Initial perturbations play a crucial role in determining the performance of an ensemble prediction system (EPS). This study comprehensively compares three types of initial perturbations—generated by ensemble data assimilation (EDA), singular vectors (SV), and their hybrid EDA–SV—using the China Meteorological Administration (CMA) global forecast model. In addition to conventional ensemble verification metrics, diagnostic tools such as kinetic energy (KE) spectrum analysis and spatial filtering are employed. Compared with the SV-based perturbations currently used operationally in the CMA global EPS (CMA-GEPS), EDA-based perturbations exhibit more smaller-scale structures and higher global perturbation KE, particularly in the Tropics. Ensemble forecasting experiments reveal that the EDA method provides superior ensemble spread and perturbation KE in the Tropics during the early forecast period. However, the SV method performs better in the extratropics throughout the forecast period and in the Tropics during the mid-to-late forecast period. The EDA–SV approach improves the overall performance of CMA-GEPS, yielding better spread–error relationships and enhanced forecast skill compared with SV and EDA methods. Results from spatially filtered ensembles further show that EDA–SV combines the subsynoptic-scale and mesoscale advantages of EDA-based perturbations in the Tropics, with the large-scale and synoptic-scale strengths of SV-based perturbations across the globe. This synergy leads to superior performance across spatial scales and lead times in both tropical and extratropical regions. Consequently, the EDA–SV method is planned for implementation in the next upgrade of the CMA-GEPS.

Madden–Julian oscillation (MJO) is one of the dominant sources of extended-range atmospheric predictability. However, many atmospheric numerical models fail to predict MJO owing to limitations in representing air–sea interaction processes. This study investigated the role of sea surface temperature (SST) forcing in improving MJO predictability using China Meteorological Administration uncoupled Global Ensemble Prediction System (CMA-GEPS), extending forecasts from 16 to 35 days for the first time by applying different SST forcing schemes. Three SST forcing schemes are examined: (1) CTL SST, which uses fixed SST forcing, (2) E-Folding SST, which adjusts initial analyses toward observed climatology, relying primarily on the historical observations, and (3) Two-Tiered SST, which combines the bias-corrected SST analysis and predictions from the CMA coupled prediction system. The results confirmed that each SST forcing scheme is reasonably designed, with the Two-Tiered scheme providing better SST forcing information and more realistic equatorial wave features. For MJO prediction, the Two-Tiered scheme exhibits superior skill, particularly with strong MJO initialization, extending predictability to 20.6 days, i.e., 4.7 days longer than the CTL run and 3.4 days longer than the E-Folding test. Furthermore, the Two-Tiered scheme outperforms CTL and E-Folding over the western Indian Ocean, but it underperforms over the Maritime Continent. It exhibits poorer eastward propagation characteristics and larger MJO errors due to larger SST forcing errors, which mainly from the analysis-related errors of the coupled model. Larger SST forcing errors increase the water vapor biases and amplify MJO prediction errors, indicating the need for optimized SST forcing. Additionally, the Two-Tiered scheme significantly enhances anomaly correlation coefficient (ACC) skills for 500-hPa geopotential height, with the improvement closely linked to the enhancement in MJO forecast skill. Overall, even the one-way SST forcing that includes multi-faceted information of the coupled model can provide guidance for extended-range prediction with atmospheric-only numerical models.

Extreme temperature events have significant impacts on human society and economic activities, yet the prediction still involves considerable uncertainties, making the use of ensemble forecasting methods crucial. The Pangu-Weather Global Ensemble Prediction System (PGW-GEPS) was developed by integrating the Pangu-Weather (PGW) with perturbed initial conditions of the China Meteorological Administration Global Ensemble Prediction System (CMA-GEPS). Using the 2022 extreme heat wave event in Zhejiang and the 2024 cold wave event in Inner Mongolia as two cases, the forecasting performances of PGW-GEPS and CMA-GEPS on these two extreme temperature events are evaluated and compared based on multiple assessment metrics. The results indicate that, for both the Zhejiang heat wave and Inner Mongolia cold wave event, PGW-GEPS exhibits forecast accuracy and uncertainty representation capabilities comparable to CMA-GEPS. Both systems effectively capture the increase in 2 m air temperature forecast uncertainty with longer lead times and its subsequent decrease as the forecast initialization approaches the observation period. However, for the Zhejiang heat wave, PGW-GEPS shows deficiencies in forecasting the shear line and exhibits larger forecast errors in the medium range. A comparative analysis of the kinetic energy spectra of these two events further reveals that PGW-GEPS exhibits an attenuation phenomenon below the sub-synoptic scale. In summary, the AI (Artificial Intelligence)-based PGW-GEPS demonstrates its forecasting capability for extreme temperature events. Its forecast accuracy for 3—10 d extreme temperature prediction is comparable to that of CMA-GEPS, while its computational speed is advantageous. However, PGW-GEPS still faces challenges in capturing rapidly evolving meso-micro scale weather systems, and further improvement in forecasting sub-synoptic systems is required. This study provides valuable insights into the application of artificial intelligence models in ensemble forecasting.

Extreme Forecast Index (EFI) provides an effective tool to extract extreme weather information from ensemble forecasts. To improve the ability of the CMA global ensemble prediction system (CMA-GEPS) for extreme weather forecast and address the difficulty of reasonably calculating the model climate distribution due to small samples of historical forecasts by CMA-GEPS and the lack of re-forecast data, this study develops a method to build the model climate distribution required by EFI using insufficient samples of deterministic forecasts. Based on the CMA global high-resolution (0.25°×0.25°) deterministic operational forecast data from 15 June 2020 to 22 July 2022, the model climate distributions are constructed for each month at different forecast lead times (1—10 d) that match the lower-resolution (0.5°×0.5°) CMA-GEPS forecast model version through extending the forecast samples in both time and space. By employing the operational forecast data of CMA-GEPS and the ERA5 reanalysis data, the forecast ability of CMA-GEPS for extreme high temperature in four representative regions both domestic and abroad for the summer of 2022 (June to August) is evaluated. Results from the relative operating characteristic curve show that the CMA-GEPS EFI has the ability to detect extreme high temperature within the short- and medium-range forecast lead times of 1—10 d. Taking the maximum TS score as the criterion, the critical threshold of EFI for issuing warning signals of extreme high temperature is determined. The forecast ability of EFI decreases with increasing forecast lead time, and different performances exhibit in different regions: the forecast ability for extreme high temperature in the middle and lower reaches of the Yangtze river in China is higher than that in North China for all lead times; the forecast ability of EFI in western Europe is better than that in central Europe for the 1—7 d lead times, yet the EFI forecast ability in central Europe for the 8—10 d lead times is better. Above results are related to the variation of ensemble forecast quality of 2 m temperature with forecast lead time and spatial location. Evaluation results from the economic value model reveal that risk decisions based on the EFI forecast information demonstrate certain economic values and reference values. Analysis results from a case study further indicate that the CMA-GEPS EFI can provide early warnings of extreme high temperature in the medium forecast range.

This paper reviews the development of ensemble weather forecast and the primary techniques employed in themain ensemble prediction systems (EPSs) designed by China and other countries. Here, the emphasis is placed on theadvancements in the China Meteorological Administration (CMA) global and regional ensemble prediction systems (i.e., CMA-GEPS and CMA-REPS), with particular attention to operational technologies such as initial and modelperturbation methods and the applications of ensemble forecast. Through comparative verification with EPSs fromother leading international numerical weather prediction (NWP) centers, CMA's EPSs demonstrate forecast skillscomparable to its global counterparts. As EPSs progress to convective scales and coupled systems between sea, land,air, and ice, the paper addresses some key challenges in ensemble forecast technologies across the aspects of opera-tion, science, integration of artificial intelligence (AI), merging of weather and climate models, and challenging userrequirements. Finally, a summary of conclusions and future perspectives on ensemble forecast are provided.

This paper reviews the development of ensemble weather forecast and the primary techniques employed in the main ensemble prediction systems (EPSs) designed by China and other countries. Here, the emphasis is placed on the advancements in the China Meteorological Administration (CMA) global and regional ensemble prediction systems (i.e., CMA-GEPS and CMA-REPS), with particular attention to operational technologies such as initial and model perturbation methods and the applications of ensemble forecast. Through comparative verification with EPSs from other leading international numerical weather prediction (NWP) centers, CMA’s EPSs demonstrate forecast skills comparable to its global counterparts. As EPSs progress to convective scales and coupled systems between sea, land, air, and ice, the paper addresses some key challenges in ensemble forecast technologies across the aspects of operation, science, integration of artificial intelligence (AI), merging of weather and climate models, and challenging user requirements. Finally, a summary of conclusions and future perspectives on ensemble forecast are provided.

The traditional model perturbation method of ensemble prediction is usually used to describe random errors of physical processes, but the model inevitably has systematic bias. Therefore, in order to reduce the impact of systematic bias on ensemble prediction, the CMA-GEPS is employed to obtain systematic bias tendency using the empirical orthogonal function (EOF) method. In the integration process, the systematic bias correction method and the traditional Stochastically Perturbed Parameterization Tendency (SPPT) are combined to build a model perturbation method (Bias correction of bias tendency based on SPPT, SPPT-B) that combines systematic bias and random errors of ensemble forecast. Ensemble forecasting experiments are designed and carried out to explore the impact of SPPT-B on global ensemble forecasting. The conclusions are as follow: (1) The first EOF mode of the systematic bias can reflect the main characteristics of the systematic bias well. It shows that basically the systematic bias in the upper troposphere is larger than that in the middle and lower troposphere and increases linearly with forecast lead time. (2) The systematic bias correction method and SPPT-B can effectively reduce the systematic bias in upper and lower levels in the southern and northern Hemispheres and in the tropics, and SPPT-B can significantly improve Spread in the tropics. (3) The effect of the two schemes on the improvement of ensemble prediction skill in the upper troposphere is better than that in the lower troposphere. The above results indicate that the model perturbation method that considers both systematic bias and random errors can effectively improve global ensemble forecasting skill, and can provide a scientific basis for the development of global ensemble forecasting model perturbation method considering both systematic bias and random errors.

Evaluating whether an ensemble prediction system (EPS) can accurately represent forecast uncertainty is a key aspect of model development and ensemble forecast applications. In this study, a four-dimensional diagnostic analysis model for assessing ensemble forecast uncertainty is proposed, by analyzing the relationship between the ensemble spread and root-mean-square error (RMSE) of the ensemble mean in terms of their temporal evolution (one-dimensional) and spatial distribution (three-dimensional), together with use of the linear variance calibration (LVC) method. Based on this model and the daily operational forecast data of the China Meteorological Administration (CMA) global EPS (CMA-GEPS) in December 2022–November 2023, characteristics of the CMA-GEPS forecast uncertainty are diagnosed and analyzed, and compared against the state-of-the-art operational global EPS of ECMWF. Generally, there is a deficiency in CMA-GEPS, which underestimates the forecast uncertainty, especially in the tropics. However, at certain initialization times in some seasons and over some locations, the spread appears greater than the RMSE, indicating an overestimation of forecast uncertainty. Moreover, CMA-GEPS performs better in capturing the forecast uncertainty of lower-level variables than upper-level variables; and in comparison with the mass and thermal fields, the forecast uncertainty of the dynamic field is better represented. Diagnostic analysis using the LVC method reveals that the relevance between the ensemble variance and the ensemble mean error variance of CMA-GEPS increases with forecast lead time, and the problem of underestimated forecast uncertainty is continuously alleviated. In addition, ECMWF EPS behaves distinctly better than CMA-GEPS in representing the forecast uncertainty and its growth process, the reasons for which are discussed and elucidated from the perspective of shortcomings in the methods to generate the initial and model perturbations, the ensemble size, and the forecast model adopted by CMA-GEPS.

Given the chaotic nature of the atmosphere and inevitable initial condition errors, constructing effective initial perturbations (IPs) is crucial for the performance of a convection-allowing ensemble prediction system (CAEPS). The IP growth in the CAEPS is scale- and magnitude-dependent, necessitating the investigation of the impacts of IP scales and magnitudes on CAEPS. Five comparative experiments were conducted by using the China Meteorological Administration Mesoscale Numerical Weather Prediction System (CMA-MESO) 3-km model for 13 heavy rainfall events over eastern China: smaller-scale IPs with doubled magnitudes, larger-, meso-, and smaller-scale IPs; and a chaos seeding experiment as a baseline. First, the constructed IPs outperform unphysical chaos seeding in perturbation growth and ensemble performance. Second, the daily variation of smaller-scale perturbations is more sensitive to convective activity because smaller-scale perturbations during forecasts reach saturation faster than meso- and larger-scale perturbations. Additionally, rapid downscaling cascade that saturates the smallest-scale perturbation within 6 h for larger- and meso-scale IPs is stronger in the lower troposphere and near-surface. After 9–12 h, the disturbance development of large-scale IPs is the largest in each layer on various scales. Moreover, thermodynamic perturbations, concentrated in the lower troposphere and near-surface with meso- and smaller-scale components being dominant, are smaller and more responsive to convective activity than kinematic perturbations, which are concentrated on the middle–upper troposphere and predominantly consist of larger- and meso-scale components. Furthermore, the increasing magnitude of smaller-scale IPs enables only their smaller-scale perturbations in the first 9 h to exceed those of larger- and meso-scale IPs. Third, for forecast of upper-air and surface variables, larger-scale IPs warrant a more reliable and skillful CAEPS. Finally, for precipitation, larger-scale IPs perform best for light rain at all forecast times, whereas meso-scale IPs are optimal for moderate and heavy rains at 6-h forecast time. Increasing magnitude of smaller-scale IPs improves the probability forecast skills for heavy rains during the first 3–6 h.

How to construct appropriate perturbations for convection-permitting ensemble prediction systems (CPEPSs) is a critical issue awaiting urgent solutions. As two common perturbations, initial perturbations (IPs) and lateral boundary perturbations (BPs) interact with each other, affecting the model error growth, especially in mesoscale models. Using the China Meteorological Administration (CMA)-CPEPS, this study tries to elucidate how BPs interact with matched and mismatched IPs under varied large-scale weather conditions/forcings. Seven groups of experiments were conducted for strong-forcing and weak-forcing weather regimes over southern China: three with single IPs, one with single BPs, and three with combined perturbations. It is found that the perturbation magnitudes were dominated by meso-α-scale components, and IPs under weak forcing exhibited more pronounced effects than under strong forcing; whereas BPs exerted more pronounced effects under strong forcing than weak forcing regimes. Furthermore, it lasts longer for high-level variables when the perturbation energy from BPs is higher than that from IPs, compared to low-level variables. Moreover, for precipitation and dynamic variables, IPs and BPs can mutually reinforce. The source of these perturbations, and their specific vertical levels, do not alter the extent of their interactions. Nevertheless, the weather regime and the scales of the perturbations influence the strength of their mutual reinforcement. In particular, the weak-forcing regimes exhibit a more pronounced reinforcing effect, and meso-α-scale perturbations are more conducive to fostering interactions compared to meso-β-scale ones. Ultimately, it is the perturbation magnitude inherent in the initial perturbation itself that determines the interactions between IPs and BPs.