Climate anomalies affect social and economic development worldwide. Accurate predictions of climate anomalies are imperative for preventing and mitigating climate associated disasters. Despite the tremendous societaldemand, short-term climate prediction spanning from 10 days to a few seasons is always a hot but difficult issue in both scientific and technical senses. In particular, China is located in the East Asian monsoon region where the climate variation is dominated by complex spatiotemporal climate variability with multi-scale interactions, which gives rise to a great challenge for the operational short-term climate prediction in China. Therefore, it is considerably important to carry out operational short-term climate prediction in China, and to improve the prediction skill by keeping abreast with the advancements in climate prediction research and practice around the world.

At present, dynamical prediction using numerical climate models is one of the most important approaches for short-term climate prediction, and it has been widely used in operational predictions of global climate and climate variability (e.g., Saha et al., 2006, 2014; Weisheimer et al., 2009; Wang et al., 2015; Ren et al., 2017). Dynamical prediction methodology has been developing over the past three decades from the two-tier approach proposed by Bengtsson et al. (1993) in the early 1990s for seasonal forecasting, which has a physical basis under slowly varying forcing of the lower boundary (ocean, land use, etc.), to the one-tier approach that uses coupled general circulation models (GCMs) or climate system models to effectively describe the interactions among various components in the climate system, which has been available since approximately 2000 (Latif et al., 2001; Davey et al., 2002; Schneider et al., 2003). Although climate scientists have made great progress in dynamical prediction, the skills of model-based prediction remain limited, and difficulties arise mainly from uncertainties of the initial conditions and model parameterizations of unresolved sub-grid processes.

One of the widely used methods for overcoming these difficulties is multi-model ensemble (MME) prediction proposed by Krishnamurti et al. (1999) and Palmer et al. (2000). The MME method has been designed to quantify the uncertainties in climate predictions and is considered as an effective way to improving climate prediction skill (Peng et al., 2002; Doblas-Reyes et al., 2005; Hagedorn et al., 2005; Kirtman and Min, 2009; Lavers et al., 2009). For a single-model, the method used to construct ensemble prediction members can be realized by perturbing either the initial conditions or model parameterization schemes. The single-model ensemble prediction constructed by the initial condition perturbation has been widely used in operational weather and climate prediction worldwide; however, the multi-initial-condition ensemble prediction system generally has the problem of insufficient spread among ensemble members (Buizza et al., 1998). In fact, compared with weather forecast, climate prediction may be more sensitive to model errors, especially the model uncertainties caused by parameterization schemes of sub-grid physical processes. Teixeira and Reynolds (2008) suggested that the single-model ensemble prediction using the initial condition perturbation and the MME approach can better reflect and measure the uncertainties of dynamical climate prediction.

In recent years, several major research and operational centers, such as the ECMWF, the Asia–Pacific Economic Cooperation Climate Center (APCC), the International Research Institute for Climate and Society (IRI), and the NCEP have developed their own MME prediction systems, in which the MME method has been implemented and employed for dynamical seasonal climate prediction. It has been confirmed that MME prediction is usually superior to predictions made by any single model (Wang et al., 2009; Becker et al., 2014; Min et al., 2014). To fill the blank in the field of operational MME prediction in China and make full use of the achievements in model development and dynamical climate prediction by Chinese scientists, the National Climate Center (NCC) of the China Meteorological Administration (CMA) has devoted considerable efforts to developing an MME prediction system to produce improved and well-validated monthly–seasonal forecasting of climate and climate variability modes, such as the western Pacific subtropical high (WPSH), El Niño–Southern Oscillation (ENSO), as well as climate state variables of temperature, precipitation, and so on for research and operational purposes, based on a combination of several domestic operational climate models and imported prediction data. Recently, the China multi-model ensemble prediction system version 1.0 (CMMEv1.0) was established and applied to real-time climate prediction at the NCC/CMA.

In this paper, we present the recent progress in the development of the CMMEv1.0 system, including an introduction to its technical framework, model setup, key methods, some of the main products, and its application to the 2018 flood season (June–August) prediction started from March 2018, as well as verification of related prediction skills. The reminder of this paper is organized as follows. Section 2 introduces the technical framework and model setup. Section 3 describes data and methods. Section 4 shows the real-time forecasts for the 2018 flood season and related prediction skills. Summary and discussions are provided in Section 5.

Based on several domestic operationally-run climate models and internationally imported prediction data, the NCC/CMA has established CMMEv1.0 for monthly–seasonal forecasting. Figure 1 shows its technical framework. As shown, the CMMEv1.0 consists of several modules, such as data collection, model initialization, time integration, post-processing, and the product output. Currently, CMMEv1.0 includes four domestic climate models and two imported prediction datasets. The domestic climate models are BCC-CSM1.1m from NCC/CMA (Wu et al., 2014), FGOALS-f2 from the State Key Laboratory of Numerical Modeling for Atmospheric Sciences and Geophysical Fluid Dynamics (LASG) of Institute of Atmospheric Physics (IAP) of Chinese Academy of Sciences (CAS) (Bao et al., 2019), FGOALS-s2 from LASG/IAP/CAS (Bao et al., 2013), and NZC-PCCSM4 from the Nansen–Zhu International Research Centre (NZC) of IAP/CAS (Ma and Wang, 2014). The imported prediction data are from ECMWF-SYSTEM4 (ECMWF-S4) (Molteni et al., 2011) and the NCEP Climate Forecast System version 2 (NCEP-CFSv2) (Saha et al., 2014).

Figure 1 Framework of the China multi-model ensemble prediction system version 1.0 (CMMEv1.0)

Figure options

Table 1 lists detailed information on these models in CMMEv1.0. The atmospheric component of BCC-CSM1.1m has a T106 horizontal resolution and 26 vertical levels; the horizontal resolution of the land component is the same as the atmospheric component; the oceanic component has a resolution of 1.0° latitude × 1.0° longitude × 40 vertical levels and has the same resolution as the sea ice component. The atmospheric component of FGOALS-f2 is FAMIL with a 1.0° × 1.0° horizontal resolution and 32 vertical levels; the land component is CLM4.0 (Oleson et al., 2010); the oceanic component of FGOALS-f2 is POP2, which uses two displaced-pole grids centered at Greenland, at a nominal 1° (gx1v6) horizontal resolution and 60 vertical levels; the horizontal resolution of the sea ice component CICE4 (Holland et al., 2012), is the same as the oceanic component. The atmospheric component of FGOALS-s2 is SAMIL2 version 2.4.7, which is a spectral model with a R42 horizontal resolution (approximately 1.66° × 2.81°) and 26 vertical levels; the land component is CLM3.0 (Oleson et al., 2004); the horizontal resolution of the oceanic component (LICOM2, Liu et al., 2004) is increased in the tropics (from 1.0° × 1.0° to 0.5° × 0.5°) with 30 vertical levels; the sea ice component is CSIM5 (Briegleb et al., 2004) and has the same horizontal resolution as the oceanic component. The atmospheric component (CAM4) of NZC-PCCSM4 has a 2.5° × 1.9° horizontal resolution and 26 vertical levels; the horizontal resolution of the land component (CLM4, Lawrence et al., 2011) is the same as that of CAM4; the oceanic component is a mixed-layer model (SOM) with a resolution of 1.0° × 1.0°; the horizontal resolution of the sea ice component CICE4 used is the same as the oceanic component.

These models use the nudging method to assimilate atmospheric and oceanic reanalysis data for initialization. The atmospheric assimilation data include the meridional and zonal winds, air temperature, and geopotential height from the Japanese 55-yr reanalysis (JRA55, Kobayashi et al., 2015) for FGOALS-f2, FGOALS-s2, and NZC-PCCSM4, and from the NCEP-I reanalysis for BCC-CSM1.1m. The oceanic assimilation data of the four models are the 3D ocean temperature from the NCEP Global Ocean Data Assimilation System (GODAS, Behringer and Xue, 2004). The initial condition perturbation is used to generate ensemble members for each mo-del. The number of ensemble members of BCC-CSM1.1m, FGOALS-f2, FGOALS-s2, and NZC-PCCSM4 are 24, 35, 4, and 8, respectively. The real-time forecasts begin on the 20th and 21st days of each month for FGOALS-f2, FGOALS-s2, and NZC-PCCSM4 and on the 1st day of each month for BCC-CSM1.1m. Recently, CMMEv1.0 has generated a monthly mean hindcast dataset for the period of 1991–2016 and a real-time forecast dataset since 2017. All of the model outputs were interpolated into a unified horizontal resolution of 1.0° × 1.0°.

The prediction skill (temporal correlation coefficient, TCC) of CMMEv1.0 is calculated based on the monthly mean hindcast for the period of 1991–2016. To verify the prediction performance, the anomalies of prediction are derived by subtracting the hindcast climatology from the original prediction data for each model in CMMEv1.0, where the model climatology is calculated by using the monthly hindcast data during 1991–2010 as a function of the initial calendar month and lead month, and observational climatology is obtained for the same period. All of the observational SST indices are calculated by using the Optimum Interpolation SST version 2 (OISSTv2) (Reynolds et al., 2007). The atmospheric verification data are computed by using NCEP–I reanalysis data for air temperature, 500-hPa geopotential height, and 850-hPa zonal wind (Kalnay et al., 1996) and the NOAA Climate Prediction Center (CPC) Merged Analysis of Precipitation (CMAP, Xie and Arkin, 1997). The summer (June–July–August, JJA) average surface (2 m) air temperature and precipitation for China are from the daily climate data of Chinese stations for global exchange (V3.0).

CMMEv1.0 has carried out the real-time forecasts for the 2018 flood season (June–August) based on the outputs of FGOALS-f2, FGOALS-s2, and NZC-PCCSM4 starting on 20 and 21 February, BCC-CSM1.1m starting on 1 March, ECMWF-S4 starting on 10 March, and NCEP-CFSv2 starting on 18 March 2018. In this paper, we show the monthly mean predictions during March–August and the JJA average predictions in 2018, including the global SST, ENSO, Indian Ocean Dipole (IOD), North Atlantic SST triple (NAST), WPSH, East Asian summer monsoon (EASM), surface air temperature (SAT), and precipitation. Table 2 shows definitions of the indices used for ENSO, IOD, NAST, WPSH, and EASM predictions.

First, Figs. 2a–f show the TCC skill of the CMMEv1.0 hindcast for the global SST anomalies (SSTAs) beginning from March. CMMEv1.0 exhibits significantly high prediction skills over most of the ocean region, especially in the mid and low latitudes. In Figs. 2g–l and 2m–r, there is a good similarity between the real-time SSTA forecasts from March 2018 and their corresponding observations, in which CMMEv1.0 has accurately predicted the gradual warm-up of the cold SSTAs in the tropical central–eastern Pacific. During springtime, pattern correlation coefficients (PCCs) of global SSTAs are high (0.60 for March, 0.57 for April, and 0.42 for May), and CMMEv1.0 successfully reproduced the weakening of the cold SSTAs in the eastern Indian Ocean and the pattern of the NAST but with a relatively weaker intensity than observation. During summertime, the PCCs are relatively low, and the predicted SSTAs are overall warmer than observation, especially in North Atlantic and the Southern Ocean.

Figure 2 (a–f) TCC skills for the global monthly (March–August) mean SSTA predicted by CMMEv1.0 initiated in March over 1991–2016; and monthly mean of the global SSTA (°C) for March–August 2018 (g–l) predicted by CMMEv1.0 and (m–r) derived from OISSTv2. The dotted areas in (a–f) are statistically significant at the 95% confidence level with a Student t-test. The numbers at the top right of (g–l) are the pattern correlation coefficients between monthly mean predictions and observations of the global SSTA.

Figure options

ENSO is a dominant mode of interannual variability with remarkable climate impacts worldwide (e.g., Li, 1990; Zhang et al., 1996; Wang et al., 2000; Chen, 2002; Zhang et al., 2011, 2012; Ren et al., 2018). Accurate and reasonable prediction of ENSO is considered as one of the primary indicators for the performance of climate prediction (e.g., Jiang et al., 2013a; Ren et al., 2017). Figure 3 shows the TCC skills for Niño3.4, Niño4, Niño3, N_CP, and N_EP indices (Ren and Jin, 2011, 2013) from March to August and JJA season over 1991–2016. As shown in Figs. 3a–e, the five ENSO indices are significantly skillful and the scores of the MME for both Niño3.4 and Niño4 indices are larger than 0.8 at the six-month leadtime, which is more skillful than any single model. We further compare the real-time forecasts of CMMEv1.0 and the corresponding observations for these indices in Fig. 3. For Niño3.4 index, the CMMEv1.0 prediction accurately captured the trend of the ENSO transition from cold phase to neutral phase and then the warm phase during spring–summer of 2018, and the MME prediction almost coincides with observation (Fig. 3f).

Figure 3 Left panels: TCC prediction skills for (a) Niño3.4, (b) Niño4, (c) Niño3, (d) NCP, and (e) NEP indices from monthly (March–August; hollow bars) and seasonal (JJA; hatched bars) mean predictions by CMMEv1.0 initiated in March over 1991–2016. Right panels: monthly evolutions of (f) Niño3.4, (g) Niño4, (h) Niño3, (i)NCP, and (j) NEP indices (°C) derived from OISSTv2 for March 2017–August 2018 (OBS; denoted by red/blue bars before March 2018 and by the thick black line after March 2018) and from CMMEv1.0 for March–August 2018 (forecast; denoted by red/blue hatched bars for the ensemble mean and by color lines for individual models M1–M6).

Figure options

There are still some major differences in the predictions among the different models. The forecast by FGOALS-f2 is closest to observation. NZC-PCCSM4 has predicted gradual warming of SSTA, which is similar to that of NCEP-CFS2, whereas ECMWF-S4 predicts relatively rapid warming. BCC-CSM1.1m predicts that the Niño3.4 index would more rapidly warm up and exceed 0.5 in summer, likely representing an El Niño event. The forecast by FGOALS-s2 is quite different from the others and is characterized by changing SSTAs with an initial increase and then a decrease. Both predictions of the Niño4 and Niño3 indices are consistent with that of the Niño3.4 index; i.e., the MME forecast is similar to observation despite some differences among the models. The above results further indicate the superiority of MME prediction to single model prediction.

As is known, the impacts of EP (Eastern Pacific)-type and CP (Central Pacific)-type El Niño on the climate are quite different from each other (e.g., Weng et al., 2007, 2009; Zhang et al., 2011, 2012). In the development summer of El Niño, the influence of CP-type El Niño on East Asian precipitation is more significant than that of EP-type El Niño (Yuan and Yang, 2012; Ren et al., 2018). Thus, in addition, we predicted evolutions of the EP- and CP-type ENSO indices (N_EP and N_CP) using CMMEv1.0. The skills of the spring–summer predictions that start in March for both N_EP and N_CP indices are larger than 0.6. The real-time forecast of CMMEv1.0 has shown no CP-type El Niño signal occurring in the 2018 flood season, which is consistent with observations. It has been known that the EP (CP) type of ENSO features a spring (summer) persistence barrier (Ren et al., 2016) and thus, is subject to the spring (summer) prediction barrier (Ren et al., 2019b), which could be indicated by the different timings of the maximum decline speeds of the TCC scores in Fig. 3. Our MME predictions show that the EP-type ENSO has a higher prediction skill than the CP-type in terms of their indices, although the dynamic models still have difficulties in distinguishing the two types in terms of their SSTA center positions (Ren et al., 2019a), which could be statistically corrected (Liu and Ren, 2017). Indeed, the difficulty of the dynamic models in predicting the two ENSO types is still formidable and further research is needed.

Both the IOD and NAST play important roles in influencing global and local climate variations (Birkett et al., 1999; Ashok et al., 2001; Li S. L. et al., 2003; Saji and Yamagata, 2003; Yang et al., 2007; Jia et al., 2011; Gitau et al., 2015; Tan et al., 2017; Lu et al., 2018). Figures 4a, c show the prediction skills, real-time forecasts, and related observations of the IOD index, respectively. As Fig. 4a shows, the TCC scores for IOD beginning from March are limited. This is presumably due to its winter prediction barrier, where model prediction of the IOD is usually subject to a significant decline in skill from January to March; the causes for this remain unclear (Wajsowicz, 2005; Luo et al., 2007; Feng et al., 2014). The results from CMMEv1.0 indicate that the simulation and prediction of IOD in the current climate models need to be improved further. For the real-time forecast, the six ensemble model members predict a near-normal IOD state in spring and a weak positive anomaly in summer, which are consistent with observations; however, the predicted intensity is significantly weakened.

Figure 4 Left panels: TCC prediction skills for (a) IOD and (b) NAST indices for monthly (March–August; hollow bars) and seasonal (JJA; hatched bars) mean predictions by the CMMEv1.0 initiated in March over 1991–2016. Right panels: monthly evolutions of (c) IOD and (d) NAST indices (°C) derived from OISSTv2 for March 2017–August 2018 (OBS; denoted by red/blue bars before March 2018 and by the thick black line after March 2018) and from CMMEv1.0 for March–August 2018 (forecast; denoted by red/blue hatched bars for the ensemble mean and by color lines for individual models M1–M6).

Figure options

Figure 4b shows that CMMEv1.0 has a higher prediction skill for NAST than for IOD. In March–June, the TCC scores for NAST are all larger than or close to 0.6 and then become lower in July–August, whereas in the JJA season the prediction skill is close to 0.6. As seen in Fig. 4d shows, the predictions of CMMEv1.0 for NAST and those of other models are very close to each other except in March, and their MME mean prediction is quite consistent with observations except for weaker positive anomalies of NAST.

The WPSH is one of the most important circulation patterns affecting the East Asian weather and climate. The variations of its area, intensity, and location have important impacts on the flood season in China (Wu et al., 2002). Accurate prediction of these characteristics of the WPSH can provide important reference information for predicting the transition and variation of summer rainbelts in China. Figures 5a–c show the prediction skills for the area index, the intensity index, and the western ridge point index of the WPSH (Liu et al., 2012), respectively. It is shown that the WPSH area, intensity, and western ridge point in CMMEv1.0 are significantly skillful, especially for the JJA season. For the 2018 flood season, real-time forecasts of the WPSH indices by CMMEv1.0 models (except for ECMWF-S4), as shown in Figs. 5e–g, are quite consistent with observations. Forecast results also reveal that both the area and intensity of the WPSH in the summer of 2018 tend to be near normal and the western ridge point is significantly eastward.

Figure 5 Left panels: TCC prediction skills for (a) WPSH area (b) WPSH intensity, (c) western ridge point, and (d) EASM indices for monthly (March–August; hollow bars) and seasonal (JJA; hatched bars) mean predictions by the CMMEv1.0 initiated in March over 1991–2016. Right panels: monthly evolutions of the indices of (e) WPSH area, (f) WPSH intensity, (g) western ridge point, and (h) EASM derived from NCEP-I for September 2017–August 2018 (OBS; denoted by red/blue bars before March 2018 and by the thick black line after March 2018) and from CMMEv1.0 for March–August 2018 (forecast; denoted by red/blue hatched bars for ensemble mean and by color lines for individual models M1–M6).

Figure options

The EASM is considerably important for the global atmospheric circulation systems. Precipitation in most regions of China, especially eastern China, is strongly affected by the intensity of the EASM, and the EASM index is significantly correlated with precipitation in central and eastern China (Jiang et al., 2013a). Successful EASM prediction will definitely help to forecast precipitation anomalies in China. Previous studies have indicated that dynamic prediction can capture the main spatial pattern and features of the EASM-related precipitation and large-scale circulation several months in advance (e.g., Yang et al., 2008; Jiang et al., 2013a). As Fig. 5d shows, CMMEv1.0 has a quite high prediction skill for the EASM intensity index (Zhang et al., 2003) in JJA season. During summer 2018, the observed EASM is stronger than climatology, and the predictions of each model and MME in CMMEv1.0 are similar to the observations but tend to be slightly weaker (Fig. 5h).

Figure 6 presents the prediction skill, the 2018 real-time forecasts of CMME v1.0, and related observations for the JJA SAT and precipitation anomalies in China. Clearly, Fig. 6a shows that the summer SAT prediction in most areas of China is skillful in CMMEv1.0 at the three-month leadtime, though TCCs are relatively low for the Huang–Huai area and small parts of western China. Comparison of Fig. 6b and 6c reveals that the real-time SAT forecast for summer 2018 is quite consistent with observation; i.e., the SAT in most regions of China is warmer than climatology with relatively larger anomalies in northern China and smaller anomalies in central and southern China; however, the predicted intensity is slightly weaker than the observation in most regions.

Figure 6 TCC prediction skills for seasonal (JJA) mean predictions of (a) SAT and (d) precipitation in China by CMMEv1.0 initiated in March over 1991–2016; and the seasonal mean of (b, c) SAT (°C) and (e, f) precipitation anomaly (%) for JJA 2018 (b, e) predicted by CMMEv1.0 initiated in March 2018 and (c, f) derived from OISSTv2. The dotted area in (a, d) denotes the statistical significant values at the 95% confidence level with the Student t-test.

Figure options

In contrast to SAT, the CMMEv1.0 has limited TCC skills for predicting summer precipitation in China for a leadtime of three months. As Fig. 6d shows, there are no significant skills in most regions of China except northwestern and eastern China. Comparison of the real-time forecast and observation for precipitation anomalies in summer 2018, as shown in Figs. 6e, f, indicates an overall good performance of CMMEv1.0. Observations show that the main rainbelt in the 2018 flood season exhibits a “positive–negative–positive” distribution in the north–south direction; i.e., precipitation in the northern and southern China is more than climatology but less than climatology over the Yangtze River basin. This rainfall anomaly pattern in summer 2018 has been well captured by the CMMEv1.0 prediction; however, the latter shows much more precipitation in the middle and lower reaches of the Yangtze River than the observation.

The multi-model ensemble (MME) prediction has become one of the most popular and useful approaches for short-term climate prediction and for development of current climate models towards a higher spatiotemporal resolution, more complicated physical processes, and seamless dynamical prediction. A number of previous studies have demonstrated that MME mean prediction is not only theoretically superior to single model prediction but can also significantly reduce the impact of systematic model errors on climate prediction. Because of the highly complex climate conditions in East Asia mixed with the uncertainties of the model itself and its initial values, short-term climate prediction based on a single model can no longer meet the increasing demands of climate services. Therefore, it is necessary to develop an operational MME climate prediction system and improve the prediction skill, particularly for East Asia and China.

Based on some advanced domestic and international climate models, the NCC/CMA has designed a unified initialization and post-processing framework and established CMMEv1.0, which has thus far been particularly used for monthly–seasonal climate forecasting. The CMMEv1.0 carries out its operational applications to predicting some primary modes of climate variability (such as the ENSO, IOD, NAST, WPSH, and EASM) and climate elements (SAT and precipitation). According to the evaluation of the 1991–2016 hindcast, CMMEv1.0 skillfully predicts global SSTA, especially for ENSO in terms of its various indices including the types. Our results show that TCC scores for both Niño3.4 and Niño4 indices are larger than 0.8 at the six-month leadtime, and the other El Niño indices and the NAST index are also accurately predicted by CMMEv1.0, whereas the accuracy of the IOD index starting from March is relatively low. The prediction skills for the area, intensity, and western ridge point of WPSH, and the intensity of EASM, especially in the boreal summer, are high. Moreover, CMMEv1.0 exhibits significant high skill for summer SAT for most regions of China for a three-month leadtime; however, it is only skillful for summer precipitation in Northwest and East China.

We have also produced real-time forecasts using CMMEv1.0 of the climate variability modes and climate elements for the 2018 flood season (June–August) starting from March 2018. Our results show that CMMEv1.0 has accurately captured the transition of ENSO from cold phase to neutral and warm phases and the weak positive anomalies of NAST. Due to the winter prediction barrier, the predicted IOD is weaker than observed in summer 2018. The WPSH is observed to be near-normal in intensity and eastward in summer 2018, which has been successfully captured by CMMEv1.0; however, the EASM is slightly weaker in the prediction than in the observation. CMMEv1.0 has also accurately predicted the spatial distribution of warmer temperatures in most parts of China. The main rainbelts in North China and southern coastal China in summer 2018 are well captured; however, CMMEv1.0 has shown much more precipitation in the middle and lower reaches of the Yangtze River than observation.

The analysis and evaluation results reported in this paper indicate that the MME mean predictions from CMMEv1.0 are more accurate than predictions of any single model. Overall, CMMEv1.0 can skillfully predict global SST anomalies (especially in the equatorial central and eastern Pacific), the WPSH area and intensity, and the EASM intensity. However, many aspects of CMMEv1.0 still need further improvement, such as its relatively lower performance for IOD prediction. Jiang et al. (2013b) mentioned that the CFSv2 has a low prediction skill for IOD, which is related to the annul cycle of IOD, the systematic errors of the Indian Ocean simulations, and the model biases in the IOD response to ENSO (Li T. et al., 2003; Fischer et al., 2005; Shi et al., 2012). Because of the relatively low ability of climate models to simulate and predict the East Asian climate, precipitation prediction remains challenging. Therefore, the MME mean shows limited improvement in precipitation prediction. Presently, the performance of CMMEv1.0 for predicting precipitation still cannot sufficiently meet the demands of operational climate prediction, and methods for improving its prediction skill need to be explored further. One option for consideration is to take advantage of the empirical correction of the prediction error, statistical downscaling, and other advanced methods for extracting predictable information and reducing model prediction error (e.g., Ren and Chou, 2006, 2007; Kug et al., 2008). Overall, research and development on the MME prediction system are still at an initial stage in China, and it is necessary to make full use of achievements in both domestic and international research in the field of MME prediction to improve the level of short-term climate prediction in China.

Acknowledgments. The authors are grateful to the three anonymous reviewers for their insightful comments, which have helped improve the quality of the paper.