Meteorological large model-based set-variation hybrid assimilation system and method

Abstract

The invention belongs to the technical field of atmospheric science, and discloses a set-variational hybrid assimilation system and method based on a meteorological large model, which adopts a physical constraint mode (RANDOMMV) to perform random disturbance on an atmospheric initial field, quickly generates set members based on the strong deduction capability of the artificial intelligence meteorological large model, and improves the set-variational hybrid assimilation efficiency. The member diversity is ensured; and the sampling error is reduced. On the basis, a set-variation hybrid assimilation method is adopted, and a set-variation hybrid assimilation system is constructed by combining a dynamic set background error covariance and a static background error estimated through an NMC method, so that effective assimilation of the multi-source observation data is realized. According to the method, the physical consistency and stability of the AI ensemble forecast members can be ensured. Long-term average static error information derived from a difference field of historical forecast data and flow dependent background error information of ensemble forecast are combined; and dynamic and static combination enhances the robustness of the hybrid assimilation system.

Description

Technical Field

[0001] The present invention belongs to the field of atmospheric science technology, and in particular relates to an ensemble-variational hybrid assimilation system and method based on a large meteorological model. Background Art

[0002] The accuracy of weather forecasts depends on the accuracy of the forecast model on the one hand, and on the quality of the initial values ​​on the other. How to obtain the optimal initial field has always been a key and difficult point in research and business. Operational numerical forecast centers often use data assimilation technology to combine model prior information and observation information to find the optimal estimate of state variables.

[0003] Currently, common assimilation methods can be divided into two categories: variational assimilation and ensemble assimilation. The variational assimilation method is convenient for assimilating multiple observational data and has low computational cost, but the background error covariance it uses is static and isotropic, and cannot accurately reflect the actual model error. Ensemble assimilation, taking the ensemble Kalman filter method as an example, uses multiple ensemble members to reflect the uncertainty of the initial field. However, the ensemble method is greatly affected by sampling errors; especially for convective systems with rapid temporal and spatial evolution, large model errors will lead to large deviations in the ensemble error covariance, limiting the assimilation effect of observational data.

[0004] The ensemble-variational hybrid assimilation method combines the advantages of ensemble assimilation and variational assimilation methods. It not only retains the advantages of variable correlation and "flow dependence" in the EnKF method, but also uses the static background error covariance to better deal with sampling error and model error problems. It is widely used in numerical weather forecasting of global and regional models.

[0005] The current problems and defects of ensemble-variation hybrid assimilation technology are: it requires a large number of ensemble members to more accurately capture the flow-dependent characteristics of the background error covariance and characterize the uncertainty of the initial value in the atmospheric system, reduce random errors, and improve the accuracy of the background error covariance, thereby improving the assimilation effect. However, more ensemble members also mean greater computing resources and time costs, affecting the efficiency of operational forecasts.

[0006] To address this problem, some current solutions include: using localization technology (Houtekamer and Mitchell, 2001), ensemble resampling technology (Anderson et al. 2007), and time-lagged ensemble technology (Wang et al. 2017) to alleviate the sampling error problem caused by insufficient ensemble members; using parallel computing and high-performance computing technology (Michalakes and Vachharajani, 2008) to speed up the generation of ensemble members. Although these methods have optimized computing resources and efficiency to a certain extent, with the improvement of forecast accuracy and resolution requirements, the increase in ensemble members is still inevitable, and the computing bottleneck problem still exists.

[0007] Therefore, the technical problem that the existing technology urgently needs to solve is how to quickly obtain a large number of reliable ensemble members in the hybrid assimilation method to more accurately characterize the flow dependence and uncertainty characteristics of the background error covariance, thereby improving the assimilation effect.

[0008] In the industrial application of ensemble-variation hybrid assimilation technology, computing resources and time costs bring about the following major technical problems, which limit the actual application efficiency and promotion of this technology:

[0009] 1. Excessive consumption of computing resources: Hybrid assimilation methods rely on a large number of ensemble members to accurately estimate the background error covariance and capture the uncertainty in the atmospheric system. However, the process of generating and updating a large number of ensemble members requires huge computing resources, especially in high-resolution forecasts and fast-updating business applications. This high computing demand significantly increases the operating cost, which is difficult for many companies or research institutions to bear.

[0010] 2. Insufficient real-time and timeliness: In operational forecasting, real-time and timeliness are crucial. However, the existing hybrid assimilation method requires a lot of computing time, resulting in delays in data assimilation and forecast output, which reduces the response speed of the forecast system. This is particularly challenging in fields that require rapid response, such as weather forecasting, climate simulation, and aerospace, and directly affects the effectiveness of operational forecasts.

[0011] 3. Computational bottlenecks affect system scalability: As the accuracy and resolution requirements of forecast models increase, the number of ensemble members also increases. Even with the use of optimization strategies such as localization, ensemble resampling, and parallel computing, computational bottlenecks still exist and become more serious as the complexity of the model increases. Existing technologies are difficult to balance the relationship between forecast accuracy, resolution, and computational cost, which limits the scalability of this technology in high-precision forecast systems.

[0012] 4. High hardware requirements and high costs: Solving computing bottlenecks requires high-performance computing clusters or advanced parallel computing technologies, which have high hardware requirements, especially when it comes to large-scale industrial applications, which means that a large number of servers and computing resources are required, increasing initial investment and maintenance costs. This is a significant obstacle for organizations with limited budgets.

[0013] 5. The complexity of the method increases the technical threshold: Although the techniques such as ensemble resampling and time-lag ensemble are effective, they increase the complexity of the system and require high professional and technical personnel. This high threshold makes it more difficult to apply hybrid assimilation technology in enterprises, hindering the widespread promotion of this technology. Summary of the invention

[0014] In view of the problems existing in the prior art, the present invention provides an ensemble-variational hybrid assimilation system and method based on a large meteorological model.

[0015] The present invention is implemented in such a way that an ensemble-variation hybrid assimilation method based on a large meteorological model comprises:

[0016] S1, initial value perturbation of global model reanalysis data;

[0017] S2, deduce the initial value after disturbance using a large meteorological model to generate an ensemble forecast;

[0018] S3, calculate the ensemble disturbance and count the ensemble background error covariance;

[0019] S4, prepare the model background field, observation data, static background error covariance, and implement assimilation analysis based on the ensemble-variation hybrid assimilation framework;

[0020] S5, weather forecasting based on physical models or machine learning models.

[0021] Further, S1 specifically includes: preparing global model analysis or forecast fields, such as ERA5 released by the European Center for Medium-Range Weather Forecasts ECMWF, GFS jointly released by the National Center for Environmental Prediction NCEP and the National Center for Atmospheric Research NCAR, JRA-55 released by the Japan Meteorological Agency, and CRA reanalysis data released by the National Meteorological Information Center of the China Meteorological Administration. By perturbing the initial value, a large number of initial ensemble members can be generated.

[0022] The specific initial value perturbation scheme is as follows: for a set of ensemble forecasts derived from an AI meteorological model, the ensemble mean is regarded as the true value, and the ensemble member minus the ensemble mean is regarded as the ensemble forecast error, thereby estimating the ensemble error covariance and obtaining the co-correlation between different model variables. These ensemble forecasts can be the forecast results of the AI ​​meteorological model at the same time with different forecast time periods driven by GEFS. The initial field is perturbed according to the ensemble error covariance matrix calculated above, and the random perturbation on each grid point will affect the surrounding grids and other variables through the rules in the ensemble error covariance matrix, thereby ensuring the physical consistency of the ensemble members after the disturbance. This method is similar to the RANDOMCV method, and both imply physical balance constraints. The difference is that the RANDOMCV method is based on the spatial background error covariance of the control variable (CV, control variable) for perturbation, while the method proposed in the present invention is based on the forecast error covariance of the model variable (MV, model variable) space for perturbation. Therefore, the invention will refer to the perturbation method as the RAMDOMMV method in the following.

[0023] Furthermore, S2 specifically includes: deducing the initial value after the disturbance to generate a future ensemble weather forecast; the deduction model is an AI meteorological large model, including Huawei's Pangu model, Google DeepMind's WeatherBench, NVIDIA's FourCastNet, and DeepMind's GraphCast.

[0024] Further, the S3 specifically includes: obtaining a background field error covariance having flow dependence and multivariate variable correlation characteristics by means of an extended control variable method;

[0025] Furthermore, S4 specifically includes: (1) preparing the model background field, which may be a global model reanalysis; (2) preparing observation data, including conventional observation data from GTS, such as sounding, buoy, aircraft report, ship report, ground station data, as well as meteorological satellite, radar, etc.; (3) preparing the estimated static background error covariance file; (4) based on the background error covariance provided by S3 to reflect the flow-dependent information of the actual weather, the hybrid assimilation analysis increment is obtained by minimizing the cost function, so that the observation information is better coordinated and transmitted between the model space and the multivariate under the action of the collective flow-dependent background error covariance, that is,

[0026]

[0027] Among them, the total assimilation analysis increment is, δx1 is the analysis increment associated with the static background error covariance, is the kth set disturbance, K is the number of set members, α is the extended control variable; A is the diagonal matrix of the constrained set extended control variable α, which localizes the set error covariance; β1 and β2 are the weights of the static background error covariance (B) and the set background covariance, respectively, satisfying The third term on the right side of the equation is the observation-related term.

[0028] The background field superimposed analysis increment is the analysis field.

[0029] Furthermore, S5 specifically includes: using the analysis field obtained in S4 as the initial field to perform weather forecasting, which may be performed using a traditional numerical weather forecasting model or using an AI large model.

[0030] Furthermore, the calculation of the static background error covariance matrix in S4 may be performed by using historical forecast samples (forecast fields with different forecast time periods but at the same time) as background error samples, or by using ensemble forecast samples as background error samples, and estimating through the GEN_BE software.

[0031] Another object of the present invention is to provide an ensemble-variation hybrid assimilation system based on a large meteorological model, comprising:

[0032] Initial value perturbation module, used to make initial value perturbations to global model reanalysis data;

[0033] The generation module is used to deduce the initial value after the disturbance and generate the ensemble forecast;

[0034] Statistics module, used to calculate the ensemble disturbance and the statistical ensemble background error covariance;

[0035] Assimilation analysis module, used to prepare model background fields, observation data, static background error covariance, and implement assimilation analysis based on the ensemble-variation hybrid assimilation framework;

[0036] Weather forecast module, used for weather forecast based on physical model or machine learning model.

[0037] Another object of the present invention is to provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the ensemble-variational hybrid assimilation method based on the large meteorological model.

[0038] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to perform the steps of the ensemble-variational hybrid assimilation method based on a large meteorological model.

[0039] Another object of the present invention is to provide an information data processing terminal, which is used to implement the ensemble-variational hybrid assimilation system based on the large meteorological model.

[0040] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0041] First, the present invention provides an ensemble-variational hybrid assimilation system and method based on an AI meteorological large model. By perturbing the initial field of the meteorological large model and generating an ensemble forecast through deduction, the flow-dependent background error covariance calculated from the ensemble forecast samples deduced from the meteorological large model is combined with the modeled static background error covariance of the three-dimensional variation, so that the ensemble-variational hybrid assimilation can be completed more efficiently to meet the needs of real-time forecasting.

[0042] (1) The ensembles in previous assimilation schemes all come from numerical weather forecasts. This patent generates ensemble members through an AI large model, which has a fast deduction speed and can make up for the lack of efficiency in traditional ensemble forecast calculations. It can generate larger ensemble samples within the allowed lag time, thereby meeting the real-time assimilation forecast requirements of the business and improving the accuracy of assimilation.

[0043] (2) This patent combines the physical constraint perturbation method with the large meteorological model deduction to ensure the physical consistency and stability of AI ensemble forecast members.

[0044] (3) The long-term average static background error information is derived based on the difference field of historical forecasts; the AI ​​ensemble samples can provide dynamic flow-dependent ensemble error covariance; the combination of dynamic and static enhances the robustness of the hybrid assimilation system.

[0045] Second, the expected benefits and commercial value of the technical solution of the present invention after transformation are:

[0046] (a) Improve forecast accuracy and efficiency: By combining AI meteorological big models with ensemble-variational assimilation systems,

[0047] Improve the accuracy of extreme weather forecasts, optimize resource consumption, and reduce computing costs.

[0048] (b) Service industry applications: High-precision forecasts can serve industries such as agriculture, transportation, and energy that have demand for accurate weather forecasts, thereby increasing market value and competitiveness.

[0049] (c) Promote technological innovation: Promote the application and development of AI in the field of meteorology, establish technological barriers, form world-leading intellectual property rights, and increase industrial added value.

[0050] (d) Supporting government disaster prevention and mitigation: Providing reliable forecast data to governments and public institutions, optimizing decision-making support, and bringing social and economic benefits.

[0051] The technical solution of the present invention fills the technical gap in the industry at home and abroad:

[0052] (a) Adding perturbations based on physical constraints obtained from historical samples to the initial values ​​of the AI ​​large model;

[0053] (b) Apply the ensemble forecast of the AI ​​meteorological model to ensemble-variational hybrid assimilation.

[0054] The technical solution of the present invention solves a technical problem that people have been eager to solve but have never succeeded in solving: the nonlinearity of the large meteorological model is very high, and the random disturbance in the initial field will cause the forecast field to be inconsistent with the laws of physics. In response to this technical problem, this patent will add physical constraints to the ensemble forecast of the large meteorological model to more accurately characterize the uncertainty of the initial value and the flow-dependent characteristics of the background error covariance in the atmospheric system, thereby establishing an ensemble-variation hybrid assimilation system based on the large meteorological model to achieve rapid update of the assimilation forecast.

[0055] Third, the technical solution of the present invention solves the following key problems in the prior art in industrial applications through an ensemble-variational hybrid assimilation method based on a large meteorological model, and has achieved significant technical progress:

[0056] 1. Solve the problem of rapid generation of ensemble samples in hybrid assimilation methods

[0057] Hybrid assimilation methods require a large number of ensemble samples to reasonably characterize the ensemble background error covariance. In existing hybrid assimilation methods, ensemble forecasts all come from numerical forecast models, which have high computational costs and affect operational forecast efficiency.

[0058] The technical progress of the present invention is as follows: by generating ensemble members through the AI ​​large model, the deduction speed is fast, which can make up for the lack of efficiency in traditional ensemble forecast calculations, generate larger ensemble samples within the allowed lag time, greatly improve the assimilation efficiency, and meet the time requirements of real-time assimilation forecasting for the business.

[0059] 2. Solve the problem of lack of physical constraints in ensemble forecasts of large AI models

[0060] In existing data-driven AI large models, the initial perturbations of the ensemble lack physical equilibrium constraints. The perturbation methods that lack physical constraints easily generate some ensemble members that do not conform to the principles of atmospheric dynamics or thermodynamics, affecting the accuracy and reliability of the ensemble forecast and making it difficult to meet the needs of data assimilation.

[0061] The technical progress of the present invention is as follows: by adopting the forecast error covariance estimation of the AI ​​ensemble forecast model variable space and combining the constraints of the physical model to perform initial value perturbation, the disturbed ensemble members have higher physical consistency, thereby improving the accuracy of the ensemble error covariance estimation.

[0062] 3. Significant technological progress in industrial applications

[0063] Improve forecast accuracy: Through the ensemble-variation hybrid assimilation method of the present invention, the initial value quality and forecast uncertainty estimation of the forecast system are significantly improved, and the overall forecast accuracy is improved.

[0064] Optimizing the meteorological assimilation system: This invention innovatively proposes a perturbation scheme based on pattern variables, which has stronger adaptability and higher computational efficiency in large-scale data assimilation and ensemble forecasting, and meets the real-time application requirements of modern meteorological forecasting.

[0065] Wide application: The technical solution of the present invention is not only applicable to conventional weather forecasts, but can also be applied to extreme weather events, climate change monitoring and meteorological risk prediction in related industries, and has broad application prospects.

[0066] The present invention solves important problems in the prior art in the field of weather forecasting, achieves significant technical progress, and has high industrial application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a flow chart of an ensemble-variational hybrid assimilation system and method based on a large meteorological model provided by an embodiment of the present invention.

[0068] Figure 2 This is the technical route of the ensemble-variational hybrid assimilation system based on the large meteorological model provided by the embodiment of the present invention.

[0069] Figure 3 It is a conceptual diagram of the ensemble-variational hybrid assimilation method provided in an embodiment of the present invention.

[0070] Figure 4 It is a structural block diagram of an ensemble-variational hybrid assimilation system based on a large meteorological model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0071] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0072] like Figure 1 As shown, an ensemble-variational hybrid assimilation method based on a large meteorological model provided by an embodiment of the present invention includes the following steps:

[0073] S1, initial value perturbation of global model reanalysis data;

[0074] S2, deduce the initial value after disturbance and generate ensemble forecast;

[0075] S3, calculate the ensemble disturbance and count the ensemble background error covariance;

[0076] S4, prepare the model background field, observation data, static background error covariance, and implement assimilation analysis based on the ensemble-variation hybrid assimilation framework;

[0077] S5, weather forecasting based on physical models or machine learning models.

[0078] The present invention is based on the ensemble-variation hybrid assimilation method of the meteorological large model, and improves the accuracy and stability of meteorological forecasts by fusing ensemble forecasts and variational assimilation techniques in multiple steps. First, in step S1, the global model reanalysis data is perturbed with initial values ​​to generate multiple initial value schemes. The process of initial value perturbation is to simulate different initial conditions by applying small random perturbations to the initial state of the atmosphere, thereby capturing the uncertainty of the atmospheric state. This step provides the initial conditions for subsequent ensemble forecasts and lays the foundation for predictions under different scenarios.

[0079] Next, in step S2, the initial value after disturbance is input into the meteorological model for deduction to generate a series of different ensemble forecasts. Each ensemble member represents the deduction result of a disturbance initial state, reflecting the weather evolution process under different initial conditions. Through the deduction of multiple ensemble members, the change of weather state under different paths can be obtained, forming a probabilistic prediction of future weather, and providing necessary data support for the calculation of ensemble background error covariance.

[0080] In step S3, the system performs statistical analysis on the results of the ensemble forecast, calculates the ensemble perturbation and calculates the ensemble background error covariance. The background error covariance reflects the difference in atmospheric state between different ensemble members and can characterize the error propagation characteristics between different regions and different variables in the atmosphere. Through the calculation of this error covariance, the system can better capture the uncertainty information in the atmospheric background field and provide dynamic background error covariance support for the assimilation analysis of hybrid assimilation.

[0081] Step S4 is the core of the method, which realizes assimilation analysis through the ensemble-variational hybrid assimilation framework. In this step, the system prepares the model background field, observation data, and static background error covariance, and introduces the dynamic background error covariance obtained based on the ensemble forecast into the variational assimilation process. Through the hybrid assimilation framework, the system can effectively fuse the observation information and model background field, correct the errors in the ensemble forecast, and obtain a more accurate analysis of the atmospheric state. The hybrid assimilation process effectively reduces the limitations of the traditional single assimilation method by combining dynamic and static covariances.

[0082] In the final step S5, the system uses physical models or machine learning models to forecast future weather based on the results of assimilation analysis. Physical models are deduced based on the laws of atmospheric dynamics and thermodynamics, while machine learning models use historical data and model outputs for fitting and optimization to enhance the ability to capture nonlinear and complex weather conditions. This combination makes the forecast results more accurate, capturing the overall trend of the weather as well as identifying subtle local changes.

[0083] The ensemble-variation hybrid assimilation method of the present invention achieves high-precision meteorological forecasting through the introduction of multi-step assimilation and dynamic background errors.

[0084] As attached Figure 2 The ensemble-variational hybrid assimilation system based on the meteorological large model in the present invention firstly uses a physically constrained method (RANDOMMV) to randomly perturb the initial atmospheric field, and based on the powerful deduction ability of the artificial intelligence meteorological large model, quickly generates ensemble members to ensure member diversity and reduce sampling errors. On this basis, the ensemble-variational hybrid assimilation method is adopted, combined with the dynamic ensemble background error covariance and the static background error estimated by the NMC method, to construct an ensemble-variational hybrid assimilation system to achieve effective assimilation of multi-source observation data.

[0085] Specific steps:

[0086] S1 provided in the embodiment of the present invention specifically includes: preparing global model analysis fields, such as ERA5 released by the European Center for Medium-Range Weather Forecasts ECMWF, GFS jointly released by the National Center for Environmental Prediction (NCEP) and the National Center for Atmospheric Research (NCAR) of the United States, JRA-55 released by the Japan Meteorological Agency, CRA reanalysis released by the National Meteorological Information Center of the China Meteorological Administration, etc. This patent will be explained using ERA5 as an example. Perform initial value perturbations on the reanalysis data.

[0087] The specific initial value perturbation scheme is as follows: for a set of ensemble forecasts derived from an AI meteorological model, the ensemble mean is regarded as the true value, and the ensemble member minus the ensemble mean is regarded as the ensemble forecast error, thereby estimating the ensemble error covariance and obtaining the co-correlation between different model variables. These ensemble forecasts can be the forecast results of the AI ​​meteorological model at the same time with different forecast time periods driven by GEFS. The initial field is perturbed according to the ensemble error covariance matrix calculated above. The random perturbation on each grid point will affect the surrounding grids and other variables through the rules in the ensemble error covariance matrix, thereby ensuring the physical consistency of the ensemble members after the disturbance. This method is similar to the RANDOMCV method, both of which imply physical balance constraints. The difference is that the RANDOMCV method is based on the spatial background error covariance of the control variable (CV, control variable) for perturbation, while the method proposed in the present invention is based on the prediction error covariance of the model variable (MV, model variable) space for perturbation. Therefore, the perturbation method is named the RAMDOMMV method.

[0088] S2 provided in the embodiment of the present invention specifically includes: deducing the initial value after the disturbance to generate the future ensemble weather forecast. The deduction model is an AI meteorological large model, such as Huawei's Pangu model, Google DeepMind's WeatherBench, NVIDIA's FourCastNet, DeepMind's GraphCast, etc. Note that for different large models, the required input grids are not the same. Therefore, the reanalysis data needs to be grid interpolated. Taking the Pangu model as an example, the initial ensemble members after the disturbance are sequentially used as the input of the Pangu model, and the ensemble forecast for the future time with the same number of members will be output.

[0089] S3 provided in the embodiment of the present invention specifically includes: obtaining the background field error covariance having the characteristics of flow dependence and multivariate correlation by means of an extended control variable method.

[0090] S4 provided in the embodiment of the present invention specifically includes: (1) preparing a model background field, which can be a global model analysis field; (2) preparing observation data, which can come from conventional observation data of GTS, such as sounding, buoy, aircraft report, ship report, ground station data, as well as meteorological satellites, radars, etc.; (3) preparing a static background error covariance, which can be generated by the NMC method, which has been introduced in the introduction of the NMC method S1; (4) based on the background error covariance provided by S3 to reflect the flow-dependent information of the actual weather, the hybrid assimilation analysis increment is obtained by minimizing the cost function, so that the observation information is better coordinated and transmitted between the model space and the multivariate under the action of the collective flow-dependent background error covariance, that is,

[0091]

[0092] Among them, the total assimilation analysis increment is, δx1 is the analysis increment associated with the static background error covariance, is the kth set disturbance, K is the number of set members, α is the extended control variable; A is the diagonal matrix of the constrained set extended control variable α, which localizes the set error covariance; β1 and β2 are the weights of the static background error covariance (B) and the set background covariance, respectively, satisfying The third term on the right side of the equation is the observation-related term.

[0093] The background field superimposed analysis increment is the analysis field.

[0094] S5 provided in the embodiment of the present invention specifically includes: using the analysis field obtained in S4 as the initial field to perform weather forecasting, which can be performed using a traditional numerical weather forecasting model or using an AI large model.

[0095] In this embodiment, the background field in step 4 may be a model forecast field predicted by S5 at the previous moment, or may be a numerical model analysis / forecast field from other sources at the same moment.

[0096] The calculation of the static background error covariance matrix in step 4 can be estimated by using the historical forecast samples (forecast fields with different forecast time periods but at the same time) as the background error samples, or by using the ensemble forecast samples as the background error samples through the GEN_BE software.

[0097] like Figure 4 As shown, an ensemble-variation hybrid assimilation system based on a large meteorological model provided by an embodiment of the present invention includes:

[0098] Initial value perturbation module, used to make initial value perturbations to global model reanalysis data;

[0099] The generation module is used to deduce the initial value after the disturbance and generate the ensemble forecast;

[0100] Statistics module, used to calculate the ensemble disturbance and the statistical ensemble background error covariance;

[0101] Assimilation analysis module, used to prepare model background fields, observation data, static background error covariance, and implement assimilation analysis based on the ensemble-variation hybrid assimilation framework;

[0102] Weather forecast module, used for weather forecast based on physical model or machine learning model.

[0103] Another object of the present invention is to provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the ensemble-variational hybrid assimilation method based on the large meteorological model.

[0104] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to perform the steps of the ensemble-variational hybrid assimilation method based on a large meteorological model.

[0105] Another object of the present invention is to provide an information data processing terminal, which is used to implement the ensemble-variational hybrid assimilation system based on the large meteorological model.

[0106] 1. Typhoon path and intensity forecast: This invention can be used to improve the forecast of typhoon path and intensity. By combining the high-precision ensemble forecast members generated by the AI ​​meteorological large model with the variational assimilation method, multi-source observation data can be quickly assimilated. This not only improves the accuracy of the typhoon path, but also more accurately predicts the changes in typhoon intensity, providing accurate path and intensity information for disaster prevention and mitigation departments, and supporting timely emergency response.

[0107] 2. Urban short-term heavy rainfall warning: In view of the rapid development of urban rainstorms, this invention can improve the early warning capability of short-term heavy rainfall events. Using the ensemble forecast quickly generated by AI, the system can quickly assimilate high-resolution observation data, accurately forecast the amount and distribution of precipitation in the next few hours, and provide real-time early warning data for urban drainage systems and traffic management departments, thereby effectively preventing waterlogging and reducing losses caused by heavy rainfall.

[0108] These two examples show that the ensemble-variational hybrid assimilation method based on large meteorological models is not only important in meteorological forecasting, but also plays a key role in other fields such as urban management and transportation. By adding physical constraints to the AI ​​large model ensemble forecast and combining advanced ensemble-variational hybrid assimilation, this technology can provide more accurate forecast support and help make research and actual operation in related fields more efficient and safer.

[0109] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0110] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. An ensemble-variational hybrid assimilation method based on a large meteorological model, characterized in that: The following steps are involved: S1, initial value perturbation of global model reanalysis data; S2, deduce the initial value after disturbance and generate ensemble forecast; S3, calculate the ensemble disturbance and count the ensemble background error covariance; S4, prepare the model background field, observation data, static background error covariance, and implement assimilation analysis based on the ensemble-variation hybrid assimilation framework; S5, weather forecasting based on physical models or machine learning models.

2. The ensemble-variational hybrid assimilation method based on a large meteorological model as claimed in claim 1, characterized in that: Said S1 further comprises: Prepare global model analysis or forecast fields, including ERA5 released by the European Centre for Medium-Range Weather Forecasts (ECMWF), GFS jointly released by the National Center for Environmental Prediction (NCEP) and the National Center for Atmospheric Research (NCAR) of the United States, JRA-55 released by the Japan Meteorological Agency, and CRA reanalysis data released by the National Meteorological Information Center of the China Meteorological Administration; The reanalysis data are perturbed with initial values ​​to generate multiple initial ensemble members.

3. The ensemble-variational hybrid assimilation method based on a large meteorological model as claimed in claim 1, characterized in that: The initial value perturbation specifically adopts the RANDOMMV method, Specifically, for a set of ensemble forecasts derived from AI meteorological models, the ensemble mean is regarded as the true value, and the ensemble member minus the ensemble mean is regarded as the ensemble forecast error, so as to estimate the ensemble error covariance and obtain the co-correlation between different model variables; these ensemble forecasts are the forecast results of the AI ​​meteorological model at the same time with different forecast time periods driven by GEFS; the initial field is perturbed according to the ensemble error covariance matrix of the model variable space, and the random disturbance at each grid point will affect the surrounding grids and other variables through the ensemble error covariance matrix, which implies a physical balance constraint.

4. The ensemble-variational hybrid assimilation method based on a large meteorological model as claimed in claim 1, characterized in that: The S2 specifically includes: deducing the initial value after the disturbance to generate the future ensemble weather forecast; the deduction model is an AI meteorological large model, including Huawei's Pangu model, Google DeepMind's WeatherBench, NVIDIA's FourCastNet, and DeepMind's GraphCast.

5. The ensemble-variational hybrid assimilation method based on a large meteorological model as claimed in claim 1, characterized in that: The S3 specifically includes: obtaining the background field error covariance with flow dependence and multivariate correlation characteristics by means of an extended control variable method.

6. The ensemble-variational hybrid assimilation method based on a large meteorological model as claimed in claim 1, characterized in that: The S4 specifically includes: (1) preparing model background fields, global model analysis fields or forecast fields; (2) preparing observation data, which are conventional observation data from GTS; (3) preparing static background error covariance; (4) based on the background error covariance provided by the ensemble perturbation, reflecting the flow-dependent information of the actual weather, and obtaining the hybrid assimilation analysis increment by minimizing the cost function, so that the observation information can be better coordinated and transmitted in the model space and multivariate reduction under the action of the ensemble flow-dependent background error covariance, that is, Among them, the total assimilation analysis increment is, δx1 is the analysis increment related to the static background error covariance, x e k is the kth set disturbance, K is the number of set members, α is the extended control variable; A is the diagonal matrix of the constrained set extended control variable α, which localizes the set error covariance; β1 and β2 are the weights of the static background error covariance (B) and the set background covariance, respectively, satisfying The third term on the right side of the equation is the observation-related term. The background field superimposed analysis increment is the analysis field.

7. The ensemble-variational hybrid assimilation method based on a large meteorological model as claimed in claim 1, characterized in that: The S5 specifically includes: using the analysis field obtained in S4 as the initial field to perform weather forecasting, using a traditional numerical weather forecasting model to perform forecasting, or using an AI large model to perform forecasting.

8. An ensemble-variational hybrid assimilation system based on a meteorological large model for implementing the ensemble-variational hybrid assimilation method based on a meteorological large model as described in any one of claims 1 to 7, characterized in that: The ensemble-variation hybrid assimilation system based on the large meteorological model includes: Initial value perturbation module, used to make initial value perturbations to global model reanalysis data; The generation module is used to deduce the initial value after the disturbance and generate the ensemble forecast; Statistics module, used to calculate the ensemble disturbance and the statistical ensemble background error covariance; Assimilation analysis module, used to prepare model background fields, observation data, static background error covariance, and implement assimilation analysis based on the ensemble-variation hybrid assimilation framework; Weather forecast module, used for weather forecast based on physical model or machine learning model.

9. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the ensemble-variational hybrid assimilation method based on a large meteorological model as described in any one of claims 1 to 7.

10. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the ensemble-variation hybrid assimilation system based on the large meteorological model as described in claim 8.

Images