Wind field data joint distribution modeling method and system based on standardized flow model

Abstract

The invention discloses a wind field data joint distribution modeling method and system based on a standardized flow model, and relates to the field of wind engineering and the field of artificial intelligence, and the method specifically comprises the steps: obtaining wind field monitoring data, and carrying out the preprocessing; a standardized flow model framework is constructed, and bidirectional reversible mapping between an original data space and a hidden variable space is established through a training module; training the standardized flow model framework by using the preprocessed data to obtain a trained standardized flow model; and based on the standardized flow model, calculating the multi-dimensional joint distribution probability density of any point in the original data space through forward transformation of bidirectional reversible mapping and a variable transformation formula. According to the method, a related structure between variables does not need to be assumed in advance, the internal probability dependence characteristics are directly learned based on the actual monitoring data, the characterization capability of the joint probability model on the actual wind field random characteristics is greatly improved, and a more accurate uncertainty quantification basis is provided for subsequent wind load and engineering structure reliability analysis.

Description

Technical Field

[0001] This application relates to the fields of wind engineering and artificial intelligence, and in particular to a method and system for joint distribution modeling of wind field data based on a standardized flow model. Background Technology

[0002] Currently, accurate modeling of environmental variables such as wind direction and wind speed in wind engineering and structural reliability assessment mainly relies on various univariate and bivariate probability distribution methods. Univariate modeling typically uses Gumbel, Weibull, or normal distributions for wind speed and von Mises distributions for independent description of wind direction. While simple and easy to implement, this method fails to consider the dependency between wind speed and wind direction, potentially leading to significant biases in wind load assessment results. To characterize the relationships between variables, existing technologies have developed bivariate joint distribution models, such as the Angular-linear model suitable for combinations of circular-linear variables, and construction methods based on Copula functions or finite mixture distributions. However, Angular-linear models often face high computational requirements and discontinuous probability density functions in large-scale data applications; while Copula methods and mixture models can improve representation accuracy, they are still limited by strict assumptions about marginal distributions or mixture components, their parameter identification process is complex, and the choice of Copula function directly affects model performance, lacking universal criteria. Furthermore, when the dimensionality of variables increases and the data exhibits a complex nonlinear correlation structure, the modeling difficulty of the aforementioned traditional methods increases dramatically, making it difficult to effectively extend to high-dimensional wind fields and other environmental datasets.

[0003] Therefore, how to construct a joint probability distribution modeling method for wind fields that can be flexibly extended to high-dimensional cases is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] In view of this, the present invention provides a method and system for joint distribution modeling of wind field data based on a standardized flow model, which overcomes the above-mentioned defects.

[0005] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a method for joint distribution modeling of wind field data based on a standardized flow model, the specific steps of which are as follows: Acquire wind field monitoring data and preprocess the wind field monitoring data; A standardized flow model framework is constructed, which establishes a bidirectional invertible mapping between the original data space and the latent variable space through a training module; the original data space is used to characterize the true multidimensional joint distribution of wind field monitoring data; the prior distribution of the latent variable space is assumed to be a normal distribution; the training module includes n coupling layers, each coupling layer including an affine coupling block and a random permutation layer; The standardized flow model framework is trained using preprocessed data to obtain a trained standardized flow model; Based on the standardized flow model, samples of the original data space are generated by sampling from the latent variable space and performing an inverse transformation of the bidirectional invertible mapping, and / or the multidimensional joint probability density of any point in the original data space is calculated using the forward transformation of the bidirectional invertible mapping and the variable transformation formula.

[0006] Optionally, the training steps of the standardized flow model framework are as follows: The wind field monitoring data is acquired and preprocessed to obtain the original training data matrix; The standardized flow model framework is constructed, and the standardized flow model framework is trained using the original training data matrix. The model parameters are optimized based on the loss function to obtain the trained standardized flow model.

[0007] Optionally, the steps for obtaining the original training data matrix are as follows: Obtain the real-time wind field monitoring data; The wind field monitoring data is preprocessed, and the preprocessed data with a unified format is converted into a data matrix to obtain the original training data matrix.

[0008] Optionally, the training module includes multiple sequentially connected affine coupling blocks and random permutation layers; The affine coupling block is configured to: divide the input data into a first part of data and a second part of data, perform a direct copy operation on the first part of data, and perform a reversible affine transformation operation on the second part of data; The random permutation layer is configured to randomly shuffle the dimensional order of the output data of the affine coupling block.

[0009] Optionally, the expression for the affine coupling block is: ; In the formula, This is the first part of the data; This is the second part of the data; and All are real-valued outputs of the neural network; These are the parameters of the neural network.

[0010] Alternatively, the expression for a reversible affine transformation operation is: Forward transform: ; Inverse transform: ; In the formula, This is the second part of the output data for the positive transform; For neural network parameters; For Hadama accumulation.

[0011] Optionally, the expression for the probability density of the multidimensional joint distribution is: ; In the formula, This represents the probability density of the original data space. Input data; For neural network parameters; The positive transformation function of the model; Let be the probability density of the latent variable space; It is a Jacobian matrix.

[0012] Secondly, this application provides a joint distribution modeling system for wind field data based on a standardized flow model, comprising: The data acquisition module is used to acquire wind field monitoring data and preprocess the wind field monitoring data; The model building module is used to construct a standardized flow model framework. The standardized flow model framework establishes a bidirectional invertible mapping between the original data space and the latent variable space through the training module. The original data space is used to characterize the true multidimensional joint distribution of wind field monitoring data. The prior distribution of the latent variable space is assumed to be a normal distribution. The training module includes n coupling layers, and each coupling layer includes an affine coupling block and a random permutation layer. The model training module is used to train the standardized flow model framework using preprocessed data to obtain a trained standardized flow model. The data processing module is used to generate samples of the original data space based on the standardized flow model by sampling from the latent variable space and performing an inverse transformation of the bidirectional invertible mapping, and / or to calculate the multidimensional joint probability density of any point in the original data space by performing a forward transformation of the bidirectional invertible mapping and variable transformation formulas.

[0013] According to the specific embodiments provided in this application, this application has the following technical effects: 1. The standardized flow model constructed in this application can flexibly map a simple prior distribution into a complex multidimensional joint distribution through reversible transformation, thereby accurately capturing the complex dependencies such as nonlinearity and asymmetry between key wind field parameters such as wind direction and wind speed. It does not require prior assumptions about the correlation structure between variables, but directly learns its inherent probabilistic dependency characteristics based on actual monitoring data, which greatly improves the joint probability model's ability to represent the stochastic characteristics of actual wind fields and provides a more accurate uncertainty quantification basis for subsequent wind load and engineering structure reliability analysis.

[0014] 2. Based on the standardized flow model, data sampling only needs to be performed from a simple distribution and then undergo affine transformation to quickly generate a large number of samples that are highly consistent with the original monitoring data in terms of statistical characteristics, thereby improving the efficiency and feasibility of wind load design and structural safety and reliability assessment. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating the overall method in one embodiment of this application; Figure 2 This is a schematic diagram of the principle of an affine coupling block in one embodiment of this application; Figure 3 This is a schematic diagram illustrating the principle of the random permutation layer in one embodiment of this application; Figure 4 This is a schematic diagram of bidirectional mapping in one embodiment of this application; Figure 5 This is a data statistics histogram and scatter plot in one embodiment of this application; Figure 6 This is a schematic diagram of wind speed data edge distribution modeling in one embodiment of this application; Figure 7 This is a schematic diagram of wind direction data edge distribution modeling in one embodiment of this application; Figure 8 This is a joint probability density map of three-dimensional wind speed and direction data in one embodiment of this application; Figure 9 This is a joint probability density map of two-dimensional wind speed and direction data in one embodiment of this application; Figure 10 This is a scatter plot of wind speed and direction data sampling in one embodiment of this application; Figure 11This is a schematic diagram of the cumulative probability density of wind speed under given wind direction conditions in one embodiment of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] This embodiment discloses a joint distribution modeling method for wind field monitoring data based on a standardized flow model, such as... Figure 1 As shown, specifically: S1. Acquire wind field monitoring data and preprocess the wind field monitoring data; S2. Construct a standardized flow model framework. The standardized flow model framework establishes a bidirectional invertible mapping between the original data space and the latent variable space through the training module. The original data space is used to represent the true multidimensional joint distribution of wind field monitoring data. The prior distribution of the latent variable space is assumed to be a normal distribution. The training module includes n coupling layers. Each coupling layer includes an affine coupling block and a stochastic permutation layer. S3. Train the standardized flow model framework using preprocessed data to obtain a trained standardized flow model; S4. Based on the standardized flow model, samples are generated from the latent variable space by sampling and inverse transformation of the bidirectional invertible mapping, and / or the multidimensional joint probability density of any point in the original data space is calculated by forward transformation of the bidirectional invertible mapping and variable transformation formula.

[0020] In one embodiment, the training steps for the standardized flow model are as follows: S3.1 Acquire and preprocess wind field monitoring data to obtain the original training data matrix; S3.2 Build a standardized flow model framework and train the standardized flow model framework using the original training data matrix. Optimize the model parameters based on the loss function to obtain the trained standardized flow model.

[0021] Furthermore, the standardized flow model framework is a bidirectional mapping model training framework. Then, the loss function and model training parameters are set to complete the training of the standardized flow model and obtain the bidirectional mapping relationship.

[0022] In one embodiment, the steps for obtaining the original training data matrix are as follows: S3.1.1 Obtain real-time wind field monitoring data; S3.1.2 Preprocess the wind field monitoring data, and convert the preprocessed data with a uniform format into a data matrix to obtain the original training data matrix.

[0023] Furthermore, the wind field monitoring data in S3.1.1 typically originates from sensing devices installed at various points on the building structure, and the monitoring is real-time and continuous. In this embodiment, the daily maximum wind speed and corresponding wind direction values ​​are used as data sample points, and a total of 1052 training data points are obtained after data processing.

[0024] Furthermore, S3.1.2 involves preprocessing the acquired wind field monitoring data (including but not limited to wind direction, wind speed, and other wind field monitoring data), specifically as follows: 1) Remove incomplete data points, such as: [wind direction, wind speed, temperature] = [20°, null value, 36℃]; 2) Remove outlier data points, such as removing data points like: [wind direction, wind speed, temperature] = [20°, 30m / s, 3600℃], resulting in... D Complete set of data points.

[0025] Converting the text data or other formatted data processed by 1) and 2) into a data matrix as model input yields the original training data (i.e., the original training data matrix): , ; in, , ,…, ; In the formula, The original training data is a D OK N A matrix of columns, containing N Different types of data (such as , (These can represent wind direction and wind speed data respectively). The training data for the first dimension is a D The column vector of the row, containing D Data points.

[0026] In one embodiment, the standardized flow model framework built in S2 mainly includes: the original data space X (a complex multidimensional joint distribution space) - the training module (containing multiple layers of affine coupling blocks and random permutation layers) - the latent variable space Z (a simple design distribution space).

[0027] In one embodiment, the training module includes a plurality of sequentially connected affine coupling blocks and a random permutation layer; The affine coupling block is configured to: divide the input data into a first part and a second part, perform a direct copy operation on the first part, and perform a reversible affine transformation operation on the second part; The random permutation layer is configured to randomly shuffle the dimensional order of the output data of the affine coupling block.

[0028] Furthermore, the execution steps of the standardized flow model are as follows: S2.1 The complex multidimensional joint distribution obtained in the original data space X using the method proposed in this application is denoted as... ; S2.2. In the latent variable space, the basic distribution can be designed according to the training objective. To facilitate obtaining the distribution of the original data space, the latent variable space distribution is set to a Gaussian distribution in this embodiment. ; S2.3 The training module is the core part of the standardized flow model framework, and its main core components are the affine coupling block and the random permutation layer.

[0029] The function of the affine coupling block used can be explained as follows: 1) First, divide the input data into two parts: ; In the formula, The first part of the data has the following dimensions: d , The second part of the data has the following dimensions. Dd ,and d=D / 2 ; 2) For Copy directly to the next layer output: ; In the formula, This is the first part of the data for the next layer.

[0030] 3) For After performing an affine transformation, the output is sent to the next layer: Forward transform: ; Inverse transform: ; but: ; Therefore, a first-order positive transformation is denoted as: The inverse transformation can be denoted as: ; In the formula, This is the second part of the output data for the positive transform; and The output of the neural network is the real-valued output (the two sub-neural networks in the affine coupling block can be any neural network that can be trained by gradient descent; this embodiment uses a 3-layer fully connected neural network with an input dimension of 2, an intermediate layer dimension of 64, and an output dimension of 2, using ReLU as the activation function); For neural network parameters; The Hadamard product, which involves multiplying vector elements one by one, follows the workflow of an affine coupling block as follows: Figure 2 As shown.

[0031] The random permutation layer used in this embodiment can be explained as follows: To ensure that each dimension of the input data is fully trained, a random permutation layer is superimposed after each affine coupling block, denoted as a coupling layer, to train the data. The order of dimensions is shuffled before being input into the next coupled layer for training. Let's take a simple six-dimensional example to illustrate this: 1) Initialize a random index sequence: ; 2): Output the coupling block from the previous step. Perform random permutations: The working principle is as follows Figure 3 As shown.

[0032] A complete standardized flow model framework is constructed: based on S2.1, S2.2, and S2.3, the affine coupling blocks and random permutation layers are alternately stacked to construct a framework that includes... n A complete normalized flow model framework (using 6 layers in this embodiment) is established (affine coupling block + random permutation layer). The bidirectional mapping from the original data space X to the latent variable space Z can then be expressed as: Forward transform: ; Inverse transform: ; Based on the bidirectional mapping relationship, for each Each value has a unique corresponding value. value: The same applies to the opposite. Calculate the probability density of the data based on the variable transformation formula: ; In the formula, This represents the probability density of the original data space. Let be the probability density of the latent variable space. Let be the Jacobian matrix. The principle of bidirectional mapping is as follows: Figure 4 As shown.

[0033] The specific steps for training the standardized flow model using the original training data matrix in S3.2 are as follows: S3.2.1 Determining the Loss Function: In this embodiment, the loss function is determined by minimizing the negative log-likelihood function, and the neural network is trained accordingly. Assume the original training data... Since the data points are independent and identically distributed, the log-likelihood function can be written as follows according to the variable transformation formula: ; Therefore, the model loss function is defined as: ; S3.2.2 Set appropriate model parameters, such as: neural network dimension and number of layers, learning rate, training epochs, batch size, etc. S3.2.3 During model training, 80% of the original data (i.e., the original training data matrix) is set as the training set and 20% as the validation set. The model is evaluated by the negative log-likelihood value on the validation set. The lower the negative log-likelihood value, the better the model training effect.

[0034] In one embodiment, the standardized flow model trained through the above steps is used to perform data sampling and probability density calculation of the original space based on the bidirectional mapping relationship and probability density calculation formula, thereby obtaining the multidimensional joint distribution modeling result of the wind field monitoring data. The specific steps are as follows: S4.1 Sampling: 1) First, randomly sample standard normal space samples (norm-samples) from the latent variable space Z (simple design space); 2) Input the data into the trained model for backpropagation, i.e., perform the inverse transformation. This yields the original data space samples (Data-samples).

[0035] S4.2 Joint probability density calculation: 1) Divide the domain of data distribution into I×J The grid, with each grid point The input model is propagated forward, i.e., a forward transform is performed. This will yield two parts of the model output: Data transformed from the original data space X to the latent variable space Z: ; The logarithmic value of the Jacobian determinant corresponding to the transformation: From the introduction of affine coupling blocks, we know that the Jacobian determinant after one transformation is: is a lower triangular matrix, and the Jacobian determinant of a first transformation can be obtained by multiplying its diagonal elements. After taking the logarithm: For hierarchical structures The logarithmic determinant of the overall Jacobian matrix is ​​for each layer. n The sum of the corresponding logarithmic determinants, therefore .

[0036] 2) Calculate the probability density of the original data space using the variable transformation formula: ; In the formula, For grid points in the domain; Let be the probability density of the latent variable space; The positive transformation function of the model; For neural network parameters; For affine transformation data of grid points; This represents the probability density of the original data space. It is a Jacobian matrix.

[0037] In one embodiment, taking the joint distribution modeling of two-dimensional wind speed and wind direction observation data and extreme wind speed prediction as an example, the above method is further explained: S1. The wind field data used in this embodiment is the monitoring data at the mid-span of a long-span bridge deck. The daily maximum wind speed and its corresponding wind direction are selected as the data source, and the observation period is 3 years. The acquired wind direction and wind speed data are processed by data cleaning, input format conversion, etc., to obtain the original training data (a total of 1052 data sample points): ; in, It is a data set with 1052 rows and 2 columns. For wind speed data, For wind direction data, its scatter plot and statistical histogram are as follows: Figure 5 As shown in the figure, the wind speed data distribution is relatively simple, approximating a log-normal distribution, but the wind direction data exhibits a multi-peak characteristic, making it a more complex distribution. Furthermore, the scatter plot reveals that the joint distribution of wind direction and wind speed has multimodal characteristics, with wind speed varying greatly along the wind direction, indicating a strong nonlinear correlation between the two.

[0038] S2. Construct a standardized flow model framework. In this embodiment, the training module uses 6 coupling layers, i.e., 6 affine coupling blocks and random permutation layers are alternately superimposed. The latent variable space Z is designed as a simple two-dimensional standard normal joint distribution. ; S3. In this embodiment, the loss function is set to minimize the negative log-likelihood of the observed data under the model: ; According to the probability density calculation formula of the original space, It can be represented as: ; The parameters for this embodiment are set as follows: 1000 training epochs, learning rate 0.001, batch size 512. After completing model training, a bidirectional mapping relationship between the original data space X and the latent variable space Z is obtained, namely: , ; S4. Using the model trained in S3, based on the bidirectional mapping relationship and probability density calculation formula, the model is used to sample data from the original space and calculate the probability density function to obtain the multidimensional joint distribution modeling results of the wind field monitoring data.

[0039] The edge distribution modeling results of wind speed and wind direction data are as follows: Figure 6 and Figure 7 As shown, the method in this embodiment can still capture these multi-peak features well even for wind direction data with complex multi-peak characteristics; the modeling results of the joint probability density function of wind direction and wind speed data are as follows. Figure 8 and Figure 9 As shown, the results indicate that the joint distribution of wind direction and wind speed data exhibits complex multimodal characteristics, and there is a strong nonlinear correlation between the two. The sampling results of the original data space based on the trained bidirectional mapping model are as follows: Figure 10 As shown, the results demonstrate a good correspondence between the generated data and the original data, as well as the ability to capture nonlinear and multimodal characteristics.

[0040] In summary, the method described in this embodiment can construct a reasonable joint distribution for wind field data with multi-peak characteristics and nonlinear correlation. Furthermore, this method is not limited to two-dimensional data and can be easily extended to three-dimensional or even higher-dimensional data, simply by changing the data input dimension.

[0041] In one embodiment, based on the constructed joint probability density function of wind speed and wind direction Predict extreme wind speeds under different wind directions: 1) Calculate the given wind direction cumulative probability density function of wind speed under the given conditions ; ; 2) Prediction T =Extreme wind speed with a 50-year return period, return period T With the cumulative probability density function of wind speed The relationship between them can be represented as follows: .

[0042] Figure 11 Show different wind directions wind speed Cumulative probability density function Table 1 shows the predicted extreme wind speeds under different wind directions. It can be observed that the predicted extreme wind speeds vary with different wind directions, indicating that the traditional method of using only wind speed as a single variable for extreme value prediction is unreasonable. The calculated extreme wind speeds can be used as standard load values ​​in structural design, providing key wind load input parameters and design basis for structural reliability design, thereby guiding the wind-resistant design and safety assessment of structures.

[0043] Table 1. Extreme wind speed prediction under given wind direction conditions.

[0044] On the other hand, a joint distribution modeling system for wind field data based on a standardized flow model is also disclosed, including: The data acquisition module is used to acquire wind field monitoring data and preprocess the wind field monitoring data; The model building module is used to construct a standardized flow model framework. The standardized flow model framework establishes a bidirectional invertible mapping between the original data space and the latent variable space through the training module. The original data space is used to represent the true multidimensional joint distribution of wind field monitoring data. The prior distribution of the latent variable space is assumed to be a normal distribution. The training module includes n coupling layers, and each coupling layer includes an affine coupling block and a random permutation layer. The model training module is used to train the standardized flow model framework using preprocessed data to obtain a trained standardized flow model. The data processing module is used to generate samples of the original data space based on the standardized flow model by sampling from the latent variable space and performing an inverse transformation of a bidirectional invertible mapping, and / or to calculate the multidimensional joint probability density of any point in the original data space by performing a forward transformation of a bidirectional invertible mapping and variable transformation formulas.

[0045] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0046] This embodiment uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application; at the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. In summary, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for joint distribution modeling of wind field data based on a standardized flow model, characterized in that, The specific steps are as follows: Acquire wind field monitoring data and preprocess the wind field monitoring data; A standardized flow model framework is constructed, which establishes a bidirectional invertible mapping between the original data space and the latent variable space through a training module; the original data space is used to characterize the true multidimensional joint distribution of wind field monitoring data; the prior distribution of the latent variable space is assumed to be a normal distribution; the training module includes n coupling layers, each coupling layer including an affine coupling block and a random permutation layer; The standardized flow model framework is trained using preprocessed data to obtain a trained standardized flow model; Based on the standardized flow model, samples of the original data space are generated by sampling from the latent variable space and performing an inverse transformation of the bidirectional invertible mapping, and / or the multidimensional joint probability density of any point in the original data space is calculated using the forward transformation of the bidirectional invertible mapping and the variable transformation formula.

2. The method for joint distribution modeling of wind field data based on a standardized flow model according to claim 1, characterized in that, The training steps for the standardized flow model framework are as follows: The wind field monitoring data is acquired and preprocessed to obtain the original training data matrix; The standardized flow model framework is constructed, and the standardized flow model framework is trained using the original training data matrix. The model parameters are optimized based on the loss function to obtain the trained standardized flow model.

3. The method for joint distribution modeling of wind field data based on a standardized flow model according to claim 2, characterized in that, The steps for obtaining the original training data matrix are as follows: Obtain the real-time wind field monitoring data; The wind field monitoring data is preprocessed, and the preprocessed data with a unified format is converted into a data matrix to obtain the original training data matrix.

4. The method for joint distribution modeling of wind field data based on a standardized flow model according to claim 1, characterized in that, The training module includes multiple sequentially connected affine coupling blocks and random permutation layers; The affine coupling block is configured to: divide the input data into a first part of data and a second part of data, perform a direct copy operation on the first part of data, and perform a reversible affine transformation operation on the second part of data; The random permutation layer is configured to randomly shuffle the dimensional order of the output data of the affine coupling block.

5. The method for joint distribution modeling of wind field data based on a standardized flow model according to claim 4, characterized in that, The expression for the affine coupling block is: ； In the formula, This is the first part of the data; This is the second part of the data; and All are real-valued outputs of the neural network; These are the parameters of the neural network.

6. The method for joint distribution modeling of wind field data based on a standardized flow model according to claim 5, characterized in that, The expression for a reversible affine transformation operation is: Forward transform: ; Inverse transform: ; In the formula, This is the second part of the output data for the positive transform; For neural network parameters; For Hadama accumulation.

7. The method for joint distribution modeling of wind field data based on a standardized flow model according to claim 1, characterized in that, The expression for the probability density function of the multidimensional joint distribution is: ； In the formula, This represents the probability density of the original data space. Input data; For neural network parameters; The positive transformation function of the model; Let be the probability density of the latent variable space; It is a Jacobian matrix.

8. A joint distribution modeling system for wind field data based on a standardized flow model, characterized in that, include: The data acquisition module is used to acquire wind field monitoring data and preprocess the wind field monitoring data; The model building module is used to construct a standardized flow model framework. The standardized flow model framework establishes a bidirectional invertible mapping between the original data space and the latent variable space through the training module. The original data space is used to characterize the true multidimensional joint distribution of wind field monitoring data. The prior distribution of the latent variable space is assumed to be a normal distribution. The training module includes n coupling layers, and each coupling layer includes an affine coupling block and a random permutation layer. The model training module is used to train the standardized flow model framework using preprocessed data to obtain a trained standardized flow model. The data processing module is used to generate samples of the original data space based on the standardized flow model by sampling from the latent variable space and performing an inverse transformation of the bidirectional invertible mapping, and / or to calculate the multidimensional joint probability density of any point in the original data space by performing a forward transformation of the bidirectional invertible mapping and variable transformation formulas.

Images