The evolution of probability with time

Besides, of evaluates the predictability of CReSS for the D18 event with the time-lagged strategy (units are days) as analyzed above. This study also focused on assessing the fluctuations in CReSS's predictive skills, and the evolution of probability with time as the number of members of ensemble changes.

Figure 4.15 shows the scenarios of the ensemble mean of 24-h rainfall and its spread for 09 December. In which, the ensemble mean of the last 05 members includes members executed within ranges days 0-1 before the target date. The ensemble mean of the last 09 members includes members executed within ranges days 0-2 before the target date. The ensemble mean of the mid-08 members includes members executed within ranges days 2-3 before the target date. The ensemble mean of the first-08 members includes members executed within ranges days 4-5 before the target date. The ensemble mean of the 21 members includes members executed within ranges 0-5 days before the target date. It shows that all scenarios of the ensemble mean of h rainfall predicted 24-h rainfall amount is commonly between 50 mm and 100 mm w24-hile t24-he observed rainfall amount is regular from 100 mm to 400 mm. There are much lower than in reality in comparison. Besides, compare with TRMM data show that the model captured the rain area over the coastal sea.

Spread scenarios show that the standard deviation of rainfall scenarios of the last 5 and the last 9 members is the largest, with maximum spread can reach over 160 mm.

With this spread scenario, the last 5 and the last 9 members can predict 24-h rainfall amount maximum is around 300m, it is quite closer to the observed rainfall data, with the peak rainfall observed is over 500 mm. However, this largest spread indicates the ensemble mean of the last 5 members has the lowest theoretical predict accuracy. Besides, spread scenarios of middle 8 and first 8 members are the lowest. In general, standard deviations decreasing with the reduction rainfall of rainfall scenarios with times (Fig 4.15).

24

Besides, figure 4.16 shows that the ensemble mean of the last 05 members becomes a highlight with higher quality QPFs than the rest at all rainfall thresholds.

Furthermore, only the ensemble mean of the last 05 members has skill scores at thresholds greater than 100 mm with TS = 0.1, POD = 0.1, FB = 0.1, and FAR = 0 at 200 mm.

Conversely, the ensemble mean of mid 08 members has worst-quality QPFs with TS = 0, POD = 0, FB = 0 at 75 mm. Scenarios of the ensemble mean of rest have TS = 0, POD = 0, FB = 0 at 100 mm. The FSS scores also show similar results for these analyzes, with the highest score being the last 05 members (FSS = 0.49) and the lowest being the middle 8 members (FSS = 0.14) (Fig. 4.16e). The maps of probability distributions by five different groups of ensembles in Fig. 4.17 shows the great change of forecasting probabilities with time. At 100 mm, the longest range at first 08 members (4-5 days before the target date) has the lowest probabilities (less than 20%). Meanwhile, in the shortest range at the last 5 members, the forecasting probability is highest (up to 60% at several small areas inland). These forecasting probability values decrease with the increase of rainfall thresholds.

Figure 4.18 shows the scenarios of the ensemble mean of 24-h rainfall and its spread for 10 December. In which, the ensemble mean of the last 05 members includes members executed within ranges 0-1 days before the target date. The ensemble mean of the last 09 members includes members executed within ranges days 0-2 before the target date. The ensemble mean of the mid-08 members includes members executed within ranges days 3-4 before the target date. The ensemble mean of the first-08 members includes members executed within ranges days 5-6 before the target date. The ensemble mean of the 25 members includes members executed within ranges 0-6 days before the target date. It is clear to see that the ensemble means of both the last 05 members and the last 09 members are considered similar to observed rainfall data not only the accumulate rainfall but also spatial distribution of rainfall.

Besides, the ensemble means of both the last 05 members and the last 09 members have much better quality QPFs than the ensemble mean of mid 08 members, first 08 members, and all members. For example, at 100 mm, the ensemble mean of the last 05 members have TS = 0.4, POD = 0.8, FB = 1.5, FAR = 0.5, the last 09 members have TS

= 0.5, POD = 0.8, FB = 1.4, FAR = 0.5, the ensemble mean of both mid 08 members and first 08 members has TS = 0, POD = 0, FB = 0, no skill scores for FAR. At 200 mm, the

25

ensemble mean of the last 05 members have TS = 0.2, POD = 0.4. FB = 1.4, FAR = 0.7, the last 09 members have TS = 0.3, POD = 0.5, FB = 1.2, FAR = 0.6, the ensemble mean of mid 08 members, first 08 member, and 25 members have no skill score (Fig. 4.19).

Besides, FSS scores show the last 09 members have highest score (FSS = 0.64), the middle 08 members have lowest score (FSS = 0.04), and the ensemble mean of 25 members is 0.43 (Fig. 4.19e). However, spread scenarios indicated the spread of the ensemble mean of both the last 05 members, and the last 09 members is largest. Especially, the spread of the ensemble mean of all members is very large (almost >140 mm), although the senarios of 24-h rainfall commonly is between 50 mm and 100 mm with peak rainfall is 150 mm. These larger spread show model can produce a wide range of possible rainfall scenarios. Synchronous, it also shows that these types of the ensemble mean has the lower theoretical predict accuracy than the rest (Fig. 4.18).

Figure 4.20 show probability distribution obtained from all 25 members, executed every 6 h from 1200 UTC 3 Dec to 1200 UTC 9 Dec, reach thresholds of 100, 200, 300, and 450 mm for the 24 h from 1200 UTC 9 Dec to 1200 UTC 10 Dec. it clear to see that there is 10 – 20% of first 8 members (the longest range at 5-6 days before the target date) reached 100 mm inland. Meanwhile, no one member of the middle 8 members (the medium range at 3-4 days before the target date) can reach 100 mm in the same area. At the shorter range at 0-2 days before the target date, there is 70-80 % of the number of the last 9 members reach 100 mm. There are 80- over 90% members of the last 5 members (the shortest range at 0-1 days before the target date) captured 100 mm. Besides, that is, 20-40% of members of 25 members can touch 100 mm inland. These probabilities are decreasing with increase the rainfall thresholds. At 450 mm, there is less than 20 %, 30%

the number of the last 5 and the last 9 members can reach this threshold, respectively. 0%

the number of both middle 8 and first 8 members can touch the threshold.

Figure 4.21 shows the scenarios of the ensemble mean of 24-h rainfall and its spread for 11 December. In which, the ensemble mean of the last 05 members includes members executed within ranges 0-1 days before the target date. The ensemble mean of the last 09 members includes members executed within ranges days 0-2 before the target date. The ensemble mean of the mid-08 members includes members executed within ranges days 3-4 before the target date. The ensemble mean of the first-12 members includes members executed within ranges days 5-7 before the target date. The ensemble

26

mean of the 29 members includes members executed within ranges 0-7 days before the target date. It is clear to see that the ensemble means of both the last 05 members and the last 09 members are considered similar to observed rainfall data not only the accumulate rainfall but also spatial distribution of rainfall. As analyzed above, the observed rainfall amount on 11 December is commonly between 30 mm and 100 mm.

It shows that the ensembles mean of the last 05 members, the ensembles mean of the last 09 members, and the ensemble mean of the mid 08 members are similar quality QPFs and closer to the observed rainfall data, although they executed at different lead-times. In particular, the ensemble mean of the last 05 members includes members executed within ranges 0-1 days before the target date. The ensemble mean of the last 09 members includes members executed within ranges days 0-2 before the target date. The ensemble mean of the middle 08 members includes members executed within ranges days 3-4 before the target date. However, skill scores of these ensembles mean are decrease rapidly with increasing rainfall threshold levels, and some skill scores are equal to zero at 75 mm, such as TS, POD. Besides, The FSS score also shows conformity with the above analysis by the highest score is only 0.26 (Fig. 4.22). Spread scenarios also show that the spread of these ensembles means the largest. It shows that there is a big difference between the individual members and the lower theoretical predict accuracy than the rest;

these are consistent with the analysis above. However, it also indicated that in the case of the spread is large, model can predict this rainfall. The ensemble mean of the first 12 members has worst quality QPFs due to including members that executed at long ranges (5-7 days before the target date)—leading to most individual members poor-predicted the wind surface fields.

The maps of probabilities distribution show that there is less than 20 % the number of the middle 8 members (the range at 3-4 days before the target days) and less than 10

% of 29 members (the range at 0-7 days before the target date) can reach the thresholds at 100mm in the same observed area. No one member of all 29 members can capture the thresholds 200 mm, 300 mm, 400 mm in both the longest range at 5-7 days before the target date (first 12 members) and the shortest range at 0-1 days before the target date (the last 5 members) (Figs. 4.23).

Figure 4.24 shows the scenarios of the ensemble mean of 72-h rainfall and its spread for periods of 1200 UTC 08 December to 1200 UTC 11 December. With five

27

groups of the ensemble mean evaluated. The ensemble mean of the last 05 members stands out with high-quality QPFs and closer to the observed rainfall data than the rest of the ensemble mean. The spatial distribution of rainfall in this scenario is considered similar to reality. However, the 72-h rainfall amount is lower than in reality.

In particular, Fig. 4.25 shows at 250 mm, the ensemble means of the last five members have TS = 0.5, POD = 0.6, FB = 0.9, FAR = 0.3, meanwhile, the ensemble mean of the nine members has TS = 0, POD = 0, FB = 0, FAR = 1. The ensemble means of the middle 08 members, the ensemble means of the first eight members, and the ensemble mean of 21 members has no skill scores. At 350 mm, only the ensemble means of the last five members have skill scores with TS = 0.2, POD = 0.3, FB = FAR = 0.4, and SR = 0.6.

At 500 mm, skill scores of the ensemble mean of the last five members, such as TS = POD = FB = 0, and FAR has no scores, while the observed rainfall amount recorded greater than 800 mm. Contrariness, the ensemble mean of the mid 08 members is considered worst-quality QPFs due to the skill scores is the lowest compared to the rest of ensemble means. The FSS score also shows that the ensemble mean of the last 05 members has the highest quality QPF with FSS = 0.7, and the lowest quality QPF is the middle 08 members (FSS = 0.14). Besides, FSS score of the ensemble mean of 21 members is 0.35.

Furthermore, spread scenarios show the ensemble mean of the last five members has the largest spread, this meaning that in scenario of the spread is largest, the rainfall amount predicted by the model is very closer to observed rainfall data. Besides, it is clear to see that the ensemble means of the last 09 members and the ensemble mean of all members has a very larger spread, although the 72-h rainfall scenarios are mostly lower than 200 mm. These wide ranges of the spread may be related to individual members in these groups, which did not predict rainfall well due to the incorrect predicted of the surface wind field (Fig. 4.24).

The maps of the probabilities distribution in Fig. 4.26 indicating that over inland, the probabilities that ensembles can reach the threshold at 100 mm of rain is over 70% of the last 5 members, 40-60% for the last 9 members, 30-40 % for the first 8 members over the haft of the south part of central, 20-40 % for 21 members, and just is 10-20% for the middle 8 members over a haft of the north part of the study area. For thresholds greater than 100mm, there is a 50-60 % chance for the last 5 members, 30-40 % for the last 9

28

members, and 10-20 % of 21 members reached the threshold at 300 mm of rain. However, only the last 5 and the last 9 members can reach a threshold at 500 mm of rain with probabilities is 20-30%. No one group in 5 groups of the ensemble can touch threshold at 800 mm of rain.

From the analysis of model results in many different aspects, including analysis of single runs, the ensemble mean, the evolution of forecasting probability above show that model has well predicted the second day and three days of D18 event within the lead time of 1 day, 2 days, and 3 days. Model’s skills decrease with time and with increasing rainfall thresholds.

29

CHAPTER 5.

DISCUSSION

By analysis in part 3 show that the D18 event occurred due to the combined effect of the northeasterly wind, the easterly wind, and the topography in central Vietnam. This is a complex mechanism due to a combination of many factors. Therefore, most forecasts for events such as the D18 event fail and are still a big challenge for operational weather forecast offices. Time-lagged cloud-resolving ensemble quantitative precipitation forecasts is a completely new approach, and the results are very encouraging and impressive. Especially for lead times of 1 to 3 days, the forecast scenarios of the model are very similar to reality not only in QPF but also in the spatial distribution of rainfall.

However, there are still some problems such as the model have well-predicted the rainfall of Dec 10. However, the model has not well predicted the rainfall of Dec 9 and 11 as analyzed and indicated in the previous sections. This difference in forecast quality is due to differences in the predictability of each member of 29 members. Particularly, fig. 5.1 shows some members that have the good or bad performance for every single day of the D18 event. It is clear to see the difference between good members and bad members, which is good members have well-predicted the surface wind while bad members did not do it. The good forecast of the wind field leads to the good prediction of moisture transport or place of convergence or moisture divergence, leading to a good prediction of rainfall.

This difference in individual members' forecast quality can relate to quickly changing the real turbulent atmosphere with time (unit is an hour), leading to much difference in the initial data. Meanwhile all members of this study are executed every 6h.

These can be seen in fig. 5.2. Besides, fig. 5.2 show also that good members have well-predicted the initial conditions than bad members. Particularly, the good members predicted the wind field and the moisture convergence and divergence at 925 hPa better than the band members.

Besides, ensemble sensitive analysis in chapter 6 shows that the 24-h accumulated rainfall is strongly sensitive to initial conditions, and is strongest sensitive on Dec 10. Furthermore, as we know, the computational errors will arise at every time step of the integration and will build up cumulatively. Hence, the result of each member is the difference from each other.

30

By analyze above and previous sections, which indicates that the differences in initial data, thermodynamic conditions at individual day, and the forecast range, have a different impact on the forecasting skills of the model. Therefore, in order to improve the ability to predict heavy rain using numerical modeling, we need to improve the quality of the initial data.

31

CHAPTER 6.

SENSITIVE OF THE ENSEMBLE

As the analysis in chapter 3, the D18 event is caused by combined effectively between the atmospheric disturbances at lower levels, such as cold surge, easterly wind, and topography. Hence, in this study, ensemble sensitivity analysis (ESA) has been applied for variables at surface, near-surface, and mid-tropospheric levels to assess the sensitivity of initial conditions to the predictability of the rainy field. In which there are a total of 21 members for December 9, 25 members for December 10, and 29 members for December 11.

a) Horizontal wind components.

Figure 6.1 shows sensitive to components of surface wind fields for every single day during the D18 event. In general, the sensitivity of the U component of wind is weak sensitive on 09 and 11 December. Stronger sensitivity on December 10 with strong negative sensitivity over the south part and strong positive over mid-central of the study area. Meanwhile, the V component of wind is considered more sensitive than the U component of wind. Moreover, there is an antagonistic that places that the component U has the strongest negative sensitivity, then component V has the strongest positive sensitivity.

At upper levels as 1476m, ensemble sensitivity is considered similar to surface level. However, the magnitude of ensemble sensitivity is weaker than at the surface level (Fig. 6.2). These are indicating that the ensemble is more sensitive with V components of northeasterly wind and southeasterly wind.

At 5424 m, there is weaker sensitivity on December 9 for both U and V components of wind. On December 10, the U component has different sensitivity with positive sensitivity over the south part and negative sensitivity over the central of mid-central Vietnam. These differences are relevant to the different wind direction, between southwesterly wind over north and central and southeasterly wind over the south part of the study area (refer to fig. 3.7). Besides, the sensitivity of the V component is negative in inland and positive at the coastal line. On December 11, ensemble sensitivity is negative for the U component and weakness for the V component of wind (Fig. 6.3).

Overall, forecast rainfall is more sensitive for U component than V component of wind.

32 b) Vertical velocity.

Figure 6.4 shows the sensitivity of the forecast precipitation to the vertical wind field. In general, it is clear to see that the strong positive and negative sensitivity areas are interweave. Furthermore, increasing density with scale-up of altitude over study areas, which indicates the strong activity of the vertical motion plays an important role in the precipitation process in this area caused by due to transportation of moisture air to precipitation.

c) Equivalent potential temperature.

Figure 6.5 shows the sensitivity of the forecast precipitation to the equivalent potential temperature field. It can be seen that the positive sensitivity areas increase. The sensitivity magnitudes also increase with scale-up of altitude, which indicates the higher equivalent potential temperature in this region related to more precipitation during three days of the D18 event caused by warm-wet air contributes to precipitation.

d) Water vapor mixing ratio.

Considers the sensitivity of forecast 24-h precipitation to the water vapor mixing ratio show that on December 9, over the mid-central, the sensitivity is slightly positive at lower levels (surface and 1476 m) and strong negative at 5424 m. besides, a strong positive is in the northern part of the SCS. On December 10, the sensitivity is the positive at both lower levels and 5424 m. On the next day, there is still positive sensitivity.

However, the magnitude of sensitivity significantly decreased in comparison with the previous day. All these shows most moisture exists in the northern part of the SCS on September 9 and then concentrates mainly in the mid-central region on October 10. By the 11th, the moisture still exists mainly in this area. However, the source of moisture has decreased sharply (Fig. 6.6). Combine with the sensitivity of the equivalent potential temperature field (Fig. 6.5) and vertical motion fields (Fig. 6.4), indicating that high

However, the magnitude of sensitivity significantly decreased in comparison with the previous day. All these shows most moisture exists in the northern part of the SCS on September 9 and then concentrates mainly in the mid-central region on October 10. By the 11th, the moisture still exists mainly in this area. However, the source of moisture has decreased sharply (Fig. 6.6). Combine with the sensitivity of the equivalent potential temperature field (Fig. 6.5) and vertical motion fields (Fig. 6.4), indicating that high