Method and system for constructing collective standardized members based on strong convection identification

Abstract

The present invention discloses a method for constructing a set of standardized members based on the identification of strong convection, comprising the steps of: acquiring historical meteorological data according to the forecast area and forecast period, and performing preprocessing to construct a data training set; using the data training set to construct and train a classification recognition model for identifying the type of strong convection; constructing a set member generation model, and training based on different types of strong convection, respectively, to construct a set member generation mode for each type of strong convection; using the classification recognition model to classify the weather of the pre-generated set members, and then selecting the corresponding set member generation mode to generate the set standardized members according to the classification result. The present invention can effectively classify strong convection, and generate set members in combination with historical strong convection information of the same type. The present invention does not need to integrate the mode, and can effectively improve the discreteness, and reduce the demand for computing resources when generating set members.

Description

Technical Field

[0001] The invention belongs to the field of atmospheric science research, and in particular relates to a method for constructing collective assimilation members based on strong convection identification. Background Art

[0002] The forecast of severe convective weather often relies on numerical models. However, the current numerical models still have significant errors in forecasting severe convection. In addition to the complexity of severe convective weather, another major reason is that the numerical models used to forecast severe convection have non-negligible model errors and initial field errors.

[0003] In order to overcome this problem, the commonly used method is to use the ensemble data assimilation method to improve the model initial field. The most popular ensemble data assimilation method is the mixed data assimilation method. The key point of this type of method is to generate reasonable ensemble members. The main method of generating ensemble members is to perturb the observation, model initial field or model, or to perturb the parameters or tendencies of the physical process and parameterization process of the numerical model. The perturbation method is generally to use a certain form of noise for perturbation, such as Gaussian noise, or to transform and calculate the current meteorological conditions through certain mathematical methods.

[0004] The above perturbation methods can only target the weather situation during this period. It is difficult to include historical case information similar or close to the current weather situation during perturbation, and it is impossible to generate ensemble forecast members related to severe convective weather types. Moreover, perturbations to patterns often require integration of the patterns, so generating ensemble members requires a lot of computing resources. 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] One object of the present invention is to provide a method for constructing ensemble standardized members based on strong convection identification, aiming to generate ensemble standardized members by utilizing historical cases with consistent or similar strong convection types, thereby reducing the demand for computing resources when generating ensemble standardized members and improving the model forecasting effect.

[0007] Another object of the present invention is to provide a collection assimilation member construction system based on strong convection identification, aiming to use the system to implement the above method.

[0008] In order to achieve the above-mentioned object, the present invention provides a method for constructing a collection of standardized members based on strong convection identification, comprising the following steps:

[0009] According to the forecast area and forecast period, historical meteorological data is obtained and preprocessed to construct a data training set;

[0010] Using the data training set, constructing and training a classification recognition model for identifying the type of severe convection;

[0011] Constructing a collection member generation model, and training it based on different types of severe convection, and constructing a collection member generation model for each type of severe convection;

[0012] The classification recognition model is used to classify the weather of the pre-generated ensemble members, and then according to the classification results, the corresponding ensemble member generation mode is selected to generate the ensemble assimilated members.

[0013] A further preferred technical solution of the present invention is that, according to the forecast area and forecast period, historical meteorological data is obtained and preprocessed to construct a data training set; the specific method is:

[0014] Collect existing ensemble forecast information according to region and forecast period, including one deterministic forecast and several ensemble forecast results;

[0015] Conduct meteorological data analysis on the collected ensemble forecast information, screen and remove redundant or useless variables, and add auxiliary variables as needed;

[0016] A training set is constructed based on the collected ensemble forecast information and the selected relevant variables, and the data are labeled according to the category of severe convection.

[0017] Preferably, the data training set is used to construct and train a classification recognition model for identifying the type of severe convection;

[0018] The classification and recognition model introduces a segmentation module based on storm helicity features to focus the model's attention on key areas of severe convection.

[0019] Preferably, the calculation formula of the segmentation module based on the storm helicity feature is:

[0020]

[0021] in, is the sigmoid function, is the separation coefficient, which takes a value between 0 and 1. The larger the value, the heavier the weight of severe convection in the artificial intelligence classification; and Represent the average pooling features and the maximum pooling features respectively; Used to smooth the grid points between severe convection and no weather;

[0022] H is the storm helicity, which is defined as:

[0023]

[0024] in, For the environmental wind field, is the velocity of the severe convective system in the training set, is the horizontal vorticity vector, is the height of the highest floor during calculation, is the integral of height;

[0025] Based on storm helicity Separation module , make the storm spiral greater than 250 and less than 250 Separate.

[0026] Preferably, the classification and recognition model adopts one of the network architectures of ALexNet, ResNet or DenseNet, and introduces the segmentation module based on the storm helicity feature between the convolution blocks of the network.

[0027] Preferably, the set member generation model is constructed, and training is performed based on different types of strong convection, and a set member generation model is constructed for each type of strong convection; the specific method is:

[0028] Construct the current weather input field Mapping to target meteorological output field The main generative model ;

[0029] Construct a reverse generation model to generate the meteorological input field from the ensemble members, i.e. the output field ;

[0030] In general, and It can be composed of an encoder and a decoder. The encoder is composed of multiple two-dimensional convolutional network blocks, and the decoder is composed of multiple transposed convolutional blocks. The encoder and the decoder can be connected by multiple residual blocks. The number of convolutional network blocks, transposed convolutional blocks, and residual blocks can be determined according to the situation, and other modules can be added in between to improve the model usability.

[0031] Construct the first judgment model for judging whether the generated set members are of the same type of severe convection as the input field ;

[0032] Construct a second judgment model for judging whether the reverse generated meteorological input field is of the same severe convection type as the original input field ;

[0033] In general, and It consists of several layers of fully convolutional layers.

[0034] and The training goal is to generate similar types of severe convection ensemble members. and The goal is to determine the differences between the generated set members and historical examples. is the main training goal, , To get better Auxiliary model.

[0035] As a preference, a convolutional neural network is used to design an auxiliary classifier , , and ,improve , , and The mathematical expression of the auxiliary classifier for the identification of different strong convections is:

[0036]

[0037]

[0038]

[0039]

[0040] in, is the sigmoid function; The kth activation map of the convolution in the meteorological input field is located at The value of is the corresponding weight; The kth activation map of the convolution in the meteorological output field is located at The value of is the corresponding weight; The kth activation map of the convolution in the first judgment model is located at The value of is the corresponding weight; The kth activation map of the convolution in the second judgment model is located at The value of is the corresponding weight;

[0041] Auxiliary classifier , Insert in and Between the encoder and decoder, if there is a residual block in the middle, the auxiliary classifier , Placed after the residual block. Auxiliary classifier and Insert in and After convolutional network.

[0042] The model training objective contains two loss functions, expressed as:

[0043]

[0044] And add consistency judgment to the loss function, expressed as:

[0045]

[0046] in, is the mathematical expectation of the model and target field, is the mathematical expectation of the model and source field, is an adjustable parameter;

[0047] The model training objective expression is:

[0048]

[0049] Finally, we get the type of each severe convection Model.

[0050] Preferably, the classification recognition model is used to classify the weather of the pre-generated set members, and then according to the classification result, the corresponding set member generation mode is selected to generate the set assimilated members; the specific method is:

[0051] For the weather of the pre-generated set member, classify the weather using the classification recognition model;

[0052] Based on the type of strong convection, after giving a random number seed, use the trained corresponding category The model generates ensemble members, which are directly used as model forecast results or as ensemble assimilated members.

[0053] In another aspect, the present invention provides a system for constructing collective standardized members based on strong convection identification, comprising:

[0054] The data acquisition module is used to obtain historical meteorological data according to the forecast area and forecast period, and perform preprocessing to build a data training set;

[0055] A classification recognition model building module is used to use the data training set to build and train a classification recognition model for identifying the type of severe convection;

[0056] The ensemble member generation model building module is used to build the ensemble member generation model, and to train the ensemble member generation model based on different types of severe convection, and to build a ensemble member generation model for each type of severe convection;

[0057] The ensemble standardized member generation module is used to classify the weather of the pre-generated ensemble members using a classification recognition model, and then select the corresponding ensemble member generation mode to generate ensemble standardized members based on the classification results.

[0058] Beneficial effects: The present invention can effectively classify strong convection and generate set members by combining historical strong convection information of the same type. Compared with the original GEFS set system, the present invention does not need to integrate the pattern and can effectively improve the discreteness and reduce the demand for computing resources when generating set members. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 Schematic diagram of the location of the classification model selected in the embodiment of the present invention and the segmentation module based on the storm helicity feature.

[0060] Figure 2 It is a flow chart of generating set assimilated members by using trained generation patterns in an embodiment of the present invention.

[0061] Figure 3 Comparison of classification accuracy of severe convection by the classification and recognition model according to an embodiment of the present invention.

[0062] Figure 4 A comparison diagram of the discreteness and RMSE ratio of corresponding variables of the set homogenized members generated by the embodiment of the present invention and the set members of the GEFS mode. DETAILED DESCRIPTION

[0063] 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.

[0064] Combine the following Figure 1-Figure 4 The invention describes a method and system for constructing collective standardized members based on strong convection identification.

[0065] Embodiment 1: This embodiment provides a method for constructing collective homogenous members based on strong convection identification.

[0066] This method is a set member generation method of joint machine learning classification developed to address the problems that existing technologies have difficulty in including historical information similar to the current weather situation and require a lot of computing resources. This method can reasonably classify the current weather situation and combine it with historical data, and can generate set members without pattern integration.

[0067] The method mainly includes two parts: the construction of artificial intelligence models for strong convection identification and set and member generation. Taking the construction of strong convection set assimilation members in June in East China as an example, the specific steps include:

[0068] S1. Select the region and forecast period. According to the need to generate the ensemble assimilation members of severe convection in June in East China, first select 15°N - 47°N and 105°E -137°E as the construction area of ​​ensemble assimilation members based on severe convection identification; and according to the determined forecast period, collect the ensemble forecast information from June 1st to June 30th every year from 2017 to 2022. This embodiment is obtained by cutting the results of the Global Ensemble Forecast System (GEFS) and ERA5-Ensemle according to the selected region.

[0069] In order to increase the number of samples in the data set, first, using the situation that GEFS reports once every 6 hours, the result of reporting 24 hours before the current time and forecasting 24 hours is taken as the first result, and the result of reporting 18 hours before the current time and forecasting 18 hours is taken as the second result, and so on to obtain 5 results. GEFS itself is an ensemble forecast product of 20 members (there is also a control member, but it is discarded), so 100 ensemble forecast members can be obtained. Secondly, using the ERA5-Ensemble ensemble forecast product, take its products two hours before the current moment, one hour, the current moment, one hour after the current moment, and two hours after the current moment as the current moment product. Because ERA5-Ensemble itself is an ensemble forecast product of 10 members, 50 ensemble members are obtained.

[0070] Through the above method, 150 64 grid points × 64 grid points forecast data will be obtained at a certain moment.

[0071] The Pearson correlation coefficient is calculated for all meteorological variables contained in these data in pairs. If the correlation coefficient of two variables is greater than 0.8, only one of the two variables involved in the calculation is retained. After the correlation analysis, the retained meteorological variables are determined and auxiliary variables are added in combination with the actual needs of the meteorological analysis and the actual output of the meteorological numerical model. In this embodiment, 2 meters of temperature, 10 meters of wind field, ground pressure, 850 hPa, 700 hPa, 500 hPa and 300 hPa of temperature, potential height, wind field and relative humidity are selected as variables involved in the calculation in subsequent steps.

[0072] S2, create a data folder according to the date and time of the data and the member name for the time, area and variable data involved in the calculation selected in step S1, and save the data in this folder. According to the type of severe convective weather, mark the weather corresponding to the date covered by this embodiment with a corresponding label.

[0073] S3. Build an artificial intelligence classification and recognition model.

[0074] In order to enhance the classification of severe convection by machine learning, this embodiment adds a segmentation module based on the storm helicity feature to the traditional artificial intelligence classification model, focuses the "attention" of the classification model on the key areas of severe weather, and then combines the more mature recognition and classification model ResNet or other popular classification models to classify severe convection. This embodiment uses the 18-layer ResNet model as the base model for classification.

[0075] S31. First, define the storm helicity used in this embodiment:

[0076]

[0077] in, For the environmental wind field, is the velocity of the severe convective system in the training set, is the horizontal vorticity vector, is the integral of height; For the height of the highest floor during calculation, the present invention takes .

[0078] S32, then define the separation module , make the storm spiral greater than 250 and less than 250 The purpose of this separation module is to let the classification model focus on the strong convection area and ignore less important weather information. Figure 1 As shown in the figure, this module is placed in the residual block between the convolution blocks in ResNet. The calculation formula is:

[0079]

[0080] in, is the sigmoid function, is the separation coefficient, which takes a value between 0 and 1. The larger the value, the heavier the weight of severe convection in the artificial intelligence classification. and Represent the average pooling features and the maximum pooling features respectively; This is to smooth the grid points between severe convection and no weather to prevent the values ​​from changing too much.

[0081] In this embodiment, the separation module is placed in the convolutional blocks of the last three layers of ResNet. In other embodiments, the The modules are placed in different convolutional blocks, and it is also possible to place the modules in different blocks (layers) of the model.

[0082] S4. Use the data set and classification results generated in step S2 to train the classification model. In this embodiment, 90% of the 2017-2021 data set is used as the training set to train the modified ResNet, and the remaining 10% is used as the test set. The 2022 data is used as the data for the verification model result and is not used in the training. During training, the parameters Test every 0.1 from 0.5 to 0.9 and select the optimal parameters ,Finally, the complete severe convection classification model is trained.

[0083] S5. Construct an artificial intelligence set member generation model.

[0084] The purpose of this embodiment is to train a model for a certain type of severe convection. This model can use the historical data of severe convection classification to input the current meteorological field Mapping to target meteorological output field , that is, the members of the set.

[0085] In order to achieve this purpose, in this embodiment, in addition to the corresponding type of main mode In addition, you need to train a Mode and two discrimination modes and .

[0086] in Used to inversely generate meteorological input fields from ensemble members, combined with Determine whether the input field generated in reverse is the same type of strong convection as the original input field. This mode also has the function of keeping the mode stable to prevent always generating the same results. It is used to identify the generated ensemble forecast members (i.e., output fields ) is a severe convection member of the same category as the input field. and The goal is to continuously generate set members similar to the same type of strong convection in history, minimizing the difference between the generated results and historical cases. and The goal is to continuously judge the differences between the generated set members and historical examples, and try to find the differences between the generated set members and historical examples. Therefore, in the training process, the two are combined to make and The technology is constantly improving, and finally reaches and The result is indistinguishable. The ensemble forecast members can be generated according to the type of current severe convection and the information of similar severe convection in history.

[0087] In this embodiment, and It consists of an encoder, a decoder, and a residual block connecting the two. The encoder contains three layers of convolution. The initial layer receives 20-dimensional input (the numerical value of the selected variable), and the third layer outputs 256 dimensions. Instance normalization is performed after each convolution layer, and the batch size is 1. The decoder is an upsampler that reconstructs the features into ensemble forecast outputs. It is constructed through two transposed convolutional layers, and the encoder output dimension is reduced to 32 dimensions. The final layer is a convolutional layer that receives 32-dimensional input and converts it into a 20-dimensional ensemble forecast output. The encoder and decoder are connected through residual blocks. This design is to solve the gradient vanishing problem in deep learning models. Four residual blocks are used, each containing two convolutional layers. The encoder output is passed to the residual layer, and then to the decoder after the residual layer.

[0088] Discriminator and It consists of five fully convolutional layers, which are responsible for classifying the 64x64-sized forecasts to determine whether they belong to the current severe convection category. The filter uses a 3x3 size, receives a 20-dimensional ensemble forecast input field, and gradually converts it to 32, 64, 128, 256 dimensions, and finally regresses to 20 dimensions.

[0089] In order to reflect the role of strong convection classification, an auxiliary classifier is added , , and , which allows the model to learn where improvements are needed. The mathematical expression of the auxiliary classifier is:

[0090]

[0091]

[0092]

[0093]

[0094] in, is the sigmoid function; The kth activation map of the convolution in the meteorological input field is located at The value of is the corresponding weight; The kth activation map of the convolution in the meteorological output field is located at The value of is the corresponding weight; The kth activation map of the convolution in the first judgment model is located at The value of is the corresponding weight; The kth activation map of the convolution in the second judgment model is located at The value of is the corresponding weight;

[0095] Auxiliary classifier , Insert in and After the residual block and before the decoder. Auxiliary classifier and Insert in and After convolutional network.

[0096] The model training objective contains two loss functions, expressed as:

[0097]

[0098] in, is the mathematical expectation of the model and target field, is the mathematical expectation of the model and source fields;

[0099] Theoretically, an ensemble forecast member should have consistent characteristics with the current forecast that generated it, so a consistency judgment is added to the loss function to make the generated ensemble member close to the current forecast member.

[0100]

[0101] In order to adjust the focus of the model during learning, such as whether the model is closer to historical members or more focused on consistency before and after generation, the These 9 adjustable parameters can freely adjust the weights between different parameters to obtain better training results.

[0102] Finally, the final target expression is:

[0103]

[0104] After training the historical data of each type of severe convection classification, we can get the type of each severe convection Model.

[0105] S6. According to experience, the adjustable parameters of this embodiment are , , Take 1 for the first parameter and 10 for the other 6 parameters. According to the classification of the data set, train the models for thunderstorm, heavy rain, tornado, and hail respectively. Model. In different embodiments, different classifications can be used to train different Model.

[0106] S7: When generating sets and members, first classify the weather for which sets and members are to be generated using the model generated in step S4, and then use the trained corresponding category The model generates sets and forecast members. When generating set forecast members, different set forecast members can be generated given different random number seeds. In this embodiment, the random number seed is set to the current time of the computer clock (POSIX format). The process is as follows Figure 2 The members generated at this time can be directly used as model forecast results, or they can be aggregated and assimilated or integrated again.

[0107] In order to verify the effect of this embodiment, a continuous month of June 2022 was selected for verification. First, the 20 members of the GEFS pattern were identified and classified using ResNet18. Then, the segmentation module based on the storm helicity feature was placed in ResNet18 as a strong convection isolation module for classification. The classification results were compared with those of the traditional ResNet18 network. The results are as follows: Figure 3 As shown, the classification method of this embodiment has a higher recognition rate for each type of strong convection than the traditional ResNet18 network.

[0108] Then, 20 ensemble members are generated using the method of this embodiment. The method for judging the quality of ensembles and members is the ratio of dispersion to RMSE. The closer the ratio is to 1, the better. This means that a good ensemble forecast member needs to have a higher dispersion and a smaller difference from the true value. Figure 4 As shown, compared with the GEFS ensemble forecast, it is found that the dispersion and RMSE ratio of the ensemble forecast members generated by the present invention are closer to 1.

[0109] Embodiment 2: This embodiment provides a system for constructing collective assimilation members based on strong convection identification, including:

[0110] The data acquisition module is used to obtain historical meteorological data according to the forecast area and forecast period, and perform preprocessing to build a data training set;

[0111] A classification recognition model building module is used to use the data training set to build and train a classification recognition model for identifying the type of severe convection;

[0112] The ensemble member generation model building module is used to build the ensemble member generation model, and to train the ensemble member generation model based on different types of severe convection, and to build a ensemble member generation model for each type of severe convection;

[0113] The ensemble standardized member generation module is used to classify the weather of the pre-generated ensemble members using a classification recognition model, and then select the corresponding ensemble member generation mode to generate ensemble standardized members based on the classification results.

[0114] Embodiment 3: This embodiment provides a non-transitory computer-readable storage medium, on which computer instructions are stored, the computer instructions causing the computer to execute a method for constructing a collection of homogenized members based on strong convection identification, the method comprising the following steps:

[0115] According to the forecast area and forecast period, historical meteorological data is obtained and preprocessed to construct a data training set;

[0116] Using the data training set, constructing and training a classification recognition model for identifying the type of severe convection;

[0117] Constructing a collection member generation model, and training it based on different types of severe convection, and constructing a collection member generation model for each type of severe convection;

[0118] The classification recognition model is used to classify the weather of the pre-generated ensemble members, and then according to the classification results, the corresponding ensemble member generation mode is selected to generate the ensemble assimilated members.

[0119] Embodiment 4: 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 constructing a collection of assimilated members based on strong convection identification, and the method includes the following steps:

[0120] According to the forecast area and forecast period, historical meteorological data is obtained and preprocessed to construct a data training set;

[0121] Using the data training set, constructing and training a classification recognition model for identifying the type of severe convection;

[0122] Constructing a collection member generation model, and training it based on different types of severe convection, and constructing a collection member generation model for each type of severe convection;

[0123] The classification recognition model is used to classify the weather of the pre-generated ensemble members, and then according to the classification results, the corresponding ensemble member generation mode is selected to generate the ensemble assimilated members.

[0124] 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.

[0125] Embodiment 5: This embodiment provides a computer program product, which includes a computer program. The computer program can be stored in a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute a method for constructing a collection of assimilated members based on strong convection identification, which includes the following steps:

[0126] According to the forecast area and forecast period, historical meteorological data is obtained and preprocessed to construct a data training set;

[0127] Using the data training set, constructing and training a classification recognition model for identifying the type of severe convection;

[0128] Constructing a collection member generation model, and training it based on different types of severe convection, and constructing a collection member generation model for each type of severe convection;

[0129] The classification recognition model is used to classify the weather of the pre-generated ensemble members, and then according to the classification results, the corresponding ensemble member generation mode is selected to generate the ensemble assimilated members.

[0130] 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.

[0131] 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.

[0132] 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 constructing collective homogenized members based on strong convection identification, characterized in that: The steps include: According to the forecast area and forecast period, historical meteorological data is obtained and preprocessed to construct a data training set; Using the data training set, constructing and training a classification recognition model for identifying the type of severe convection; The classification and recognition model adopts one of the network architectures of ALexNet, ResNet or DenseNet, and introduces a segmentation module based on the storm helicity feature between the convolution blocks of the network to focus the model's attention on the key areas of severe convection; The calculation formula of the segmentation module based on the storm helicity feature is: Among them, σ is the sigmoid function, μ is the separation coefficient, and the coefficient ranges from 0 to 1. The larger the value, the heavier the weight of severe convection in artificial intelligence classification; and Represent the average pooling features and the maximum pooling features respectively; Used to smooth the grid points between severe convection and no weather; H is the storm helicity, which is defined as: Where V is the ambient wind field, C is the speed of the severe convective system in the training set, ω h is the horizontal vorticity vector, h is the height of the highest layer during calculation, and dz is the integral over the height; The separation module M(F) based on the storm helicity H separates the storm helicity greater than 250m 2 ·s" 2 and less than 250m 2 ·s" 2 separate; Construct a collection member generation model, and train it based on different types of severe convection, and construct a collection member generation model for each type of severe convection; including: Construct the main generation model G that maps the current meteorological input field X5 to the target meteorological output field X6 5→6 ; Construct a reverse generation model G that reversely generates the meteorological input field from the set members, that is, the output field 6→5 ; G 5→6 and G 6→5 It consists of an encoder and a decoder. The encoder consists of multiple two-dimensional convolutional network blocks, and the decoder consists of multiple transposed convolutional blocks. The encoder and decoder are connected by multiple residual blocks. Constructing a first judgment model D5 for judging whether the generated set members are of the same severe convection type as the input field; Constructing a second judgment model D6 for judging whether the reversely generated meteorological input field is of the same severe convection type as the original input field; D5 and D6 consist of several fully convolutional layers; G 5→6 and G 6→5 The training goal is to generate similar types of severe convection ensemble members. The goals of D5 and D6 are to judge the differences between the generated ensemble members and historical cases. 5→6 As the main training goal, G 6→5 , D5 and D6 are to obtain better G 5→6 Auxiliary model of Use convolutional neural networks to design auxiliary classifiers η5, η6, and To improve G 5→6 , G 6→5 , the recognition effect of different strong convection in D5 and D6, the mathematical expression of the auxiliary classifier is: Among them, σ is the sigmoid function; is the value at (i, j) in the kth activation map of the convolution in the meteorological input field, is the corresponding weight; is the value at (i, j) in the kth activation map of the convolution in the meteorological output field, is the corresponding weight; is the value at (i, j) in the kth activation map of the convolution in the first judgment model, is the corresponding weight; is the value at (i, j) in the kth activation map of the convolution in the second judgment model, is the corresponding weight; Auxiliary classifiers η5 and η6 are inserted in G 5→6 and G 6→5 If there is a residual block between the encoder and the decoder, the auxiliary classifiers η5 and η6 are placed after the residual block; the auxiliary classifiers and Inserted after the D5 and D6 convolutional networks; The model training objective contains two loss functions, expressed as: And add consistency judgment to the loss function, expressed as: in, is the mathematical expectation of the model and target field, is the mathematical expectation of the model and the source field, λ R ~λ b is an adjustable parameter; The model training objective expression is: Finally, we get G for each type of severe convection 5→6 Model; The classification recognition model is used to classify the weather of the pre-generated ensemble members, and then according to the classification results, the corresponding ensemble member generation mode is selected to generate the ensemble assimilated members.

2. The method for constructing collective homogenized members based on strong convection identification according to claim 1 is characterized in that: According to the forecast area and forecast period, historical meteorological data is obtained and preprocessed to construct a data training set; the specific method is: Collect existing ensemble forecast information according to region and forecast period, including one deterministic forecast and several ensemble forecast results; Conduct meteorological data analysis on the collected ensemble forecast information, screen and remove redundant or useless variables, and add auxiliary variables as needed; A training set is constructed based on the collected ensemble forecast information and the selected relevant variables, and the data are labeled according to the category of severe convection.

3. The method for constructing collective homogenized members based on strong convection identification according to claim 1, characterized in that: The classification recognition model is used to classify the weather of the pre-generated ensemble members, and then the corresponding ensemble member generation mode is selected to generate the ensemble assimilated members according to the classification results; the specific method is as follows: For the weather of the pre-generated set member, classify the weather using the classification recognition model; Based on the type of strong convection, after giving a random number seed, the trained G 5→6 The model generates ensemble members, which are directly used as model forecast results or as ensemble assimilated members.

4. A system for constructing collective homogenized members based on strong convection identification, characterized in that: include: The data acquisition module is used to obtain historical meteorological data according to the forecast area and forecast period, and perform preprocessing to build a data training set; A classification recognition model building module is used to use the data training set to build and train a classification recognition model for identifying the type of severe convection; The classification and recognition model adopts one of the network architectures of ALexNet, ResNet or DenseNet, and introduces a segmentation module based on the storm helicity feature between the convolution blocks of the network to focus the model's attention on the key areas of severe convection; The calculation formula of the segmentation module based on the storm helicity feature is: Among them, σ is the sigmoid function, μ is the separation coefficient, and the coefficient ranges from 0 to 1. The larger the value, the heavier the weight of severe convection in artificial intelligence classification; and Represent the average pooling features and the maximum pooling features respectively; Used to smooth the grid points between severe convection and no weather; H is the storm helicity, which is defined as: Where V is the ambient wind field, C is the speed of the severe convective system in the training set, ω h is the horizontal vorticity vector, h is the height of the highest layer during calculation, and dz is the integral over the height; The separation module M(F) based on the storm helicity H separates the storm helicity greater than 250m 2 ·s" 2 and less than 250m 2 ·s" 2 separate; The ensemble member generation model building module is used to build the ensemble member generation model, and to train the ensemble member generation model based on different types of strong convection, and to build a ensemble member generation model for each type of strong convection; it includes: Construct the main generation model G that maps the current meteorological input field X5 to the target meteorological output field X6 5→6 ; Construct a reverse generation model G that reversely generates the meteorological input field from the set members, that is, the output field 6→5 ; G 5→6 and G 6→5 It consists of an encoder and a decoder. The encoder consists of multiple two-dimensional convolutional network blocks, and the decoder consists of multiple transposed convolutional blocks. The encoder and decoder are connected by multiple residual blocks. Constructing a first judgment model D5 for judging whether the generated set members are of the same severe convection type as the input field; Constructing a second judgment model D6 for judging whether the reversely generated meteorological input field is of the same severe convection type as the original input field; D5 and D6 consist of several fully convolutional layers; G 5→6 and G 6→5 The training goal is to generate similar types of severe convection ensemble members. The goals of D5 and D6 are to judge the differences between the generated ensemble members and historical cases. 5→6 As the main training goal, G 6→5 , D5 and D6 are to obtain better G 5→6 Auxiliary model of Use convolutional neural networks to design auxiliary classifiers η5, η6, and To improve G 5→6 , G 6→5 , the recognition effect of different strong convection in D5 and D6, the mathematical expression of the auxiliary classifier is: Among them, σ is the sigmoid function; is the value at (i, j) in the kth activation map of the convolution in the meteorological input field, is the corresponding weight; is the value at (i, j) in the kth activation map of the convolution in the meteorological output field, is the corresponding weight; is the value at (i, j) in the kth activation map of the convolution in the first judgment model, is the corresponding weight; is the value at (i, j) in the kth activation map of the convolution in the second judgment model, is the corresponding weight; Auxiliary classifiers η5 and η6 are inserted in G 5→6 and G 6→5 If there is a residual block between the encoder and the decoder, the auxiliary classifiers η5 and η6 are placed after the residual block; the auxiliary classifiers and Inserted after the D5 and D6 convolutional networks; The model training objective contains two loss functions, expressed as: And add consistency judgment to the loss function, expressed as: in, is the mathematical expectation of the model and target field, is the mathematical expectation of the model and the source field, λ R ~λ b is an adjustable parameter; The model training objective expression is: Finally, we get G for each type of severe convection 5→6 Model; The ensemble standardized member generation module is used to classify the weather of the pre-generated ensemble members using a classification recognition model, and then select the corresponding ensemble member generation mode to generate ensemble standardized members based on the classification results.

Images