l h cycling



Cycling, all data

Shanshan. Moreover, the temporal distribution of the soundings data is not homogeneous. It varies from zero at 0600 UTC 13 September to 13 soundings at 1600 UTC 13 September. A total of seven experiments are conducted (see Table 1).

The schematic diagram of all dataassimilation experiments is shown in Fig. 3, and the experimental domain is shown in Fig. 2(a) with the size of 222 x 128 in a 45 km grid distance. There are a total of 45 full vertical model levels with the n values of 1.0, 0.995, 0.988, 0.98, 0.97, 0.96, 0.945, 0.93, 0.91, 0.89, 0.87, 0.85, 0.82, 0.79, 0.76, 0.73, 0.69, 0.65, 0.61, 0.57, 0.53, 0.49, 0.45, 0.41, 0.37, 0.34, 0.31, 0.28, 0.26, 0.24, 0.22, 0.2, 0.18, 0.16, 0.14, 0.12, 0.10, 0.082, 0.066, 0.052, 0.04, 0.03, 0.02, 0.01, 0.0. The model top is 30hPa, and the time step is 180 s. The moist physics include the new Kain-Fritsch cumulus parametrization scheme and the WSM 5-class explicit scheme. The YSU PBL scheme and the Monin-Obukhov scheme are used as the boundary and surface layers' physics, and the Noah land-surface model is selected for the low-boundary condition over the land. The long-wave and short-wave radiation schemes, which are called every 30 min during the model integration, are the RRTM scheme and the Goddard scheme, respectively. The first experiment (NODA) is a no-data-assimilation experiment. The initial condition for this experiment is obtained from the NCEP AVN global analysis at 0000 UTC 13 September. The purpose of the NODA experiment is to serve as a benchmark for illustrating the impact of data assimilation.

In the data-assimilation experiments with WRF-Var (see the section on WRF-Var at the link users / tutorial/tutoriaLpresentation.htm), the

Figure 3. Schematic diagram illustrating the design of data-assimilation experiments.

background error statistics (BES) are obtained by interpolation from a 41-level CV Option-5 BES file derived based on the three months' (August, September, and October 2006) forecast data over the same domain with the NMC method. The observations are obtained from CWB (referred to as "CWB-obs"), and include the upper-air soundings, PILOT, surface data (SYNOP, SHIPS, METAR, BUOY), aircraft data (AIREP), and the satellite-derived or -retrieved data (satellite wind, QuikScat wind, SATEM, etc.). In addition to CWB-obs, the bogus data are also provided by CWB [Fig. 2(c)].

Two sets of data-assimilation experiments are conducted. The first set is cold-start experiments. For this set, data assimilation is performed at only one time, which is 0000 UTC 14 September. COLDNBNG is an experiment that does not assimilate bogus data or COSMIC GPSRO data. For CWB operations, two types of bogus data are used. The first is called "global bogus." To minimize "systematic drifting" of regional analysis, CWB extracts a set of bogus soundings from their global analysis, and treats them as synthetic sounding data. These soundings are located at regular intervals. Based on past experience, the assimilation of these global bogus soundings has been effective in maintaining consistence between regional and global analysis and in preventing regional analysis from drifting away from global analysis. When a typhoon is found within the analysis domain, CWB also creates a set of typhoon bogus soundings. These consist of wind soundings in the vicinity of the typhoon. For the Typhoon Shanshan case, a total of 134 global bogus soundings and 40 typhoon bogus soundings are available at 6 h intervals from 0000 UTC 13 to 0000 UTC 14 September 2006 [see Fig. 2(c)]. For the COLDNBNG (meaning cold-start, no-bogus, no-GPSRO soundings) experiment, only CWB-obs data are assimilated. Neither the bogus nor the GPSRO soundings are assimilated.

For the COLDNB experiment, global and typhoon bogus data are not assimilated. However, a total of 7 COSMIC GPSRO soundings (available within a 1 h interval centered at 0000 UTC 14 September) are assimilated, in addition to CWB-obs data. In COLDALL, the 134 global and 40 typhoon bogus soundings are assimilated, in addition to the 7 COSMIC GPSRO soundings and CWB-obs data.

The second set of data-assimilation experiments involves cycling data assimilation experiments. In these experiments, continuous assimilation is performed from 0000 UTC 13 to 0000 UTC 14 September at 1h intervals. Basically, the 1 h forecast is used as the first guess for the next analysis cycle. Then data that fall within +/— 30 min of the particular hour are assimilated. This procedure is repeated over the 24 h period (see Fig. 3). Obviously, the cycling experiments will be able to assimilate a lot more data than the cold-start experiments. For CYCLNBNG, neither bogus nor COSMIC GPSRO soundings are assimilated. However, more CWB-obs data are assimilated, in comparison with COLDNBNG. For CYCLNB and CYCLALL, they are similar to COLDNB and COLDALL, with the exception of continuous cycling. CYCLNB assimilates a total of 110 COSMIC GPSRO soundings, and CYCLALL assimilates an additional 870 (174 x 5) bogus soundings (which are available at 6 h intervals) in addition to the GPSRO soundings. With the assimilation of a much larger number of soundings, we would expect the cycling dataassimilation experiments to have a larger impact on the forecast.

Figure 4 shows the track and central pressure of Typhoon Shanshan in no-data assimilation, and cold-start series of WRF 3D-Var experiments. The best track and the observed central pressure are also shown in the figure. The no-data-assimilation (NODA) experiment has the worst performance, as it is essentially a 24-96 h forecast using the NCEP AVN analysis at 0000 UTC 13 September 2006 as the

Atmospheric Science Letters
Figure 4. (a) Track (the letters A, B, C, D, etc. indicate the 6h storm positions, and the contours are SLP from NODA at 0000 UTC 14 September 2006) and (b) central pressure of cold-start WRF 3D-Var experiments.

initial condition. The averaged track forecast error over the period of 0300 UTC 14 to 0000 UTC 17 September is 273 km. With the assimilation of CWB-obs data, the COLDNBNG

experiment had a noticeable improvement over the NODA experiment. The three-day averaged track forecast error is reduced from 273 km to 233 km. The assimilation of seven COSMIC

GPSRO soundings at 0000 UTUC 14 September had only very minor impact on the track and intensity forecasts. The results of COLDNB are almost identical to those of COLDNBNG. However, the assimilation of 134 global bogus soundings and 40 typhoon bogus soundings at 0000 UTC 14 September in the COLDALL experiment has a major impact on the forecast. The three-day averaged track forecast error is reduced to 140 km, almost half of those in the COLDNB experiment. Figure 4 also clearly shows that the track in COLDALL is much closer to that of the best track and the intensity closer to the observation. These results suggest that, for the cold-start experiment, the assimilation of bogus data (particularly the typhoon bogus data) has a positive and profound influence on typhoon track forecasting.

The results of cycling data-assimilation experiments are shown in Fig. 5. In comparison with the cold-start experiments, the continuous assimilation improves the results considerably. For example, the three-day averaged track errors for CYCLNBNG and CYCLNB are 144 and 111 km, while they are 233 and 232 km for the COLDNBNG and COLDNB experiments. Figure 5 also shows that the tracks of Typhoon Shanshan in CYCLNBNG and CYCLNB are much closer to the best track compared with their counterparts in the cold-start experiments. It is interesting to note that the assimilation of 110 COSMIC GPSRO soundings has produced a noticeable impact. The three-day averaged track error is reduced by 33km (from 144 km to 111km) with WRF 3D-Var. This is a 23% improvement. Also, the assimilation of COSMIC GPSRO soundings produces a storm with stronger intensity (by about 10 mb), which is closer to the observation. These results suggest that in order for the COSMIC GPSRO soundings to have an impact, it is essential that a continuous assimilation approach is used.

The incorporation of bogus data in the cycling experiments produces improvements for the first two days. However, for day 3, the

CYCLALL experiment performs worse than without the assimilation of bogus data. Figure 5 shows that the storm track is biased westward after one day. The storm then moves much slower than the observation, ending with a larger track error on day 3. The exact reasons for this poorer performance with the assimilation of the bogus data are not known. However, we suspect that this might be related to the details of the bogusing procedure and the background error statistics used in WRF 3D-Var, etc. For the CWB typhoon bogusing procedure, 40 wind soundings are extracted from a symmetric Rankine vortex plus the large-scale environmental flow. These soundings are assimilated into the system without further data quality control. For a real typhoon, the storm would often develop asymmetric structures. The assimilation of the bogus soundings would destroy these real asymmetric structures, and the model storm must recreate them. A continuous assimilation of bogus data implies that these counteracting procedures are repeated throughout the assimilation windows. One cannot make a general statement that the assimilation of a bogus vortex does not improve the forecast. However, it is clear that improvement of the typhoon bogusing procedure is needed to improve WRF 3D-Var performance, particularly in cycling experiments.

2.3. Comparison of WRF 3D-Var and WRF/DART ensemble assimilation

Over the past several years, a Data Assimilation Research Testbed (DART), a community dataassimilation facility for geosciences, has been developed at NCAR. DART includes a wide variety of ensemble filter assimilation algorithms (Anderson, 2003) which can be applied to a wide range of geosciences problems, including those of the atmosphere, oceans, atmospheric chemistry, and ionosphere. Details on DART can be found at

Figure 5. Same as Fig. 4 but for 3D-VAR cycling run experiments.

DART makes it easy to implement deterministic ensemble filter data-assimilation approaches with various types of numerical models. A number of models have already been implemented with DART, including the WRF model and the CCSM-3 Atmospheric Model (CAM-3). Both CAM-3/DART and WRF/DART have been used for the assimilation of GPSRO data. For example, using the CAM-3/DART system, Liu et al. (2007) demonstrated the importance of forecast error multivariate correlations, between specific humidity, temperature, and surface pressure, for the assimilations of GPSRO data. Liu et al. (2008) used the WRF/DART system to compare the performance of a nonlocal observation operator (Sokolovskiy et al., 2005) and of a local refractivity observation operator in the assimilation of GPSRO soundings. Since both the WRF 3D-Var and the WRF/DART ensemble filter data-assimilation system are available for the assimilation of GPSRO soundings, it would be desirable to compare the performance of these two data-assimilation systems for the same Typhoon Shanshan case. For such a comparison, we try to make the two systems as compatible as possible. For example, they use the same local refractivity observation operator, the same observational data sets, and the same model domain and grid configurations. For this article, WRF/DART assimilated only the key observation types, including the radiosonde, satellite wind, QuikScat wind from CWB, and GPSRO data from COSMIC/CDAAC. The WRF/DART system is also set up for continuous 1 h cycling, in ways similar to that of WRF 3D-Var cycling experiments. For the WRF/DART system, we do not assimilate the global and typhoon bogus data.

Figure 6 compares the tracks and central pressures of Typhoon Shanshan in the WRF/ DART and WRF 3D-Var experiments. The track map shows that the WRF/DART experiments (DARTNBNG and DARTNB) follow the best track closely. This is also reflected in Table 2. For a three-day average, DARTNBNG and DARTNB have track errors of 75 and 62 km, respectively. They represent almost a 50% improvement over the WRF 3D-Var experiments (e.g. CYCLNBNG and CYCLNB), which have 144 and 111km, respectively. The central pressure time series also indicates that the WRF/DART experiments produce stronger typhoons, although they differ only by less than 5mb initially (at 0000 UTC 14 September). By 36 h, they differ by more than 20 mb, with

WRF/DART producing typhoons with more realistic intensity. It is interesting to note that in terms of track errors, WRF/DART without the assimilation of global and typhoon bogus data performs better than WRF 3D-Var which assimilates everything, including global and typhoon bogus data, after one day [as shown in Table 2]. Figure 6(c) also indicates that the assimilation of COSMIC GPSRO soundings with the WRF/DART system improves the track forecasts. The three-day averaged track error is reduced from 75km to 62km (Table 2), a 17% improvement.

The more realistic simulation of Typhoon Shanshan by WRF/DART experiments is illustrated in Fig. 7. Here we show vertically integrated cloud water (which serves as surrogate cloud fields) for CYCLNB, CYCLALL, and DARTNB experiments, together with the observed IR satellite images. Even with a horizontal resolution of 45 km, the DARTNB experiment clearly shows an eye for Typhoon Shanshan and the eyewall clouds. In contrast, the corresponding WRF 3D-Var dataassimilation experiment, CYCLNB, does not show an eye. The CYCLALL experiment gives the hint of an eye, with less cloud water in the center of the storm. However, its position is biased considerably westward when compared with that of the observed storm.

The more realistic structure of Typhoon Shanshan after one day of data assimilation with WRF/DART is illustrated in Fig. 8, which shows the potential temperature and tangential winds along a north-south cross section (along 125.8°E) that cuts across the center of the storm. The WRF/DART experiment without the assimilation of GPSRO produces a vortex, with a radius of maximum wind of about 250 km. The maximum tangential wind exceeds 15ms-1. With the assimilation of GPSRO data from COSMIC, the tangential winds are increased, and the radius of maximum winds is decreased slightly. In contrast, WRF 3D-Var produces a much weaker


Was this article helpful?

0 0

Post a comment