Dynamic Downscaling of Summer Precipitation Prediction over China in 1998 Using WRF and CCSM4

Although global climate models are the main tool for seasonal climate simulation and prediction, their coarse horizontal resolution limits their usefulness on smaller scales. In particular, their poor performance in capturing high-frequency features makes it difficult to use them in regional climate research. As computational constraints ultimately limit the use of very high-resolution global climate models for ensemble seasonal predictions, one remedy for this is to use a downscaling approach for regional-scale climate predictions.

Statistical and dynamical downscaling methods have traditionally been the most widely used methods (Wilby and Wigley, 1997; Diez et al., 2005). In statistical downscaling, a model is developed based on the statistical relationship between large-scale climate variables and local variables under historical conditions (Friederichs and Hense, 2007; Liu et al., 2011; Chen et al., 2012; Liu and Fan, 2012; Sun and Chen, 2012). In dynamical downscaling, a high-resolution regional climate model (RCM) is run for the region of interest, forced by the global climate model at the lateral boundaries (Giorgi, 1990; Nobre et al., 2001). Because of the higher resolution, RCMs have been shown to more accurately capture near-surface winds and temperatures over complex terrain and coastlines (Feser et al., 2011).

Due to the complexity of the East Asian summer monsoon, seasonal climate prediction in China has long challenged meteorologists (Wang et al., 2012a), especially with respect to extreme climate events (e.g. heavy rainfall, intense snowfall, tropical cyclone activity, droughts, and cold surge activity) that are currently more difficult to predict (Wang et al., 2012b). Global climate models produce large biases over China resulting from their coarse resolution, so it seems reasonable to expect that RCMs driven by global models could improve climate predictions. (Liu et al., 2005) carried out a 10-year hindcast experiment by nesting the regional climate model (RegCM_NCC) and a global atmosphere-ocean coupled model. They argued that RegCM_NCC had some ability to predict summer precipitation over China, though the differences between the two models were not specified in their study. (Ju and Lang, 2011) nested RegCM3 and the IAP (Institute of Atmospheric Physics, Chinese Academy of Sciences) grid-point nine-layer atmospheric general circulation model (IAP9L-AGCM) for summer climate prediction over China. Regional characteristics were a significant feature in their results. The Weather Research and Forecasting (WRF) model has been widely used in climate research, especially in studies of heavy rainfall (Sun and Zhao, 2003; Ha and Lee, 2012; Liu, 2012). The present reported work used the WRF model and the Community Climate System Model (CCSM) for a seasonal climate prediction experiment, to study whether or not WRF can improve the CCSM results.

The present work studies a flood that struck the middle-lower reaches of the Yangtze River and the Nen River valley, China, in 1998, causing severe social and economic costs. This event was chosen as it has been a typical case study in climate variance mechanism analysis and short-term climate prediction in China (Huang et al., 1998; Wang et al., 2000; Tao et al., 2001). Furthermore, the influences of a strong El Ni\ no event in pre-winter, the spring snow cover anomaly, and spring atmospheric initial anomalies have been proven to play important roles in this extreme flood event (Huang et al., 1998; Tao et al., 1998; Wang et al., 2000; Zhang and Tao, 2001). The flood over the middle-lower reaches of the Yangtze River was successfully predicted due to the high predictability associated with the strong signal in pre-winter and spring (Lin et al., 1998). In this paper we set out to discuss not the known results, but whether or not a high-resolution RCM could have been even more effective in this case. Specifically, we applied a dynamic downscaling method in seasonal climate prediction.

The global climate model used in this study was version 4.0 of CCSM (Gent et al., 2011), developed by the National Center for Atmospheric Research (NCAR, US) to predict the global climate on a seasonal scale. The horizontal grid increment was 1.9^° (lat) × 2.5^° (lon) and there were 26 hybrid vertical layers with a finite volume (FV) dynamical core. The regional climate model used in this study was the "advanced research" version 3.2 of the WRF model (Skamarock et al., 2008). Its performance as a climatology and extreme weather model has previously been assessed (Yu et al., 2010; Wang et al., 2011).

We chose June, July, and August (JJA) of 1998 as the study period. For both the CCSM and WRF, seven-member ensembles were integrated. The corresponding initial conditions were chosen from 0000 UTC 28 April to 4 May at 24-hour intervals for the CCSM. The retrospective experiment design was the same as that of (Ma and Wang, 2014). The Community Atmospheric Model (CAM) and the Community Land Model (CLM) were initialized with the NCEP Reanalysis 1 data (Kalnay et al., 1996) and NCEP's Climate Forecast System Reanalysis (CFSR) (Saha et al., 2010) respectively. The ocean component was initialized with the Global Ocean Data Assimilation System (GODAS) dataset (Behringer and Xue, 2004), using only mixed layer information. The ice model was initialized with the model's climatology.

The WRF ensemble members were generated by initializing and forcing at 6-hour intervals with each ensemble member of the CCSM. A one-way nesting method was used in this study. Both the models were integrated from 0000 UTC 1 May to 0000 UTC 1 September. The first month of the integration was neglected in the analysis, as it was included to allow the model to spin up.

Figure 1 shows the WRF domain. There were 132 cells in the west-east direction and 112 cells in the south-north direction spaced 60 km apart, with the central point at (30^°N, 110^°E). The parameterizations of the main physical processes were as follows: the Kain-Fritsch cumulus parameterization (Kain, 2004); the rapid radiation transfer model for longwave radiation of (Mlawer et al., 1997); the Dudhia shortwave radiation model (Dudhia, 1989); the WRF single-moment 3-class microphysics model (Hong et al., 2004); the Noah land surface model (Chen and Dudhia, 2001); the Yonsei University model for the planetary boundary layer (Hong et al., 2006); and the MM5 similarity theory for the surface layer.

Figure 1.  Surface heights in m used in (a) the CCSM and (b) the WRF model; the map has the Lambert conformal conic projection. Panel (b) also indicates the eight continental sub-regions of China used in this study.

Figures 1a and b show the topography used in the CCSM and WRF models. Some details are visible in the relatively high-resolution topography of the WRF, e.g. the Sichuan basin, the hills of Southeast China, and the "two basins sandwiched between three mountains" in Xinjiang.

Precipitation from the CN05 dataset was used in this study. As per (Xu et al., 2009), the CN05 dataset was constructed using the "anomaly approach" during the interpolation but with more observations (nearly 2400 stations). A gridded climatology was calculated, and then a gridded daily anomaly was added to obtain the final dataset (Wu and Gao, 2013). This approach has been widely used in climate research in China, due to the resulting high spatial and temporal resolution and the ability to include data from more stations. The near-surface wind came from the CFSR dataset, with a horizontal resolution of 0.5^° by 0.5^°.

We compared the JJA mean results for each model followed by a discussion of the inter-seasonal variance between the two models. The results are presented in this section as an ensemble mean over all seven seasonal simulations for the two models. Figure 2 shows the observed precipitation, and that from the CCSM and WRF hindcasts. The observation values are presented as JJA cumulative values on the 0.5^° latitude-longitude grid. On account of the higher resolution of the WRF grid relative to that of the CCSM, there are more small-scale features visible in Fig. 2c. Both models had a wet bias over North China. However, the unrealistic precipitation over the eastern part of the Tibetan Plateau in the CCSM simulation was improved upon by the WRF simulation. Furthermore, the increase in precipitation predicted by the WRF model over the south of the Yangtze River agreed better with the observations. The global model predicted unrealistically low precipitation and had a banded structure in that region, probably as a result of numerical instability arising from the Gibbs phenomenon near the steep Tibetan Plateau. The WRF model improved the precipitation simulation over the downstream of the Tibetan Plateau south of 30^°N, in the altitude range 500-1500 mm, compared to the near-dry conditions (< 500 mm) predicted by the CCSM simulation. The spatial correlations between the CCSM and WRF models and the observations were 0.42 and 0.81 respectively.

Figure 2.  Accumulative precipitation in JJA 1998 from (a) CN05 observation data, (b) CCSM4, and (c) WRF over China (units: mm).

Considering anomaly metrics relative to the climatology is more important for climate prediction research. However, simulating climatology requires a lot of computational resources. For this reason, the interannual increment was chosen instead of the anomaly in this study. (Wang et al., 2010) argued that the interannual increment is in fact better suited to climate prediction in China. Figure 3 shows the 1998 JJA mean precipitation increments, defined as the difference of the JJA mean precipitation between 1998 and 1997. It is apparent that the WRF model produced a superior simulation of the precipitation increments. The exclusive rainfall over the middle-lower reaches of the Yangtze River was also in excellent agreement with CN05. However, the flood area was not well located by the CCSM model, placing it too far north. The flood over Northeast China was missed by both models.

Figure 3.  The JJA mean precipitation difference between 1998 and 1997 from (a) CN05 observation data, (b) CCSM4, and (c) WRF (units: mm d^-1).

From the above analysis, simulation of the summer precipitation in 1998 over China was greatly improved using a downscaling method, albeit with obvious regional differences. With this in mind, the results between sub-regions was further examined. The sub-regions are shown in Fig. 1b. As Table 1 shows, downscaling resulted in an obvious simulation improvement over the south of the middle-lower reaches of the Yangtze River (SYR), South China (SC), Northwestern China (NW), and Tibet (TB) for the regional mean of the JJA mean precipitation simulation (MEAN). The MEANs from the WRF model were in closer agreement with the observations over the above regions. However, the changes in spatial correlation coefficients (CCs) and root-mean-square error (RMSE) were inconsistent with the MEAN. The CCs increased across most of China, except SYR. The RMSE reduced over most of China, except Northeast China (NE), North China (NC) and the Huai River valley (HR). Note, however, that the spatial correlation coefficient has some limitations for model behavior evaluation (Wang and Zhu, 2000). Figures 2 and 3 show that the precipitation predictions over SYR and SC were undoubtedly improved after downscaling.

The superiority of the WRF simulations was also obvious after analysis of the JJA mean precipitation increment. Improvements in both the MEAN and RMSE were found in NC, HR, SYR, and Southwestern China (SW). Over SYR and SW, where CN05 gave a change in the MEAN precipitation that was more than the previous year but CCSM gave a change that was less, downscaling made the precipitation increments for CCSM positive.

Figure 4 shows the daily precipitation along 110^°-120^°E. There were two rainy periods in summer 1998 in CN05 (Fig. 4a). The first period covered the last 10 days of June, corresponding to precipitation caused by the Mei-yu front. This rain band then moved northwards associated with the westward extension of the subtropical high. However, during late July, a high precipitation period appeared near the middle-lower reaches of the Yangtze River, which was called the occurrence of the second period of the Mei-yu rainfall (Tao et al., 1998; Tao et al., 2001). As Fig. 4b shows, the highest rainfall occurred in early July, in the north of the observed flood area in CCSM. After downscaling, the daily rainfall from the WRF model was in better agreement with the observation (Fig. 4c). The two occurrences of heavy rainfall and their locations also agreed with CN05.

Figure 4.  Daily precipitation (mm d^-1) along 110^°E-120^°E from (a) CN05 observation data, (b) CCSM4, and (c) WRF.

The improvements in predicting precipitation are closely related to how accurately changes in circulation are represented. (Feser et al., 2011) showed that RCMs produce better simulations for near-surface winds. In this study, the simulation results for near-surface (10 m) winds are shown in Fig. 5. There is a larger bias in the CCSM compared to the CFSR results, e.g. the exclusively strong south wind east of 105^°E or the uniform northeast wind over NW. The result from the WRF model was better. The main features of the near-surface circulation agreed well with those of the CFSR. In addition, divergence at 10 m was calculated. WRF also produced better simulations compared to CCSM for this parameter. Many small-scale convergence areas were consistent with the reanalysis. The pattern correlation between CFSR and CCSM (WRF) was 0.19 (0.40). However, no significant improvements were found in the circulation predictions from the mid-top layers (figures not shown). This may indicate the subtle influence of the terrain.

Figure 5.  The JJA mean wind (vectors, units: m s^-1) and divergence (shading, green indicates convergence) at 10 m height in 1998 from (a) CFSR data, (b) CCSM4, and (c) WRF.

An attempt at dynamic downscaling prediction for JJA precipitation over China in 1998 was made using a WRF model driven by CCSM. The results of the two models were compared. The ensemble mean results showed that the WRF model was able to improve the JJA accumulative precipitation prediction over China. Although the wet bias over North China was preserved, the dry bias over south of 30^°N was removed. Additionally, the unreasonable rainy center in the east of the Tibetan Plateau was also absent in the WRF results. To account for climate variance, the interannual precipitation increment was examined. These results also showed that improvements were made using the WRF model. The intensity of the severe flooding over the middle-lower reaches of the Yangtze River in JJA 1998 was under-predicted by the WRF model. However, the flood location in the CCSM simulations was north of the Yangtze River and had weaker intensity. Neither model was capable of flood prediction over Northeast China, which also flooded severely in JJA 1998. The better performance of the WRF model may arise from the higher horizontal resolution and more realistic topography, as argued by Gao et al. (2006a, 2006b, 2012).

Focusing on the flood over the middle-lower reaches of the Yangtze River, daily precipitation from the two models was compared to evaluate the prediction skill for sub-seasonal precipitation variances. Even though some biases still existed, the location of rainy centers and the timing of heavy rain agreed better with the CN05 data in the WRF results. In addition, the two occurrences of Mei-yu rainfall were resolved using WRF, but not CCSM. Because the direct circulation systems leading to heavy rainfall at the Mei-yu front were mainly mesoscale convective systems (Tao, 1980), the high horizontal resolution of the WRF model helped to capture these types of heavy rainfall compared to the CCSM model. The differences in physical progresses related to the precipitation between the two models may also have contributed to the improvements.

There was a big improvement in circulation prediction at the near-surface layer, but not the mid-high layers after downscaling. This difference indicates that the influence of topography was more significant for the lower-level layers. Both the better topography and higher resolution of the WRF model contributed to these improvements.

The prediction of the East Asian summer climate is very challenging. We chose the extreme summer flood of 1998 as our case study, and our encouraging results suggest that the dynamic downscaling method can be used for seasonal climate prediction over China to improve the skill of a global model. However, our conclusion is so far limited to a single case and consequently many questions remain unresolved. The changes in precipitation after downscaling were not uniform over China. There was no obvious improvement over NE and NC. The two models both gave incorrect hindcasts. The near-surface circulation was improved in WRF over NE, but the precipitation increment trend was still opposite to the observations. The JJA precipitation results over NE and NC of the WRF model followed those of the CCSM with similar spatial patterns and bias. The poor performances over NE were related to the low predictability at mid-high latitudes. Previous studies have argued that the flood over NE in 1998 was related to the blocking systems at higher latitudes and the northeast cold vortex variability (Li et al., 2001; Sun and An, 2001). Due to the stronger internal chaos effect of the atmosphere in mid-high latitudes, and the absence of some important factor (e.g. snow cover and sea ice) in the model initialization, the climate model has had low predictive skill in mid-high latitudes until now (Wang, 1997; Ma and Wang, 2014). Moreover, many studies have suggested that errors can be imported into RCMs from global climate models via poor lateral boundaries (Denis et al., 2002). The domain, the nesting method, and the parameterizations can all also affect the performance of RCMs. Several questions remain. It is not known which factors are primarily responsible for the poor performance over NE and NC. The method used in this paper may or may not improve the predictions for other abnormal years. There may be further differences between neutral years and abnormal years. (Yuan and Liang, 2011) demonstrated the equitable threat score differences in winter precipitation between CFS and the climate extension of the WRF model (CWRF) nesting approach were larger in ENSO-neutral years than in strong anomalous years during 1982-2008 over the United States. But the conclusion may be different for East Asian summer precipitation. The work in this paper is very preliminary and we hope to address these questions in our future work.