Growth and Interactions of Multi-Source Perturbations in Convection-Allowing Ensemble Forecasts

Due to the unique climatic background and the complex underlying surface of the Yangtze–Huai River Basin (YHRB) in East China, extreme rainfall and associated flash floods occur frequently during the Meiyu season, which poses a great risk to lives and property (Ding ^[1]; Sun and Zhang ^[2]; Luo et al. ^{[3, 4]}; Luo and Chen ^[5]; Min and Fang ^[6]; Shen et al. ^[7]). There are various forcing backgrounds in the YHRB, such as the synoptic-scale forcing brought by the Meiyu front and the southwest low-level jet and the complex multi-scale topography in the warm zone south of the front (Zhao et al. ^[8]; Zhang and Zhang ^[9]; Fu et al. ^[10]; Fu et al. ^[11]; Wang et al. ^[12]; Zheng et al. ^[13]; Chen et al. ^[14]; Zhang et al. ^[15]; Zhuang et al. ^[16]). Under such diverse forcing backgrounds, the numerical weather prediction of these events is challenging because simply increasing the model resolution cannot improve the forecast accuracy (Mass et al. ^[17]; Walser and Schär ^[18]; Lean et al. ^[19]; Johnson et al. ^[20]; Judt et al. ^[21]; Shen et al. ^[22]). One way to resolve this issue is to use convection-allowing ensemble forecast (CAEF) systems (Peralta et al. ^[23]; Bouttier et al. ^[24]; Hagelin et al. ^[25]; Clark et al. ^[26]; Zhang ^[27]; Schwartz et al. ^[28]), a practical way to estimate forecast uncertainties and produce probabilistic forecasts (Toth and Kalnay ^[29]; Raynaud and Bouttier ^[30]; Zhang et al. ^[31]; Zhuang et al. ^[32]; Xu et al. ^[33]).

Ensemble forecasts estimate forecast uncertainties by adding perturbations that conform to error distribution characteristics. A successful ensemble forecast system should consider all of the uncertainty sources, including initial conditions (ICs), lateral boundary conditions (LBCs), and the model physical parameterization (MO), to obtain sufficient ensemble dispersion (Xue et al. ^[34]; Gebhardt et al. ^[35]; Vié et al. ^[36]; Romine et al. ^[37]; Zhang ^[38]; Xu et al. ^[39]). Therefore, understanding the mechanism of perturbation growth from different perturbation sources and their interactions is vital to generating reasonable perturbations for CAEFs (Johnson et al. ^[20]; Zhuang ^[40]).

Previous studies have comprehensively investigated the perturbation evolution characteristics of a single perturbation source in CAEFs. IC perturbations prevail in the first 12 h of the forecast, and LBC perturbations gradually dominate as the forecast time increases (Gebhardt et al. ^[35]; Vié et al. ^[36]; Hohenegger et al. ^[41]; Peralta et al. ^[42]; Kühnlein et al. ^[43]; Zhang et al. ^[44]). Compared to IC and LBC perturbations, MO perturbations mainly affect temperature and humidity related to cloud and radiation processes, while they have a relatively small effect on the wind field in the middle and upper troposphere (Fujita et al. ^[45]). Combining IC, LBC, and MO perturbations yields larger ensemble dispersion than that generated by one or two perturbation sources (Romine et al. ^[37]; Peralta et al. ^[42]; Surcel et al. ^[46]). However, the combination may suppress the diversity of precipitation prediction (Bouttier et al. ^[24]; Baker et al. ^[47]; Zhang ^[48]). Despite these studies, the optimal design and application of convection-allowing ensembles remain largely unknown (Frogner et al. ^[49]).

Perturbation growth also depends on the synoptic-scale forcing and the geographic regions. Vié et al. ^[36] found that IC perturbations have a smaller (larger) impact when the synoptic-scale forcing is stronger (weaker) compared to LBC perturbations in autumn Mediterranean precipitation events. However, the effects of IC and LBC perturbations are similar under different forcing backgrounds in the warm season precipitation of central Europe (Kühnlein et al. ^[43]; Keil et al. ^[50]), while the MO perturbations are more important for weak-forcing (WF) cases than strong-forcing (SF) cases in Germany (Keil et al. ^[50]; Keil and Craig ^[51]). Recent studies revealed that the growth of multi-scale perturbation varies according to forcing backgrounds. For convective events in the United Kingdom, Flack et al. ^[52] reported a rapid upscale perturbation growth and less temporal variability in synoptic-scale forcing cases. In the southeastern United States, the synoptic-scale forcing cases were found to be insensitive to the scale of the IC perturbations, while the WF cases showed greater sensitivity to small-scale rather than large-scale IC perturbations (Weyn and Durran ^[53]).

Recently, some Chinese scholars have also started conducting research in this field, mainly focusing on the heavy-rainfall cases with different synoptic-scale forcings over South China. Zhang ^[48] examined the case dependence of the IC and MO perturbations and their interactions and found that MO perturbations showed larger influences in the WF case, leading to smaller dispersion reduction. Yang et al. ^[54] found that the IC and LBC greatly influence the location of SF rainfall, while the MO has a great impact on convection triggering of WF rainfall. Since the impact of the synoptic-scale forcing on the perturbation growth varies with the geographic regions, it is worth investigating the perturbation growth for the YHRB, where the forcing background is complex (Sun and Zhang ^[2]; Luo et al. ^[3]; Zhao et al. ^[8]; Zhang and Zhang ^[9]; Fu et al. ^[10]; Zhang et al. ^[15]; Zhuang et al. ^[16]; Shen et al. ^[55]).

As a preliminary study for the above concerns, we selected two typical convective events to represent the strong-forcing (SF) and weak-forcing (WF) backgrounds, respectively, and designed several ensemble experiments involving different perturbation sources. Through the two cases, this study will determine uncertainty sources of precipitation under different forcing backgrounds in the YHRB and investigate the added value and interactions of varying perturbation sources in the CAEFs. Our results will promote the understanding of the growth and interactions of multi-source perturbations varying with forcing backgrounds and guide the selection of the optimal joint perturbations for different types of severe convection in the YHRB.

The remainder of this paper is organized as follows. Section 2 provides an overview of the two selected cases and introduces the data, model configurations, experiments design and diagnostic methods used in this study. Sections 3 and 4 present a comprehensive analysis, including analysis of multi-source perturbation evolutions, uncertainty sources, and their added values. Finally, the summary and conclusions are given in section 5.

Two severe convection cases were selected to represent different forcing backgrounds in the YHRB during the Meiyu period. Case 1 was a typical Meiyu frontal rainfall (SF case), while Case 2 belonged to locally warm-sector rainfall (WF case). This study selected a forecast period between 00:00 UTC June 29, 2015 and 12:00 UTC June 30, 2015 for the SF case and a forecast period between 00:00 UTC July 26, 2018 and 12:00 UTC July 27, 2018 for the WF case.

Figure 1 shows the spatial distribution of accumulated precipitation for the two cases. Two precipitation areas in the SF case were investigated in detail, considering precipitation near the regional boundary (Area A1) and the frontal rain band located in the center (Area B1) (Figs. 1a and 1b). In this case, a typical Meiyu frontal system existed, accompanied by the rainband moving southward, and the 850hPa low-level jet appeared to the south of the rainband (Figs. 2a1–2f1). In the WF case, precipitation scattered throughout Jiangsu (Area B2) and Anhui provinces (Area A2) (Figs. 1c and 1d). There was a convergence line along the junction of Jiangsu and Anhui provinces, which determined the occurrence and evolution of precipitation, but no synoptic-scale systems (e.g., fronts and jets) (Figs. 2a2–2f2) were observed in this case.

Figure 1.  The accumulated precipitation distribution of (a, c) 00–18 h and (b, d) 18–36 h in the (a, b) SF case and (c, d) WF case (red square: precipitation area, black square: non-precipitation area, and black solid line: cross-section line).

Figure 2.  The maximum reflectivity (dBZ, fill color), 850 hPa equivalent temperature (K, red contours from 344 to 348), and the wind field (m s^–1, arrow) in (a1–f1) SF case and (a2–f2) WF case on the (a1, a2) 03 h, (b1, b2) 06 h, (c1, c2) 12h, (d1, d2) 18 h, (e1, e2) 24 h and (f1, f2) 36 h.

The advanced research core of the Weather Research and Forecasting (WRF) model version 4.0 (Skamarock et al. ^[56]) was used, and two one-way nested domains were employed in this study. Based on the nested settings adopted in previous literature (Zhang et al. ^[57]; Wang et al. ^[58]), the convection-allowing domain (d02) was composed of 440 × 440 horizontal grid points at 3 km grid spacing covering the YHRB, nested within an outer parameterizing domain (d01) of 480 × 360 grid points at 15 km grid spacing (Fig. 3). The d02 domain was the analysis region. All domains contained 50 terrain-following hydrostatic-pressure vertical levels topped at 50 hPa. The major physics parameterization schemes were the WRF single-moment 6-class microphysics scheme (WSM6; Hong et al. ^[59]), Kain-Fritsch cumulus scheme (which was only applied to the outer domain, Kain et al. ^[60]), Yonsei University (YSU) planetary boundary layer (PBL) scheme (Hong et al. ^[61]), Rapid Radiative Transfer Model longwave radiation scheme (Mlawer et al. ^[62]), and Goddard shortwave scheme (Chou et al. ^[63]).

Figure 3.  (a) Configuration of model domains, and the d02 domain is the analysis region. (b) Terrain height (m) in the analysis region.

The ICs and LBCs were extracted from the fifth generation European Centre for Medium-Range Weather Forecasts (ECMWF) global reanalysis data (ERA5) with a horizontal resolution of 0.25°×0.25° (Yang et al. ^[54]; Zhang et al. ^[64]; Zhu et al. ^[65]). Hourly Climate Prediction Center (CPC) morphing technique (CMORPH) data, with a resolution of 0.1°×0.1° provided by the China National Information Center, were used for the precipitation forecast evaluation (http://data.cma.cn/).

The growth and interactions of multi-source perturbations were examined under different forcing backgrounds. As in Zhang ^[38], three independent perturbation experiments were conducted considering the single errors from the IC, LBC, and MO perturbations, respectively, and four joint perturbation experiments with different perturbation combinations were also designed (Table 1).

Both the IC and LBC perturbations adopted the dynamic downscaling (DOWN) method. In this method, the analysis perturbations were added and interpolated into the inner and outer domains at the beginning of the forecast to imitate the IC and LBC uncertainties (Zhuang et al. ^[66]; Zhang ^[67]). The LBC perturbations of the inner domain were provided by the forecasts of the outer domain. The perturbation variables were U, V, T, and Q_v, and were calculated by removing the mean state from the first 16 members of the ECMWF global ensemble forecast products (https://apps.ecmwf.int/datasets/). The IC and LBC perturbations were verified to be reasonable and reliable at convective-allowing scale in some researches (Wu et al. ^[68]) The MO perturbations, with reference to Zhang ^[66], were generated by combining the stochastically perturbed parameterization tendency (SPPT; Bouttier et al. ^[47]), multi-physics (MP; Xue et al. ^[34]), and perturbed parameters (PP; Gebhardt et al. ^[35]) schemes. Table 2 gives the configurations of each ensemble member of PPMP (i.e., the combination of PP and MP). In each experiment, the ensemble forecasts contained 16 members with a duration of 36 h for two different forcing backgrounds.

The perturbation was quantified using the difference total energy (DTE; Zhang et al. ^{[69, 70]}), which was defined as:

where ∆ indicates the difference of u, v and T between each ensemble member and the unperturbed control run, T_r is the reference temperature of 287 K, and c_p is specific heat capacity in dry air at constant pressure (c_p=1004 J kg^–1 K^–1). The occurrence and development of severe convective weather are affected by the environmental fields at different height levels and scales, while the traditional DTE can resolve the forecast error at a single height level alone. Given this situation, Nielsen and Schumacher ^[71] applied a vertical function to calculate the two-dimensional root mean vertically integrated DTE (RMDTE):

where n indicates the number of ensemble members, l stands for the vertical levels, p is the value of pressure, and n_x and n_y are the numbers of the grid points in the meridional and zonal directions, respectively.

In this section, we first compared the growth characteristics of single IC, LBC, and MO perturbations under the two forcing backgrounds.

Figure 4 shows the time series of RMDTEs in the selected areas. The RMDTEs of the precipitation areas reached their maximum at the peak of precipitation for both the SF and WF cases, confirming the importance of moist convection in perturbation growth. For the SF case, significant disparities were observed in the evolution of perturbations between the IC and LBC ensembles across the two precipitation areas. Specifically, within Area A1, the RMDTE from the LBC ensemble increased swiftly, surpassing that of the IC ensemble after a mere 6-h forecast period, thus both reached saturation almost simultaneously at 9 h (Fig. 4a). Conversely, in Area B1, the forecast errors arising from the LBC ensemble progressed at a slower pace compared to the IC ensemble, resulting in a longer time to saturation in the LBC ensemble (Fig. 4c). The difference in the perturbation evolution of IC and LBC ensembles can be more clearly seen from the spatio-temporal distribution of the RMDTE (Fig. 5). The RMDTEs from the IC ensemble initially manifested within the central frontal rainband of the region and progressed eastward, maintaining synchronicity with the advancement of the synoptic-scale system (Figs. 5a1–5f1). In contrast, the RMDTEs from the LBC ensemble first appeared in the precipitation area near the boundary and then gradually propagated towards the central precipitation area (Figs. 5a2–5f2). Unlike the IC and LBC perturbations, the forecast errors of the MO perturbations emerged and increased almost synchronously in the boundary and central precipitation areas (Figs. 5a3–5f3). Notably, the time to saturation for MO perturbations lagged behind that observed for IC and LBC perturbations (Figs. 4a and 4c), which indicated that the MO perturbations were less sensitive to the precipitation location.

Figure 4.  Temporal evolution of the RMDTE in the (a–d) precipitation areas and (e–f) non-precipitation areas in the (a, c, e) SF case and (b, d, f) WF case. The gray dotted lines represent the corresponding area-averaged precipitation from the control forecast.

Figure 5.  Distribution of the RMDTE of different forecast times for (a1–f1) ICP, (a2–f2) LBP and (a3–f3) MOP in the SF case.

By contrast, for the WF case, the perturbation evolution trends and saturation times from different sources were similar but exhibited different magnitudes: the RMDTEs from the LBP ensemble were relatively modest in magnitude and did not grow rapidly even in the precipitation area (Figs. 4 and 6). This was because the environmental wind was small under the weak-forcing background, which made the forecast errors introduced by the LBC perturbations propagate slowly, thus leading to a relatively weaker influence. However, the perturbations from the MO ensemble were dominated and evolved into a broader area (Figs. 4 and 6), which indicated that the MO perturbations had a larger impact than the other types of perturbations in the WF case.

Figure 6.  Distribution of the RMDTE of different forecast times for (a1–f1) ICP, (a2–f2) LBP, and (a3–f3) MOP in the WF case.

The above results demonstrate that MO perturbations can generate large forecast uncertainties across various precipitation areas and under different forcing backgrounds, indicating that MO perturbations were more sensitive to convection triggering. Consequently, the MO perturbations could induce forecast uncertainties even when no precipitation occurs (Figs. 4e, and 4f). On the contrary, the growth of IC and LBC perturbations was greatly affected by the occurrence and location of precipitation, thus leading to disparities in perturbation growth between the precipitation and non-precipitation areas.

Figure 7 presents the vertical cross-sections along the bold lines shown in Fig. 1 in the precipitation areas. The frontal system can be seen in the vertical section of the SF case, as well as the confrontation between the warm, moist airflow on the south side of the front and the cold air mass invading at a lower level on the north side (Fig. 7a). The red dashed rectangle shown in Fig. 7b marks the range of the synoptic low-level jet, where the southwesterly wind exceeded 12 m s^–1. The main precipitation areas were located around the low-level jet and the terrain. Additionally, a distinct layer of vertical wind shear was observed near the 850 hPa level, characterized by a notably weak horizontal wind component.

Figure 7.  The vertical cross-section of (a, c) the wind vector (m s^–1, arrow, the vertical wind is enlarged by a factor of 10) and equivalent temperature (K, fill color), (b) the horizontal wind speed (m s^–1, fill color) and reflectivity (contour), and (d) the horizontal divergence (10^–5 s^–1, shading) and reflectivity (contour) along the bold lines in Fig. 2 at (a, b) 00:00 UTC June 30, 2015, and (c, d) 00:00 UTC July 27, 2018.

By contrast, the blocking effect of Dabie Mountain played an important role in the WF case (Figs. 7c and 7d). On the one hand, the wind speed on the leeward slope of the mountain was weakened due to the influence of the topography, which led to the convergence of the wind. On the other hand, the downdraft on the leeward slope of the mountain carried cold air and formed a cold pool near the ground, which escalated instability in the atmosphere (Figs. 7c and 7d). The occurrence of precipitation was mainly caused by the superposition of the two factors.

Figure 8 shows the vertical cross-sections of the RMDTE and ensemble mean wind field from the three perturbation experiments, aiming to determine the sources of forecast uncertainties through the distribution of perturbations and further compare the estimations of uncertainties from different perturbations. Regarding the SF case, the distributions of RMDTEs correspond to the front system and the jet core, indicating large forecast uncertainties in the synoptic-scale systems. Notably, the RMDTE of IC perturbations was mainly distributed in the low-level jet, while the MO perturbations were located in the vertical wind shear layer. Compared with the two, the LBP ensemble can capture the uncertainties from both the low-level jet and the vertical wind shear. Therefore, from the result of single source perturbation, the LBC perturbations offered the most inclusive characterization of the forecast uncertainties caused by the synoptic-scale systems under the SF background.

Figure 8.  The vertical cross-section of the ensemble mean the wind vector (m s^–1, arrow, the vertical wind is enlarged by a factor of 10) and the RMDTE (J kg^–1, shading) along the bold lines in Fig. 2 at (a1–c1) 00:00 UTC June 30, 2015 and (a2–c2) 00:00 UTC July 27, 2018 in (a1, a2) the ICP, (b1, b2) the LBP, and (c1, c2) the MOP ensemble.

For the WF case, the RMDTEs had a strong relationship with the updraft area, and the RMDTEs of the lower level appeared downstream of the Dabie Mountain (Figs. 8a2–8c2). However, the uncertainties from the boundary layer were characterized differently by the perturbations from different sources, with the MO perturbations yielding the largest forecast uncertainties while the LBC perturbations generated the smallest. The result showed that MO perturbations provided a reliable estimation of the uncertainties present in the boundary layer variable field under the WF background, while LBC perturbations substantially underestimated these uncertainties.

Through comparative analysis of the results of these two typical cases, we found that the forecast uncertainties of precipitation and the performances of ensemble perturbations were different, which confirms that the perturbation growth affected by the forcing backgrounds to some extent, especially for the LBC perturbations.

The preceding section demonstrates the growth and performance of single perturbations through two forcing backgrounds. However, interactions between perturbations have been found in previous studies (Zhang ^[38]; Baker et al. ^[47]). To explore the optimal combination of perturbations under the two forcing backgrounds, this section will further analyze the net effect of introducing different perturbations to the joint perturbations, namely the added value.

Figures 9a1 and 9a2 show the RMDTE relative increments when the IC perturbations are introduced to different joint experiments. The added values of the IC perturbations in two cases were obvious in the early forecast period, corresponding to the rapid growth from the IC perturbations in the early stage, and then a gradual decrease with the forecast time, especially in the ICMOP and ALL experiments. By contrast, the IC perturbations generated larger increments in the ICLBP experiment, which can be more clearly shown from the vertical distribution of RMDTE increments (Fig. 10). Because of the different sources of uncertainties in the two forcing backgrounds, the added values of IC perturbations varied in two cases. In the SF case, the RMDTE increments of IC perturbations mainly appearred in the low-level jet region (Figs. 10a1–10c1), suggesting that introducing IC perturbations can be beneficial to capture the uncertainties related to the low-level jet. Whereas in the WF case, the introduction of IC perturbations strengthened the ascending motion of the lee slope of the Dabie Mountain, consequently leading to an increase of the RMDTE. Notably, the added values of the IC perturbations were mainly apparent in the ICLBP experiment (Fig. 10b2), revealing the positive effect of IC perturbations on LBC perturbations under the two forcing backgrounds to some extent.

Figure 9.  The RMDTE relative increments in different joint perturbation experiments of (a1, a2) the ICP, (b1, b2) the LBP, and (c1, c2) the MOP in (a1, b1, c1) the SF case, and (a2, b2, c2) the WF case.

Figure 10.  The vertical cross-section of the RMDTE relative increments (shading) and ensemble mean wind fields (vector) along the black lines in Fig. 2 at (a1–c1) 00:00 UTC June 30, 2015 and (a2–c2) 00:00 UTC July 27, 2018 from (a1, a2) the ALL–LBMOP, (b1, b2) the ICLBP–LBP, and (c1, c2) the ICMOP–MOP.

In contrast to IC perturbations, the RMDTE increments of LBC perturbations increased gradually from 0 with the forecast time (Figs. 9b1 and 9b2). Fig. 11 shows the vertical distribution of RMDTE increments of LBC perturbations. For the SF case, the added values were manifested in the different joint experiments and were particularly noticeable in the vicinity of 850 hPa (Figs. 11a1–11c1), corresponding to the synoptic-scale frontal system (as shown in Fig. 7a). This result indicates that the introduction of the LBC perturbations played an important role in improving the estimation of the synoptic-scale uncertainties under the SF background. However, for the WF case, the RMDTE increments of LBC perturbations were small and lost the fluctuation related to moist convection, even though the evolution of the isolated LBC perturbations had the characteristics of moist convection (Fig. 9b2). Thus, the added value of LBC perturbations was mainly manifested in the SF case rather than the WF case, which further revealed the case dependence of LBC perturbations.

Figure 11.  As in Fig. 10, but from (a1, a2) the ALL–ICMOP, (b1, b2) the ICLBP–ICP and (c1, c2) the LBMOP–MOP.

Similarly, the RMDTE increments of MO perturbations also increased gradually from 0 with the forecast time (Figs. 9c1 and 9c2). The vertical cross-section of the RMDTE increments of MO perturbations in different joint experiments is given in Fig. 12. The introduction of MO perturbations significantly improved the description of uncertainties in the wind shear layer caused by the frontal system under the SF background (Figs. 12a1–12c1). Under the WF background, due to the dominance of the MO perturbations, large RMDTE increments can be generated throughout the forecast period, particularly in the LBMOP experiment (Fig. 9c2). Furthermore, the added values were mainly manifested downstream of the Dabie Mountain (Fig. 12c2), indicating the introduction of MO perturbations helped capture the uncertainties of wind field convergence (dynamic) and surface cold pool (thermal) caused by topography. This result shows that the MO perturbations had different positive contributions under the two forcing backgrounds when IC, LBC, and MO were synchronously perturbed.

Figure 12.  As in Fig. 10, but from (a1, a2) the ALL–ICLBP, (b1, b2) the ICMOP–ICP and (c1, c2) the LBMOP–LBP.

To construct the optimal perturbations for different types of severe convection in CAEFs, it is essential to understand the perturbation growth and interactions resulting from different-source perturbations. Therefore, in this study, seven 36 h ensemble forecast experiments were conducted to investigate the growth and added value from IC, LBC, and MO perturbations in different joint perturbations. Two cases over the YHRB in East China initialized at 00:00 UTC on June 29, 2015, and 00:00 UTC on July 26, 2018, were used as representative SF and WF cases, respectively, to assess the sources of uncertainty related to perturbation growth under different forcing backgrounds.

The perturbation growth from the IC, LBC, and MO ensemble under different forcing backgrounds was different. For the SF case, the growth of IC and LBC perturbations was strongly affected by the location of precipitation. Thus, there was a great difference in perturbation growth between the central and the boundary precipitation areas. By contrast, the perturbation growth exhibited obvious characteristics of moist convection under the WF background, in which the growth of MO perturbations was dominant, while the growth of LBC perturbations was limited. Furthermore, the MO perturbations were more sensitive to convection triggering. Thus, the MO perturbations can also produce forecast uncertainties in the non-precipitation area.

The sources of uncertainties under the two forcing backgrounds were different and the performances of perturbations also varied. For the SF case, the synoptic-scale uncertainties were mainly derived from the low-level jet and vertical wind shear. The LBC perturbations had a better estimation for the synoptic-scale uncertainties under the SF background, which was related to the large-scale information introduced by the LBC perturbations. However, the uncertainties of the WF case were mainly derived from the local vertical motion and the surface cold pool effect caused by Dabie Mountain. The performance of MO perturbations was the best, and that of LBC perturbations was the worst, revealing that the LBC perturbations have a certain case dependence.

Additionally, the added values of different source perturbations were analyzed to obtain the optimal perturbation method under different forcing backgrounds. The added values of different perturbations were reflected in different forecast periods and different source uncertainties under the SF background. For example, the added values of the IC perturbations were apparent in the early forecast stage and the low-level jet area, while that of the LBC and the MO perturbations were mainly reflected in the late forecast time and the synoptic frontal system (vertical wind shear layer). Therefore, the combined perturbations could make up for the deficiency of the single perturbations and achieve a better ensemble performance in the SF case. However, the added values generated by different source perturbations differed substantially under the WF background. The effect of the MO perturbations was dominant, while the added values of LBC perturbations lost the fluctuation of moist convection. Thus, the MO perturbations should be given priority in the WF case.

This study focused on the perturbation growth and added values caused by the introduction of different perturbations. In the future, the added values of MO perturbations compared to the IC and LBC perturbations need to be further explored, particularly for the cases under weak forcing. In addition, the impacts of different perturbation sources on the precipitation forecast performance have yet to be investigated in this study. The sensitive factors, such as topography and low-level jet, affecting precipitation occurrence and development in the SF and WF cases are not involved. We are interested in exploring all of these factors and investigate more cases under different forcing backgrounds in the YHRB in the future.