Midsummer precipitation prediction over eastern China by the dynamic downscaling method

This study assesses the midsummer precipitation prediction over eastern China by the dynamic downscaling method. Based on the Climate Forecast System version 2 of the National Centers for Environmental Prediction and the Weather Research and Forecasting Model, the prediction performance of global and regional models on the July precipitation over eastern China is further analyzed by hindcast experiments from 1982 to 2010 and prediction experiments from 2011 to 2021. The results suggest that the regional model forced by the global model can noticeably improve the prediction skill for precipitation in eastern China, especially in the region from the South of North China to the Yangtze River Basin, referred as the northern China in this paper. In addition, we perform a diagnostic analysis of the reason for the improvement of the model prediction skill. The results indicate that the high resolution of the regional model and the refinement of physical process parameterizations contribute to improving the simulation ability for the East Asian atmospheric circulation pattern, heat flux, especially for the meridional teleconnection pattern in East Asia and the sensible heat flux in the northern China, thus further improving precipitation prediction.

processes, initial value errors, and limited spatial resolution, global climate models (GCMs) have limited ability to simulate and predict the key characteristics of some atmospheric circulation systems and local climate elements (Li et al., 2005;Wu et al., 2014).Especially, precipitation prediction is not improved obviously by GCMs (Saha et al., 2014).Therefore, it is a vital topic for downscaling of precipitation based on high-skill variables of GCMs, including statistical downscaling and dynamic downscaling methods (Brogli et al., 2019;Gu et al., 2011;Huth, 1999;Sun et al., 2018).Using the initial and boundary conditions provided by the GCMs to drive regional climate models (RCMs) can effectively study and predict local climate characteristics and variations when the computer power permits (Xu et al., 2019;Xue et al., 2014).Currently, the widely used RCMs include the MM4, RegCM4 (Giorgi & Bates, 1989;Cheng et al., 2020), and the Weather Research and Forecasting Model (WRF, Skamarock et al., 2005, Skamarock & Klemp, 2008, Qiu & Im, 2021).
Many researchers have analyzed the simulation and prediction effects of the WRF model (Gao et al., 2015;Pan et al., 2012;Xu & Yang, 2012).The initial field and lateral boundary conditions of the RCMs are provided by the GCMs.Through higher horizontal resolution (Caron et al., 2011;Chan et al., 2013;Xue et al., 2014), finer regional complex terrain (Antic et al., 2006;Liu et al., 2022;Zhang et al., 2013) and finer physical processes (Jankov et al., 2005;Nasrollahi et al., 2012;Xia et al., 2019), the RCMs may improve the GCM simulation to some extent.Due to the coarse resolution of GCMs, there are large biases in the simulation and forecast of the climate in China.Previous studies have attempted to use WRF downscaling to improve the simulation and forecasting skills of precipitation over China (Bao et al., 2015;Huang & Gao, 2018;Ma et al., 2015;Tang et al., 2016).For instance, Ying et al. (2022) utilized the global Climate Forecast System of Nanjing University of Information Science and Technology (NUIST-CFS1.0)data to drive the WRF, and the results showed that this model could better predict five of the seven summer extreme precipitation events over the Yangtze River Basin in China from 1982 to 2020.
Numerous studies have shown that the WRF model can further improve climate simulations and predictions.However, there is still a lack of insight into the performance of the WRF model on midsummer precipitation predictions in eastern China.Eastern China is located in a typical East Asia monsoon region.With the northward advance of the East Asian summer monsoon, the main rainbelt moves from South China and the Yangtze River to North China, and July is a crucial stage of the flood season.The understanding and prediction accuracy of climate anomalies in July basically determine the predictability of summer precipitation.In order to improve the prediction skill of precipitation in July, the WRF model (Skamarock & Klemp, 2008) is used to run dynamic downscaling prediction experiments.The initial and boundary fields of WRF model are driven by the hindcast and real-time forecast data from the Climate Forecast System version 2 of the National Centers for Environmental Prediction (NCEP-CFSv2, Saha et al., 2014).By comparing and analyzing the precipitation prediction skills in July of global and regional models, we further investigate the possible reasons for the improvement of precipitation skills.Section 2 describes the models, data, and methods used in this study.Section 3 shows the evaluation of the dynamic downscaling method on precipitation prediction in eastern China.Section 4 summarizes the main conclusion of this study and provides some further discussions.

| MODELS, DATASETS AND METHODS
The GCM used in this research is the NCEP CFSv2 which has upgraded to nearly all aspects of the data assimilation and forecast model components of the system.A coupled reanalysis was made over a 32-year period , which provided the initial conditions to carry out a comprehensive reforecast over 29 years .And there are four control runs per day from the 0000, 0600, 1200, and 1800 UTC of the CFSv2 real-time (2011-2020) data assimilation system.
Based on the 6-hourly CFSv2 outputs, we further run the WRF V3.8 to do dynamic downscaling tests.All the downscaled hindcast experiments were run at a horizontal resolution of 30 km with 38 vertical levels.Following Gao et al. (2011), the WRF computational domain in this study (Figure 3c) is based on the Lambert conformal map projection centered at (35 N, 105 E), with a total of 180 Â 210 grid points.Table 1 gives the list of the parameterization schemes of the main physical processes that were adopted in the WRF model downscaling experiment.Zhu et al. (2014) selected eight parameterization schemes of cloud microphysical processes in WRF V3.4 to comparatively study their impact on rainstorm simulations in South China.They found that the WRF Single-Moment 3-class scheme (WSM3) performs the best on light to heavy rain and large rainstorm and the worst on the rainstorm simulations, while the WRF Single-Moment 5-class scheme (WSM5) had the best simulations of rainstorms.In this research, we compare the results from the hindcast experiments based on the WSM3 and WSM5 schemes and find that the overall prediction effect of the WSM3 scheme is better than that of the WSM5.Therefore, the WSM3 scheme is selected for verification and analysis of predictions in this study.
The initial and lateral boundary conditions of the region model are provided by the CFSv2 model, and they are input into the WRF every 6 hours.For the historical hindcast experiment of the regional model, the 6-hour hindcast data of the CFSv2 at 0000 UTC on June 30 during 1982-2010 are used to drive the WRF, and the integration time is 31 days.For the real-time prediction experiment, the 6-hour forecast data of the CFSv2 at 0000 UTC, 0600 UTC, 1200 UTC, and 1800 UTC on June 30 during 2011-2021 are utilized to drive the WRF model for the 31-day integration, and the ensemble average results of four samples are employed for analysis.
In order to evaluate the hindcast and prediction performance of the dynamical model on precipitation and main circulations in eastern China, we adopt CN05.1 precipitation data (Wu & Gao, 2013) and ERA5 (Hersbach et al., 2023) atmospheric reanalysis data.The horizontal resolution of these two datasets is 0.25 Â 0.25 , and the period is 1982-2021.To facilitate the comparison with observations, we interpolate the data from the CFSv2 and WRF models to the same horizontal resolution as that of the CN05.1 and ERA5 data.The anomaly correlation coefficient (ACC) and temporal correlation coefficient (TCC) are selected as the evaluation methods.

| EVALUATION OF THE DYNAMIC DOWNSCALING METHOD ON PRECIPITATION PREDICTION IN EASTERN CHINA
Figure 1 shows the TCCs of precipitation anomaly percentage between the model predictions (CFSv2 and WRF) and observations in July over eastern China.The results suggest that the TCCs are high in the south and low in the north of China.Specifically, the high correlation areas for the CFSv2 model predictions are mainly located in the lower reaches of the Yangtze River Valley and the eastern part of South China (Figure 1a), while the WRF model shows high prediction skills in the south of North China, the coastal areas of East China, and southern China (Figure 1b).Compared with the CFSv2, the WRF can obviously improve the precipitation prediction in eastern China, and the regions with obviously improved prediction skills are located from the south of North China to the north of the Jianghuai region.In addition, there is a certain improvement in the west and southwest of the Jiangnan region (Figure 1c).This result indicates that compared with the global model, the dynamic downscaling results improve the prediction skill of precipitation in most of eastern China, especially in the sub-domain from the south of North China to the Yangtze River Basin (30 N-40 N, 110 E-120 E) which is chosen and referred as the northern China in this paper.This region is selected to further analyze the precipitation prediction skills of two models from 1982 to 2021.The ACC of the two models for the July precipitation over the region (Figure 1d) indicates that the forecast skills of the CFSv2 and WRF have significant and similar inter-annual variability.For the 40-year precipitation forecasts, the average ACC value of the WRF in the region is 0.156, higher than that of the CFSv2 (0.126), and the WRF performs markedly better than the CFSv2 in most years.The ACC values of the models (Figure 1d) reveal that the prediction skill of the dynamic downscaling of the regional model (line in Figure 1d) relies heavily on the prediction results of the GCM (column in Figure 1d), but the more refined physical processes in the regional model can further improve the precipitation prediction.
Figure 2 presents the distribution of the precipitation anomaly percentage in July over eastern China between observations and model predictions in 3 years.The results indicate that in 2019, the main rainy areas in eastern China are located from the Jiangnan region to the north of South China, while other areas are controlled by less rainfall (Figure 2a).The location of the southern rainy belt predicted by the CFSv2 model is more northward than that of the observations, and the precipitation intensity is obviously underestimated (Figure 2b).From observation, the abnormal rainfall is above 80% less than normal over some regions of north of the Yangtze River, and above 80% more than normal in the south of the Yangtze River.However, the precipitation anomalies by CFSv2 are around 50% at most.The distribution and magnitude of the precipitation anomaly predicted by the WRF (Figure 2c) are closer to the observations, and the rainy areas predicted by the WRF are located from the Jiangnan region to South China, slightly southward T A B L E 1 Physical parameterization schemes adopted in the WRF V3.8 downscaling experiments.

Physical parameterization schemes Options
Cumulus parameterization Kain-Fritsch

Shortwave radiation Dudhia
Land surface model Noah

Planetary boundary layer YSU
Near-surface layer Monin-Obukhov compared with the observations.The precipitation anomalies by WRF model are around 80%, has the same magnitude with that of observation.In 2020, the data suggests that the rainy belt in eastern China is in the Yangtze River Basin, and the rainfall is abnormally excessive (Figure 2d).The rainy areas predicted by the CFSv2 are in the north of the Yangtze River, and the precipitation anomaly predictions in the Huaihe River Basin and the lower reaches of the Yangtze River are opposite to the observations (Figure 2e).Nevertheless, the location of the rainy belt predicted by the WRF (Figure 2f) is more consistent with the observations than that of the CFSv2, the precipitation anomaly center is in the Yangtze River Basin, and the rainfall intensity is also closer to the observations.The observation data show that the precipitation anomaly in most areas of the middle and lower reaches of the Yangtze River is more than 80%, while the precipitation anomaly predicted by the CFSv2 in the same area is less than 50%.The WRF predicts that the precipitation anomaly in the middle reaches of the Yangtze River can reach more than 80%, which is closer to the reality.In 2021, the CFSv2 and WRF both predict more rainfall in the lower reaches of the Yellow River, the Huaihe River Basin, and the lower reaches of the Yangtze River (Figure 2h, i).In July 2021, the extremely heavy rainfall causing severe disasters occurred in Henan Province, with above 100% anomalously more monthly precipitation (Figure 2g).Both the CFSv2 and WRF models predict the trend of more precipitation in Henan Province.But the CFSv2 predicts less than 80% rainfall anomaly.The WRF predicts the intensity of above 100% precipitation in most of the Huaihe River Basin and the lower reaches of the Yangtze River (Figure 2i), which is closer to the observation.Overall, based on the initial and boundary conditions provided by the CFSv2, the dynamic downscaling of the WRF remarkably improves the prediction of July precipitation anomalies over eastern China in the 3 years, which includes the prediction of the location of the rainy belt and rainfall magnitude in the eastern China.From the observation of the 3 years, the precipitation abnormally is 80% greater than normal in some areas in eastern China.However the prediction by the CFSv2 is 50% less than normal.The prediction by the WRF is 80% more than normal in most of regions, which close to the observation obviously.
Compared with the predictions of the GCM, the dynamic downscaling obviously improves the prediction skill for precipitation in eastern China, especially for the northern China (30 N-40 N, 110 E-120 E).In the subsequent content, the possible reasons for the improvement of the precipitation forecast skill over the northern China are analyzed in the perspectives of dynamical and thermal conditions.Figure 3 shows the distributions of correlation coefficients between the precipitation anomaly percentage over the northern China and U200, H500, and the moisture flux divergence at the 850 hPa in July.In terms of the dynamical conditions from atmospheric circulations, the East Asia-Pacific teleconnection pattern affecting the sub-domain shows a distribution of "+ À +" from the south to the north in the real atmosphere (Wu et al., 2020), that is, obviously positive correlation at low latitudes, negative correlation from North China to southern Japan, and positive correlation in the Okhotsk Sea (Figure 3d).Correspondingly, the 200-hPa zonal wind shows a teleconnected wave train of "À + À", with the westerly jet moving northward (Figure 3a).The highly significant negative correlation region of moisture flux convergence in the lower troposphere is located in the northern China (Figure 3g).The features show that there is more precipitation in the northern China, corresponding to the stronger moisture flux convergence in the lower troposphere, the "+ À +" wave train in East Asia, and the northward westerly jet.The H500 simulations by the CFSv2 model indicate a consistent positive correlation in East Asia, without obvious wave train characteristics (Figure 3e).The H500 simulations by the WRF model in East Asia relatively perfectly reproduce the distribution features of the teleconnection (Figure 3f) in the real atmosphere.Similarly, the WRF simulation of the relationship between precipitation and 200-hPa zonal wind and 850-hPa zonal wind (Figure 3c, i) over East Asia is superior to that of CFSv2 (Figure 3b, h).This result indicates that compared with the GCM, the dynamic downscaling characterizes the relationship between precipitation and atmospheric circulations closer to the real atmosphere.
Furthermore, the sources of the predictable signals that may improve the precipitation prediction skills of models are analyzed in terms of thermal conditions.between the precipitation anomaly percentage over the northern China and the sensible heat flux in the real atmosphere and model atmosphere (Figure 4c-e).In the real atmosphere, the correlation coefficients show an obviously negative in the north of the Yangtze River, that is, the precipitation is less when the sensible heat flux is stronger (Figure 4c).For the WRF simulations, the areas with the highest correlation between the two variables are located from the Huang-Huai to the northern Jianghuai region (Figure 4e).In terms of the CFSv2 simulations, the areas with the obviously high correlation are located from the southern Huang-Huai region to the Jianghuai region (Figure 4d), more southward than those of the actual and WRF model atmosphere.
The TCC distributions of the latent heat flux in July between observations and predictions suggest that the forecast skill of the WRF for the latent heat flux in northern China is close to that of the CFSv2, with the higher correlation coefficient in the Huang-Huai region.In addition, the forecast skill in the eastern Jiangnan region is also improved by WRF.The correlation coefficients between the precipitation anomaly percentage over the northern China and the latent heat flux indicate positive values.The performance of WRF is closer to the observations than that of the CFSv2 in the south of North China (Figure S1).
Previous research results show that the improvement of surface heat flux based on RCMs is helpful to the improvement of the temperature and precipitation prediction (Chen et al., 2019;Li et al., 2016).Li et al. (2016) found that the intensity of simulated sensible heat flux over Asian continent in regional models can induce tropospheric temperature anomaly over land.The adaptive modulation of geopotential height gradients affects wind field (through geostrophic balance) simulation especially at lower levels, which subsequently affects the simulation of large-scale circulation, 2-m temperature, and monsoon precipitation as well as the model's ability.So the WRF simulation results of both sensible heat flux and latent heat flux are better than those of CFSv2, which helps to improve the prediction of precipitation.
Overall, there are two main reasons for the improvement of precipitation forecasts over the northern China based on the WRF.For dynamic conditions, the prediction skill of the regional model for meridional teleconnection circulation pattern in East Asia is obviously improved, especially at low latitudes.In terms of thermal conditions, the simulations of the sensible and latent heat fluxes in the northern China are markedly improved by the regional model, especially for the sensible heat, whose simulations are closer to the real atmosphere.

| CONCLUSIONS AND DISCUSSION
Based on the ERA5 reanalysis data and the WRF regional model driven by the NCEP-CFSv2 data, we conducted hindcast and prediction experiments in this study.Furthermore, the prediction ability of the global and regional models for July precipitation over eastern China was analyzed.
The results from the WRF driven by the CFSv2 suggested that the dynamic downscaling method dramatically improved the prediction skill of the precipitation over eastern China, especially from the south of North China to the Yangtze River Basin.This spatial distribution of the prediction skill is similar to that of the simulated precipitation for 2011-2020 obtained by Ying et al. (2022) using the WRF driven by the NUIST-CFS1.0 data.Furthermore, we analyzed the sources of the predictable signals which may help to improve the precipitation prediction skill of models in the northern China.Comparing the relationship between the precipitation and circulations in the real atmosphere and model atmosphere, and the relationship of the precipitation with the sensible and latent heat fluxes, we found that the regional model improved the simulation ability for the atmospheric circulation pattern and the sensible and latent heat fluxes in East Asia, especially for the meridional teleconnection pattern in East Asia and the sensible heat flux in the northern China.The experiments conducted in this research demonstrated that the high resolution of the regional model and the refinement of physical process parameterizations contribute to improving the prediction skill for midsummer precipitation over the northern China in terms of dynamic and thermal conditions, especially for the spatial characteristics and magnitude of the precipitation anomaly in rainy regions.
This study has proven the improvement prediction of precipitation in eastern China by using the regional climate model, but there is still much other work to do in the future.The output of the GCM on June 30 is used to drive the regional climate model.The lead time is 0 days for the July prediction.It is necessary to use the initial and boundary conditions of different lead time to carry out more hindcast and forecast tests in order to get more predictability information.The variance distribution features and probability information of ensemble predictions can also be evaluated by increasing the number of regional climate model samples.In addition, this paper only focuses on the prediction skill of monthly precipitation, the skills of sub-seasonal precipitation process are worthy of further studies.Li et al. (2020) indicated the skill of CWRF in the Yangtze-Huaihe River basin is relatively low, which shows the similar information in Figure 1c.The failure of dynamic downscaling in these region is worth of further studies.Ying et al. (2022) also pointed out that dynamic downscaling with finer resolutions and updated physics as well as statistical downscaling should be explored in future studies.The machine learning that can better capture the complex nonlinear interactions among different phenomena may play an important role in improving the downscaling forecasts of summer precipitation in China.These ideas offer prospects for further improvement of prediction capabilities.

F
I G U R E 1 The temporal correlation coefficient (TCC) values of the precipitation anomaly percentage between observations and the (a) CFSv2 and (b) WRF simulations.(c) The difference of the TCC values over eastern China between CFSv2 and WRF.(d) The anomaly correlation coefficient (ACC) of the models (column for the CFSv2 results, line for the WRF results) for the precipitation over the northern China in July from 1982 to 2021.The dotted grid indicates the TCCs passing the significance test at the 95% confidence level.The black rectangle shows sub-domain for northern China.
Figure 4a, b presents the TCC values of the sensible heat flux in July between the observations and the predictions.Compared with the simulations by the CFSv2 model, the WRF model dramatically improves the prediction skill for the sensible heat flux in the northern China, with correlation coefficients in most areas passing the significance test at the 95% confidence level.In addition, the sensible heat flux simulations in the southern Jiangnan region are improved.Furthermore, we analyze the correlations F I G U R E 3 Distributions of the correlations between the precipitation anomaly percentage over the northern China and (a)-(c) U200, (d)-(f) H500, and (g)-(i) the moisture flux divergence at the 850 hPa for the (a), (d), (g) ERA5, (b), (e), (h) CFSv2 and (c), (f), (i) WRF model data in July from 1982 to 2021.The dotted grid indicates the correlations passing the significance test at the 95% confidence level.

F
I G U R E 4 The TCC values of the sensible heat flux anomaly between observations and (a) CFSv2 and (b) WRF simulations.The correlation between the precipitation anomaly percentage over the northern China and the sensible heat flux anomaly for the (c) ERA5 data, (d) CFSv2 model data, and (e) WRF model data in July from 1982 to 2021.The dotted grid indicates the correlations passing the significance test at the 95% confidence level.