HTML
-
One target of this study is to investigate the impacts of multiple options on the precipitation over TP and the surrounding regions. Considering configurations are highly correlated with each other; therefore, based on land static data optimization and the optimized Noah-MP, 4 categories of 19 sensitive experiments are comprehensively examined to further discuss the model performance, including the nudging setting, 5 CUs, 2 PBLs, and 6 MPs. All these numerical experiments are carried out in 2015 because more in situ observation and the new land use over TP from the TPDC are available during this period. Table 1 shows the summary for each experiment.
Categories Experiment Description Quantity Nudging Spectral nudging The wave number17, 7. 2 Grid nudging Turn on / off nudging q. 4 Nudging q coefficients is 0.0001/0.0002/0.0003, respectively. Cumulus parameterization (CUs) GD Grell-Devenyi Ensemble Scheme, multiple cumulus schemes and variants running within each grid box. The results are averaged with typically 144 sub-grid members. 5 G3 Grell-3 Scheme, an improved version of the GD scheme, is more suitable for fine mesh grids (≤ ~10 km). KF Modified Kain-Fritsch Scheme, using the mass flux with downdrafts and CAPE removal time scale for deep and shallow convection sub-grid scheme. MKF Compared with the KF scheme, Multi-Scale Kain-Fritsch Scheme, due to the introduction of new functions such as resolution-dependent parameters, has the function of adapting to the time scale of the convective process. NoCU Without cumulus parameterization. Planetary boundary layer (PBLs) MYNN2.5 Mellor-Yamada-Nakanishi-Niino (MYNN) Levels 2.5, the local TKE (Turbulent Kinetic Energy) scheme is suitable for the stable environment and complex terrain in the polar region. 2 YSU The Yonsei University PBL, a revised vertical diffusion package with a nonlocal turbulent mixing coefficient, is an explicit treatment of entrainment processes at the top of the PBL. The MKF can only be combined with YSU in this model version. Microphysics (MPs) M100** Morrison double-moment scheme, good performance in the polar region. The default value of the droplet concentration is reduced to 100 cm-3 in this study. Hydrometeor: five species including cloud droplets, cloud ice, rain, snow, and graupel/hail.
Ice-Phase Processes: Yes
Mixed-Phase Processes: Yes6 P51** Predicted Particle Property scheme, one ice category that represents a combination of ice, snow and graupel, using the double moment rain, ice and cloud water setting.
Hydrometeor: 3 species including cloud droplets, cloud ice and rain.
Ice-Phase Processes: Yes
Mixed-Phase Processes: YesTPA** Thompson aerosol-aware, water- and ice-friendly aerosols.
Hydrometeor: 5 species including cloud water, cloud ice, rain, snow, and graupel/hail.
Ice-Phase Processes: Yes
Mixed-Phase Processes: YesWDM6** WRF Double-Moment 6-class scheme, double moment warm-rain but is otherwise like WSM6.
Hydrometeor: 6 species including water vapor, cloud, rain, ice, snow, and graupel.
Ice-Phase Processes: Yes
Mixed-Phase Processes: NoWSM5* WRF Single-Moment 5-class scheme allows for mixed-phase processes and super-cooled water.
Hydrometeor: 5 species including water vapor, cloud, rain, ice, and snow.
Ice-Phase Processes: Yes
Mixed-Phase Processes: NoWSM6* WRF Single-Moment 6-class scheme, added graupel based on WSM5, suitable for high-resolution simulations.
Hydrometeor: 5 species including water vapor, cloud, rain, ice, snow, and graupel.
Ice-Phase Processes: Yes
Mixed-Phase Processes: No** double-moment, *sigle-moment Table 1. Summary of the characteristics of each experiment.
Nudging can force the model simulation towards a series of analyses grid-point by grid-point (Skamarock et al. [81]), which is used to reduce model drift (e.g., Xue et al. [65]; Bromwich et al. [67]; Hines et al. [79]; Glisan et al. [99]), particularly for simulations of extended periods (e.g., months). Inherited from the WRF, Polar WRF can afford to nudge u, v, t, and q from driving fields with spectral and grid nudge methods. Previous studies have shown that model skills are very sensitive to nudge settings when long-term runs or dynamic aspects need to be addressed (Xue et al. [66]; Cha et al. [100]; Liu et al. [101]; Otte et al. [102]; von Storch et al. [103]). Therefore, the 6 experiments (nudging category in Table 1) for methods and nudging coefficients are investigated to increase the knowledge of nudging settings.
Several excellent investigations have identified that CUs schemes have a considerable impact on the simulation of precipitation over TP and nearby areas (Gao et al. [3]; Gao et al. [48]; You et al. [49]; Lin et al. [51]; Ou et al. [53]; Wu et al. [56]; Yu et al. [58]; Gao et al. [97]; Chen et al. [104]; Gao et al. [105]; Qian et al. [106]; Zheng et al. [107]). Referring to these findings, here we mainly study the influence of CUs settings by the following 3 typical kinds of simulation (a total of 5 experiments). (1) The Grell-Devenyi Ensemble Scheme (GD) and its improved version Grell-3 scheme (Grell et al. [108]). The GD combines multiple schemes and variants run within each grid point and gives an average feedback result. Currently, each grid box has 144 members in this scheme. Grell-3 (G3) is based on GD and is especially suitable for high-resolution simulations that should not be coarser than 10 km (Skamarock et al. [81]). (2) Kain-Fritsch schemes, including the Modified Kain-Fritsch (KF) (Kain [109]) and Multi-Scale Kain-Fritsch (MKF) (Zheng et al. [107]). KF using the mass flux with downdrafts and CAPE removal time scale for deep and shallow convection sub-grid scheme. The MKF stemmed from the original KF scheme and introduced new functions, such as resolution-dependent that adapt to the time scale of the convective process. (3) The NoCU represents the CUs that are not used, which also achieves good results at the "gray zone" grid spacing (~4–10 km resolution) (Ou et al. [53]; Qian et al. [106]).
For the PBLs, the Mellor-Yamada-Nakanishi-Niino (MYNN) Levels 2.5 (MYNN2.5) (Nakanishi et al. [110]) and the Yonsei University schemes (YSU) (Hong et al. [111]) are examined here. MYNN2.5 is tested by the Polar Meteorology Group, which is outstanding for polar environmental features and orographic influences (e.g., Xue et al. [65]; Bromwich et al. [67]; Hines et al. [78, 79]). Moreover, the improvement of the reduction in the downward shortwave radiation bias has been verified as a significant improvement of Polar WRF V 4.1.1 (Xue et al. [65]; Olson et al. [112]). For the TP, YSU also has been frequently used (Ou et al. [53]; Gao et al. [97]; Chen et al. [104]; Gao et al. [105]; Qian et al. [106]). More importantly, when the MKF scheme is selected for CUs, the PBLs matched with YSU; that is, only the combination of MKF +YSU can be used in the model right now (Skamarock et al. [81])
MPs dealing with cloud and precipitation particles must be parameterized in the numerical weather and climate models. The parameterizing of microphysics remains highly challenging because of the complexity of the underlying physics and the lack of understanding of these coupling processes (Bauer et al. [113]; Benjamin et al. [114]; Randall et al. [115]). Hence, a reasonable selection and modification of MPs are expected to achieve better simulation results of clouds and precipitation. Here, MPs are studied based on the research advances that have been extensively tested in the Arctic, Antarctica, and TP (Gao et al. [48]; You et al. [49]; Xue et al. [65, 66]; Bromwich et al. [68]; Hines et al. [78]; Hines and Bromwich [80]; Gao et al. [97]; Qian et al. [106]; Zhu et al. [116]; Li et al. [117]; Ma et al. [118]; Shen et al. [119]; Yang et al. [120]; Zhang et al. [121]). Modify the liquid water droplet concentration in Morrison 2-mom to 50 cm-3 which has been verified to significantly improve the performance of polar clouds such as supercooled water in clouds (Xue et al. [66]; Hines et al. [78]; Hines and Bromwich [80]). TP, similar to the Antarctic and Arctic, is a colder region, and the liquid water of clouds more easily exists in the form of supercooled water. For the TP, the Morrison 2-mom specified droplet concentration of 100 cm–3 (referred to as M100) because of a greater amount of could particles (probably twice for maximum) and more prone to supercooled water than the Arctic (Yan et al. [122]). P3 (Morrison et al. [123]) used a new approach to predict the transformation of ice particles, named "one ice category, " which represents a combination of ice, snow, and possibly graupel. The setting of double-moment for rain, ice, and cloud water and supersaturation-dependent activation is applied in this study (P51). Thompson aerosol-aware (Thompson et al. [124]) has been modified for mid-latitude convective, orographic, and snowfall conditions and considers water- and ice-friendly aerosols. WRF Double-Moment 6-class scheme is the double moment for warm-rain, and other processes are basically the same as WRF Single-Moment 6-class scheme (WSM6) (Lim et al. [125]). The condensation nuclei (CNN), such as cloud water and rain, adopt the "prediction method." It performs reasonably well for high-resolution simulation studies in China (Zhu et al. [116]; Li et al. [117]; Shen et al. [119]; Zhang et al. [121]). Compared with the two-moment schemes, single-moment schemes such as WRF Single-Moment 5-class scheme(WSM5) and WSM6 are often used as they can maintain a good balance between model skills and computational resources. And the WSM6 is also used and tested by the GRAPES_Meso operational weather forecasting model (Shen et al. [119]; Zhang et al. [121]; Huang et al. [126]; Ma et al. [127]). Table 1 shows a summary of the microphysics mentioned above, the hydrometeors, and whether ice-phase/mixed-phase processes are also included.
-
Considering the limited performance of observation and reanalysis over TP, multiple datasets are used to evaluate the simulations. The in situ observation obtains the daily precipitation dataset located in the TP and surrounding areas supplied by the TPDC (1979–2015, about 184 stations Fig. 1) (China Meteorological Data Network [128]). Satellite precipitation estimation products used the Tropical Rainfall Measuring Mission (TRMM) 3B42 V7 dataset with 3-hourly and 0.25° × 0.25° resolution (Huffman et al. [129]). CRA-RA/Land, the first generation global atmosphere and land reanalysis released by the China Meteorological Administration. Probably assimilation of various conventional and satellite observations over East Asia, CRA-RA/Land is in good agreement with and even outperforms other reanalyses such as ERA5 datasets in near-surface variables over China (Li et al. [130]; Shen et al. [131]; Yang et al. [132]; Yu et al. [133]; Zhang et al. [134]). Precipitation estimates from CRA-Land with a horizontal resolution of ~34 km and a 6-hour temporal interval. The ERA-Interim precipitation has been investigated as a benchmark and is also used in this study. The data is provided every 3 hours with a horizontal resolution of 0.75°×0.75°.
Due to the different spatial resolutions, the grid data are extracted and horizontally interpolated onto the domain by the WRF Preprocessing System (WPS) (e.g., Xue et al. [65]; Bromwich et al. [67]) with the same re-grid method. It is beneficial for comparisons using the same resolution (10 km) and grid points through the utilization of bilinear interpolation to interpolate grid data to station locations when comparing data between grid and in situ.
3.1. Numerical experiment design
3.2. Validation data and methods
-
In this section, the multiple aspects of the model option for precipitation over the TP nearby areas are examined by the sensitivity experiments. Considering the main targets of this study and successor work, the variations of monthly and daily precipitation between multiple model settings are evaluated against in situ, satellite observations, and reanalyses. It focuses on the spatiotemporal comparison of large-scale and average conditions and does not analyze the specific rainfall processes, such as comparing rainfall forecasts by the Threat Scores (Jiang et al. [135]; Mesinger [136]; Yang and Tung [137]). Here, January and July, the two months with the largest difference in a year are used to investigate. Since the precipitation is mainly occupied in the summertime, the analysis of the precipitation characteristics of each experiment in July 2015 is discussed below.
-
Figure 2 shows the results of precipitation in July 2015 under different nudging settings and the references with ERA-Interim (EI), CRA-Land (CRA), and TRMM 3B42 (TRMM), respectively. Overall, the model well captures the main characteristics of precipitation distribution over the domain. Precipitation decreases from southeast to northwest, and there is a clear borderline along the edge of the plateau (such as an altitude of 3000 m). The Bay of Bengal has the most precipitation amount, with a maximum value of more than 800 mm, whereas the Tarim Basin is the region with the least, with most areas less than 10 mm. At the same time, it can also be seen that EI has the largest precipitation, with most areas on the plateau having more than 50 mm during this month. CRA and TRMM are roughly similar, and the covering of precipitation greater than 50mm is significantly smaller than EI. For TRMM, we also noticed that the precipitation in the northern part of the domain (north of 35ºN, such as the Tianshan area) is insufficient. It is likely to be related to the low coverage of TRMM data in extratropical regions.
Figure 2. Precipitation of July 2015 over the model domain. (a)-(c) the model results with specified nudging settings. (a) Spectral nudging wave number was specified as 17, (b) the wave number was specified as 7, and (c) grid nudging; (d)-(f) ERA-Interim (EI), CRA-Land (CRA), TRMM 3B42 (TRMM). The black solid line indicates the altitude of 3000 m.
Further analysis shows that the results of spectral nudging (Fig. 2 (a) and (b)) located a very weak anomalous precipitation belt along the southern side of the Himalayas. This belt is also accompanied by an abnormally strong surface short-wave down and weak long-wave radiation (not shown), which indicates that the low cloud cover and the convective process might be restricted. In addition, the tests also found that the intensity of this abnormal precipitation belt is related to the wave number used. For instance, when the wavenumbers are selected as 17, 7, and 3 (the cutoff wavelength is roughly equal to ~350 km, ~800 km, and ~1800 km, respectively), the belt gradually blurred with the wavenumber decreases. In other words, the model has more freedom as the wave number decreases, and the impacts of nudging on medium scales are also weakened. That means the modulation for overestimation precipitation is suppressed. The grid nudging result is different from spectral nudging, which is consistent with EI, CRA, and TRMM maintaining a strong precipitation belt on the south side of the Himalayas. Some possible reasons can be inferred as follows. The spectral nudging skills are determined by the nudge scale, which is related to the model domain and the scale of driving fields (Skamarock et al. [81]; Liu et al. [101]; Miguez-Macho et al. [138]). The southern foothills of the Himalayas are more sensitive to scale due to the complex local topographical and convective variations. Thus, the skill of spectral nudging really penalizes here. However, the option of grid nudging can avoid "scale selection" and good performance with precipitation and moisture adjustment (e.g., Bowden et al. [139]; Wootten et al. [140]). Therefore, it shows better performance.
Nudging toward the water vapor mixing ratio is believed to be beneficial for the simulation of precipitation and moisture (e.g., Mai et al. [141]; Spero et al. [142]). Therefore, the impacts of nudge q and its strength are examined. First of all, sensitive test results indicate nudging q can significantly improve precipitation on the plateau. In the central and northern parts of the plateau, where precipitation is overestimated (such as exceeding 50 mm), there is a significant decrease (not shown). More details are shown in Fig. 3. It gives the difference in monthly precipitation between model results with specified grid nudging settings and EI, CRA, and TRMM, respectively. Although the simulation precipitation minus each dataset shows more variation, it is consistent that the superfluous precipitation can be reduced by nudging q, and the greater the intensity of the nudging coefficient, the more significant the effect on the reduction of precipitation. Compared with EI, the model mainly shows a negative difference, especially along the western and southern mountain margins, where there is a precipitation belt with negative 80–100 mm. This is probably caused by the overestimation of precipitation over these areas from EI. Both the differences in the model minus CAR and TRMM indicate the negative values in the west while positive in the central and eastern parts of the plateau. The maximum positive region is located at the east edge of the plateau and the Hengduan Mountains. It is consistent with the general inability of current models to simulate precipitation in this region (Chen et al. [45]). It is worth noting that with increasing the intensity coefficient of nudging q, the positive trend on the central and eastern TP will decrease, and the changes are intensifying correspondingly in the western plateau. Thence, trade-offs for the obvious regional features possessed by nudging q. The intensity coefficient is 0.0002 for q and other settings, including nudging u, v, t, and ph, using the same value of 0.0003. All the nudging applied above the vertical 35 layers higher than the top of the PBL reasonably improved the model performance at the upper layers (Xue et al. [65]; Mai et al. [141]).
Figure 3. Precipitation difference in July 2015 with specified grid nudging settings. Model minus ERA-Interim (EI), CRA-Land (CRA), TRMM 3B42 (TRMM), respectively. The columns from left to right are turned off nudging q, nudging q coefficients as 0.0001, 0.0002 and 0.0003, respectively. The black solid line indicates the altitude of 3000 m.
-
Since EI significantly overestimates precipitation in the southern TP and TRMM underestimates in the north, Fig. 4 only shows the monthly precipitation difference of the model minus CRA for July 2015. At present, MKF can only be combined with YSU. Therefore, the suite of cumulus with MYNN2.5 and YSU are tested, respectively. It is clear that the model simulated with CUs usually produced too much precipitation, but NoCUs seems precipitation is not enough conversely. Furthermore, the impacts of different CUs on precipitation are also evident. Fig. 4a-c shows that G3 has the most significant positive difference in precipitation, and KF produces too much precipitation covering large areas such as the east and south of TP (the Hengduan Mountains). We also noticed that the G3 scheme produced more precipitation than GD (Fig. 4b-c), which had been investigated by the simulation analysis of the precipitation over the southern slopes of the TP (Wu et al. [56]). The MKF performed much superior to KF (Fig. 4d-e). It means the scale adaptive cumulus scheme shows better skills for precipitation over the TP. In other words, it is that making use of scale-aware performs superior for solving interaction/uncertainty problems, especially for simulations forced by complex terrain and low confidence in initial and lateral boundary conditions. Therefore, MKF is suggested for priority use.
-
Figure 4 also remarkably indicates that GD +MYNN2.5 and MKF+YSU are better at simulating precipitation. And MKF+YSU seems more sophisticated as over the eastern side of the TP (such as west of Sichuan Basin), the central part (the Three-River headwater region), and northern India (figure omitted), the precipitation performed by GD+MYNN2.5 more superfluous than that of MKF+YSU. Due to the different setting suites, KF+YSU and KF+MYNN2.5 served as supplementary to determine MKF and YSU, which one improved model skills more significantly. Compared with MKF+YSU (Fig. 4d), KF+YSU (Fig. 4e) overestimated precipitation at central and eastern TP. It indicates that MKF has better performance than KF. Correspondingly, the results of KF+MYNN2.5 (Fig. 4a) and KF+YSU (Fig. 4e) revealed that MYNN 2.5 shows the ability to improve precipitation slightly compared to TP. However, as the MKF performs greater modeling skills, the MKF+YSU performs much better. In other words, compared with MKF, both YSU and MYNN2.5 could have a limited impact on precipitation simulation.
-
Based on the optimized settings for nudging, CUs, and PBL, microphysics options can be further examined to expand the knowledge for a "better-performing" configuration suite. Fig. 5 compares the spatial distribution of the precipitation difference from the model with specified microphysics minus CRA. Significant differences indicate that the model configured with different microphysics offers a wide variety of capabilities or limitations. Among the four double-moment schemes, M100, P51, and TPA carried similar results. Compared with CRA, precipitation is overestimated in the eastern slope of TP (such as the western Sichuan Basin), the southern foot of the Nyainqentanglha Mountains, and the western slope of the Hengduan Mountains, etc. On the contrary, in the Tianshan Mountains and other places in the west of the TP (west of ~85ºE), the modeling precipitation is less than that of the CRA. Specifically, M100 is close to P51. However, the negative values of M100 in the west of the plateau are smaller, and it is much closer to CRA. TPA is slightly poor, and too much precipitation is produced in the southeast, such as Yunnan and Guizhou. For WDM6, it is evident precipitation is underestimated in the east of TP while slightly overestimated in places such as the Kunlun Mountains in the northwest of TP. In the two single-moment schemes, WSM5 is similar to WDM6 but with significantly less precipitation. WSM6, which extended the WSM5 by adding additional processes related to graupel (Hong [143]), performs much better than WSM5 and is roughly comparable to M100 and P51.
Figure 5. Precipitation difference in July 2015 from the model with specified MPs. Model minus CRA-Land. The black solid line indicates the altitude of 3000 m.
To evaluate the finer differences, the RMSE (Root Mean Square Error) and precipitation difference from model data against in situ observations were calculated. Fig. 6 shows the RMSE value and spatial distribution from the model with specified MPs accompanied by EI, CRA, and TRMM that are used as references. First, RMSE exhibits increasing from the northwest (values are usually less than 4 mm) to the southeast (most are greater than 10 mm). In particular, the maximum RMSE is mainly distributed in the Hengduan Mountains and on the edge of the Himalayas. One aspect is likely related to the higher absolute precipitation that produces larger biases. On the other hand, it is still challenging for the model to simulate precipitation under the complex terrain and strong water vapor transport conditions over the TP and surroundings. Compared with different MPs, M100 and WDM6 show less RMSE than the other two double-moment schemes at the southeast of TP. The single-moment schemes indicate a similar spatial pattern to that of double-moment schemes, and WSM5 shows less RMSE than WSM6. Furthermore, we analysed the differences in monthly precipitation. Although WDM6 and WSM6 perform better RMSE, precipitation is systematically underpredicted compared to in situ observation (Fig. 7), which is consistent with the previous findings compared with CRA (Fig. 5). By comparison with in situ dataset, Fig. 6 and Fig. 7 also verify that the CRA shows the best performance in the 3 kinds of grid precipitation products.
Figure 6. Precipitation RMSE (colored dots) in July 2015 with from the model with specified MPs and ERA-Interim (EI), CRA-Land (CRA), and TRMM 3B42 (TRMM), respectively. Model/Reanalysis minus in situ observation. The black solid line indicates the altitude of 3000 m.
Figure 7. Precipitation difference (colored dots) in July 2015 from the model with specified MPs and ERA-Interim (EI), CRA-Land (CRA), and TRMM 3B42 (TRMM), respectively. Model/Reanalysis minus in situ observation. The black solid line indicates the altitude of 3000 m.
Considering the significant spatial variation of precipitation differences through the TP, the region is divided into four sub-regions along 34ºN (Li [144]) and 95ºE (Gao et al. [3]). That is, Northwest (NW), Northeast (NE), Southeast (SE), and Southwest (SW), and the number of in situ stations in each sub-region are 24, 53, 68, and 29, respectively. Fig. 8 gives the whole TP and its sub-regions of station-averaged precipitation in July 2015 from the model with specified MPs. It can be seen that among the four sub-regions, the precipitation in the SE is the largest while the NW is the smallest, which is about one-tenth of SE. NE and SW are equivalent, probably half of the SE. These characteristics are consistent with the spatial distribution discussed above that the precipitation decreases from southeast to northwest over TP. The difference is mainly in the southern TP (south of 34°N), especially in the SE and SW, where MPs perform significant impacts on precipitation. For instance, in SE where is the most significant precipitation difference, the maximum (TPA 201 mm) is more than 2.5 times larger than the minimum (WSM5 77 mm). From Fig. 8, the overarching findings are that the M100, WSM6, P51, and TPA are relatively overestimated precipitation, whereas WSM5 and WDM6 are underestimated precipitation. It also can be seen that M100 is superior to others and is closest to CRA and in situ observation, respectively.
-
Standing on the "better-performing" configuration investigated above, the performance of precipitation simulation using Polar WRF can be confidently further discussed in the following sections. The refined model parametrizations are as follows: grid nudging is used, where the q intensity coefficient is 0.0002 while the rest are 0.0003; the MKF cumulus parameter scheme; the YUS boundary layer parameter scheme and the M100 as the microphysics. Here, as mentioned in the introduction, a full calendar year simulation and compared with multiple reanalyses and observations are carried out.
-
Figure 9 shows the annual precipitation over TP and the surrounding regions from simulation, EI, CRA, TRMM, and in situ observations. All the results show an evident similar pattern in that precipitation decreases from southeast to northwest, as discussed above. Further comparison with station observations, EI exhibits more precipitation over the domain while TRMM does not have enough precipitation north of ~35ºN. For CRA, most areas are very close to the observation except for some stations on the edge of 3000 m of the southeast. Although the precipitation on the western slope of the Hengduan Mountains and the eastern slope of TP is relatively heavy, the distribution of M100 performs consistently with observations. By examining the annual precipitation frequency, the abnormally high frequency of precipitation can be seen in the EI, CRA, and M100. But M100 shows much better performance than the two reanalyses, such as the western slopes of the Hengduan Mountains, the eastern slopes of TP, and the Tianshan Mountains (Fig. 10). Corresponding to these features, the annual precipitation intensity from models and reanalyses is also weaker than observations. M100 presents superior to the two reanalysis data, especially in the southeast of TP (Fig. 11).
Figure 9. The distribution of annual precipitation over the Qinghai-Tibet Plateau and the surrounding regions. The black solid line indicates the altitude of 3000 m.
-
Monthly and daily precipitation can describe more spatial and temporal variations and have been used to examine precipitation performance (e.g., Bai et al. [61]; Li et al. [130]; Jeworrek et al. [145]). Fig. 12 shows monthly station-averaged precipitation, including the whole TP and its sub-regions from January to December 2015. Precipitation has significant annual temporal variation and regional spatial characteristics. The precipitation is mainly concentrated in the summer half year. The amounts from May to September account for 70%–80% of the annual precipitation. Among them, in situ observation (OBS) is 80.0%, M100 is 74.4%, EI is 72.2%, CRA is 79.5%, and TRMM is 80.3%. Moreover, the precipitation from April to October contributed more than 90% of the annual precipitation (only EI accounted for 87.4%, and the rest ranged from 90.2% to 92.9%). This intra-annual variability is in step with the outbreak, maintenance, and retreat of the Asian summer monsoon.
Figure 12. Monthly station-averaged precipitation from January to December in 2015. (a) NW, 24 stations; (b) NE, 57 stations; (c) SE 68, stations; (d) SW, 29 stations; (e) TP, 178 stations.
For the three precipitation products, CRA and TRMM are relatively close to those observed by the station, and CRA is slightly larger. The precipitation from TRMM in the northern TP (NW, NE) is somewhat underestimated, whereas that in the southern TP (SE, SW) is slightly overestimated. This is also the reason why TRMM has the smallest statistical precipitation difference between the whole TP and station observations. Obviously, EI shows too much precipitation, especially in the sub-regions of SE and SW. Two reasons might be related to this. One, EI has relatively large precipitation in these areas, so the ability to represent precipitation in complex terrain areas such as the south is limited. The two horizontal resolution of EI is the lowest in the four grid data used here; it may produce larger errors than others after interpolation.
The model's monthly precipitation is evidently better than EI and generally close to CRA and TRMM, even particularly slightly superior to CRA and TRMM in the NW and SW sub-regions. However, over the SE, the monthly precipitation trend is quite well depicted by the model, but the simulated precipitation is still too overestimated. This is probably due to the sparse observations and the limited direct observation data available for data assimilation at the NW and SW. CRA failed to show a better advantage, while the downscaling of the model has certain advantages. In particular, the model optimizes the ice-snow-atmosphere interaction and performs better in the area with large ice and snow coverage in the west of TP. On the contrary, observations in the eastern region are relatively abundant, and CRA performs better in this region as richer observations have been assimilated (Li et al. [5]; Li et al. [130]). In addition, since the initial and lateral boundary conditions derived from EI, it will also lead to large positive abnormalities where precipitation is overestimated dramatically, such as in the southeast of TP. These findings also suggest the importance of data assimilation (improvement of initial conditions) and model-driving data for the regional model.
Unifying the mode results in the daily temporal resolution as the in situ, Fig. 13 displays the time series of daily station-averaged precipitation in 2015. The model shows an excellent ability to capture the daily precipitation evolution during the annual cycle. It skillfully describes the precipitation characteristics over TP and nearby areas. In the view of temporal distribution, precipitation is mainly produced from April to October (90–300 days on the X-axis in Fig. 13), which is consistent with the above finding. The daily precipitation also shows good agreement between the model and observations. The disadvantage is that the peak of daily precipitation simulated by the model is smaller than the observation. It indicates that the model is limited in simulating some heavy/extreme precipitation, which also caused the shortcoming of weak intensity for precipitation mentioned in 4.2.1. In addition, it can also be seen clearly that the daily precipitation in the SE from April to July (90–210 days on the X-axis in Fig. 13) is greater than the observation. Which further verifies the results discussed above.
Figure 13. The time series of daily station-averaged precipitation in 2015. The grey solid line and black dashed line indicate observation and model output, respectively. (a) NW, 24 stations; (b) NE, 57 stations; (c) SE, 68 stations; (d) SW, 29 stations; (e) TP, 178 stations.
To combine multiple data into a comprehensive evaluation, Fig. 14 shows the daily precipitation of the Taylor diagram for M100, CRA, and TRMM, respectively. The daily precipitation simulated by the model (black dot) is at the forefront among the results, and the correlation of each region is in the range of 0.75–0.86 (Table 2). which is better than others. Also, the distribution of each region exhibited in the figure is relatively concentrated, which indicates the model skills in various regions of TP are relatively stable. In general, the model performance is comparable to that of CRA, especially in the NW and SW, which is better than the former. This is also supported by a quantitative comparison of the correlation, bias, and RMSE that are shown in Table 2. Although the monthly precipitation from TRMM is the closest to the observation, the analysis of daily precipitation shows that it is not superior to others. The results generally show a smaller correlation and larger RMSE (the blue box in Fig. 14). This suggests that TRMM is limited for characterizing the evolution of the refined precipitation. EI (the red triangle in Fig. 14) performs worst, and the ability to describe the daily precipitation is extremely limited. Moreover, the distribution of each point is too widespread to draw the SW whose results are beyond the scope of the illustration
Figure 14. Daily precipitation of Taylor diagram for Model (M100), ERA-Interim (EI), CRA-Land (CRA) and TRMM precipitation (TRMM) compared with what was observed in 2015. The direction of the azimuth cosine is the correlation, the distance to the azimuth is the normalized variance ratio, and the distance to the reference point (OBS) represents the RMSE of the observation. The poor results of EI in the SW area are not shown in the figure; that is, only 4 values are displayed.
TP NW SW NE SE CORR M100 0.86 0.75* 0.76* 0.75 0.81 CRA 0.89* 0.62 0.75 0.76* 0.84* Bias M100 0.44 –0.05* 0.39* 0.05* 0.99 CRA 0.15* 0.07 0.41 0.09 0.14 RMSE M100 0.95 0.37* 1.15* 1.10 2.08 CRA 0.73* 0.54 1.30 1.06* 1.52* * Indicates the better one Table 2. The correlation, bias and RMSE of M100 and CAR-Land against in situ observation for daily precipitation.