors responsible for the extreme rainstorm.Note that for the GOOD members in TRAMS9km-EPS, a locally semiclosed cyclonic system, namely, the MβCV, northwest of the Guangzhou key area was embedded in a much larger-scale cyclonic circulation pattern over South China at 12:00 UTC on 21 May (Figure 6a).To the south, consistent southwesterlies were present along the coastline, with strong water vapor greater than 16 g kg −1 converging toward the Guangzhou key area to support the subsequent 12-hr rainfall equal to 50-75 mm over the key area and its eastern region (Figure 6a).Similar circulations and moisture transport also appeared in the OTHER group and POOR group of TRAMS9km-EPS (Figures 6c and 6e).However, the MβCVs in the above two groups were much weaker during this period, especially for the POOR members, in which there was only cyclonic wind shear instead of a closed MβCV north of the Guangzhou key area (Figures 6c and 6e).As a result, weakened cyclonic circulation associated with MβCV will inevitably lead to weakening of the subsequent local convergence process over the key area, thus producing an accumulated rainfall amount of only 25-50 mm for the GOOD members and less than 25 mm for the POOR members in TRAMS9km-EPS (Figures 6c and 6e).
In contrast, closed cyclonic MβCVs were recognizable upstream of the key area for all three groups in TRAMS3km-EPS at 12:00 UTC on 21 May (Figures 6b-6f), accompanied by strengthened LLJs and high-value water vapor transport on the southeast side of the MβCV.The only differences were the intensity and direction of the LLJ across the Guangzhou key area between the GOOD members and the other two groups (Figures 6b-6f), with many more converging and stronger LLJs in the GOOD members resulting in a reinforced rainfall intensity greater than 150 mm and a significant rainstorm center over the Guangzhou key area (Figure 6b).Although the rainfall intensities in the OTHER and POOR members were relatively weaker and their rainstorm centers were also slightly out of position (Figures 6d and 6f), the two groups still performed better than all the members in TRAMS9KM-EPS.
To further compare the differences in the convective evolution in relation to the MβCVs, Figure 7 depicts the time-longitude cross sections of the composite combined radar reflectivity along the Guangzhou key area for the GOOD and POOR groups of the two EPSs.Since radar-based convective initiation (CI) identification is usually declared to be the first occurrence of radar reflectivity that is at least 35 or 40 dBZ (Bai et al., 2020), in this study, we define the first occurrence of a convective cell (>35 dBZ) as a CI event to discriminate between convective and stratiform rain.Note in Figures 7a and 7b that the composite combined radar reflectivities show similar evolutions from 12:00 UTC 20 to 12:00 UTC 21 May between the POOR and GOOD members of TRAMS9KM-EPS, with reflectivities greater than 15 dBZ propagating persistently eastward to cover the range of the Guangzhou key area.As discussed above, such strong radar reflectivities during this period reflected an earlier occurrence of the rainfall process in comparison to observations.This phenomenon was much more pronounced in the POOR members of TRAMS9KM-EPS when CI events occurred over the western and central Guangzhou key area after 00:00 UTC on 21 May (Figure 7b).In addition, there was no recognizable radar reflectivity during the Guangzhou 5.22 rainstorm period for this group, which was remarkably different from the other three groups.In contrast, radar reflectivities mainly appeared after 06:00 UTC on 21 May for the GOOD and POOR members of TRAMS3KM-EPS (Figures 7c and 7d), with the organized reflectivities developing eastward and keeping almost stagnant for 12 hr within the Guangzhou key area.Note also that the intensity of convective activities as well as the radar reflectivities in POOR members of TRAMS3KM-EPS was apparently weaker than that in GOOD members.This finding was probably related to the structure of MβCV and large scale environmental fields.The ensemble-based sensitivity analysis (EAS) was further conducted to distinguish the differences among ensemble members for certain models (i.e., TRAMS9km-EPS and TRAMS3km-EPS).Note in Figure 8a that the high intensity of the rainstorm process over the Guangzhou key area in TRAMS9km-EPS was markedly related to southerly wind to its southwestern direction based on the ESA.These correlated southerly wind vectors directly extended to the key area and then immediately turned left to result in a cyclonic pattern over the northwest Guangzhou key area.Such a correlation field actually reflected the intensity differences in cyclonic circulation development on the east side of the MβCV among different members.This feature can be further manifested by the correlation between the low-level geopotential height field and the rainfall process (Figure 8a), with apparent negative correlations greater than −0.5 concentrated on the west side of the key area, corresponding exactly to the position of the MβCV center.This result further implies that the simulation performance of the Guangzhou 5.22 rainstorm largely depends on the MβCV in TRAMS9km-EPS.
For TRAMS3km-EPS, apparent positive correlations between the previous low-level southerly wind and the Guangzhou 5.22 rainstorm show that the development of the rainstorm was closely related to the correlated wind vectors with southwesterlies and southeasterlies converging toward the Guangzhou key area (Figure 8b).Such a correlation pattern of the wind field demonstrates that LLJ and the associated convergent flows over the key area were important factors in triggering the Guangzhou 5.22 rainstorm among different members in TRAMS3km-EPS.Therefore, negative correlations of the 700-hPa geopotential height appeared not only over the position of the MβCV but also over the key area (Figure 8b), corresponding well to the cyclonic converging southerly wind.
The above discussions suggest that the performances of TRAMS9km-EPS among different members were significantly sensitive to the preliminary circulation structure in relation to the MβCV.Furthermore, there was weak or even no convective motion over the key area during the Guangzhou 5.22 rainstorm for the POOR members of TRAMS9km-EPS and TRAMS3km-EPS (cf., Figure 7).To unify the number of good-performing members between the two EPSs and minimize other forcing factors, only GOOD members in the 9 and 3 km versions of the EPSs were selected in the following study to explore the effect of the horizontal resolution on the simulation of MβCV, and distinguish the different dynamic and thermodynamic mechanisms of the MβCV evolution in the two EPSs with different resolutions.

Coupling Effect With LLJ
The above analyses indicate that the propagation and variation in MβCV and LLJ in the lower troposphere had a great impact on the Guangzhou 5.22 rainstorm.In combination with the method for defining the surface cyclone (Wernli & Schwierz, 2006) and Tibetan Plateau Vortex (Zhang et al., 2021), the present study defines the MβCV center as a local maximum relative vorticity in cyclonic flow of the 850-hPa circulation within closed or semiclosed contours of the of the 850-hPa geopotential height fields.The lower-level circulation in association with the LLJ as well as the centers of MβCV for the ensemble mean of GOOD members in the two EPSs are shown in Figure 9.
Note in Figures 9a1 and 9a2 that the original MβCVs for the two EPSs were both generated north of Guangxi Province at 04:00 UTC on 21 May, with significantly closed cyclonic circulations around them.Due to the significant convective development of the MβCVs and converging effect by the cyclonic circulation, moderate rainfall greater than 4 mm hr −1 was generated over the northern and eastern sides of the MβCVs.In addition, large-range southwesterlies prevailed in southern Guangdong Province and the coastal lines, and then immediately veered left downstream to generate apparent rainfall locally (Figures 9a1 and 9a2).Notably, there was a regional LLJ center of at least 16 m s −1 on the southwestern side of the key area in TRAMS9km-EPS, and distinct heavy rainfall was distributed to the left front side of the LLJ center.This finding also corresponded to the high radar reflectivity as shown in Figure 7b, which was apparently different from that in TRAMS3km-EPS.
The MβCVs in the two EPSs persistently propagated southeastward to app roach the Guangzhou key area by 12:00 UTC on 21 May (Figures 9b1-9c1 and 9b2-9c2).Although the intensity of the background LLJ along the coastlines in TRAMS9km-EPS was relatively stronger than that in TRAMS3km-EPS, the LLJ in the two EPSs prominently accelerated the southwesterlies on the southeast side of the MβCVs to induce stronger rainfall (Figures 9b1-9c1 and 9b2-9c2).In turn, due to the reinforced horizontal pressure gradient force caused by the low-pressure system, the moving MβCV in each of the EPS significantly modulated the structure of the LLJ to form a new center south of the key area (Figures 9c1 and 9c2).The horizontal wind speed gradient around the LLJ center led to more reinforced convergence over the local area, resulting in the occurrence of the concentrated and extreme rainstorm.This finding reflects the coupling effect between the LLJ and the moving MβCVs.Therefore, within the next 8 hr, as the MβCVs propagated further downstream, stronger rainfall with a magnitude greater than 12 mm hr −1 mainly occurred in the area near the LLJ centers (Figures 9d1-9e1 and 9d2-9e2).
However, note that the centers of the MβCV and the resultant rainstorm in TRAMS9km-EPS rapidly moved southeastward to the coastal zone during the period from 12:00 UTC to 20:00 UTC on 21 May.The cyclonic structure of the MβCV also gradually incorporated into a southwest-northeast extended wind shear line.This result was obviously different from the evolution of the MβCV in TRAMS3km-EPS, in which the MβCV was maintained in a quasisymmetric cyclonic structure with little movement, resulting in continuous accumulation of rainfall in the Guangzhou key area.Consequently, the abnormal reinforced LLJ over key area and the rapidly MβCV evolution in TRAMS9km-EPS should be further studied separately.

Topographic Effect
As discussed above, to explain the development of the convective motion and LLJ responsible for the earlier occurrence of rainstorms over the key area in TRAMS9km-EPS, the difference field of the lower level moisture flux at 04:00 UTC 21 May between TRAMS9km-EPS and TRAMS3km-EPS is depicted in Figure 10.In comparison with TRAMS3km-EPS, TRAMS9km-EPS exhibited stronger low-level moisture flux over South China Sea, as well as the west side of the key area with an anomalous magnitude equal to 9 × 10 −3 kg m −1 Pa −1 s −1 (Figure 10).Referring to the contrastive circulation distribution shown in Figures 9a1 and 9a2, this phenomenon was obviously related to the regional enhancement of LLJ.In addition, there were discrete distributions of negative moisture flux anomalies north of 24°N with anomalous southward moisture flux vectors aligning along the mountain regions.This result denotes that the convergence of the moisture flux on the west side of the key area in TRAMS9km-EPS was much more prominent, which was more conducive to the development of local convection.Therefore, to further explore the relationship between the differences in moisture flux and the development of convection, a meridional cross section along the western side of the key area is provided in Figure 11.11a and 11b that reinforced southerlies prevailed in the middle and lower troposphere south of the mountain regions at 04:00 UTC 21 May for both two EPSs.However, these southerlies in TRAMS9km-EPS were significantly uplifted in front of the mountains (Figure 11a), with a relatively large-scale dipole pattern of the low-level divergence field.In turn, the strong moisture flux was further accelerated at the entrance area of the  lifting motion due to the vertical pumping effect and resulted in a high-value center below 800 hPa near the surface.The distinct convergence field as well as the high moisture flux in lower level resulted in a background convective instability with the equivalent potential temperature decreasing with height (figure not shown), and was obviously facilitated the triggering of CI.In contrast, there were recognizable, small-scale disturbances of vertical motion in TRAMS3km-EPS when the low-level southerlies passed over the mountain terrain (Figure 11b).Note also that a series of small-ranging convergence and divergence centers appeared as a result of the turbulent vertical motion (Figure 11b), exhibiting a typical gravity wave disturbance phenomenon.The above indicates that there are certain differences in the response and feedback mechanisms of wind fields to topography in different resolution EPSs.In the lower resolution EPS, the uplift effect of the topography on the LLJ appeared to be amplified, resulting in an abnormal large-scale convergence process and accompanying convective motion in comparison to that in the higher resolution EPS.The differences in such topographic effect directly lead to the earlier outbreak of convection and rainstorm processes in TRAMS9km-EPS by 04:00 UTC on 21 May (Figures 4a and 4b).At the same time, the compensatory low-level convergence development resulting from the CI will also affect the circulation structure of the upstream MβCV to a certain extent, as discussed in the following sections.

Structure and Propagation of MβCV
Since the movement and structure of the MβCV directly affected the subsequent rainstorm process over the Guangzhou key area, Figure 12 first illustrates the track paths of the MβCVs for the composite of good members in TRAMS9km-EPS and TRAMS3km-EPS, with the time evolution showing their main meteorological element variations around the MβCV centers.As shown in Figure 12a, the MβCVs in the two EPSs both propagated southeastward at an approximate speed from the northern Guangxi Province in early stage before 1200 UTC on 21 May.However, the initial structure of MβCV in relation to the low-level convergence and ascending motion was apparently weaker in TRAMS9km-EPS than in TRAMS3km-EPS (Figure 12b).As discussed in the previous sections, this is because the strong convective motion over the key area at 04:00 UTC 21 May caused the southwesterlies on the eastern side of the MβCV to compensate for the downstream convergence in TRAMS9km-EPS (cf., Figure 9a1).Therefore, at this time, the low-level moisture transported by the southwesterlies as well as the surface temperature influenced by latent heat release were also lower in comparison to that in TRAMS3km-EPS (Figure 12b).
As the MβCVs propagated and evolved southeastward, the structural differences in the MβCVs between the two EPSs gradually decreased (Figure 12b), and the thermodynamic processes reflected by the evolution of specific humidity and surface temperature in TRAMS9km-EPS even became stronger after 10:00 UTC on 21 May.Nevertheless, when the coupling processes occurred between the eastward-moving MβCVs and the LLJ after 12:00 UTC on 21 May (cf., Figures 9c1 and 9c2), the MβCV in TRAMS9km-EPS rapidly propagated across the key area in the next 8 hr (Figure 12a).The MβCV in TRAMS3km-EPS, however, moved steadily and slowly to persistently influence the key area during this time.In addition, the differences in the thermodynamic and dynamic structures of the MβCVs between the two EPSs also prominently increased again (Figure 12b).
To further specify how and why the differences in the propagation of the MβCVs occurred between the two EPSs, the low-level PV tendency and its forcing terms were recalculated quantitatively in the four directions of the moving MβCV center from a Lagrangian perspective based on Equation 2, as shown in Figure 13.Since PV is a synthetic meteorological quantity that can represent both dynamic and thermodynamic properties of atmospheric systems, the propagation variation in MβCV can be obviously reflected by the evolution of PV to a certain extent.Additionally, due to the small contribution of vertical PV advection forcing in the lower troposphere, only the horizontal PV advection [(F 1 ＋ F 2 ) in Equation 2], horizontal (F 4 ) and vertical diabatic heating (F 5 ) terms are diagnosed in the following sections.Note in Figure 13 that the PV tendencies in the four directions of the MβCV center exhibited similar evolution before 12:00 UTC on 21 May for both EPSs, with distinct negative PV tendencies in the northwest direction and positive PV tendencies in the southeast direction.This finding manifested the immediate increase of high PV at the southeast side of the MβCV center, verifying the southeastward propagation for the early stage of the MβCVs life cycle (cf., Figure 12a).However, after 12:00 UTC on 21 May, the negative PV tendency in the northwest direction in TRAMS3km-EPS began to weaken immediately (Figure 13a1) and even turned to positive, which was mainly caused by the reinforcement of the horizontal and vertical diabatic heating terms (Figures 13c1 and 13d1).The opposite situation occurred in the southeast direction, in which the suddenly weakened positive PV tendency (Figure 13a4) was attributed to the attenuation of diabatic heating terms after 12:00 UTC on 21 May (Figures 13c4 and 13d4).On the other hand, the horizontal PV advection term in this direction for TRAMS9km-EPS turned rapidly from negative to positive during this time (Figure 13b4).Accordingly, the PV tendency in TRAMS9km-EPS maintained strong positive anomalies in the southeast direction of the MβCV (Figure 13a4), facilitating for its persistently southeastward movement.Actually, such different PV evolutions in the two EPSs were due to the position of high PV increase and its relative movement in relation to the MβCV center.Form a Lagrangian perspective, the main generation source for absolute PV tendency (dP/ dt) in atmosphere is diabatic heating based on Equation 2. Note in Figure 13 that the positive PV generation in TRAMS9km-EPS was mainly located to the south side of the MβCV center, and was significantly stronger than that in TRAMS3km-EPS after 12:00 UTC (Figures 13c3 and 13c4).Such newly generated high PV was immediate transported downstream by the strong LLJ (cf., Figures 9c1-9e1), reflecting as the increase of positive PV advection at the southeast side of MβCV center.This allowed the quickly evolution of PV in the southeast direction of MβCV in TRAMS9km-EPS (Figure 13a2) and the rapid MβCV movement.
However, the vertical diabatic heating term in northeast direction in TRAMS3km-EPS increased dramatically after 1200 UTC (Figure 13c2).Although such PV development was largely offset by the advection term (Figure 13b2), the positive PV tendency in TRAMS3km-EPS was significantly stronger than that in TRAMS9km-EPS (Figure 13a2).Therefore, the PV evolution around the center of MβCV in TRAMS3km-EPS exhibited strong positive tendency in the northeast and southwest directions (Figures 13a2 and 13a3), with weak variation in the northwest and southeast directions (Figures 13a1 and 13a4).These results indicate that the MβCV in TRAMS3km-EPS tended to stretch and strengthen on its northeast and southwest side, while the displacement of its center was relatively small.In contrast, the remarkable differences in PV tendency between north and south side of the MβCV in TRAMS9km-EPS lead to southward evolution of PV immediately and rapid movement of the MβCV.Such a phenomenon is similarly discussed by Zheng et al. (2013), who found that the positive PV tendency caused by the diabatic latent heating ahead of a vortex not only intensifies the local vertical vorticity but also affects the migrating direction of the vortex.However, PV advection also played a important role in such MβCV movement as discussed above.
After 16:00 UTC on 21 May, the MβCV in TRAMS3km-EPS mainly propagated eastward (Figure 12a) due to the increase of the negative PV tendency on its western side (Figures 13a1 and 13a3) and positive PV tendency on the eastern side (Figures 13a2 and 13a4).This is because that the relatively weak LLJ due to its cyclonic variation (as discussed further in the following section) resulted in a weak negative PV advection in southeast direction in TRAMS3km-EPS (Figure 13b4), leading to the increase of net PV tendency in the southeast direction after 1600 UTC (Figure 13a4).The MβCV in TRAMS9km-EPS still kept southeastward movement (Figure 12a).These facts verified that the joint effect of diabatic heating and advection process on the unsymmetrical PV tendency around MβCV determines the movement and evolution law of MβCV.

Small-Scale Activities in Relation to MCSs
According to the above discussion, the diabatic heating process acted as an important generating term for the PV systems, and the subsequent advection of these manufactured PV exerted great importance in the movement and structural evolution of the MβCV.To specify the differences in the location and intensity of diabatic heating occurrence between the two EPSs, Figure 14 depicts the respective composite distributions of the 850 hPa PV and associated circulations around the MβCV centers.Note that in the early stage of MβCVs development for the two EPSs, there were similar spiral high-value PV structures, which were accompanied by strong convective activities over the northeastern sections of the MβCVs (Figures 14a1 and 14a2).However, with the evolution of the MβCVs, several distinct isolated PV centers with extremely high magnitudes greater than 4 PVU began to appear around the MβCV center and over the downstream areas in TRAMS3km-EPS (Figure 14b2).Such high-value PV centers basically corresponded to strong radar reflectivity, demonstrating the existence of regionally smaller MCSs development with stronger small-scale cyclonic circulations in higher resolution EPS.In comparison to TRAMS3km-EPS, the high-value PV in TRAMS9km-EPS exhibited a continuously wide-ranging distribution with a magnitude of only 1-3 PVU (Figure 14b1).In addition, the southerlies over the southeast side of the MβCV in lower resolution EPS were weaker, leading to an unsymmetrical cyclonic MβCV with eastward stretched high-PV.Such different MβCV structures between the two EPSs became more apparent by 12:00 UTC on 21 May when the MβCV in TRAMS9km-EPS already turned into a zonally high-PV band within a larger-scale wind shear (Figure 14c1).However, the MβCV in TRAMS3km-EPS was still maintained in a spiral cyclonic structure with more small-scale strong PV centers generated along the high-PV spiral arms (Figure 14c2).
To further examine the impact of the small-scale high-value PV and the MCSs on the development of the cyclonic circulations, especially on the southerlies east of the MβCV in higher resolution EPS, two vertical cross sections at 12:00 UTC on 21 May along the downstream area of the MβCVs were cropped.The PV diagnoses in relation to PV advection and diabatic heating process are shown in Figure 15 to confirm the evolution mechanism of the PV system.For each of the two EPSs (Figures 15a and 15b), the PV tendency indicated by shading was consistent with the relative vorticity variation denoted by contours, suggesting that the PV tendency influenced by diabatic heating can reflect the dynamic characteristics of the MβCVs.Note that the PV tendency in TRAMS9km-EPS exhibited a single dipole pattern in lower level caused mainly by the horizontal PV advection (Figure 15c).In contrast, the distinct positive and negative PV tendency anomalies alternating over the downstream area of the MβCV in TRAMS3km-EPS (Figure 15b) denotes that these PV anomalies in higher resolution EPS were mainly manufactured by distinct diabatic heating processes and transported by horizontal advection (Figure 15d).
Notably, the contribution of the vertical diabatic heating processes in relation to the latent heat release was basically positive though it was offset by the contribution of the horizontal diabatic heating process to a certain extent (Figure 15d).Form a Lagrangian perspective, the reinforced positive PV anomalies caused by vertical diabatic heating immediately induced the generation of small-scale cyclonic disturbances to influence the regional convergent flow (Figure 15b).In turn, the anomalous convergent flow further enhanced the individual MCS under favorable moist environment, releasing more latent heat due to the reinforced rainstorm.As a result, the MCS and diabatic heating formed a positive feedback process so that the number and intensity of small-scale high-value PV systems in TRAMS3km-EPS gradually increased (Figures 14a2-14c2).On the other hand, the existence of the more reinforced MCSs and high-PV systems resulted in a stronger cyclonic structure of the MβCV in TRAMS3km-EPS (cf., Figure 12b).This was obviously favorable for the north-south exchange of moisture and anomalous PV transport north of the strong LLJ, triggering more MCSs in higher latitude.In this manner, the PV system in TRAMS3km-EPS was thus developed more symmetrically, especially over the downstream of MβCV, affecting the subsequent development and movement of the MβCV.On the contrary, weak MCSs occurrence significantly weakened the cyclonic development of MβCV in TRAMS9km-EPS.Thus, convergence and the moisture flux transported by the LLJ were mainly concentrated at the south side of the MβCV, resulting in the rapid southeastward movement and weakening of the MβCV under the combined effect of diabatic heating and advection process, as discussed in Section 4.3.
The effect of the small-scale disturbances in TRAMS3km-EPS can be further confirmed by the time averaged eddy flux quantities.Generally, a time averaged eddy flux can be obtained from the time mean of the product of In such manner, a comparative distribution of eddy flux quantities in relation to the 850 hPa circulations for the two EPSs from 10:00 UTC to 14:00 UTC on 21 May are showed in Figure 16.Note that during the period when MβCVs coupled with the LLJ in front of the key area, wide-ranging high-value eddy flux quantities in TRAMS9km-EPS extended from the center of the MβCV to the downstream with distinct mean cyclonic wind shear (Figure 16a).This phenomenon indicates that the larger-scale PV advection as well as the weak diabatic heating resulted in the large-ranged variation of the downstream circulation.This process was not favorable for the stable development of MβCV.In contrast, small-scale, concentrated high-value eddy flux quantities were distributed within just the circular sectors of the MβCV in TRAMS3km-EPS (Figure 16b).The development of the southerlies as well as the dense eddy flux centers within the MβCV actually reflected the reinforcement of the cyclonic circulation in relation to the kinetic energy, supporting the steady cyclonic structure of the MβCV, which was obviously affected by the MCSs.

Summary and Discussion
This study investigated the effect of horizontal resolution on a meso-β-scale vortex (MβCV) simulation involved in an unprecedented transient rainstorm process over Guangzhou city during the summer of 2020 by comparing the results of two EPSs with different resolutions based on the CMA-TRAMS model (i.e., TRAMS9km-EPS and TRAMS3km-EPS).We examined the sensitivity of the convective rainfall simulation to uncertainties in the structure of the MβCV and LLJ in each EPS, with special attention being given to the different evolution and developed mechanisms of the MβCV between the two EPSs.The physical processes that affect the performance of the MβCV simulation were separately analyzed based on different-resolution EPSs.The major findings are summarized as follows: The results show that TRAMS3km-EPS can generally capture the extreme rainstorm process over the Guangzhou key area, including the location and duration of the rainfall process similar to the observations.However, the rainstorm burst in most members of TRAMS9km-EPS was approximately 20 hr earlier than the actual rainfall, and the rainfall intensities were also weaker.Contrastive analysis with the ERA-5 reanalysis data shows that the different performances between the two EPSs depend largely on the simulation of the vertically integrated moisture flux divergence field in terms of the intensity and time evolution.The performance differences among the ensemble members of the two EPSs were compared by subgroup ensemble analysis, and three ensemble subgroups (GOOD, POOR, and OTHER) were classified based on the ETSs of the members in each EPS.A robust feature of the good-performing members that reasonably simulated the rainstorm is the presence of quasiclosed cyclonic circulations in relation to the MβCV upstream of the Guangzhou key area before the rainstorm.Such a characteristic was confirmed by the ESA, which further reveals that the simulation performance of the Guangzhou 5.22 rainstorm among different members in TRAMS9km-EPS was more sensitive to the circulation structure of the MβCV as well as the LLJ, demonstrating the important role of the evolution of MβCV among different members as well as between the two EPSs with different resolutions.GOOD members were examined to clarify the differences in the developed structure and propagation of MβCVs for the two EPSs.The composite evolution and track paths of the MβCVs in the two EPSs show that both MβCVs could trigger the rainstorm process downstream by modulating the structure of the LLJ to converge toward the key area.Prior to the rainstorm, however, the dynamic uplift effect induced distinct vertical motion when the LLJ passed through the mountains in TRAMS9km-EPS, while the topographic effect in TRAMS3km-EPS tended to act as a gravity wave disturbance.The differences in such topographic effect directly lead to the early outbreak of convection and rainstorm processes over the Guangzhou key area in TRAMS9km-EPS, and they also affected the upstream MβCV intensity due to the compensatory low-level convergence development.
The MβCV propagation in the two EPSs were analyzed and quantitatively diagnosed based on the PV equation by decomposing the total PV tendency into four direction parts around the moving MβCV.The results show that the rapid propagation of MβCV in TRAMS9km-EPS was attributed to the unsymmetrical high-PV tendency around the MβCV, which was caused concurrently by the diabatic heating and PV advection process.However, the regionally dispersed, high-value PV systems accompanied by the development of MCSs were important in the good performance of MβCV simulation in higher resolution EPS.The positive feedback effect between the MCSs and diabatic heating reinforced the entire cyclonic structure of the MβCV.Such enhanced structure of MβCV was favorable for the north-south exchange and transmission of moisture and energy north of the LLJ, leading to a symmetrically developed PV system within the MβCV.Thus, the symmetrically developed PV system allowed the enhanced cyclonic circulation to persistently affect the rainstorm over the key area with little movement of the MβCV center, inducing extreme rainstorms over Guangzhou city similar to the observations.Notably, although the direct factor causing the different simulation results of the extreme rainstorm event between the two EPSs is attributed to the eastward-moving MβCV in this study, other synoptic systems, such as the LLJ (Xiao et al., 2021), as well as the large-scale systems in the mid-and upper troposphere where the convergent/ divergent field was apparently amplified (cf., Figure 11), may also influence the rainstorm simulation.In addition, the correspondence between the evolution of high-PV and CI is only qualitatively analyzed in this study.Since there were mutual influences and feedback effects between the dynamic and thermodynamic processes, it is still unclear how and to what extent the dynamic or the thermodynamic processes affect the simulation of MβCV and CI in different resolution models, respectively.Therefore, some EPS experiments based on the dynamic and thermodynamic disturbances, especially near the underlying surface topography, will also be conducted separately in the future work.This research was sponsored by the National Key R&D Program of China (Grant 2021YFC3000902), the National Natural Science Foundation of China (Grant 42205065 and 41975136), and the Guangzhou Municipal Science and Technology Planning Project of China (Grant 202103000030).

Figure 1 .
Figure 1.Mean distribution of the 10 m wind (vectors, m s −1 ), 12-hr accumulated rainfall (color shading with respect to the bottom-right color bar, mm), and 500 hPa vertical velocity (dotted shading with respect to the bottom-left color bar, Pa s −1 ) from 12:00 UTC 21 to 00:00 UTC 22 May based on CMPA and ERA-5 reanalysis data.The thick purple curves represent the provincial boundaries.The red triangle indicates the position of Guangzhou city and the red rectangle denotes the Guangzhou key area (22.8°-24°N, 112.8°-114.3°E,which are the same below in Figures 3,6,8,10,12, and 16).

Figure 3 .
Figure 3. (a) Twelve-hour accumulated rainfall averaged from the 30 members of TRAMS9km-EPS for the period of 1200 UTC 21-00:00 UTC 22 May (color shading, mm) and the corresponding standard deviation of the accumulated rainfall for the 30 members (purple contours, mm).The gray shading shows a terrain altitude greater than 500 m.(c) Time series of the observed hourly rainfall (bars, mm) and the hourly rainfall for the 30 members from TRAMS9km-EPS (red curves, mm) over the Guangzhou key area from 12:00 UTC 20 to 12:00 UTC 22 May.The thick red curve denotes the ensemble mean hourly rainfall, and the red shading shows the standard deviation of the 30 members.(b) and (d) are the same as (a) and (c), respectively, except for TRAMS3km-EPS.

Figure 5 .
Figure 5. (a) Equitable threat scores (ETSs) of TRAMS9km-EPS for the 12-hr accumulated rainfall from 12:00 UTC 21 to 00:00 UTC 22 May over Guangdong Province (bars, with pale yellow, green, and blue bars denoting the ETSs of rainfall greater than 10 mm, 25 and 50 mm, respectively).The yellow, green, and red dashed lines show the mean score of the 30 members for rainfall greater than 10 mm, 25 and 50 mm, respectively.The solid and hollow stars represent the five selected GOOD members and the five POOR members, respectively.(b) Same as (a) except for the ETS scores of TRAMS3km-EPS.

Figure 6 .
Figure 6.(a) Composite distribution of the 12-hr accumulated rainfall (color shading, mm) from 12:00 UTC 21 to 00:00 UTC 22 May and the 850 hPa wind (vectors, m s −1 ) superimposed by the specific humidity (shading superimposed on the vectors, g kg −1 ) at 12:00 UTC 21 May for the GOOD members of TRAMS9km-EPS.(b) Same as (a) except for TRAMS3km-EPS.(c) and (d) are the same as (a) and (b) except for the OTHER members.(c) and (d) are the same as (a) and (b) except for the POOR members.The gray shading shows a terrain altitude of 500 m.

Figure 7 .
Figure 7. Time-longitude cross sections of the composite combined radar reflectivity averaged along the latitude of Guangzhou key area (shading, dBZ) for the GOOD and POOR groups of TRAMS9km-EPS and TRAMS3km-EPS.The combined radar reflectivities with areas greater than 35 dBZ along the longitudinal direction over the Guangzhou key area are highlighted with purple dots.

Figure 8 .
Figure 8.(a) Forecast sensitivity of ensemble members between the 12-hr accumulated rainfall from 12:00 UTC 21 to 00:00 UTC 22 May and 700 hPa geopotential height field at 12:00 UTC 21 May for TRAMS9km-EPS.The stippling denotes the regions where the correlation coefficient is statistically significant at the 95% confidence level.Vectors show the forecast sensitivity of the accumulated rainfall and the 850 hPa circulation field (only the regions where the correlation coefficient is statistically significant at the 95% confidence level are plotted).The gray shading shows a terrain altitude greater than 500 m.(b) Same as (a) except for TRAMS3km-EPS.

Figure 10 .
Figure 10.Difference fields of the 925 hPa moisture flux (vectors, 10 −3 kg m −1 Pa −1 s −1 ) and its magnitude (gridded color shading with respect to the bottom-left color bar, 10 −3 kg m −1 Pa −1 s −1 ) between the GOOD members of TRAMS9km-EPS and TRAMS3km-EPS at 04:00 UTC 21 May.The gray shading shows the terrain altitude at 200 m, 400 and 600 m with respect to the bottom-right color bar.The black dashed lines denote the location of the cross sections used in Figure 11.

Figure 11 .
Figure 11.(a) Composite pressure-latitude cross sections of the relative vorticity (color shading, 10 −4 s −1 ), horizontal moisture flux (contours, 10 −3 kg m −1 Pa −1 s −1 ), and meridional-vertical circulation (vectors, zonal wind in m s −1 and vertical motion (multiplied by a factor of 50) in Pa s −1 ) along the black dashed lines given in Figure 10 for the GOOD members of TRAMS9km-EPS.(b) Same as (a) except for the TRAMS3km-EPS.

Figure 12 .
Figure 12.(a) Track paths of the MβCV during the period from 04:00 UTC to 20:00 UTC 21 May for the good members of TRAMS9km-EPS (blue lines) and TRAMS3km-EPS (red lines), with the dots indicating the centers of the MβCV.(b) Composite time series of the differences in 850-925 hPa mean divergence (10 −4 s −1 ), vertical velocity (Pa s −1 ), specific humidity (g kg −1 ), and 2-m temperature (K) averaged over a 2° × 2° area around the centers of the moving MβCVs between the good members of TRAMS9km-EPS and TRAMS3km-EPS.

Figure 13 .
Figure 13.Composite time series of the 850-925 hPa mean PV tendency (10 −5 PVU s −1 ) and its forcing terms (10 −5 PVU s −1 ) due to horizontal advection, vertical and horizontal diabatic heating for the GOOD members of TRAMS9km-EPS (blue lines) and TRAMS3km-EPS (red lines), respectively, averaged over a 1.5° × 1.5° area to the northwest (Top left panel), northeast (Top right panel), southwest (Bottom left panel), and southeast (Bottom right panel) orientations of the centers of the MβCVs.

Figure 14 .
Figure 14.Composite distributions of the 850 hPa PV (color shading, PVU) and wind (vectors, m s −1 ) around the centers of the MβCVs for the GOOD members of TRAMS9km-EPS and TRAMS3km-EPS at 04:00, 08:00, and 12:00 UTC on 21 May.The purple dots indicate the area where the combined radar reflectivity is greater than 35 dBZ.The thick black lines (a-b) and (c-d) denote the locations of the cross sections used in Figure 15.

Figure 15 .
Figure 15.Composite vertical cross sections of the PV tendency (color shading, 10 −5 PVU s −1 ), relative vorticity tendency (contours, 10 −8 s −2 ), and the vertical-horizontal circulations (vectors, horizontal wind in m s −1 and vertical motion (multiplied by a factor of 50) in Pa s −1 ) along the black line A-B for (a) the GOOD members of TRAMS9km-EPS and line C-D for (b) the GOOD members of TRAMS3km-EPS given in Figure 14.The 850-925 hPa mean PV tendency (black lines, 10 −5 PVU s −1 ) and its forcing terms (10 −5 PVU s −1 ) due to horizontal advection (black box) and horizontal (red box) and vertical (purple box) diabatic heating for the GOOD members of (c) TRAMS9km-EPS along line A-B and (d) TRAMS3km-EPS along line C-D.

Figure 16 .
Figure 16.Composite distribution of the time averaged eddy flux quantities (   ′  ′ , color shading, m 2 s −2 ) and mean wind (vectors, m s −1 ) on 850 hPa from 10:00 UTC to 14:00 UTC 21 May for the GOOD members of TRAMS9km-EPS.(b) Same as (a) except for the TRAMS3km-EPS.