Key Deviation Analysis of Initial Fields on Ensemble Forecast in South China During the Rainfall Event on May 21, 2020

Ensemble forecasts can quantify the uncertainty of the atmospheric state in numerical weather forecasts and provide the probability density function quantitatively (Bowler ^[1]). Ensemble forecasts offer an opportunity to improve the short-term prediction of local rainstorms ^[2-6]. Several studies have verified that an ensemble forecast is more accurate than an individual model realization for numerical weather prediction (NWP) ^[7-9]. Previous studies have shown that an ensemble simulation using a realistic model can give a good simulation of the meteorological variables during the period of precipitation. It is increasingly important in numerical prediction, and the research and application related to it has been rapidly developed (Li et al. ^[10-14]; Zhang ^[15-17]).

The atmosphere is a chaotic system, and therefore, even small errors in the initial condition of any NWP model are amplified as the forecast evolves. There is great uncertainty in precipitation numerical prediction, which is greatly affected by the initial field of numerical model. For the convection-permitting NWP models, the uncertainty of the initial value greatly affects the accuracy of precipitation forecasting ^[18-19]. A large number of studies show that the initial value error affects the occurrence and development of heavy precipitation and convection by influencing the description degree of mesoscale system. Reducing the initial error will reduce the prediction error accordingly (Nielsen and Schumacher ^[20]).

The predictability of rainstorm and severe convection is a research hotspot in the field of weather, and many scholars have studied the predictability of weather systems with different scales. Melhauser and Zhang ^[21] studied a squall line process and found that the small bifurcation of ensemble members at the high trough and surface pressure is an important environmental factor to determine whether the isolated convective cell can scale up and develop into a squall line. The large differences of the members emphasize the limitation of the rapid growth of error on predictability in the process of wet convection. Zhang et al. ^[22] studied the predictability of a tornado thunderstorm event based on convective scale and found that small disturbances will propagate rapidly in the boundary layer. Wu et al. ^[23] carried out convective scale ensemble forecast of a typical rainstorm in the warm coastal zone of south China. Through the comparison of good and bad members and disturbance sensitivity test, the actual and internal predictability of the process were discussed, and it was found that the internal predictability of the process was 6-12 h.

Key synoptic-scale factors of extreme rainfall events have been examined in many regions using ensemble-based sensitivity analysis, with the accumulated precipitation serving as a forecast metric. Distinct rainfall characteristics and dynamical precipitation mechanisms have been noticed in different rainfall stages of persistent rainfall events. For instance, Schumacher ^[24] examined the factors contributing and inhibiting the development of heavy rainfall in the warm season in the United States based on the ensemble forecast from the European Centre for Medium-Range Weather Forecasts (ECMWF) Ensemble Prediction System (EPS). Using operational global ensemble forecasts from the ECMWF, Zhang and Meng ^[25] found that the precipitation (in Guangdong Province, during 29-31 March 2014) was significantly correlated with mid-level trough, low-level vortex, and particularly the low-level jet on the southeast flank of the vortex and its associated moisture transport. Previous studies have shown that heavy rainfall frequently occurs over south China during the warm season under the influence of Mei-Yu fronts (Chen^[26]), jets (e. g., Chen et al. ^[27]; Du and Chen ^[28]), land-sea breezes (Chen et al. ^[29]), vortices that develop in southwestern China, and cold pools produced (Wang et al. ^[30]). Liang et al. ^[31] found that was by the blocking of an eastward moving low-level trough over southwestern China by a stable anticyclone over eastern China.

This study mainly aims to determine the key deviation of ensemble forecasts for a warm-season heavy rainfall event in south China. Compared with previous studies, the precipitation forecast time of this study was relatively short (6 h), and the horizontal resolution was higher. Section 2 provides an overview of the case and describes the experimental design of the study. In Section 3, key factors and ensemble-based sensitivity analysis for different forecast skills are presented. A summary and conclusions are presented in Section 4.

The heavy rainfall case examined occurred in south China between May 20 and 22 2020, which was the strongest precipitation of the year and was characterized by intense short-term rain, a wide range of heavy rain, and large accumulated rainfall. Heavy to torrential rain occurred in the Pearl River Delta, northern Guangdong, and eastern Guangdong. In some areas of Guangzhou, Dongguan, Huizhou, Shenzhen Shantou Special Cooperation Zone, and Jiangmentai Mountain, heavy rain of more than 250 mm occurred. Moderate to heavy rain and local rainstorms occurred in other cities and counties.

In this study, the research day was selected as 0000- 1200 UTC on May 21, 2020. During this period, the 6 h operational accumulated precipitation forecast (Fig. 1) was not good compared with the observation at 1200 UTC on May 21, 2020. The observation data used is grid precipitation observation, which comes from the NC file issued by the China Meteorological Administration. Moreover, the data is the fusion of satellite, radar, and ground station; therefore, the observed precipitation may exist at sea. Fig. 1c shows the observed precipitation was mainly concentrated in three regions: the South China Sea, Pearl River Delta, and northern Guangdong, all of which were east of the Pearl River Delta. However, the 6 h operational accumulated precipitation forecast of 0.125° European Centre for Medium Range Weather Forecasts (ECMWF) and 0.03° Global/Regional Assimilation and Prediction System developed by the Guangzhou Institute of Tropical and Marine Meteorology (CMA-GD) were mainly concentrated in northeast Guangdong (Fig. 1a and 1b), which was quite different from the observed rainfall area. Table 1 displays TS scores of the above two models, showing that the forecasts of light rain, moderate rain, heavy rain, and rainstorm were all unsatisfactory.

Figure 1.  6h accumulated precipitation distribution of operational forecast (a, b) and observation (c) at 1200 UTC on 21 May 2020. (a) and (b) are derived from 0.125° ECMWF and 0.03° CMA-GD, respectively.

The environment of this event was examined synoptically using 0.25°ECMWF analysis fields at 0000 UTC on May 21 (Fig. 2). The 500-hPa geopotential height revealed a short trough east of Guangdong (Fig. 2a). A cyclonic vortex at 850 hPa (Fig. 2b) was located in South China. Abundant water vapor was transported to the South China coast, and there was a vortex in northeast Guangdong, Jiangxi, and Fujian Province at 925 hPa (Fig. 2c), which was conducive to the triggering of precipitation in south China.

Figure 2.  The 500 hPa geopotential height (a), 850 hPa wind stream and divergence (b), and 925 hPa vapor flux (c) at 0000 UTC on May 21, 2020.

The experiment was performed using the version 3.4.1 Weather and Research Forecasting (WRF) model. In this study, the ECMWF 0.1° analysis data was used for ensemble initialization. The horizontal grid spacing was 3 km, with 50 vertical levels. WRF single-moment (WSM) 6-class microphysics scheme with graupel (Hong et al. ^[32], the Yonsei State University scheme (Noh et al. ^[33]) for planetary boundary layer processes and the Grell-Devenyi cumulus scheme (Grell and Devenyi ^[34]) were used in this research. The 60 ensemble members for initiation were generated by adding perturbations from 0000 UTC to 1200 UTC on May 21, 2020, which were randomly sampled from the default"cv3"background error covariance option in the WRF-3DVar package (Barker et al. ^[35]). The WRF-based Ensemble Kalman filter (EnKF) system was first developed for regional-scale data assimilation by Meng and Zhang ^[36-37]. The control variables, the perturbed variables, the horizontal (vertical) length of the covariance localization and the covariance relaxation method all followed that of Zhu et al. ^[38]. The perturbed variables included the horizontal wind components (u, v), potential temperature and mixing ratio for water vapor, and standard deviations of 2 m s^-1 for wind, 1 K for temperature, and 0.5 g kg^-1 for mixing ratio (Meng and Zhang ^[36-37]). Similar perturbations were also used to represent the boundary condition uncertainties of the ensemble. The covariance relaxation method proposed by Zhang ^[36] was used to inflate the background error covariance with a relaxation coefficient of 0.8. According to Zhu et al. ^[35], the prognostic variables of perturbation potential temperature (T), vertical velocity (W), horizontal wind components (U and V), mixing ratio for water vapour (QVAPOR), cloud water (QCLOUD), rainwater (QRAIN), perturbation geopotential (PH), perturbation dry air mass in column (MU), surface pressure (PSFC), and perturbation pressure (P) were updated.

Table 2 displays TS scores of ensemble mean, eastward and westward members, which can be found that the light rain, moderate rain and heavy rain scores of the ensemble mean were better than those of the above operational forecast. However, the ensemble mean performance was mediocre (Fig. 3a), but the intensity of precipitation was obviously weaker than the observation (Fig. 1c). Although the precipitation area was obviously larger than the observation, it was still significantly better than the operational forecast (Fig. 1a and 1b). In order to find out the key factor of precipitation deviation, the ensemble was further examined by comparing ensemble members with different performances. Based on the rainfall distribution and pattern, two types were selected for analysis compared with the observed rainfall (Fig. 1c): eastward and westward. Each type consisted of three members. Members 32, 43, and 46 were chosen as eastward members, while members 1, 11, and 12 were chosen as westward members. The rainfall distributions of these members are shown in Fig. 3. The rainfall of eastward members was located in the east of the Pearl River Delta, and the precipitation area was larger than that observed. The rainfall of westward members predicted a widespread rainfall area and covered the Pearl River Delta and its western region. It should be pointed out that the above classification can't give an objective and quantitative definition, but only rely on subjective discrimination. Table 3 shows different variables at 925 hPa of ensemble spread were moderate, the ensemble didn't diverge excessively.

Figure 3.  6h accumulated precipitation distribution of ensemble mean (a) contoured with the observation (above 50mm), eastward (b, d, f) and westward (c, e, g) members at 1200 UTC on 21 May 2020.

First, the difference of the thermodynamic fields at 0000 UTC of the two types was examined. Fig. 4 compares the wind field and equivalent potential temperature of eastward and westward members at 700 hPa. The main difference between the two types lay in the level of the equivalent potential temperature in the Pearl River Delta. The equivalent potential temperature of west members was significantly lower than that of east members. However, the equivalent potential temperature distribution at 925hPa was just the opposite in the west of the Pearl River Delta (Fig. 5). The equivalent potential temperature of west members was significantly higher than that of east members. For the westward members, this configuration of low-level warm and medium-level cold represents that the instability of stratification, which was conducive to the triggering of precipitation in the west of the Pearl River Delta. The difference between the two types of wind fields was also obvious at 925hPa (Fig. 5). The southwest wind speed of eastward members was significantly stronger than that of westward members, which would promote the rain to fall easterly. It can be seen that the difference of thermodynamic force in the initial field significantly affects the position of convection trigger, thus affecting the falling area of precipitation.

Figure 4.  700 hPa wind field (m s^-1) and equivalent potential temperature (K) of eastward (a, c, e), and westward (b, d, f) at 0000 UTC May 21, 2020. Wind speed is contoured (12 to 16 by 2) and equivalent potential temperature is shaded.

Figure 5.  925 hPa wind field (m s^-1) and equivalent potential temperature (K) of eastward (a, c, e), and westward (b, d, f) at 0000 UTC May 21, 2020. Wind speed is contoured (14 to 16 by 2) and equivalent potential temperature is shaded.

Then, the difference of initial field further expanded with time goes by, and the thermodynamic fields of the two types displayed at 0600 UTC. The equivalent potential temperature of westward members was higher than that of eastward members in the west of Guangdong and the east of Guangxi (Fig. 6), which was conducive to the sustainable development of convection in western Guangdong for the westward members. Next, the divergence field was checked (Fig. 7). The eastward members had obvious wind convergence in eastern Guangdong, while the westward members had evident wind convergence in the west of the Pearl River Delta, which was conducive to the further development of convection in the corresponding areas. Hence, the initial deviation of thermal and dynamic conditions was very important for precipitation distribution.

Figure 6.  925 hPa wind field (m s^-1) and equivalent potential temperature (K) of eastward (a, c, e), and westward (b, d, f) at 0600 UTC May 21, 2020.

Figure 7.  925 hPa divergence (s^-1) and wind field (m s^-1) of eastward (a, c, e), and westward (b, d, f) at 0600 UTC May 21, 2020.

Finally, correlation coefficients between equivalent potential temperature and accumulated precipitation at 700 hPa (Fig. 8a) and 925 hPa (Fig. 8b) were calculated to provide more physical insight into sensitive areas, to better reveal the critical factors for the distribution of heavy rainfall and determine the cause of the key deviation. The calculated area of precipitation was the area shown in the box (Fig. 1c), which was also the area where the observed precipitation relatively concentrated. As can be seen from Fig. 8a, there was a significant positive correlation in the west of Guangdong, indicating that the higher equivalent potential temperature, the greater precipitation of the boxed area. Fig. 8b shows a more significant positive correlation between equivalent potential temperature and accumulated precipitation at 925 hPa. Furthermore, the higher equivalent potential temperature in the Pearl River Delta, the greater precipitation in the boxed area. The high correlation between equivalent potential temperature and precipitation was also consistent with the above statement (Fig. 4-6). The correlation coefficient between precipitation and wind field was relatively low, which was not shown in this paper. Therefore, the deviation of eastward and westward members of precipitation was mainly caused by the difference of thermal field.

Figure 8.  Correlation coefficient between 6h area-averaged accumulated precipitation at 1200 UTC May 21, 2020 and 700 hPa (a), and 925 hPa (b) equivalent potential temperature at 0000 UTC May 21, 2020.

In this study, the stage-dependent rainfall forecast skill in persistent heavy rainfall in south China and the underlying synoptic-scale features were investigated using an ensemble forecast. From May 20 to 22, 2020, heavy rainfall was clearly identified in Guangdong from rain gauge. This study mainly aims to find out the key deviation of ensemble forecasts for this heavy rainfall, then, the ensemble was further examined by comparing ensemble members with different performances. Based on the rainfall distribution and pattern, eastward and westward types were selected for analysis compared with the observed rainfall. Through the comparison of the thermal and dynamic fields in the middle and lower layers, it can be found that the thermal difference between the middle and lower layers (especially the 925 hPa) was an important factor which generated the deviation of precipitation distribution. The dynamic factors also had some effects on the precipitation area, but they were not as significant as the thermal factors. By correlating accumulated precipitation to atmospheric state variables, different key factors that influenced rainfall forecast uncertainty were identified. The results showed that precipitation was positively correlated with the equivalent potential temperature, further corroborating the above conclusion. This study suggests that the uncertainty of the thermal and dynamic factors in the numerical model can exert an important influence on the quantitative forecast skills of heavy rainfall in South China.