Flow-dependent background field error covariance dynamic adjustment method based on machine learning

Abstract

The invention discloses a flow-dependent background field error covariance dynamic adjustment method based on machine learning, which is characterized in that background field key variables are mapped to set dispersion through a machine learning model, so that background field uncertainty is quantified, and original static error covariance is dynamically adjusted on the basis. The method can flexibly reflect the flow dependence error characteristics under different weather background conditions, and can effectively capture high-dimensional, nonlinear and anisotropic error distribution characteristics even under the condition that ensemble forecasting is not needed. The method has important application value in the aspects of background field optimization of an assimilation system, dynamic adjustment of a background field error covariance matrix, flow dependence error analysis and the like, and is suitable for the fields of numerical weather forecast and the like.

Description

Technical Field

[0001] The present invention relates to the fields of atmospheric science and numerical weather forecasting, and in particular to a method for dynamically adjusting flow-dependent background field error covariance based on machine learning. Background Art

[0002] For numerical weather forecasting, the performance of the data assimilation system directly affects the accuracy and reliability of the forecast. Data assimilation is to generate an initial field that is closer to the actual state by combining the observed data with the background field information of the numerical model. Therefore, the background error covariance (BEC) plays a vital role in the data assimilation system, determining how the observed data corrects the background field.

[0003] However, in the variational framework, such as three-dimensional variational assimilation (3D-Var) and four-dimensional variational assimilation (4D-Var), most of them use static error covariance matrices. These BEC matrices remain unchanged throughout the assimilation cycle and fail to fully consider the flow-dependent characteristics of the atmospheric state. Static error covariance often cannot accurately capture the rapidly changing dynamics and physical processes in the atmosphere, limiting the accuracy of numerical weather forecasting.

[0004] Methods such as Ensemble Kalman Filter (EnKF) can introduce flow dependence to a certain extent by using ensemble forecasts to calculate the background error covariance. However, due to the limited number of ensemble forecast members, the estimation of the error covariance is uncertain, and problems such as over-filtering or noise amplification may occur. In addition, the estimation of the error covariance by ensemble forecasts is often subject to the limitation of computing resources, and cannot provide a sufficiently accurate covariance matrix for high-dimensional and complex numerical forecast systems. Summary of the invention

[0005] The present invention aims to solve one of the technical problems existing in the related art at least to a certain extent.

[0006] An object of the present invention is to provide a method for dynamic adjustment of flow-dependent background field error covariance based on machine learning, which estimates the background field uncertainty through a machine learning method and dynamically adjusts the spatial distribution of the error covariance, thereby generating a flow-dependent analysis and improving the accuracy of the data assimilation system.

[0007] In order to achieve the above-mentioned object, the present invention provides a method for dynamically adjusting the flow-dependent background field error covariance based on machine learning, comprising the following steps:

[0008] S100, using an ensemble perturbation method to perturb the background field of the control test, generating multiple initial members, and obtaining the forecast field of the ensemble members through numerical model integration;

[0009] S200, calculating the set discreteness of the target time based on the forecast field, and quantifying the uncertainty of the background field;

[0010] S300, extracting meteorological elements from the background field of the control test as input features, establishing a nonlinear mapping relationship between the input features and the set discreteness through a trained machine learning model, and predicting the uncertainty of the background field;

[0011] S400, normalizing the uncertainty of the prediction;

[0012] S500, dynamically adjusting the static background field error covariance matrix based on the normalized uncertainty value to generate a flow-dependent error covariance matrix;

[0013] S600, inputting the adjusted background field error covariance matrix, the model background field and the observation data into the assimilation system to solve the analysis field;

[0014] S700, using the analysis field to perform deterministic forecasting, and inputting the forecasting result into the assimilation cycle as the background field of the next cycle, to perform continuous real-time optimization and adjustment of the background error covariance.

[0015] A further preferred technical solution of the present invention is that the machine learning model trained in step S300 is a machine learning model of background field uncertainty constructed based on LightGBM.

[0016] The goal of the model is to extract meteorological elements from the background field of the control experiment as input features. Predict the corresponding set dispersion , expressed as:

[0017] ;

[0018] in, For input features To set discreteness The mapping relationship;

[0019] And the model is optimized by minimizing the loss function.

[0020] As a preference, in training the machine learning model of background field uncertainty based on LightGBM, a layer-by-layer training mode is adopted, that is, the discreteness of each layer is learned using the input features of the layer; for each layer The training of the model aims to minimize the loss function, which is expressed as:

[0021] ;

[0022] in, and Respectively Layer The predicted value and true value of samples, is the total number of samples.

[0023] Preferably, the collective perturbation method in step S100 is a perturbation scheme generated by singular vector perturbation or collective Kalman filtering.

[0024] Preferably, in step S200, the set discreteness at the target time is calculated based on the forecast field, specifically, the uncertainty of the control variable is represented by the set discreteness as the target output of the machine learning model, and the control variable includes the stream function , potential function ,temperature , surface pressure and relative humidity , and calculate the standard deviation of each control variable to represent the set dispersion.

[0025] Preferably, in step S400, the uncertainty of the prediction is normalized to limit the uncertainty of the prediction to the range of [0,1], and the formula is:

[0026] ;

[0027] in, is the normalized uncertainty, is the original uncertainty, is the maximum value of uncertainty, is the minimum value of uncertainty, subscript represents different control variables.

[0028] Preferably, the meteorological elements extracted from the background field of the control test in step S300 include at least wind speed, temperature, air pressure, humidity and hydrometeoric variables. , expressed as:

[0029] ;

[0030] in, , is the wind speed component, is the temperature, is the air pressure, For humidity, , , , and They are cloud water, rain water, ice crystals, snow and graupel among the hydrometeoric variables.

[0031] Preferably, the specific method of step S500 is:

[0032] The normalized uncertainty As a weight index, it is introduced into the calculation of the background field error covariance matrix, where Each point in the three-dimensional space has an independent value, so as to dynamically adjust the spatial distribution of the covariance to make it flow-dependent. The adjustment formula is expressed as:

[0033] ;

[0034] in, is the adjusted background field error covariance matrix, is the original background field error covariance matrix, the inner product To dynamically adjust the weight matrix;

[0035] Dynamically adjust the weight matrix The calculation formula is:

[0036] ;

[0037] in, To adjust the parameters, control With uncertainty That is, when hour, ,wrong Make adjustments; when As the uncertainty increases, increases exponentially, thus flexibly amplifying the background error covariance. The size of the index can flexibly control the speed of exponential growth. The larger the uncertainty, the greater the Significantly larger; Smaller will make As uncertainty increases more gently.

[0038] Preferably, the assimilation system in step S600 is one of a three-dimensional variational assimilation system, a four-dimensional variational assimilation system or a hybrid assimilation system;

[0039] The assimilation system uses variational assimilation or hybrid assimilation methods to fuse the adjusted background field error covariance matrix, model background field and observation data.

[0040] On the other hand, the present invention provides a non-transitory computer-readable storage medium having computer instructions stored thereon, which enable a computer to execute the above-mentioned method for dynamic adjustment of flow-dependent background field error covariance based on machine learning.

[0041] Yet another aspect of the present invention provides an electronic device, comprising: a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus, and the processor calls the logic instructions in the memory to execute the above-mentioned machine learning-based flow-dependent background field error covariance dynamic adjustment method.

[0042] On the other hand, the present invention provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer executes the above-mentioned machine learning-based flow-dependent background field error covariance dynamic adjustment method.

[0043] Beneficial effects: The present invention is a method for dynamically adjusting the flow-dependent background field error covariance based on machine learning. By mapping the key variables of the background field (such as wind speed, temperature, air pressure, humidity and hydrometeor variables, etc.) to the ensemble discreteness through a machine learning model, the uncertainty of the background field is quantified, and on this basis, the original static error covariance is dynamically adjusted. The present invention can flexibly reflect the flow-dependent error characteristics under different weather background conditions, and can effectively capture the high-dimensional, nonlinear and anisotropic error distribution characteristics even without the need for ensemble forecasting. The present invention has important application value in the background field optimization of the assimilation system, the dynamic adjustment of the error covariance matrix, and the flow-dependent error analysis, and is suitable for fields such as numerical weather forecasting.

[0044] The present invention introduces machine learning technology, uses the ensemble forecast discreteness to dynamically estimate the error of the background field, and introduces an exponential weight function to adjust the static background field error covariance, so that the error covariance changes in real time with the atmospheric state, and truly reflects the atmospheric flow field characteristics under different weather backgrounds. Compared with the traditional static error covariance method, the present invention can better capture high-dimensional and nonlinear error characteristics, and can effectively improve the analysis and forecasting accuracy. Through the key features in the background field, the machine learning method is used to predict the error distribution with flow-dependent characteristics, and the background field error covariance can be adaptively adjusted according to the background field uncertainty, thereby bringing anisotropic and inhomogeneous analysis increments to the assimilation system.

[0045] The present invention realizes that the error characteristics of flow dependence can be obtained without ensemble forecast, thereby reducing the computational cost and improving the performance and efficiency of the assimilation system. It also provides a more reasonable error distribution for the data assimilation system while maintaining the computational efficiency, greatly saving computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 The present invention is a flow chart of a method for dynamically adjusting flow-dependent background field error covariance based on machine learning. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings in the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments, and they should not be understood as limitations on the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. In the description of the present invention, it should be understood that the terms used are only for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0048] Combine the following Figure 1 The present invention describes a method for dynamic adjustment of flow-dependent background field error covariance based on machine learning.

[0049] Embodiment 1: This embodiment provides a method for dynamic adjustment of flow-dependent background field error covariance based on machine learning.

[0050] Based on the problems existing in the assimilation background field error covariance estimation technology, the present invention proposes to estimate the background field uncertainty through a machine learning method, and dynamically adjust the spatial distribution of the error covariance, thereby generating a flow dependency analysis and improving the accuracy of the data assimilation system. The application of machine learning technology in the field of numerical forecasting is becoming more and more extensive. It can mine the relationships hidden in complex data through the learning of big data and high-dimensional nonlinear features. In particular, machine learning can be used to process high-dimensional data, identify complex patterns, and capture the spatiotemporal variation characteristics of atmospheric states. Applying machine learning to the estimation of the background field error covariance can more effectively capture the flow dependency characteristics of the atmospheric state, dynamically adjust the spatial distribution of the background error covariance, and thus improve the performance of the assimilation system.

[0051] This results in the technical solution of this embodiment, such as Figure 1 As shown in the figure, the general process of this solution includes:

[0052] S100, using an ensemble perturbation method to perturb the background field of the control test, generating multiple initial members, and obtaining the forecast field of the ensemble members through numerical model integration;

[0053] S200, calculating the set discreteness of the target time based on the forecast field, and quantifying the uncertainty of the background field;

[0054] S300, extracting meteorological elements from the background field of the control test as input features, establishing a nonlinear mapping relationship between the input features and the set discreteness through a trained machine learning model, and predicting the uncertainty of the background field;

[0055] S400, normalizing the uncertainty of the prediction;

[0056] S500, dynamically adjusting the static background field error covariance matrix based on the normalized uncertainty value to generate a flow-dependent error covariance matrix;

[0057] S600, inputting the adjusted background field error covariance matrix, the model background field and the observation data into the assimilation system to solve the analysis field;

[0058] S700, using the analysis field to perform deterministic forecasting, and inputting the forecasting result into the assimilation cycle as the background field of the next cycle, to perform continuous real-time optimization and adjustment of the background error covariance.

[0059] The above steps are explained in detail.

[0060] Step S100 is to generate disturbance members using the ensemble perturbation method and perform ensemble forecasting.

[0061] First, the control experiment is perturbed using the ensemble perturbation method to generate multiple initial condition perturbation members. Common ensemble perturbation methods include perturbation schemes generated by singular vector perturbation or EnKF (EnsembleKalman Filter). Through this method, 30 initial member fields are formed, and each member and the control experiment are integrated for 6 hours using the numerical weather model, and a total of 31 forecast fields are obtained.

[0062] set up To control the background field of the experiment, the initial field of the ensemble forecast member is ,Right now:

[0063] ;

[0064] in, is the disturbance term.

[0065] The numerical model is used to integrate these members and obtain the corresponding ensemble forecast results. After integration in the numerical mode, we get members at a certain time Forecast status at the moment In this embodiment, 6 hours is taken. Assuming the model kinetic equation is , it can be expressed as:

[0066] ;

[0067] in, represents the atmospheric numerical model, It's in time Moment The forecast results of each member.

[0068] Step S200 is to calculate the set discreteness. After generating the set members, the 30 set members are used to calculate the set discreteness (i.e., the target variable y) of each control variable (such as stream function, velocity potential function, temperature, relative humidity, surface pressure, etc.), quantitatively characterize the uncertainty of the background field (control test), and provide labels (output values) for the subsequent training of the machine learning model. The larger the discreteness, the higher the uncertainty of the atmospheric state in the area.

[0069] The uncertainty of the control variable is expressed by the set discreteness as the target output of the machine learning model (Output, i.e. ). The ensemble dispersion is measured by calculating the standard deviation of the ensemble forecast members. The control variables include the stream function , potential function ,temperature , surface pressure and relative humidity , and calculate the standard deviation of each control variable to represent the set dispersion.

[0070] Taking the stream function as an example, the calculation formula for discreteness is:

[0071] ;

[0072] in, For the The stream function of each member, is the average value of all members of the set, is the number of set members.

[0073] The set dispersion of other variables is calculated in a similar way.

[0074] Step S300 first extracts input variables in the background field.

[0075] Extract key variables from the background field of the control test as input to the machine learning model. At least wind speed, temperature, air pressure, humidity and hydrometeoric variables are included. Meteorological elements are extracted from the background field of the control test as input features. , expressed as:

[0076] ;

[0077] in, , is the wind speed component, is the temperature, is the air pressure, For humidity, , , , and They are cloud water, rain water, ice crystals, snow and graupel in the hydrometeor variables. By properly preprocessing these features (such as standardization and sparseness), representative input data is provided for subsequent model training.

[0078] Then build a machine learning model based on LightGBM. Using the input features extracted above (meteorological elements from background fields) and output tags (target defined by set discreteness), build and train the LightGBM model. The machine learning model of background field uncertainty built based on LightGBM (Light Gradient Boosting Machine) can efficiently process high-dimensional data and capture the nonlinear characteristics of the atmospheric state. By training a large amount of historical data or samples, the model can predict the corresponding background field uncertainty under given background field conditions, ensuring that the flow-dependent background error characteristics can be quickly estimated using machine learning methods.

[0079] The goal of the model is to extract the input features Predict the corresponding set dispersion , expressed as:

[0080] ;

[0081] in, For input features To set discreteness The mapping relationship.

[0082] In order to ensure better training effect, a layer-by-layer training mode is adopted, that is, the discreteness of each layer is learned by using the input features of each layer; The training of the model aims to minimize the loss function, which is expressed as:

[0083] ;

[0084] in, and Respectively Layer The predicted value and true value of samples, is the total number of samples.

[0085] Through layer-by-layer optimization, the model can more focusedly learn discrete features at different levels, thereby effectively reducing feature redundancy and improving prediction accuracy.

[0086] Step S400 is to normalize the uncertainty. The uncertainty predicted by the machine learning model is normalized so that its value is between [0,1], which facilitates the subsequent adjustment of the relative magnitude of the error covariance, makes the uncertainty of different physical quantities comparable, and simplifies the setting of weight distribution.

[0087] The normalization formula is:

[0088] ;

[0089] in, is the normalized uncertainty, is the original uncertainty, is the maximum value of uncertainty, is the minimum value of uncertainty, subscript Represents different control variables, including stream functions , potential function ,temperature , surface pressure and relative humidity .

[0090] Step S500 is to dynamically adjust the background field error covariance. The background error covariance matrix is ​​dynamically adjusted using the predicted uncertainty information. The predicted uncertainty is mapped to weights in exponential form, and then the static background field error covariance matrix is ​​adjusted accordingly. High uncertainty areas correspond to larger background errors, and observations in these areas are given more weight during assimilation; low uncertainty areas remain relatively stable, and the background field is more trusted during assimilation.

[0091] The specific method is:

[0092] The normalized uncertainty As a weight index, it is introduced into the calculation of the background field error covariance matrix, where Each point in the three-dimensional space has an independent value, so as to dynamically adjust the spatial distribution of the covariance, and dynamically adjust the spatial distribution of the covariance to make it flow-dependent. The adjustment formula is expressed as:

[0093] ;

[0094] in, is the adjusted background field error covariance matrix, is the original background field error covariance matrix, the inner product To dynamically adjust the weight matrix;

[0095] Dynamically adjust the weight matrix The calculation formula is:

[0096] ;

[0097] in, To adjust the parameters, control With uncertainty growth rate.

[0098] That is, when hour, ,wrong Make adjustments; when As the uncertainty increases, increases exponentially, thus flexibly amplifying the background error covariance. The size of the index can flexibly control the speed of exponential growth. The larger the uncertainty, the greater the Significantly larger; Smaller will make As uncertainty increases more gently.

[0099] Step S600 is to perform assimilation iteration using the adjusted covariance matrix.

[0100] Dynamically adjusted background field error covariance matrix Together with the model background field and observation data, it is input into the data assimilation system to solve the analysis field. The uncertainty characteristics of flow dependence have been incorporated, and the assimilation system can make better use of observational information when updating the analysis, obtaining an analysis field that is closer to the actual atmospheric state.

[0101] Step S700 uses the analysis field to carry out deterministic forecasting and iterates in a loop.

[0102] Finally, the analysis field is used for deterministic forecasting, and the forecast result is input into the assimilation cycle as the background field for the next cycle. That is, the analysis result of this cycle is used as the new background field in the next cycle, and the uncertainty estimation is performed again, which can achieve continuous flow-dependent background error covariance optimization.

[0103] The method for constructing the flow-dependent background field error covariance based on machine learning in this embodiment predicts the error distribution with flow-dependent characteristics by using the machine learning method through the key features in the background field, and can adaptively adjust the background field error covariance according to the background field uncertainty, thereby bringing anisotropic and inhomogeneous analysis increments to the assimilation system. This invention realizes the reflection of the transient characteristics of the weather system in the background error without the need for ensemble forecasts, thereby providing a more reasonable error distribution for the data assimilation system while maintaining computational efficiency, greatly saving computing resources.

[0104] Embodiment 2: This embodiment provides a non-transitory computer-readable storage medium on which computer instructions are stored, and the computer instructions enable a computer to execute a method for dynamically adjusting flow-dependent background field error covariance based on machine learning, the method comprising the following steps:

[0105] S100, using an ensemble perturbation method to perturb the background field of the control test, generating multiple initial members, and obtaining the forecast field of the ensemble members through numerical model integration;

[0106] S200, calculating the set discreteness of the target time based on the forecast field, and quantifying the uncertainty of the background field;

[0107] S300, extracting meteorological elements from the background field of the control test as input features, establishing a nonlinear mapping relationship between the input features and the set discreteness through a trained machine learning model, and predicting the uncertainty of the background field;

[0108] S400, normalizing the uncertainty of the prediction;

[0109] S500, dynamically adjusting the static background field error covariance matrix based on the normalized uncertainty value to generate a flow-dependent error covariance matrix;

[0110] S600, inputting the adjusted background field error covariance matrix, the model background field and the observation data into the assimilation system to solve the analysis field;

[0111] S700: Use the analysis field to make a deterministic forecast, and input the forecast result into the assimilation cycle as the background field of the next period to perform continuous optimization adjustment of the background error covariance.

[0112] Embodiment 3: This embodiment provides an electronic device, which may include: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus. The processor may call the logic instructions in the memory to execute a method for dynamic adjustment of the flow-dependent background field error covariance based on machine learning, and the method includes the following steps:

[0113] S100, using an ensemble perturbation method to perturb the background field of the control test, generating multiple initial members, and obtaining the forecast field of the ensemble members through numerical model integration;

[0114] S200, calculating the set discreteness of the target time based on the forecast field, and quantifying the uncertainty of the background field;

[0115] S300, extracting meteorological elements from the background field of the control test as input features, establishing a nonlinear mapping relationship between the input features and the set discreteness through a trained machine learning model, and predicting the uncertainty of the background field;

[0116] S400, normalizing the uncertainty of the prediction;

[0117] S500, dynamically adjusting the static background field error covariance matrix based on the normalized uncertainty value to generate a flow-dependent error covariance matrix;

[0118] S600, inputting the adjusted background field error covariance matrix, the model background field and the observation data into the assimilation system to solve the analysis field;

[0119] S700: Use the analysis field to make a deterministic forecast, and input the forecast result into the assimilation cycle as the background field of the next period to perform continuous optimization adjustment of the background error covariance.

[0120] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program code.

[0121] Embodiment 4: This embodiment provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can perform a flow-dependent background field error covariance dynamic adjustment method based on machine learning, and the method includes the following steps:

[0122] S100, using an ensemble perturbation method to perturb the background field of the control test, generating multiple initial members, and obtaining the forecast field of the ensemble members through numerical model integration;

[0123] S200, calculating the set discreteness of the target time based on the forecast field, and quantifying the uncertainty of the background field;

[0124] S300, extracting meteorological elements from the background field of the control test as input features, establishing a nonlinear mapping relationship between the input features and the set discreteness through a trained machine learning model, and predicting the uncertainty of the background field;

[0125] S400, normalizing the uncertainty of the prediction;

[0126] S500, dynamically adjusting the static background field error covariance matrix based on the normalized uncertainty value to generate a flow-dependent error covariance matrix;

[0127] S600, inputting the adjusted background field error covariance matrix, the model background field and the observation data into the assimilation system to solve the analysis field;

[0128] S700, using the analysis field to perform deterministic forecasting, and inputting the forecasting result into the assimilation cycle as the background field of the next cycle, to perform continuous real-time optimization and adjustment of the background error covariance.

[0129] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0130] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for dynamic adjustment of flow-dependent background field error covariance based on machine learning, characterized in that: The following steps are involved: S100, using an ensemble perturbation method to perturb the background field of the control test, generating multiple initial members, and obtaining the forecast field of the ensemble members through numerical model integration; S200, calculating the set discreteness of the target time based on the forecast field, and quantifying the uncertainty of the background field; S300, extracting meteorological elements from the background field of the control test as input features, establishing a nonlinear mapping relationship between the input features and the set discreteness through a trained machine learning model, and predicting the uncertainty of the background field; S400, normalizing the uncertainty of the prediction; S500, dynamically adjusting the static background field error covariance matrix based on the normalized uncertainty value to generate a flow-dependent error covariance matrix; S600, inputting the adjusted background field error covariance matrix, the model background field and the observation data into the assimilation system to solve the analysis field; S700, using the analysis field to perform deterministic forecasting, and inputting the forecasting result into the assimilation cycle as the background field of the next cycle, to perform continuous real-time optimization and adjustment of the background error covariance.

2. The method for dynamic adjustment of flow-dependent background field error covariance based on machine learning according to claim 1, characterized in that: The machine learning model trained in step S300 is a machine learning model of background field uncertainty built based on LightGBM. The goal of the model is to extract meteorological elements from the background field of the control test as input features. Predict the corresponding set dispersion , expressed as: ； in, For input features To set discreteness The mapping relationship of And the model is optimized by minimizing the loss function.

3. The method for dynamic adjustment of flow-dependent background field error covariance based on machine learning according to claim 2, characterized in that: In the training of the machine learning model of background field uncertainty built based on LightGBM, a layer-by-layer training mode is adopted, that is, the discreteness of each layer is learned by using the input features of each layer; The training of the model aims to minimize the loss function, which is expressed as: ； in, and Respectively Layer The predicted value and true value of samples, is the total number of samples.

4. The method for dynamic adjustment of flow-dependent background field error covariance based on machine learning according to claim 1, characterized in that: The collective perturbation method in step S100 is a perturbation scheme generated by singular vector perturbation or collective Kalman filtering.

5. The method for dynamic adjustment of flow-dependent background field error covariance based on machine learning according to claim 1, characterized in that: In step S200, the collective discreteness at the target time is calculated based on the forecast field, specifically, the collective discreteness is used to represent the uncertainty of the control variable as the target output of the machine learning model. The control variable includes the stream function , potential function ,temperature , surface pressure and relative humidity , and calculate the standard deviation of each control variable to represent the set dispersion.

6. The method for dynamic adjustment of flow-dependent background field error covariance based on machine learning according to claim 5, characterized in that: In step S400, the uncertainty of the prediction is normalized to limit the uncertainty of the prediction to the range of [0,1], and the formula is: ； in, is the normalized uncertainty, is the original uncertainty, is the maximum value of uncertainty, is the minimum value of uncertainty, subscript represents different control variables.

7. The method for dynamic adjustment of flow-dependent background field error covariance based on machine learning according to claim 1, characterized in that: The meteorological elements extracted from the background field of the control test in step S300 include at least wind speed, temperature, air pressure, humidity and hydrometeoric variables. These meteorological elements extracted from the background field of the control test are used as input features. , expressed as: ； in, , is the wind speed component, is the temperature, is the air pressure, For humidity, , , , and They are cloud water, rain water, ice crystals, snow and graupel among the hydrometeoric variables.

8. The method for dynamic adjustment of flow-dependent background field error covariance based on machine learning according to claim 1, characterized in that: The specific method of step S500 is: The normalized uncertainty As a weight index, it is introduced into the calculation of the background field error covariance matrix, where Each point in the three-dimensional space has an independent value, so as to dynamically adjust the spatial distribution of the covariance to make it flow-dependent. The adjustment formula is expressed as: ； in, is the adjusted background field error covariance matrix, is the original background field error covariance matrix, the inner product To dynamically adjust the weight matrix; Dynamically adjust the weight matrix The calculation formula is: ； in, To adjust the parameters, control With uncertainty growth rate.

9. The method for dynamic adjustment of flow-dependent background field error covariance based on machine learning according to claim 1, characterized in that: In step S600, the assimilation system is one of a three-dimensional variational assimilation system, a four-dimensional variational assimilation system or a hybrid assimilation system; The assimilation system uses variational assimilation or hybrid assimilation methods to fuse the adjusted background field error covariance matrix, model background field and observation data.

Images