Hybrid three-dimensional variational assimilation method fusing multi-scale ensemble forecast samples

Abstract

The invention provides a mixed three-dimensional variational assimilation method fusing a multi-scale ensemble forecast sample, and belongs to the technical field of numerical weather forecast and data assimilation, and the method comprises the steps: obtaining and processing a global ensemble forecast sample and a regional convective scale ensemble forecast sample, obtaining a large-scale ensemble disturbance sample and a small-scale ensemble disturbance sample, and carrying out the processing of the large-scale ensemble disturbance sample and the small-scale ensemble disturbance sample; inputting into a hybrid three-dimensional variational analysis system, taking a forecasting field of a global numerical weather forecasting mode of a meteorological center as a first background field, constructing a large-scale flow dependent background error covariance by using a large-scale set disturbance sample to obtain a first analysis field, taking the first analysis field as a second background field, and constructing a second analysis field by using the second background field; and constructing a small-scale flow dependent background error covariance by using the small-scale set disturbance sample to obtain a final analysis field. According to the method, the problem of inaccurate numerical weather forecast caused by insufficient information fusion of large-scale information and small-scale information, insufficient observation and static background error covariance in the existing three-dimensional variation assimilation technology is solved.

Description

A hybrid three-dimensional variational assimilation method that integrates multi-scale ensemble forecast samples Technical Field

[0001] This invention belongs to the field of numerical weather prediction and data assimilation technology, and particularly relates to a hybrid three-dimensional variational assimilation method that integrates multi-scale ensemble forecast samples. Background Technology

[0002] Three-dimensional variational assimilation is a core technology widely used in numerical weather prediction. The meteorological bureau has established a hybrid three-dimensional variational analysis (CMA-MESOHybrid-3DVar) system for numerical weather prediction. This system constructs a hybrid assimilation framework by linearly combining the static climatological background error covariance with the dynamic flow-dependent background error covariance. However, the estimation of the dynamic covariance relies on ensemble forecast samples. Existing three-dimensional variational methods have limitations in obtaining dynamic covariance: their flow-dependent background error covariance is mainly generated using random observation perturbation methods, which do not accurately describe the flow-dependent characteristics of the error structure. Furthermore, when using ensemble samples, existing methods either use only global ensemble forecast system samples or only regional ensemble forecast system samples. Global ensemble samples provide large-scale error information but lack resolution, while regional ensemble samples, although high-resolution, may have incomplete large-scale information or be affected by boundary conditions. This makes it difficult for existing methods to construct dynamic background error covariance that fully represents error information from synoptic to convective scales, thus failing to meet the needs for refined analysis of extreme weather systems such as rainstorms and typhoons. Summary of the Invention

[0003] To address the aforementioned shortcomings in existing technologies, this invention provides a hybrid three-dimensional variational assimilation method that integrates multi-scale ensemble forecast samples. This method solves the problems of inaccurate numerical weather predictions caused by insufficient information fusion between large-scale and small-scale information, insufficient observation, and static background error covariance in existing three-dimensional variational assimilation techniques.

[0004] To achieve the above objectives, the technical solution adopted in this invention is as follows: a hybrid three-dimensional variational assimilation method for fusing multi-scale ensemble forecast samples, comprising the following steps: acquiring global ensemble forecast samples and regional convective-scale ensemble forecast samples from the CMA-GEPS global ensemble forecast system; performing vertical and horizontal co-scale downscaling on the global ensemble forecast samples to obtain large-scale ensemble perturbation samples matching the resolution of the CMA-MESO regional medium-scale numerical weather prediction model; performing vertical spline interpolation on the convective-scale ensemble forecast samples to obtain small-scale ensemble perturbation samples matching the resolution of the CMA-MESO regional medium-scale numerical weather prediction model; inputting the large-scale ensemble perturbation samples and small-scale ensemble perturbation samples into a hybrid three-dimensional variational analysis system, using the forecast field of the meteorological center's global numerical weather prediction model as the first background field, constructing a large-scale flow-dependent background error covariance using the large-scale ensemble perturbation samples, and assimilating the large-scale observation data to obtain the first analysis field; using the first analysis field as the second background field, constructing a small-scale flow-dependent background error covariance using the small-scale ensemble perturbation samples, and assimilating the small-scale observation data to obtain the final analysis field.

[0005] To address the problems of existing three-dimensional variational assimilation techniques, which rely on single ensemble forecast samples from only the global or regional scales, resulting in dynamic background error covariances that cannot simultaneously represent the completeness of large-scale circulation background and mesoscale convective systems, and the inaccuracy of numerical weather predictions due to the static nature of flow-dependent background error covariances, this invention provides a hybrid three-dimensional variational assimilation method that integrates multi-scale ensemble forecast samples. First, global ensemble forecast samples and regional convective-scale ensemble forecast samples from the CMA-GEPS global ensemble forecast system are acquired. These are then subjected to vertical and horizontal co-downscaling and vertical interpolation to match the resolution of the CMA-MESO mesoscale numerical forecast model, respectively. This yields large-scale ensemble perturbation samples representing large-scale events and small-scale ensemble perturbation samples representing mesoscale error information. A sequential assimilation process, first large-scale and then small-scale, is employed, utilizing the large-scale ensemble perturbation... This method constructs a large-scale flow-dependent background error covariance sample. It sequentially assimilates large-scale observation data such as radiosonde observations, ship observations, and aircraft reports to obtain a first analysis field that contains coordinated large-scale information. Then, it uses the first analysis field as a new and better background field to construct a small-scale flow-dependent background error covariance using small-scale ensemble perturbation samples. It assimilates radar radial wind and ground observations as small-scale observation data to obtain a final analysis field that contains both an accurate large-scale environmental field and a fine small-scale structure, and is dynamically coordinated. By assimilating large-scale observation data first and then assimilating small-scale observation data with the first analysis field as the background, it solves the problem that existing hybrid assimilation methods do not fully integrate multi-scale ensemble samples. At the same time, it enables the final analysis field to retain the stability of large-scale circulation in atmospheric motion and capture the rapid changes of small and medium-scale systems, thereby improving the forecasting capability for extreme weather such as rainstorms and typhoons.

[0006] Furthermore: the global ensemble forecast samples are from 21 global ensemble forecast members generated by the global ensemble forecast system, used to characterize large-scale information; the regional convective-scale ensemble forecast samples are from the regional ensemble forecast system, used to characterize small- and medium-scale information.

[0007] The further beneficial effects mentioned above are as follows: This invention uses multiple members of the real-time global ensemble forecasting system CMA-GEPS, such as the 21 members, and the output of the regional convective-scale ensemble forecasting system CMA-REPS as data sources, ensuring that the acquired large-scale and small-scale error information is representative and real-time, making the constructed final analysis field closer to the actual weather situation, and enhancing the reliability and practicality of the three-dimensional variational assimilation analysis results.

[0008] Further: Obtaining large-scale ensemble perturbation samples that match the resolution of the CMA-MESO regional ensemble numerical weather prediction model specifically includes: based on the global ensemble forecast samples of the CMA-GEPS global ensemble forecast system, vertically adjusting the number of model surface layers of the global ensemble forecast samples to the number of model surface layers in the hybrid three-dimensional variational analysis system through vertical spline interpolation, and adjusting the spatial resolution of the global ensemble forecast samples to the spatial resolution adapted to the hybrid three-dimensional variational analysis system through horizontal bilinear interpolation, thereby obtaining large-scale ensemble perturbation samples that match the resolution of the CMA-MESO regional ensemble numerical weather prediction model.

[0009] The further beneficial effects mentioned above are as follows: This invention uses horizontal bilinear interpolation for longitude and latitude data in the horizontal direction, and uses meteorological variables such as temperature and air pressure from four neighboring grid points around the target point for linear weighting to quickly obtain a continuous horizontal distribution; vertical spline interpolation, in the vertical direction, i.e., at altitude or pressure layer, accurately calculates the variable values ​​at the target altitude by fitting a smooth cubic spline curve based on data points from multiple adjacent layers, which can better maintain the continuity and fine structure of the atmospheric vertical profile. This invention solves the problem of differences in vertical and horizontal resolution between the global ensemble forecast field of CMA-GEPS and the regional scale numerical forecast model of CMA-MESO, thus...

[0010] Further: obtaining small-scale ensemble perturbation samples that match the resolution of the CMA-MESO regional ensemble numerical weather prediction model specifically includes: based on the regional convective-scale ensemble forecast samples of the CMA-GEPS global ensemble forecast system, interpolating the regional convective-scale ensemble forecast samples to the model surface layer number of the hybrid three-dimensional variational analysis system using the vertical spline interpolation method to obtain small-scale ensemble perturbation samples that match the resolution of the CMA-MESO regional ensemble numerical weather prediction model.

[0011] The further beneficial effects mentioned above are as follows: by using the vertical spline interpolation method, the problem of vertical differences between the regional ensemble forecast field of CMA-REPS and the regional mesoscale numerical forecast model of CMA-MESO is solved, and small-scale ensemble perturbation samples can improve the subsequent analysis field's characterization of the fine structure of small and medium-scale weather systems such as rainstorm convection.

[0012] Furthermore, the expression for the first analysis field is as follows:

[0013]

[0014] in, For the first analysis field, As the first background scene, For large-scale gain matrices, For large-scale observation data, For observation operators targeting large-scale observation data, For large-scale background error covariance matrix, For observation error covariance matrix of large-scale observation data, This is the transpose symbol.

[0015] Furthermore, the expression for the final analysis field is as follows:

[0016]

[0017] in, For the final analysis field, This serves as both the first analytical field and the second background field. For small-scale gain matrices, For small-scale observation data, For observation operators targeting small-scale observation data, The small-scale background error covariance matrix, The small-scale observation error covariance matrix, This is the transpose symbol.

[0018] The beneficial effects of this invention are as follows: By sequentially fusing global large-scale and regional small-scale ensemble samples, a dynamic background error covariance that can simultaneously coordinate and characterize the flow-dependent error features at multiple scales is constructed. This enables the analysis field to more accurately depict the complete information from the synoptic-scale circulation background to the fine structure at the convective scale, solving the problem that traditional static or single-scale dynamic covariance is difficult to accurately analyze small- and medium-scale systems. For observational data of different scales, flow-dependent background error covariances containing both large-scale and small- and medium-scale information are used respectively, effectively solving the problem of weak flow dependence of the background error covariance. This allows the covariance to dynamically characterize the evolutionary features of weather systems, independent of model and parameter settings. This invention employs sequential multi-scale ensemble variational assimilation, with global ensemble samples dominating large-scale observation assimilation and regional ensemble samples dominating small-scale observation assimilation. In extreme weather forecasts such as heavy rain and typhoons, this improves the rationality of numerical forecasts such as precipitation and solves the problem of insufficient and unreasonable utilization of multi-scale observational information. Attached Figure Description

[0019] Figure 1 is a schematic diagram of a hybrid three-dimensional variational assimilation method that integrates multi-scale ensemble forecast samples; Figure 2 is a plot of the analytical bias and root mean square error of each group of experiments relative to radar radial wind (VR) at 00:00 on June 27, 2022. Detailed Implementation

[0020] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0021] Example 1, as shown in Figure 1, is a schematic diagram of a hybrid three-dimensional variational assimilation method that integrates multi-scale ensemble forecast samples. The method includes the following steps: acquiring global ensemble forecast samples and regional convective-scale ensemble forecast samples from the CMA-GEPS global ensemble forecast system; performing vertical and horizontal co-scaling on the global ensemble forecast samples to obtain large-scale ensemble perturbation samples matching the resolution of the CMA-MESO regional medium-scale numerical weather prediction model; performing vertical spline interpolation on the convective-scale ensemble forecast samples to obtain small-scale ensemble perturbation samples matching the resolution of the CMA-MESO regional medium-scale numerical weather prediction model; inputting the large-scale and small-scale ensemble perturbation samples into a hybrid three-dimensional variational analysis system, using the forecast field of the China Meteorological Administration's global numerical weather prediction model as the first background field, constructing a large-scale flow-dependent background error covariance using the large-scale ensemble perturbation samples, and assimilating the large-scale observation data to obtain the first analysis field; using the first analysis field as the second background field, constructing a small-scale flow-dependent background error covariance using the small-scale ensemble perturbation samples, and assimilating the small-scale observation data to obtain the final analysis field.

[0022] In one embodiment of the present invention, global ensemble forecast samples are acquired and input into the CMA-MESO Hybrid-3Dvar hybrid three-dimensional variational analysis system for subsequent processing. Specifically, 21 global ensemble forecast members generated by the CMA-GEPS global ensemble forecast system are used as the source of global ensemble forecast samples containing large-scale information, with a horizontal resolution of 25 km and 89 model surface standard layers. The hybrid three-dimensional variational analysis system has a horizontal resolution of 3 km and 72 model surface standard layers. The global ensemble forecast samples need to undergo vertical and horizontal co-scaling to obtain large-scale ensemble perturbation samples that match the resolution of the CMA-MESO regional scale numerical forecast model. The horizontal resolution and the number of model surface layers vary depending on the selected version of the GEPS global ensemble forecast system and can be adjusted accordingly based on the specific horizontal resolution and the number of model surface layers.

[0023] This invention first performs vertical spline interpolation on the global ensemble forecast samples, vertically adjusting the 89-layer model surface standard layer data to the 72-layer model surface standard layer of the hybrid three-dimensional variational analysis system. Then, through horizontal bilinear interpolation, the spatial resolution is improved from 25km to 3km. After vertical and horizontal co-scale downscaling, large-scale ensemble perturbation samples matching the resolution of the CMA-MESO regional mesoscale numerical prediction model are successfully obtained.

[0024] In one embodiment of the present invention, regional convective-scale ensemble forecast samples are obtained and input into the CMA-MESO Hybrid-3Dvar system for subsequent processing. Specifically, 15 regional ensemble members generated by the China Meteorological Administration's Regional Ensemble Forecasting System (CMA-REPS) are used as the source of regional convective-scale ensemble forecast samples containing small- and medium-scale information. The horizontal resolution is 3 km, and the vertical layer is a 51-layer model surface standard layer. This differs from the 72-layer model surface standard layer in the vertical direction of the CMA-MESO regional medium-scale numerical forecast model. Therefore, the regional convective-scale ensemble forecast samples output by CMA-REPS need to be interpolated onto the 72-layer model surface standard layer of the hybrid 3Dvar system using a vertical spline interpolation method to obtain small-scale ensemble perturbation samples that match the resolution of the CMA-MESO regional medium-scale numerical forecast model.

[0025] In one embodiment of the present invention, sequential multi-scale ensemble variational assimilation is employed in a hybrid three-dimensional variational analysis system to fuse large-scale and small-scale ensemble perturbation samples. Here, "sequential" refers to simulating the cascade process of energy and information in atmospheric motion. In atmospheric motion, large-scale circulations such as the subtropical high and monsoon circulation provide the background environment and initial perturbations for the occurrence and development of small-scale systems such as typhoons and rainstorms. Sequential assimilation first determines the large-scale background field and then assimilates small-scale observations under this background field, physically simulating the actual development process of weather systems. The large-scale analysis field provides a more physically coordinated and stable atmospheric environment for small-scale assimilation. During the cascade process, the sources of observational data and background error covariance scale are clearly distinguished. Conventional methods, such as 2D... Independent multiscale assimilation methods, represented by the DCT method, treat large-scale and small-scale as completely independent parts, assimilate them separately, and then simply superimpose them. Traditional methods implicitly assume that the error covariance between different scales is zero, that is, the error of the large-scale analysis field will not affect the small-scale and vice versa. However, in the actual atmosphere, the interaction between scales is strong and ubiquitous. The implicit assumptions of conventional methods introduce physical inconsistencies, resulting in spurious or dynamically unbalanced structures in the final analysis field after superposition. Moreover, the assimilation of traditional methods all use the same original background field. When assimilating small-scale observations, the flow-dependent error covariance includes correlation components from the large scale that may be too long, making it difficult to accurately locate and assimilate local observation information.

[0026] The sequential multi-scale ensemble variational assimilation of this invention specifically involves: using the forecast field of the global numerical weather prediction model of the meteorological center as the first background field; constructing a large-scale flow-dependent background error covariance using large-scale ensemble perturbation samples; and assimilating the large-scale observation data to obtain the first analysis field. The expression of the first analysis field is as follows:

[0027]

[0028] in, For the first analysis field, As the first background scene, For large-scale gain matrices, For large-scale observation data, For observation operators targeting large-scale observation data, For large-scale background error covariance matrix, For observation error covariance matrix of large-scale observation data, This is the transpose symbol.

[0029] Using the first analysis field as the second background field, a small-scale flow-dependent background error covariance is constructed using small-scale ensemble perturbation samples. The small-scale observation data is then assimilated to obtain the final analysis field, whose expression is as follows.

[0030]

[0031] in, For the final analysis field, This serves as both the first analytical field and the second background field. For small-scale gain matrices, For small-scale observation data, For observation operators targeting small-scale observation data, The small-scale background error covariance matrix, The small-scale observation error covariance matrix, This is the transpose symbol.

[0032] The large-scale observation data includes three-dimensional wind, temperature, and humidity fields provided by radiosonde observations; sea level pressure, wind, and humidity parameters provided by ship observations; upper-level wind and temperature fields, cloud-guided winds, GPS-derived precipitable water, GNSS radio occultation, sea surface wind speed from scatterometers, and wind profiler radar observations, among other observation data. The small-scale observation data includes radar radial wind and ground observation data, among other observation data.

[0033] After obtaining the final analysis field, it is assimilated and subjected to numerical forecast case and batch experiments. The root mean square (RMS) score of the forecast case is output and verified through analysis. In one embodiment of this invention, the CMA-MESO V6.0 system is used, with a horizontal resolution of 3 km and a vertical layer of 72 model surface standard layers. The experiment analyzes the data for 00 UTC on June 27, 2022, followed by a 24-hour forecast. The study area is a region with latitude and longitude of 31°N-43°N and 109°E-124°E.

[0034] Based on 21 large-scale ensemble samples from global ensemble forecast products and 15 small-scale ensemble samples from regional ensemble products, three-dimensional variational analysis and single-scale hybrid three-dimensional variational analysis, both using conventional techniques, were designed. Additionally, a hybrid three-dimensional variational assimilation method integrating multi-scale ensemble forecast samples, as proposed in this invention, was also designed and tested. A 3DVAR experiment assimilating all observation data served as a control experiment. The background error covariance adopted a static structure. All observation data were assimilated using a Hybrid-3DVAR system with 21 large-scale ensemble samples (21-Full Hybrid); all observation data were assimilated using a Hybrid-3DVAR system with 15 small-scale ensemble samples (15-Full Hybrid); and large / small-scale observations were sequentially assimilated using a Hybrid-3DVAR system with both large and small-scale ensemble samples (Ms-Full Hybrid), which is the core of this invention. In all Hybrid-3DVAR analyses, fully dynamic background error covariance was used for hybrid three-dimensional variational analysis. This invention first uses a Hybrid-3DVAR system with 21 large-scale ensemble samples to assimilate large-scale observations represented by radiosonde observations, ship observations, aircraft reports, cloud-guided winds, GPS inversion of precipitable water, GNSS radio occultation, scatterometer sea surface wind speed, and wind profiler radar observations. Then, using the large-scale analysis field as the background field, i.e., the first analysis field as the background field, a Hybrid-3DVAR system with 15 small-scale ensemble samples assimilates small-scale observations represented by radar radial wind and ground observation data.

[0035] Figure 2 shows the analytical bias and root mean square error (RMSE) of each group of experiments relative to radar radial wind (VR) at 00:00 on June 27, 2022. In Figure 2(a), the horizontal axis Bias Vr represents the analytical bias of radar radial wind, and the vertical axis Pressure represents the isobaric surface height. In Figure 2(b), the horizontal axis RMSE Vr represents the analytical root mean square error of radar radial wind. The black line represents the 3DVar experiment results, the red line represents the 15-Full Hybrid experiment, the blue line represents the 21-Full Hybrid experiment, and the yellow line represents the Ms-Full Hybrid experiment. Regarding the analytical bias, the Ms-Full... The bias distributions of the Hybrid experiment and the other three groups of experiments were generally similar with no significant differences. In terms of root mean square error (RMSE), the 3DVar experiment showed a significantly higher RMSE for wind field analysis than the other experiments. After introducing the full dynamic background error covariance, the RMSE for wind field analysis in the single-scale hybrid three-dimensional variational experiment was significantly reduced. Furthermore, after adopting the multi-scale hybrid three-dimensional variational experiment, the RMSE for wind field analysis was further reduced, resulting in the best analysis quality and the best agreement with radar radial wind observations. The above research shows that the hybrid three-dimensional variational assimilation method of the present invention, which integrates multi-scale ensemble forecast samples, has the most significant improvement on the analysis field, effectively reducing analysis errors and significantly improving analysis quality.

[0036] The beneficial effects of this invention are as follows: By sequentially fusing global large-scale and regional small-scale aggregate samples, a dynamic background error covariance that can simultaneously coordinate and characterize multi-scale flow-dependent error features is constructed. This enables the analysis field to more accurately depict complete information from synoptic-scale circulation background to convective-scale fine structure, solving the problem that traditional static or single-scale dynamic covariance is difficult to accurately analyze small- and medium-scale systems. By employing flow-dependent background error covariances containing both large-scale and small- and medium-scale information for different scale observation data, the problem of weak flow dependence in background error covariance is effectively solved. This allows the covariance to dynamically characterize the evolution characteristics of weather systems, without relying on model and parameter settings, and also improves physical constraint capabilities. This invention employs a sequential multi-scale ensemble variational assimilation approach, where global ensemble samples dominate large-scale observation assimilation, and regional ensemble samples dominate small-scale observation assimilation. This improves the rationality of numerical forecasts for extreme weather events such as heavy rain and typhoons, and addresses the problem of insufficient and unreasonable utilization of multi-scale observational information. Through this invention, the root mean square error of wind, humidity, and temperature fields in multi-scale hybrid three-dimensional variational experiments is significantly reduced, the ETS score for precipitation forecasts at all levels is improved, and false alarms at levels above the level of heavy rain are significantly alleviated. Typhoon paths and intensities are also closer to reality.

Claims

1. A hybrid three-dimensional variational assimilation method that integrates multi-scale ensemble forecast samples, characterized in that, Includes the following steps: Global ensemble forecast samples and regional convective-scale ensemble forecast samples from the CMA-GEPS global ensemble forecast system were acquired. Vertical and horizontal co-scaling was performed on the global ensemble forecast samples to obtain large-scale ensemble perturbation samples matching the resolution of the CMA-MESO regional medium-scale numerical weather prediction model. Vertical spline interpolation was performed on the convective-scale ensemble forecast samples to obtain small-scale ensemble perturbation samples matching the resolution of the CMA-MESO regional medium-scale numerical weather prediction model. The large-scale and small-scale ensemble perturbation samples were input into a hybrid three-dimensional variational analysis system. The forecast field from the China Meteorological Administration's global numerical weather prediction model was used as the first background field. A large-scale flow-dependent background error covariance was constructed using the large-scale ensemble perturbation samples, and the large-scale observation data were assimilated to obtain the first analysis field. The first analysis field was used as the second background field. A small-scale flow-dependent background error covariance was constructed using the small-scale ensemble perturbation samples, and the small-scale observation data were assimilated to obtain the final analysis field.

2. The hybrid three-dimensional variational assimilation method for fusing multi-scale ensemble forecast samples according to claim 1, characterized in that, The global ensemble forecast samples are derived from 21 global ensemble forecast members generated by the global ensemble forecast system and are used to characterize large-scale information; the regional convective-scale ensemble forecast samples are derived from the regional ensemble forecast system and are used to characterize small- and medium-scale information.

3. The hybrid three-dimensional variational assimilation method for fusing multi-scale ensemble forecast samples according to claim 1, characterized in that, The process of obtaining large-scale ensemble perturbation samples that match the resolution of the CMA-MESO regional ensemble numerical weather prediction model specifically includes: based on the global ensemble forecast samples of the CMA-GEPS global ensemble forecast system, vertically adjusting the number of model surface layers of the global ensemble forecast samples to the number of model surface layers in the hybrid three-dimensional variational analysis system through vertical spline interpolation, and adjusting the spatial resolution of the global ensemble forecast samples to the spatial resolution adapted to the hybrid three-dimensional variational analysis system through horizontal bilinear interpolation, thereby obtaining large-scale ensemble perturbation samples that match the resolution of the CMA-MESO regional ensemble numerical weather prediction model.

4. The hybrid three-dimensional variational assimilation method for fusing multi-scale ensemble forecast samples according to claim 1, characterized in that, The process of obtaining small-scale ensemble perturbation samples that match the resolution of the CMA-MESO mesoscale numerical weather prediction model specifically includes: based on the regional convective-scale ensemble forecast samples of the CMA-GEPS global ensemble forecast system, interpolating the regional convective-scale ensemble forecast samples to the model surface layer number of the hybrid three-dimensional variational analysis system using the vertical spline interpolation method to obtain small-scale ensemble perturbation samples that match the resolution of the CMA-MESO mesoscale numerical weather prediction model.

5. The hybrid three-dimensional variational assimilation method for fusing multi-scale ensemble forecast samples according to claim 1, characterized in that, The expression for the first analysis field is as follows: in, For the first analysis field, As the first background scene, For large-scale gain matrices, For large-scale observation data, For observation operators targeting large-scale observation data, For large-scale flow-dependent background error covariance matrix, For observation error covariance matrix of large-scale observation data, This is the transpose symbol.

6. The hybrid three-dimensional variational assimilation method for fusing multi-scale ensemble forecast samples according to claim 1, characterized in that, The expression for the final analysis field is as follows: in, For the final analysis field, This serves as both the first analytical field and the second background field. For small-scale gain matrices, For small-scale observation data, For observation operators targeting small-scale observation data, For the small-scale flow-dependent background error covariance matrix, The small-scale observation error covariance matrix, This is the transpose symbol.

Images