-
The fast-growing SVs are calculated by solving an eigenvalue problem defined by the tangent forward and adjoint model. By taking the time evolution of a small perturbation X at an initial time t0 by a tangent forward operator L of the dynamical equations, we obtain:
where X(t) is the evolved perturbation from the initial time t0 to the future time t in the tangent forward model. Equation (1) can be described as a linearized assumption of the nonlinear trajectory of the dynamic systems. Then, the fastest-growing SV is found by identifying the phase–space direction with the largest growth rate J(x) [Eq. (2)] between the evolved perturbation X(t) and initial perturbation X(t0) during the optimal time interval (OTI), so that:
where [.,.] denotes the Euclidean inner product and E is a matrix operator that defines the specific form of the inner product to transform X into a dimensionless vector in Eulerian space, where the expression of the relationship between X and
$ {\mathbf{\hat X}} $ can be represented byIn this study, the dry-TE norm (Liu et al., 2013) is defined by:
where φ denotes the latitude of the spherical coordinate, ρr denotes reference density, Tr denotes reference temperature, θr denotes reference potential temperature, Πr denotes reference dimensionless air pressure, and Cp denotes the specific heat of air at constant pressure. The u and v represent the zonal and meridional wind, respectively,
$\tilde \theta $ ($ \tilde \theta = \theta - {\theta _r} $ ) is the perturbed potential temperature, and$ \tilde \Pi $ ($ \tilde \Pi = \Pi - \Pi_r $ ) is the perturbed Exner pressure corresponding to the perturbations of the CMA-TLM, i.e.,$[{u'},{v'},{(\tilde \theta )'},{(\tilde \Pi )'}]$ [See Liu et al. (2017) for details]. Hence E can be expressed as:As a result, the CMA-TLM perturbed variables with respect to E can be expressed as
$ {\mathbf{\hat X}} $ :Thus, Eq. (2) can be computed as
According to the variational principle of the symmetric matrix, the maximization problem in Eq. (5) can be considered an SVD problem [Eq. (6)]:
where LT is the adjoint model and λ is the eigenvalue of the matrix (ELPE−1)T(ELPE−1). P is the local projection operator (LPO), making it possible to compute SVs in the target area (Buizza, 1994a). Finally, the matrix ELPE−1 is the target SV with the fastest growing rate.
Thus, two types of SVs are computed in this study (The details of these multiscale SVs are summarized in Table 1 and Fig. 1). First, the LSV is designed to represent the synoptic scale uncertainties in the troposphere. The calculation settings for the LSV are the same as the operational GEPS at the CMA. The LSV is computed by 2.5° TLM with dry linearized physical processes (subgrid-scale orographic effect and vertical diffusion), a 48-h OTI, and a dry TE norm. Most NWP centers use similar configurations (e.g., lower horizontal resolution, longer OTI, and a dry-TE norm) to compute an extratropical SV to generate initial perturbation for GEPS (Leutbecher and Palmer, 2008; Descamps et al., 2015; Yamaguchi et al., 2018). Due to the LSV being focused on synoptic scale uncertainties in the troposphere, the dry TE norm is computed by vertical integration from the 4th to 50th model level (nearly 100 m to 16 000 m) over the target areas (NH and SH).
LSV MSV Target area 30°–80°N (NH)
30°–80°S (SH)30°–80°N
30°–80°SResolution 2.5° 1.5° OTI 48 h 24 h Norm Dry TE Dry TE linearized physical processes
(dry)Subgrid–scale orographic effect,
vertical diffusionSubgrid–scale orographic effect,
vertical diffusionNo. of SVs 10 10 Table 1. Calculation settings for the multiscale SVs.
Figure 1. Forecast region of CMA-GEPS (Global). Shown are the target areas of LSV and MSV (NH and SH, blue dashed lines) and the verification regions for GEPS (NthH and SthH, orange dashed-dotted lines, details can be found in section 4.2).
However, mesoscale errors also influence the medium-range prediction skill, then the LSV is insufficient to detect mesoscale uncertainties. An MSV is set to capture meso-α- to synoptic scale uncertainties in meteorological systems over the target areas. The differences between MSV and LSV’s settings include the horizontal resolution (e.g., 1.5°) and OTI (e.g., 24 h). Previous studies have proven that increasing horizontal resolution leads to a reduction in the ability of the OTI to obtain fast-growing SVs in a reasonable manner (Buizza, 1994b; Wang et al., 2020; Ono et al., 2021).
It should be noted that, unlike previous studies, Ye et al. (2020) and Ono et al. (2021) used SVs of differing scales for REPS with higher resolution, while this study uses multiscale SVs for GEPS with lower resolution, considering that the main goal of GEPS is to provide guidance information about medium-range prediction. Therefore, the specific calculation settings of multiscale SVs differ from previous studies and are evaluated based on several sensitivity experiments.
-
Since each type of SV can only reflect uncertainty information separately at distinguished meteorological scales, it is necessary to combine all multiscale SVs to generate initial perturbations that reflect uncertainties as much as possible with finite ensemble members. Meanwhile, the focus of this study is to determine a reasonable method for combining multiscale SVs to generate initial perturbations for the GEPS, according to the probability density function.
Before the combination procedure, the LSVs and MSVs are interpolated to the same horizontal resolution as the CMA-GEPS (e.g., 0.5°). Subsequently, using a Gram-Schmidt re–orthogonalization for each of the SVs ensures the removal of duplicate information. After that, it is necessary to adjust the amplitude of each type of multiscale SV according to the analysis error (
$ {e_u},{e_v},{e_\theta } $ and$ {e_\Pi } $ ) by a rescale operator$ {\bar L^2} $ defined as follows:where the overbar represents a mean over grid-point spaces, and N1 is the total grid points in the CMA-TLM. The analysis error is estimated from the CMA's four-dimensional variational data assimilation (CMA-4DVAR) system. Each perturbed variable has a reference analysis error that is related to target areas (NH and SH) and model level and varies according to the month in question.
Hence, the LSVs and MSVs, have a different
$ {\bar L^2} $ according to Eq. (7). Therefore, the matrix of the SVs in the real state (X), and their matrix in the Eulerian state ($ {\mathbf{\hat X}} $ ), can be described by:where γ is a constant factor to control the maximum acceptable ratio between perturbations amplitude and analysis error. After the rescaling procedure, the multiscale SV initial perturbations Pj are generated by:
where NNH and NSH denote the number of SVs in the NH and SH, respectively; αj,k denotes the random coefficients according to the Gaussian distribution, j and k denote the number of initial perturbations and SVs, respectively; coefficients β and δ determine the amplitude of the LSV’s and MSV’s perturbation at their initial states, respectively.
Finally, a digital filter method is used to modify abnormal values in the multiscale SV initial perturbations according to the triple reference analysis error which is estimated from the CMA-4DVAR.
The perturbed variables listed here include potential temperature and the u and v wind speeds. Subsequently, a pair of multiscale SV initial perturbations are added to or subtracted from the control forecast to generate ensemble members for CMA-GEPS. Each ensemble member has an equal probability. The experimental settings and parameters for the initial perturbations of multiscale SVs in this study are summarized in Table 2, and the main process of the multiscale SV approach is summarized in Fig. 2.
Experiment name Details of perturbation γ β δ LSV Only use LSV perturbation 0.05 0 0 MSV Only use MSV perturbation 0.045 0 0 MLSV1 Use multiscale SV perturbations 0.068 0.75 0.75 MLSV2 Use multiscale SV perturbations 0.068 0.5 0.5 MLSV3 Use multiscale SV perturbations 0.068 0.25 0.25 Table 2. Details of multiscale SV initial perturbation experiments.
Compared with the previous study, the JMA combines all SVs determined by the variance minimum rotation method for a regional EPS (Saito et al., 2011; Ono, 2020). The Gaussian sampling with coefficients method is introduced in this study, which combines all multiscale SVs to generate initial perturbations for GEPS with finite ensemble members. The purpose of coefficient settings is to explore whether the multiscale SV initial perturbations with different amplitudes can influence the prediction skills of the GEPS.
-
This study used the CMA-GEPS to carry out global ensemble forecast experiments. The CMA-GEPS was developed independently by the CMA’s Center for Earth System Modeling and Prediction, based on the CMA Global Forecast System (CMA-GFS) and has been running operationally since 2018 (Li et al., 2019; Shen et al., 2020; Peng et al., 2022). The control member (CTL, unperturbed member) of CMA-GEPS is initialized by CMA-4DVAR (the four-dimensional variational data assimilation system of the CMA) (Zhang et al., 2019), including 720 × 360 horizontal grid points with a resolution of 0.5° and 60 vertical layers. Furthermore, the five experiments of initial perturbation are added to and subtracted from the CTL to initialize the ensemble member in CMA-GEPS.
The multiscale SV characteristics and their performance in CMA-GEPS are analyzed based on a total of 32 cases for each experiment. To reduce the impact of test cases, representative months were selected from different seasons (JAN, APR, JUL, and OCT) for experimentation and verification. For each experiment, the same set of 10-day ensemble forecasts is initialized every fourth day (day 1, 5, 9, 13, 17, 21, 25, 29) for four months (total of 32 cases) at 0000 UTC. Heavy rainfall (≥25 mm (24 h)–1) cases mainly occurred in APR, JUL, and OCT, with 23 cases in total. The ensemble forecast time is 240 h and the forecast interval is 24 h. The configurations of CMA-GEPS in this study are summarized in Table 3.
Parameter CMA–GEPS Forecast region Global Resolution 0.5°, L60 Model top 3 hPa Forecast length 240 h Output frequency 24 h Initial condition CMA–4DVAR (upscaling) Initial perturbation method Single–scale SV or multiscale SV
(mentioned in section 2.2)Ensemble members 1 unperturbed member + 20 perturbed members Model perturbation method − Table 3. Configurations of CMA-GEPS for ensemble experiments
LSV | MSV | |
Target area | 30°–80°N (NH) 30°–80°S (SH) |
30°–80°N 30°–80°S |
Resolution | 2.5° | 1.5° |
OTI | 48 h | 24 h |
Norm | Dry TE | Dry TE |
linearized physical processes (dry) |
Subgrid–scale orographic effect, vertical diffusion |
Subgrid–scale orographic effect, vertical diffusion |
No. of SVs | 10 | 10 |