Two high-impact extreme precipitation events during the Meiyu season: simulations and their sensitivity to a scale-aware convective parameterization scheme

In this study, we investigate the impact of a scale-aware convective parameterization scheme (CPS) on simulating two high-impact extreme precipitation events during the Meiyu season at a 1 km resolution. Compared with explicit resolving simulation, applying scale-aware CPS mostly affects the atmospheric environment before convection is triggered and the regions outside the convective cell (e.g., stratiform regions) by consuming convective available potential energy (CAPE), which further inhibits precipitation. For the 2019 event, since the explicit resolving simulation underestimates precipitation, applying scale-aware CPS further inhibits precipitation and reduces the skill for the heavy rainfall category. In contrast, for the 2020 event, as the explicit resolving simulation overestimates precipitation, using scale-aware CPS inhibits precipitation and improves the skill of the model. When scale-aware CPS is applied, increasing σ1, which represents the effect of horizontal resolution, the CPS precipitation is reduced and the grid-scale precipitation is increased, but the overall effect inhibits the total precipitation. However, with the increase in σ1, although the number of stations with heavy rainfall (≥50 mm) is reduced, the average precipitation among these stations is increased. Further analysis of those stations with the largest 24 h precipitation and cross sections over regions with large radar reflectivity shows that increasing σ1 reduces the CAPE consumed by CPS and provides a more favorable environment for strong convections and strengthens precipitation. The results of this study may provide useful information for operational model application and may be beneficial to society.


Introduction
During the East Asia monsoon season (May to July), a quasistationary front extends from eastern China to southern Japan, which is known as the Meiyu front (e.g., Tao and Chen 1987, Ding 1992, Sampe and Xie 2010, Xue et al 2018, Cui et al 2020).The Meiyu front dominates the warm-season precipitation in East Asia (Hong 2004).Organized mesoscale convective systems usually develop along the Meiyu front and induce abundant precipitation in eastern China (e.g., Lin et al 2022).Such precipitation is generally continuous and stable and is sometimes accompanied by thunderstorms and rain, which frequently cause extreme flooding events in densely populated cities in eastern China and result in great casualties and economic losses (e.g., Li et al 2019, Cui et al 2020, Wang and Yu 2022, Yu et al 2022).For example, precipitation during the Meiyu season in 2020 induced severe floods in eastern China (including Zhejiang Province, which is one of the most developed regions in China with a large population and complex topography) and resulted in a direct economic loss of approximately 16 billion US dollars and more than 140 casualties (Wei et al 2020, Tang et al 2022).Therefore, an accurate prediction of precipitation events during the Meiyu season is important and valuable for both the social economy and humans (e.g., Wang and Yu 2022, Yu et al 2022).
With the increase in computational power, operational regional numerical models can provide weather predictions with a horizontal grid spacing of 1-10 km (called gray zone resolution; e.g., Zhang et al 2021).Such a resolution is comparable to the size of convective systems (Wyngaard 2004); therefore, some of the assumptions (e.g., the steady-state and quasiequilibrium assumptions) in traditional convective parameterization schemes (CPSs) are no longer valid (Yano 2015).However, convections are partly resolved under that resolution, and if the CPS is not used, the convective updrafts in subgrids are forced by grid-scale forcings, and the simulated or forecasted precipitation is biased at the gray zone resolution (Deng and Stauffer 2006, Han and Hong 2018, Reszler et al 2018, Zhang et al 2021).To address this issue, scale-aware techniques have been introduced, meaning that convective and grid-scale precipitation depend on grid spacing at gray zone resolution (e.g., Grell and Freitas 2014, Zheng et al 2016, Kwon and Hong 2017).Previous studies have shown that the simulated precipitation under gray zone resolution is improved with the benefit of scale-aware techniques (e.g., Zheng et al 2016, Zhang et al 2021).Sims et al (2017) showed that adding a scale-aware parameter to the Kain-Fritsch scheme can improve the convection timing of the mesoscale convection phenomenon.Kwon and Hong (2017) used scale-aware parameters in the simplified Arakawa-Schubert scheme and showed that the simulated precipitation induced by the Asian summer monsoon was improved over the Korean Peninsula.Jeworrek et al (2019) also showed that applying scale-aware CPS can improve the location and intensity of simulated precipitation under gray-zone resolution.Park et al (2022) revealed that the key modulated parameters in scale-aware CPS are important in balancing the effects of CPS and microphysics parameterization schemes (MPS) at gray-zone resolution.Currently, there are few studies that apply scale-aware CPS in precipitation forecasts in China, where heavy precipitation during warm seasons is organized by synoptic frontal systems than mesoscale convective systems (Hong 2004, Zhang et al 2021).Zhang et al (2021) conducted several 9 km resolution simulations and revealed that scale-aware CPS outperformed explicit resolution in precipitation forecasting in southwestern China, but the relative importance of grid-scale and subgrid-scale precipitation relied on geographical regions and weather regimes.Therefore, it is worthwhile to investigate whether scale-aware CPSs can improve precipitation forecasts during the Meiyu season in Zhejiang Province, which suffers greatly during the Meiyu season.
Since 2013, the Zhejiang Institute of Meteorological Science has been developing an operational numerical weather prediction system based on the Weather Research and Forecasting Model (WRF; Skamarock et al 2019) and the Advanced Data Assimilation System (ADAS; Xue et al 2003).This operational numerical model was named the Zhejiang WRF-ADAS Regional Mesoscale System (ZJWARMS) and is initialized at 00:00 and 12:00 UTC every day.Recently, ZJWARMS has critically contributed to preventing meteorological disasters and has benefited the social economy (e.g., Qiu et al 2015, Xu et al 2020).However, ZJWARMS shows poor performance in forecasting extreme precipitation events during the Meiyu season (Wang and Yu 2022, Yu et al 2022), which has attracted our attention.Yu et al (2022) investigated the impact of WRF domain size on precipitation forecasts during the Meiyu season and adjusted the domain settings of WRF.Wang and Yu (2022) tested numerous different combinations of physical parameterization schemes and showed that scale-aware CPS outperformed turning off CPS (NoCU) in 1 km resolution precipitation forecasts in Zhejiang Province during the Meiyu season.However, how scale-aware techniques impact Meiyu precipitation simulations is unclear, especially the circumstances under which scale-aware techniques can improve or reduce the performances of NoCU forecasts.
In this study, we investigated two high-impact extreme precipitation events during the Meiyu season.For one event, applying the scale-aware CPS called the Korean Institute of Atmospheric Prediction Systems (KIAPS) Simplified Arakawa-Schubert scheme (KSAS; Han and Pan 2011, Kwon and Hong 2017) outperforms the NoCU, while in the other event, the NoCU outperforms the use of the KSAS at a 1 km resolution.Then, we conducted several sensitivity experiments using KSAS to reveal how the scale-aware technique affects precipitation and the simulated atmospheric environment in Zhejiang Province and show under what circumstances scale-aware techniques can improve or reduce the performances of forecasts compared with NoCU.The results of this study not only increase our understanding of how scale-aware techniques affect forecasted precipitation during the Meiyu season but also improve forecast skill and are beneficial to society.
The rest of the paper is organized as follows.Section 2 describes the extreme precipitation events, data, methods and numerical model used in this study.Section 3 provides the results of the two experiments.Section 4 summarizes the major results of this study.

Case description
In this study, two high-impact extreme precipitation events that occurred during the Meiyu season were simulated.Precipitation of the 2019 event emerged at 10:00 UTC on 6 June.This event was the first and most disastrous extreme precipitation event in Zhejiang Province in the 2019 Meiyu season.During this event, 176 weather stations observed more than 50 mm of precipitation over 24 h, which induced severe floods in southwestern Zhejiang Province and resulted in many casualties.Precipitation of the 2020 event occurred at 08:00 UTC on 7 July.This event was a typical extreme precipitation event that occurred during the Meiyu season.During this event, 360 weather stations observed more than 50 mm of precipitation in 24 h.Heavy rainfall during this event forced the Xin'An River Reservoir to open all nine holes to discharge water, which was the first time that all holes were opened since the reservoir was built in 1959, resulting in great societal panic.More details of these two precipitation cases can be found in Wang and Yu (2022).

Evaluation methods
In this study, the model-simulated precipitation during the analysis periods was evaluated using the fractions skill score (FSS), which is more suitable than other methods in assessing Meiyu precipitation simulations in Zhejiang Province (Wang and Yu 2022).The formula for FSS is expressed as follows (Mittermaier 2021): where F and O represent the spatial fractions of locations that exceed a given threshold at a given neighborhood size n in the forecasted and observed precipitation, respectively; N x and N y denote the number of grid points in the x and y directions, respectively.In this study, n is set to 1, 3, 5, 7, 9 and 11 km, and the threshold is set to 0.1, 10.0, 25.0 and 50.0 mm following Wang and Yu (2022).

A brief review of KSAS
The KSAS scheme is developed based on the New Simplified Arakawa-Schubert scheme (NSAS; Han and Pan 2011) by introducing scale-aware techniques to improve the precipitation simulation at the gray zone resolution.The NSAS scheme is well documented in numerous studies and has demonstrated its reliability and performance in the operational National Centers of Environmental Prediction (NCEP) Global Forecast System (GFS) model for a long time (Han and Pan 2011, Lim et al 2014, Kwon and Hong 2017).In this section, only the modifications related to scale-aware techniques were briefly summarized.More details and information on the KSAS scheme can be found in Kwon and Hong (2017).In the NSAS scheme, if convection is triggered, a whole grid box will be covered by convection.Such treatment is not suitable for simulations at gray zone resolution, and scale-aware techniques should be applied.
The key to the scale-aware technique is to properly define the convective updraft fraction s in a grid box.In the KSAS scheme, there are two kinds of s that are applied simultaneously (denoted as 1 s and 2 s in the following equations): where Δx is the horizontal resolution; Δ1 and Δ2 are constants that are set to 5000 and 1000, respectively; w is the vertically average vertical velocity between the cloud top and cloud bottom in a grid box; and c w is the average convective updraft velocity in a grid box.When the horizontal resolution of a simulation is determined, during the calculation processes, 1 s is constant, while 2 s is calculated in each integration time step.When 1 2 s -equals zero, the whole grid box is covered by convection, whereas when 1 2 s -equals one, the whole grid box is free of convection.At a 1 km horizontal resolution, 1 s is near but not equal to 0.9.
Compared with the NSAS scheme, 1 s and 2 s are directly applied to three variables: Cloud-base mass flux (m b ), convective inhibition factor (CINF) and convective cloud water detrainment (DTR).The three modifications are expressed as follows: In the KSAS, when the pressure difference between the level of the originating air parcel and the level of free convection is larger (smaller) than that in the CINF, convection is inhibited (triggered).The convective cloud water detrainment affects the calculation of cloud water and cloud ice in the microphysics scheme (MPS), and the detrained hydrometeors are added to corresponding MPS processes and affect the precipitation produced by MPS (Kwon and Hong 2017).Generally, when the horizontal resolution becomes finer, scale-aware techniques inhibit the cumulus scheme by modifying the above variables since grid-scale saturation becomes more active (Kwon and Hong 2017).

Model setup and experimental design
The numerical experiments in this study were conducted using the Weather Research and Forecasting (WRF) model version 4.0.2 with three one-way nesting domains (9 km, 3 km and 1 km; figure 1(b)).A total of 51 vertical levels in sigma coordinated from the surface to 10 hPa were used.The initial and lateral boundary conditions were provided by the NCEP GFS at a 3 h interval and 0.25°horizontal resolution.For the 2019 precipitation case, a 24 h WRF simulation was initialized at 00:00 UTC on 6 June using the GFS forecast data initialized at 00:00 UTC on 6 June.For the 2020 precipitation case, a 24 h WRF simulation was initialized at 00:00 UTC on 7 July using the GFS forecast data initialized at 00:00 UTC on 7 July.Both simulations were initialized at 00:00 UTC because the nearest initialization time of the real-time forecasting system was 00:00 UTC.Data assimilation was not performed during the simulation.There were several hours for model spin-up.
The CPS used in WRF was the KSAS scheme (Han andPan 2011, Kwon andHong 2017).The other physics used also followed Wang and Yu (2022), including the WRF single-moment six-class microphysics scheme (WSM6; Hong et al 2004, Hong andLim 2006), the Rapid Radiative Transfer Model for Global Climate Models (RRTMG) scheme (Iacono et al 2008) for longwave and shortwave radiation calculations, the Revised MM5 Monin-Obukhov surface layer scheme (Jiménez et al 2012), the Unified Noah land-surface model scheme (Chen andDudhia 2001a, 2001b), and the Yonsei University (YSU) planetary boundary layer scheme (Hong et al 2006).
Two kinds of numerical experiments were conducted in this study.One was established to compare the performances of the KSAS scheme and the NoCU in precipitation simulations in Zhejiang Province.The other experiment was a sensitivity experiment for the KSAS scheme, where we aimed to show the influence of changes in the scale-aware technique key parameter on the KSAS scheme simulated precipitation.In this experiment, we set the 1 s parameter in the KSAS scheme to 0.1 (denoted as TEST-01), 0.5 (TEST-05) and 0.9 (TEST-09).Other settings remained the same as those in the former experiment.Since 1 s was set manually, the simulated 1 2 s represented the convective free fraction.
In the numerical experiments in this study, large-scale circulations and synoptic patterns were mostly constrained by the initial and lateral boundary conditions (Qian et al 2018) and were similar among all abovementioned experiments with negligible differences (figures not shown).Therefore, the sensitivity of scaleaware techniques and variation in 1 s to simulated precipitation mostly manifests at meso and smaller scales.In this study, the analysis was focused on 1-km domain simulations during a 24-h period from 00:00 UTC 6 June (00:00 UTC 7 July) to 00:00 UTC 7 June (00:00 UTC 8 July) in 2019 (2020).

Observed and simulated precipitation
For the 2019 event, the core rainband (50 mm) was in southwestern Zhejiang Province near the boundary of QuZhou and LiShui (figure 2(a)).For the 2020 event, the core rainband was in western Zhejiang Province at the boundary of HangZhou, QuZhou and JinHua (figure 2(b)).The largest 1 h precipitation of both events occurred at 20:00∼21:00 (figure 3).Twenty-one stations observed more than 200 mm of accumulated precipitation, and two stations observed more than 300 mm of accumulated precipitation over 24 h during the 2019 event, while no stations observed over 200 mm precipitation during the 2020 event.Compared with the 2019 event, the precipitation in the 2020 event was moderate, but the core rainband covered a larger area (figure 2).Both KSAS and NoCU captured the main feature of precipitation distribution in the 2019 event, but both produced a much weaker and narrower core rainband, especially KSAS (figures 4(a), (b)).No stations had over 200 mm precipitation in KSAS and NoCU.The simulated largest 1 h precipitation also occurred at 20:00∼21:00 for KSAS and NoCU (figures 3(b) and (c)), which was consistent with the observations (figure 3(a)).For the 2020 event, KSAS and NoCU also captured the main characteristics of the precipitation distribution (figures 4(c), (d)), but both produced a much wider and stronger core rainband compared with the observations (figure 2(b)).Note that 10 stations and 12 stations had over 200 mm of 24 h accumulated precipitation in the KSAS and NoCU, respectively.The simulated largest 1 h precipitation also occurred at 20:00∼21:00 (figures 3(e) and (f)), which was consistent with the observations (figure 3(d)).Further analysis classified the simulated precipitation in the KSAS experiment as convective precipitation and grid-scale precipitation, which were provided by the CPS and MPS, respectively.The results showed that for both the 2019 and 2020 events, most precipitation was produced by MPS (figures 4(e)-(h)).
To quantitatively evaluate the performance of the KSAS and NoCU, the FSS skill score of 24 h accumulated precipitation was calculated and presented in figure 5.For the 2019 event, both the KSAS and NoCU performed well under thresholds of 0.1 mm and 25 mm.The KSAS outperformed the NoCU under a threshold of 10 mm, but NoCU performed better than KSAS under a threshold of 50 mm (figures 5(a) and (b)).This result suggested that NoCU performed better than KSAS in simulating the core rainband, while KSAS performed better in simulating precipitation outside the core rainband.Different from the 2019 event, the KSAS scheme outperformed the NoCU under thresholds of 10 mm, 25 mm and 50 mm in the 2020 event simulation (figures 5(c) and (d)), implying that the KSAS scheme had a better performance in both the core rainband and the outside precipitation simulation.The number of stations for the 0.1 mm, 10 mm, 25 mm and 50 mm thresholds are shown in S figure 1 for both observations and simulations.For the 2019 event, NoCU produced fewer stations with 50 mm precipitation than the observations, and KSAS produced fewer 50 mm stations than NoCU (S figure 1(a)).The KSAS simulated average precipitation over those 50 mm stations in the 2019 event was smaller than that of NoCU, and both KSAS and NoCU were much smaller than the observations (S figure 2).For the 2020 event, KSAS and NoCU produced similar amounts of 50 mm stations, but much more than the observations (S figure 1b).However, the KSAS and NoCU simulated average  precipitation over those 50 mm stations in the 2020 event was larger than the observations, and KSAS was closer to the observations (S figure 2).
In summary, for the 2019 event, both the KSAS and NoCU produced less precipitation than the observations.In contrast, for the 2020 event, both the KSAS and NoCU produced more precipitation than the observations.The KSAS produced less precipitation than the NoCU for both the 2019 and 2020 events, especially the 2019 event.The FSS skill score showed that the KSAS performed better than the NoCU under thresholds 0.1 mm and 10 mm but worse than the NoCU under thresholds 25 mm and 50 mm for the 2019 event, while the KSAS performed better than the NoCU under all thresholds for the 2020 event.In the next section, to further investigate the impact of scale-aware techniques on precipitation simulations, a series of sensitivity experiments were conducted and analyzed.

Precipitation
With the increase in , 1 s the simulated 24 h accumulated precipitation of the 2019 event decreased (figures 6(a)-(c)).In contrast to the 2019 event, when 1 s increased, the 24 h accumulated precipitation of the 2020 event first decreased but then remained nearly stable (figures 7(a)-(c)).For both events, the precipitation magnitude in the core rainband remained stable.Further analysis showed that convective precipitation was reduced (figures 6(d)-(f), 7(d)-(f)), while grid-scale precipitation was increased (figures 6(g)-(i), 7(g)-(i)) with increasing .
1 s This result suggested that with the increase in , 1 s the KSAS scheme was inhibited, while the MPS was more active.In summary, with the increase in , 1 s the CPS was inhibited while the MPS was more active in both extreme precipitation events, and the total precipitation relied on the increase/decrease magnitude of convective and grid-scale precipitation.
The FSS skill of the sensitivity experiments is shown in S figure 3.For each precipitation threshold category of FSS skill, we also accounted for the corresponding number of stations (S figure 4).For the 2019 event, with increasing σ 1 , the FSS skill under the 10 mm threshold increased, the FSS skill under the 25 mm threshold increased first but then decreased, and the FSS skill under the 50 mm threshold decreased (S figure 3).The simulated number of stations of each threshold category supported this result (S figure 4).For the 2020 event, with increasing σ 1 , the FSS skill under the 10 mm and 50 mm thresholds all increased, and the FSS skill under 25 mm increased first but then slightly decreased (S figure 3).Similar to the 2019 event, the simulated number of stations of each threshold category was consistent with this result.Overall, for both events, with increasing σ 1 , the number of stations in each threshold category decreased, which was consistent with the results shown in figures 6 and 7. S figure 5 shows the average precipitation over stations under the 50 mm threshold category for both events.With the increase in σ 1 , the average precipitation of stations with a 50 mm threshold increased (S figure 5), indicating that although fewer stations had 50 mm precipitation (S figure 4), the average precipitation strengthened with increasing σ 1 .Further analysis of those stations with the largest 24 h precipitation and cross sections over regions with large radar reflectivity in the following sections showed that increasing 1 s reduced the CAPE consumed by CPS and provided a more favorable environment for strong convections and strengthened precipitation, which is consistent with the increase in 50 mm threshold category average precipitation.

Parameterized convections
To further investigate how 1 s affects the KSAS scheme, , 2 s cloud-base mass flux, convective adjustment time scale and some other related variables were analyzed and are presented in figures 8 and 9.For both events, with the increase in ,  difficult to trigger with the increase in 1 s (figures 8(j)-(l), 9(j)-(l)) since the pressure difference between the level of the originating air parcel and the level of free convection was larger than the CINF.In addition, the vertically integrated hydrometers from the MPS were increased in association with increasing 1 s (figures 8(m)-(o), 9(m)- (o)), suggesting that the MPS was more active.In summary, with the increase in , 1 s three aspects of the CPS were inhibited: 1.The fraction of a grid box covered by parameterized convection was reduced in the CPS; 2. The intensity of parameterized convection was weakened in the CPS; 3. Parameterized convection became more difficult to trigger in the CPS.Accompanied by the inhibited CPS, the MPS was promoted, and more hydrometeors were produced.

Cross section and stational analysis
To further clarify how 1 s affected the simulation, figures 11 and 12 analyze the vertical cross sections of the convective tendencies.The analysis time was 20:00 UTC for both events (when the largest 1 h precipitation occurs), and the vertical cross sections were selected according to the places with the largest 1 h precipitation (20:00∼21:00) and radar reflectivity (see figure 10).For both the 2019 and 2020 events, with the increase in ,    cloud mixing ratio and ice mixing ratio were obtained by averaging variables over Zhejiang Province, and the target 24 h period also supported these conclusions (Sfig.6 and Sfig.7).For both 2019 and 2020, with the increase in , 1 s the water vapor mixing ratio and the potential temperature tendency were reduced (Sfig.6a, 6d and Sfig.7a, 7d), suggesting that there was less vapor contributing to the formation of convective precipitation or cloud/ice hydrometeors and the release of latent heat in the CPS.The cloud mixing ratio and ice mixing ratio tendency profile also reached their maxima when 0.5 1 s = (Sfig.6 and Sfig.7).The profile of the cloud mixing ratio tendency was quite similar between 0.1 1 s = and 0.9 1 s = (Sfig.6b and Sfig.7b), while the ice mixing ratio tendency when 0.9 1 s = was smaller than that when 0.1 1 s = (S figures 6(c) and 7(c)).In addition, the convective water vapor, potential temperature, cloud mixing ratio and ice mixing ratio tendency were relatively limited in the convective cell when compared with other regions in the convection system (figures 11 and 12), especially the stratiform regions (e.g., figures 11(a), (b), 12(a), (b)).This result suggested that the scale-aware CPS tended to affect the atmospheric environment outside the convective cells.
To further understand how the 1 s affected the atmospheric environment, we found the weather station with the largest 24 h accumulated precipitation in TEST-01, TEST-05 and TEST-09 (locations are shown in figure 10)  effect before convection was triggered, and the CPS mainly reduced the CAPE by consuming water vapor and heating the atmosphere (figures 14(g)-(l)).With the increase in , 1 s the CPS was inhibited, and less CAPE was consumed.For the 2020 event, similar to the 2019 event, most precipitation (figure 13(b)) in TEST-01, TEST-05 and TEST-09 was provided by MPS (figure 13(j)).The precipitation from the CPS was also reduced with increasing 1 s for the 2020 event (figure 13(h)).Two convections occurred for the 2020 event in all three experiments.One occurred between 01:00∼06:00 UTC, and the other occurred after approximately 17:00 UTC (figures 15(a)-(c)), while the latter produced much more precipitation.Therefore, we focused on the latter convection.Before the precipitation became stronger after 17:00 UTC, with the increase in , 1 s the simulated CAPE was increased (figure 13(d)), while the convective inhibition energy (CIN) was reduced (figure 13(f)), which favored convection.The convective tendency of the water vapor mixing ratio and potential temperature at the weather stations showed that the CPS mainly took effect before convection was triggered, and the CPS mainly reduced the CAPE by consuming water vapor and heating the atmosphere (figures 15(g)-(l)).With the increase in , 1 s the CPS was inhibited, and less CAPE was consumed.Although the atmospheric environment before convection was triggered was affected by the CPS, the simulated spatial-temporal average vertical profiles of the moist static energy, saturated moist static energy, relative humidity, horizontal wind speed and potential temperature were nearly identical for all experiments in both the 2019 and 2020 events (S figure 6 and S figure 7), suggesting that the simulated average thermodynamic structure of the atmosphere and baroclinicity were similar between different 1 s values.

Summary and discussion
In this study, we first compared the performance of the KSAS and NoCU in simulating two high-impact extreme precipitation events during the Meiyu season and then investigated the sensitivity of the simulation to the KSAS.Overall, both KSAS and NoCU can capture the major features of the precipitation distribution of the two events.For the 2019 event, both KSAS and NoCU produced less precipitation than the observations, while for the 2020 event, both KSAS and NoCU produced more precipitation than the observations.Compared with KSAS and NoCU, for the 2019 event, KSAS produced less precipitation and showed a lower FSS skill under a threshold 50 mm than NoCU.In contrast, for the 2020 event, KSAS produced less precipitation but showed a higher FSS skill under almost all thresholds.These results suggested that KSAS overall tends to inhibit precipitation compared with NoCU.As a result, when NoCU produces more precipitation than observations, using KSAS may help to improve the skill of the simulation, but when NoCU produces less precipitation than observations, applying KSAS tends to further reduce the skill of the simulation.
A series of sensitivity experiments were conducted by changing the key parameter 1 s in the KSAS scheme.
For both the 2019 and 2020 events, with increasing , 1 s the convection-free fraction in a grid box in KSAS increased.Moreover, the parameterized convection intensity in KSAS was reduced, and the parameterized convection became more difficult to trigger in KSAS.As a result, the KSAS was inhibited, and parameterized convective precipitation was reduced.In contrast, since grid-scale saturation became more active, the MPS became more active, and grid-scale precipitation increased.However, the impact of changing 1 s on the mixing ratio is hydrometeor dependent.Further analysis revealed that for both the 2019 and 2020 events, the KSAS mainly affected the atmospheric environment before convection was triggered and the regions outside the convective cell (e.g., stratiform regions).With the increase in , 1 s the convective tendency of the water vapor mixing ratio and potential temperature decreases, which suggests that there is less CAPE consumed by KSAS.Overall, with the increase in , 1 s the simulated CAPE increases while the CIN decreases, which means that the simulated atmospheric environment is more favorable for strong convection, which is consistent with the previous analysis.In summary, compared with turning off CPS, applying scale-aware CPS mainly affects the atmospheric environment before convection is triggered and the regions outside the convective cell (e.g., stratiform regions) by consuming CAPE, which further inhibits precipitation.Therefore, for the 2019 event, since NoCU underestimated precipitation, applying KSAS further inhibited precipitation and reduced the FSS skill for the heavy rainfall category.In contrast, for the 2020 event, as NoCU overestimated precipitation, using scale-aware CPS inhibited precipitation and improved the FSS skill of the model.When scale-aware CPS is applied, changing the key parameter 1 s can balance the parameterized saturation and grid-scale saturation.With increasing , 1 s the CPS precipitation is reduced, and the MPS precipitation is increased, but the overall effect is inhibiting the total precipitation.However, with the increase in , 1 s although the number of stations with heavy rainfall (50 mm) is reduced, the average precipitation for these stations is increased.Further analysis of those stations with the largest 24 h precipitation and cross sections over regions with large radar reflectivity shows that increasing 1 s reduces the CAPE consumed by CPS and provides a more favorable environment for strong convections and strengthens precipitation, which is consistent with the increase in 50 mm threshold category average precipitation.Parameterization is the process by which the important physical processes that cannot be resolved directly by numerical models are represented.There are always physical processes and scales of motion that cannot be fully represented by a numerical model, regardless of the resolution (Stensrud 2009).In parameterization, those unresolved variables are related to some combination of known/predicted variables using experimental data, empirical relationships or simplified fundamental concepts (Pielke 2013).A parameterization does not necessarily have to simulate the physical processes it is representing to be a realistic representation of these unknown terms (Stensrud 2009, Pielke 2013).Therefore, some key parameters (such as 1 s ) are empirical and can be tuned, and the predicted variables can be regulated by those key parameters (Stensrud 2009).However, since we cannot know what kind of rainfall event will happen in advance, it is unsuitable to use different 1 s for different rainfall events.Typically, an optimal 1 s value is determined by conducting long-term retrospective experiments (e.g., Zhang et al 2021).Sensitivity experiments by changing key parameters in simulating some special events can help us to better understand the parameterizations.
The results of this study provide some guidance to apply scale-aware CPS to improve the performance of ZJWARMS in forecasting high-impact extreme precipitation events during the Meiyu season.For example, applying KSAS always tends to inhibit precipitation compared with NoCU.Further investigation is needed to understand how to make the KSAS suitable for those cases when convective-permitting simulations produce less precipitation than the observations.In future, when ensemble forecasts techniques are introduced to ZJWARMS, forecast results of applying KSAS and turning off CPS should be ensemble members.A more suitable 1 s for Zhejiang Province is another issue that needs to be addressed.The reason why simulated cloud and ice mixing ratio first increase and then decrease needs further investigation.Furthermore, in our future work, we will try to expand the simulation time to cover the whole Meiyu season and conduct several retrospective forecast experiments.

Figure 1 .
Figure 1.(a) The topography and cities of Zhejiang Province.(b) The three nested simulation domains.The color shading represents the topography, and red dots denote the locations of weather stations.

Figure 2 .
Figure 2. Observed spatial distribution of 24 h accumulated precipitation for (a) 2019 and (b) 2020 precipitation events.

Figure 5 .
Figure 5.The FSS skill scores of the (a), (c) KSAS and (b), (d) NoCU experiments in simulations of the (a), (b) 2019 and (c), (d) 2020 events.The x-axis represents the given neighborhood size, and the y-axis denotes the precipitation thresholds.

1 s the simulated 1 2 s
-increased, which meant that a greater fraction in the grid box was free of convection (figures 8(a)-(c), 9(a)-(c)).The simulated cloud-base mass flux was decreased (figures 8(d)-(f), 9(d)-(f)), and the convective adjustment time scale was increased (figures 8(g)-(i), 9(g)-(i)), suggesting that the parameterized convection was weakened, and it took more time to remove the convective available potential energy (CAPE) after convection was triggered (e.g., Mishra 2012).Parameterized convection also became more

1 s
the convective water vapor mixing ratio (figures 11(d)-(f), 12(d)-(f)) and the potential temperature tendency (figures 11(g)-(i), 12(g)-(i)) were reduced, suggesting that less vapor contributed to the formation of convective precipitation or cloud/ice hydrometeors and the release of latent heat in the CPS.Different from the water vapor mixing ratio and potential temperature, the convective cloud mixing ratio (figures 11(j)-(l), 12(j)-(l)) and ice mixing ratio (figures 11(m)-(o), 12(m)-(o)) tendency tended to reach their maxima when 0.5. 1 s = The spatiotemporal average vertical profiles of the tendency of the water vapor mixing ratio, potential temperature,