A method for simulating stratospheric wind fields

Abstract

This application discloses a stratospheric wind field simulation method, comprising: taking stratospheric wind field sample data and a conditional control vector as input, extracting features from the stratospheric wind field sample data through a conditional encoder, and mapping them to a low-dimensional latent space vector; gradually adding random noise to the latent space vector to obtain a noisy latent space vector; taking the conditional control vector and the noisy latent space vector as input, predicting the noise through a denoising neural network model of a latent space diffusion model, performing an inverse denoising process, and obtaining a denoised latent space vector; inputting the denoised latent space vector and the conditional control vector into a conditional decoder for decoding to obtain stratospheric wind field simulation data; and taking the stratospheric wind field simulation data as input, obtaining higher-resolution stratospheric wind field simulation data through a spatiotemporal super-resolution network model. This application can efficiently generate stratospheric wind field simulation data.

Description

Technical Field

[0001] The present invention relates to the fields of deep learning and wind field simulation, and in particular to a stratospheric wind field simulation method. Background Art

[0002] As a key platform for high-resolution Earth observation systems, stratospheric aerostats possess the ability to execute long-duration, all-weather, and real-time missions, and are of strategic importance to promoting social, economic, national defense, and scientific and technological development. With the continuous advancement of artificial intelligence (AI), the flight paths of stratospheric aerostats equipped with intelligent navigation capabilities have become controllable, offering broad market application prospects and undoubtedly a leading direction for the future development of stratospheric aerostats.

[0003] Research on intelligent navigation technology for stratospheric aerostats relies on extensive flight test data. However, stratospheric aerostats are expensive, and conducting stratospheric flight tests to obtain flight data for intelligent navigation is slow and expensive. In extreme cases, a stratospheric aerostat may require weeks of maneuvering flights to recover from control decision errors. Therefore, establishing a simulation environment to train intelligent navigation controllers for stratospheric aerostats is crucial. Through flight experiments in a simulated environment, the training cost of intelligent navigation controllers can be greatly reduced, while also saving a considerable amount of time.

[0004] Simulating stratospheric winds is a critical step in constructing a flight simulation environment for stratospheric aerostats. Simulating stratospheric winds not only needs to conform to the characteristics of real stratospheric winds but also needs to exhibit randomness and temporal and spatial correlation with real stratospheric winds. Given the large number of simulated flight experiments required, generating stratospheric wind simulation data requires a balance between simulation quality and diversity while also balancing efficiency.

[0005] Currently, there are three main methods for simulating wind fields: one is the numerical simulation method based on computational fluid dynamics (CFD), the second is to construct a simulated wind field based on spectral representation, and the third is to construct a simulated wind field based on deep learning methods.

[0006] The numerical simulation method based on CFD has many advantages in simulating the fine flow field of the boundary layer under complex terrain conditions, such as short calculation time, high calculation efficiency, and the ability to simulate the characteristics of the real wind field. However, since this method needs to consider physical processes at multiple scales when simulating the wind field, it is necessary to calculate various fluid dynamics equations when using this method to simulate the wind field. The calculation complexity is high and a large amount of computing resources and time are required. Moreover, the wind field characteristics finally generated are difficult to control, and the parameters need to be adjusted repeatedly to obtain a suitable simulated wind field. The numerical simulation method based on CFD firstly requires setting initial conditions and boundary conditions, and secondly, the generation of simulated wind fields based on the numerical simulation method based on CFD usually involves a large amount of numerical calculations. In particular, when a large amount of stratospheric wind field simulation data needs to be generated, the efficiency of the numerical simulation method based on CFD is greatly reduced.

[0007] Constructing wind field simulations based on spectral representations primarily involves Fourier synthesis, orthogonal decomposition, and numerical filtering. These methods discretize the power spectrum of a specific wind field into multiple segments and then combine them with random sequence synthesis to generate a simulated wind field. This method can meet requirements for wind field simulation, such as turbulence intensity and spectral characteristics, and is particularly suitable for simulating steady wind fields. To improve the efficiency of wind field simulations, these methods often incorporate fast Fourier transform (FFT) technology. A key challenge in integrating FFT technology is effectively decoupling time-frequency coupled functions. Currently, existing methods include proper orthogonal decomposition (POD), wavelet decomposition, and non-negative matrix factorization. However, as the number of simulation points increases significantly, the computational complexity of these decoupling processes increases dramatically. Furthermore, when the coherence functions of the simulated wind field exhibit time-varying characteristics, a significant number of spectral matrix decomposition calculations are required, further reducing the overall efficiency of the wind field simulation. Furthermore, these methods are typically based on limited wind field characteristics. Given the complex and variable nature of wind field characteristics in reality, there will always be a certain gap between the simulated wind field characteristics generated by these methods and the actual wind field characteristics. Constructing simulated wind fields based on spectral representation is mainly used to generate simulated wind fields that meet specific wind field characteristics, but it lacks universality. For the task of generating a large amount of stratospheric wind field simulation data, it is necessary not only to meet efficiency requirements but also to ensure the diversity of the generated stratospheric wind field simulation data.

[0008] Simulating wind fields using deep learning methods primarily relies on generative networks such as generative adversarial networks (GANs) or variational autoencoders (VAEs) to generate stratospheric wind data. However, GANs are prone to instability and mode collapse when generating stratospheric wind data. The GAN training process involves two networks, the generator and the discriminator, which compete with each other. This adversarial training makes the GAN training process highly unstable, prone to oscillations and non-convergence. This results in the GAN failing to consistently generate samples that meet expectations when generating stratospheric wind data. GANs can also experience mode collapse during training, where the generator is limited to generating a limited number of samples and fails to cover the entire data distribution. In stratospheric wind data simulation tasks, this can result in a lack of diversity in the generated stratospheric wind data samples, failing to accurately reflect the complexity and variability of the actual stratospheric wind field. VAEs, on the other hand, tend to learn a latent representation of the input data during training and attempt to generate samples similar to the training data. However, stratospheric wind simulation data is a complex and ever-changing system, and its data distribution can be highly complex and uncertain. VAEs may not fully capture this complex distribution, resulting in significant discrepancies between the generated stratospheric wind simulation data samples and the actual stratospheric wind field conditions. Furthermore, since VAE training aims to minimize the error between the output data and the input data, it tends to generate samples that are closer to the training data, which can lead to a lack of sufficient diversity in the generated stratospheric wind simulation data samples.

[0009] In view of the shortcomings of existing wind field simulation methods, it is of great significance to study and develop a more accurate and efficient wind field simulation method. Summary of the Invention

[0010] In response to the deficiencies of the prior art, the present application provides a stratospheric wind field simulation method that can simultaneously meet the accuracy and efficiency of stratospheric wind field simulation.

[0011] The technical solutions provided in this application are:

[0012] In a first aspect, the present application provides a stratospheric wind field simulation method, comprising:

[0013] Step S1: constructing a conditional control network, inputting conditional information for controlling the stratospheric wind field into the conditional control network, and obtaining a conditional control vector;

[0014] Step S2: Taking the stratospheric wind field sample data and the conditional control vector as input, extracting features from the stratospheric wind field sample data through the conditional encoder in the pre-trained encoder-decoder model and mapping them to a low-dimensional latent space vector; wherein the encoder-decoder model includes two parts: the conditional encoder and the conditional decoder;

[0015] Step S3: gradually adding random noise to the latent space vector (gradually diffusing the latent space vector) to obtain a noisy latent space vector; using the conditional control vector and the noisy latent space vector as input, predicting the noise through a pre-trained latent space diffusion model (Latent Diffusion Model, LDM) denoising neural network model, performing a reverse denoising process to obtain a denoised latent space vector; inputting the denoised latent space vector and the conditional control vector into a conditional decoder in a pre-trained encoder-decoder model for decoding to obtain first stratospheric wind field simulation data;

[0016] Step S4: using the first stratospheric wind field simulation data as input, and obtaining second stratospheric wind field simulation data through a pre-trained spatiotemporal super-resolution network model; wherein the resolution of the second stratospheric wind field simulation data is higher than the resolution of the first stratospheric wind field simulation data.

[0017] Based on the spatiotemporal super-resolution network model, high-resolution stratospheric wind field simulation data can be obtained from low-resolution stratospheric wind field simulation data. Therefore, the present application can obtain high-resolution stratospheric wind field simulation data.

[0018] In one possible implementation, the conditional control network in step S1 includes a numerical coding layer, a non-numerical coding layer, and three linear layers, wherein the numerical coding layer is used to encode numerical conditions, and the non-numerical coding layer is used to encode non-numerical conditions. The outputs of the numerical coding layer and the non-numerical coding layer are concatenated and input into the three linear layers to obtain the conditional control vector; among the three linear layers, the first linear layer serves as an input layer to receive the output of the numerical coding layer and the non-numerical coding layer, the second linear layer serves as a hidden layer, and the third linear layer serves as an output layer to output the conditional control vector.

[0019] Among them, numerical conditions include time, etc.; non-numerical conditions include the initial state of the wind farm.

[0020] In one possible implementation, the conditional encoder includes a plurality of groups of spatiotemporal downsampling modules and spatiotemporal attention modules connected in sequence; each group of spatiotemporal downsampling modules and spatiotemporal attention modules includes a spatiotemporal downsampling module and a spatiotemporal attention module connected in sequence;

[0021] Each spatiotemporal downsampling module includes a spatial downsampling unit and a temporal downsampling unit connected in sequence;

[0022] The spatial downsampling unit includes a plurality of three-dimensional spatial convolution layers connected in sequence, which are used to process the spatial dimension information of the stratospheric wind field sample data; the temporal downsampling unit includes a one-dimensional temporal convolution layer, which is used to process the temporal dimension information of the stratospheric wind field sample data; each convolution layer is followed by a residual layer;

[0023] Each spatiotemporal attention module includes a spatial attention layer and a temporal attention layer connected in sequence, which are used to process the spatial dimension information and temporal dimension information of the stratospheric wind field sample data respectively;

[0024] Sequential connection means that the output of the previous module, layer or unit is used as the input of the next module or layer;

[0025] Since the stratospheric wind field sample data contains both spatial dependencies and temporal dependencies, the method described in the present invention performs attention calculations on space and time respectively. The spatial attention layer only processes the spatial dimension information of the stratospheric wind field sample data, and the temporal attention layer only processes the temporal dimension information of the stratospheric wind field sample data. While simplifying the calculation complexity, it can also capture the spatiotemporal dependencies of the stratospheric wind field sample data.

[0026] The conditional encoder receives the conditional control vector and the stratospheric wind field sample data as input, and the stratospheric wind field sample data is input into the first three-dimensional spatial convolution layer in the conditional encoder; the other three-dimensional spatial convolution layers except the first three-dimensional spatial convolution layer in the conditional encoder, and the one-dimensional temporal convolution layer, take the output of the previous module, layer or unit and the conditional control vector as input;

[0027] In one possible implementation, the conditional encoder includes two sets of spatiotemporal downsampling modules and a spatiotemporal attention module. The spatial downsampling unit includes two three-dimensional spatial convolutional layers. The following details the input-output relationships between each module, layer, and unit.

[0028] The conditional encoder receives a conditional control vector and stratospheric wind field sample data as input. The stratospheric wind field sample data is first input into the first spatiotemporal downsampling module of the conditional encoder. The data is first processed by the first three-dimensional spatial convolution layer of the spatial downsampling unit and then passes through the residual layer. The resulting output, combined with the conditional control vector, is input into the second three-dimensional spatial convolution layer and the second residual layer to obtain the output of the spatial downsampling unit. The output of the spatial downsampling unit, combined with the conditional control vector, is input into the temporal downsampling unit to obtain the output of the first spatiotemporal downsampling module. The output of the first spatiotemporal downsampling module is input into the first spatiotemporal attention module of the conditional encoder. It first passes through the spatial attention layer of the first spatiotemporal attention module and then passes through the temporal attention layer of the first spatiotemporal attention module to obtain the output of the first spatiotemporal attention module. The output of the first spatiotemporal attention module is input into the second spatiotemporal downsampling module of the conditional encoder to obtain the output of the second spatiotemporal downsampling module. The output of the second spatiotemporal downsampling module is input into the second spatiotemporal attention module of the conditional encoder to obtain the output of the conditional encoder, i.e., the latent space vector. The data processing process of the second spatiotemporal downsampling module and the second spatiotemporal attention module is the same as that of the first spatiotemporal downsampling module and the first spatiotemporal attention module.

[0029] The conditional decoder includes multiple groups of sequentially connected spatiotemporal upsampling modules and spatiotemporal attention modules; each group of spatiotemporal upsampling modules and spatiotemporal attention modules includes a sequentially connected spatiotemporal upsampling module and a spatiotemporal attention module;

[0030] The spatiotemporal upsampling module includes a spatial upsampling unit and a temporal upsampling unit; the spatial upsampling unit includes a plurality of sequentially connected three-dimensional spatial convolutional layers for upsampling the spatial information of the stratospheric wind field sample data; the temporal upsampling unit includes a one-dimensional temporal convolutional layer for upsampling the temporal information of the stratospheric wind field sample data; each convolutional layer is followed by a residual layer; four-dimensional spatiotemporal stratospheric wind field simulation data is obtained through the spatiotemporal upsampling module.

[0031] In some embodiments, the conditional decoder includes two sets of spatiotemporal upsampling modules and a spatiotemporal attention module. The spatial upsampling unit includes two three-dimensional spatial convolutional layers. The following details the input-output relationships between each module, layer, and unit.

[0032] The conditional decoder accepts the conditional control vector and the denoised latent space vector as input. The denoised latent space vector is first input into the first three-dimensional spatial convolution layer of the spatial upsampling unit of the first spatiotemporal upsampling module of the conditional decoder, passes through the residual layer, and then is input into the second three-dimensional spatial convolution layer and the second residual layer in combination with the conditional control vector. It is then processed by the temporal upsampling unit to obtain the output of the first spatiotemporal upsampling module. The output of the first spatiotemporal upsampling module is input into the first spatiotemporal attention module of the conditional decoder, and the obtained output is input into the second spatiotemporal upsampling module of the conditional decoder to obtain the output of the second spatiotemporal upsampling module, which is input into the second spatiotemporal attention module of the conditional decoder to obtain the output of the conditional decoder, i.e., the first stratospheric wind field simulation data. The data processing process of the second spatiotemporal upsampling module and the second spatiotemporal attention module is the same as that of the first spatiotemporal upsampling module and the first spatiotemporal attention module.

[0033] In combination with the above embodiments, the encoder-decoder model may include two spatiotemporal downsampling modules and two spatiotemporal upsampling modules, and each downsampling / upsampling module is followed by a spatiotemporal attention module.

[0034] In one possible implementation, the pre-training process of the encoder-decoder model includes:

[0035] Step S21: constructing an encoder-decoder model and initializing parameters of the encoder-decoder model;

[0036] Step S22: inputting the condition information for controlling the stratospheric wind field into the conditional control network to obtain a conditional control vector; inputting the stratospheric wind field sample data and the conditional control vector into the encoder-decoder model to obtain stratospheric wind field simulation data generated by the encoder-decoder model;

[0037] Step S23: calculating a first loss function value based on the stratospheric wind field simulation data and the stratospheric wind field sample data;

[0038] Step S24: updating the parameters of the encoder-decoder model based on back propagation according to the first loss function value;

[0039] Step S22: Iterate steps S22 to S24 until the iteration end condition is reached to obtain a pre-trained encoder-decoder model.

[0040] In a possible implementation, the first loss function in step S23 is calculated using the following formula:

[0041]

[0042] Wherein, Los(r,x) is the first loss function, r represents the generated stratospheric wind field simulation data, x represents the stratospheric wind field sample data, mse(r,x) represents the mean square error between r and x, σ represents the variance of the generated stratospheric wind field simulation data, KL(r,x) represents the KL divergence between r and x, and γ represents the weight of the KL divergence.

[0043] The mean square error is calculated by the following formula:

[0044]

[0045] Where N represents the number of grid points in a stratospheric wind field, r i represents the wind speed value of the i-th grid point in the generated stratospheric wind field simulation data, x i Represents the wind speed value of the i-th grid point in the stratospheric wind field sample data.

[0046] The KL divergence is calculated by the following formula:

[0047]

[0048] Among them, μ r and σ r represent the mean and variance of the generated stratospheric wind field simulation data, μ x and σ x They represent the mean and variance of the corresponding stratospheric wind field sample data used for training.

[0049] The above loss function combines L2 loss (MSE) and KL divergence. L2 loss focuses on the specific differences between the simulated stratospheric wind data and the sample stratospheric wind data, while KL divergence focuses on the difference between the probability distribution of the simulated stratospheric wind data and the sample stratospheric wind data. This combination helps the model strike a balance between prediction accuracy and probability distribution similarity.

[0050] In a possible implementation, the denoising neural network model of the latent space diffusion model in step S3 includes a downsampling module, an intermediate module, and an upsampling module connected in sequence;

[0051] The downsampling module includes a plurality of downsampling units connected in sequence;

[0052] Using multiple downsampling units can gradually reduce the spatial and temporal resolution of the data while retaining key feature information for easier subsequent processing. Multiple downsampling units have the same structure but different parameters to better adapt to the characteristics of the data at different stages.

[0053] The downsampling unit includes a three-dimensional spatial convolution layer, a one-dimensional temporal convolution layer, a residual layer, and a spatiotemporal attention module connected in sequence to process the spatial and temporal dimension information of the stratospheric wind field sample data. The three-dimensional spatial convolution layer is responsible for processing the spatial dimension information, and the one-dimensional temporal convolution layer focuses on feature extraction in the temporal dimension. The residual layer is introduced to enhance the network's feature propagation capability and reduce information loss. The spatiotemporal attention module further enhances the model's feature selection capability in both spatial and temporal dimensions. The spatiotemporal attention module has the same structure as the spatiotemporal attention module in the encoder-decoder model.

[0054] The intermediate module includes a spatiotemporal attention module, and the spatiotemporal attention module is provided with residual layers before and after;

[0055] The upsampling module includes a plurality of upsampling units connected in sequence, each upsampling unit includes a four-dimensional spatiotemporal convolution module and a spatiotemporal attention module connected in sequence; the four-dimensional spatiotemporal convolution module includes a three-dimensional spatial convolution layer, a one-dimensional temporal convolution layer and a residual layer connected in sequence.

[0056] Multiple upsampling units are used to gradually restore the spatial and temporal resolution lost during the downsampling process, while maintaining or enhancing key features. Similar to the downsampling units, each upsampling unit has the same structure but different parameters to accommodate changes in data characteristics during the inverse transformation. These upsampling units enable efficient mapping from low-resolution stratospheric simulation data to high-resolution stratospheric simulation data while maintaining the spatial and temporal consistency of the data.

[0057] In one possible implementation, the downsampling module includes two downsampling units, the intermediate module includes two residual layers, and the upsampling module includes two upsampling units. The following details the input-output relationships between each module, layer, and unit.

[0058] The denoising neural network model of the latent space diffusion model takes the conditional control vector and the noise latent space vector as input. The noise latent space vector is first input into the first downsampling module of the denoising neural network model, passes through the three-dimensional spatial convolution layer, one-dimensional temporal convolution layer and residual layer of the downsampling module, and then passes through the spatiotemporal attention module to obtain the output of the first downsampling module, which is input into the second downsampling module in combination with the conditional control vector, wherein the operation process of the second downsampling module is the same as that of the first downsampling module; the output of the second downsampling module is input into the intermediate module, first passes through the first residual layer of the intermediate module, then is processed by the spatiotemporal attention module of the intermediate module, and finally passes through the second residual layer of the intermediate module to obtain the output of the intermediate module; the output of the intermediate module is input into the first upsampling module of the denoising neural network model, first passes through the three-dimensional spatial convolution layer, one-dimensional temporal convolution layer and residual layer of the upsampling module, and then passes through the spatiotemporal attention module to obtain the output of the first upsampling module; the output of the first upsampling module is input into the second upsampling module of the denoising neural network model to obtain the output of the second upsampling module, which is the noise predicted by the denoising neural network model. The process of processing data in the second upsampling module of the denoising neural network model is the same as that in the first upsampling module of the denoising neural network model.

[0059] In one possible implementation, the pre-training process of the denoising neural network model of the latent space diffusion model includes:

[0060] Step S31: constructing a denoising neural network model of the latent space diffusion model and initializing the parameters of the denoising neural network model;

[0061] Step S32: inputting the conditional information for controlling the stratospheric wind field into the conditional control network to obtain a conditional control vector, inputting the stratospheric wind field sample data and the conditional control vector into the encoder part of the encoder-decoder model to obtain a latent space vector;

[0062] Step S33: gradually adding random noise to the latent space vector to obtain a noisy latent space vector; inputting the conditional control vector and the noisy latent space vector into a denoising neural network of a latent space diffusion model to obtain predicted noise, and performing a reverse denoising process based on the predicted noise to obtain a denoised latent space vector; inputting the denoised latent space vector and the conditional control vector into the conditional decoder for decoding to obtain stratospheric wind field simulation data;

[0063] Step S34: Calculating a second loss function value based on the stratospheric wind field simulation data obtained through the denoising neural network and decoding;

[0064] Step S35: updating the parameters of the denoising neural network model based on back propagation according to the second loss function value;

[0065] Step S36: Iterate steps S33 to S35 until the iteration end condition is reached to obtain a pre-trained encoder-decoder model.

[0066] In a possible implementation, the second loss function value in step S34 is calculated using the following formula:

[0067]

[0068] Among them, Loss represents the second loss function, ε t represents the random noise added in step t in step S33, T represents the maximum number of steps, ε θ (X t , t) represents the predicted noise obtained by the denoising neural network model in step S33, Represents ε t -ε θ (X t ,t), γ represents the proportion of the difference between the stratospheric wind field simulation data and the stratospheric wind field sample data in the loss function, O represents the stratospheric wind field simulation data decoded by the conditional decoder, and Y represents the corresponding stratospheric wind field sample data used for training.

[0069] In a possible implementation, the spatiotemporal super-resolution network model in step S4 includes an initial feature extraction module, a feature fusion module and a reconstruction module, wherein the initial feature extraction module includes a first feature extraction unit, a downsampling layer and a second feature extraction unit connected in sequence; the first feature extraction unit and the second feature extraction unit both include a three-dimensional spatial convolution layer, a spatial attention layer and a residual layer connected in sequence; the feature fusion module includes a first feature fusion processing unit and a second feature fusion processing unit connected in sequence; the first feature fusion processing unit and the second feature fusion processing unit include two sub-processing units, each of which has two sub-processing units. Both include a four-dimensional spatiotemporal convolution module and a spatiotemporal attention module connected in sequence; the second feature fusion processing unit includes a three-dimensional spatial convolution layer, a spatiotemporal attention module and a residual layer connected in sequence; the output end of the initial feature extraction module is connected to the input end of the second feature fusion processing unit in the feature fusion module; the four-dimensional spatiotemporal convolution module includes a three-dimensional spatial convolution layer, a one-dimensional temporal convolution layer and a residual layer connected in sequence; the output end of the feature fusion module is connected to the input end of the reconstruction module; the reconstruction module includes a four-dimensional spatiotemporal convolution module, a spatiotemporal attention module, an upsampling layer, a residual layer and a spatiotemporal attention module connected in sequence.

[0070] In one possible implementation, the pre-training process of the spatiotemporal super-resolution network model includes the following steps:

[0071] Step S41: constructing a spatiotemporal super-resolution network model and initializing the parameters of the spatiotemporal super-resolution network model;

[0072] Step S42: inputting the three-dimensional wind field data of the first time point of the first stratospheric wind field sample data into the initial feature extraction module of the spatiotemporal super-resolution network model to obtain an initial feature vector I0;

[0073] Step S43: inputting the first stratospheric wind field sample data into the first feature fusion processing unit of the feature fusion module of the spatiotemporal super-resolution network model to obtain a second feature vector I1, concatenating the initial feature vector I0 and the second feature vector I1, and inputting the concatenated data into the second feature fusion processing unit of the feature fusion module to obtain a fused feature vector I2;

[0074] Step S44: inputting the fused feature vector I2 into the reconstruction module of the spatiotemporal super-resolution network model to obtain reconstructed third stratospheric wind field simulation data;

[0075] Step S45: calculating the L2 loss based on the second stratospheric wind field sample data corresponding to the first stratospheric wind field sample data and the third stratospheric wind field simulation data reconstructed by the spatiotemporal super-resolution network model; wherein the resolution of the third stratospheric wind field sample data is higher than the resolution of the first stratospheric wind field sample data; and the second stratospheric wind field sample data is the high-resolution stratospheric wind field sample data that actually corresponds to the first stratospheric wind field sample data;

[0076] Step S46: updating the parameters of the spatiotemporal super-resolution network model based on the L2 loss value through back propagation;

[0077] Step S47: Iterate steps S42 to S46 until the iteration end condition is reached to obtain a pre-trained spatiotemporal super-resolution network model.

[0078] In a second aspect, the present application provides an electronic device, comprising: a memory and a processor;

[0079] The memory is used to store computer programs;

[0080] The processor is configured to call the computer program to execute the method described above.

[0081] In a third aspect, the present application provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed on an electronic device, the electronic device implements the method described above.

[0082] In a fourth aspect, the present application provides a computer program product, including a computer program, which, when executed on an electronic device, enables the electronic device to implement the method described above.

[0083] The specific implementation methods of the second to fourth aspects of this application can refer to the implementation method of the first aspect above, and will not be repeated here.

[0084] The present invention is based on a latent space diffusion model for generating simulated wind fields. This model effectively combines the generation capabilities of the diffusion model with the spatiotemporal characteristics of a real wind field. By directly generating the simulated wind field using the latent space diffusion model, the complex spatiotemporal distribution characteristics of the wind field data are learned, avoiding the calculation of various fluid dynamics equations required for wind field simulation based on CFD methods. Simultaneously, the method of the present invention does not require the multiple spectral matrix decompositions and high computational complexity required when simulating a wind field based on spectral representation. By learning the complex spatiotemporal distribution characteristics of the wind field data, the complexity and computational complexity of the simulated wind field are reduced, thereby improving the simulation efficiency while meeting the accuracy and efficiency requirements of stratospheric wind field simulation.

[0085] The method described in the present invention constructs a conditional control network based on the latent space diffusion model, so that the latent space diffusion model can generate simulated wind fields in different scenarios according to different input conditions. The numerical condition information and non-numerical condition information are fused to obtain a conditional control vector, which is used to control the inverse denoising process of the encoder-decoder model and the diffusion model to generate simulated wind field data that meets the specified conditions.

[0086] The method described in the present invention constructs a spatiotemporal super-resolution network model. The resolution of the simulated wind field data generated based on the latent space diffusion model is the same as the resolution of the original stratospheric wind field sample data used for training. However, the resolution of the original stratospheric wind field sample data is generally lower. When high-resolution wind field data is required for stratospheric aerostat flight simulation, interpolation must be performed based on the simulated wind field data, which will lead to cumulative errors and affect the effectiveness of wind field simulation. Most existing super-resolution models are designed for image data and process two-dimensional data. Super-resolution operations on video data also add a time dimension to the two-dimensional image data. Existing super-resolution processing or downscaling operations on wind field data only target wind speed variables at a single altitude. Stratospheric wind field sample data includes data at different altitudes, and its data dimensions include three-dimensional space and time dimensions. The spatiotemporal super-resolution network model designed by the method described in the present invention improves the horizontal resolution of wind field data and expands the wind field data in the time dimension. It improves the resolution based on the simulated wind field data generated by the latent space diffusion model, and can obtain simulated wind fields with higher resolution and smaller time scales as needed.

[0087] The present invention's method for generating stratospheric wind field simulation data based on a latent space diffusion model does not require setting initial conditions and boundary conditions. Furthermore, the latent space diffusion model, after training, can directly generate stratospheric wind field simulation data via Gaussian noise. The complexity and diversity of the generated data can be controlled by adjusting parameters and the training process, greatly improving generation efficiency in tasks requiring the generation of large amounts of stratospheric wind field simulation data. Furthermore, the method, based on the latent space diffusion model, directly models and predicts complex four-dimensional spatiotemporal stratospheric dynamic wind field data, capturing the spatial and temporal evolution characteristics of the stratospheric wind field and learning the inherent laws and distribution patterns of the data, rather than just learning the characteristics of a specific stratospheric wind field. This method is applicable to stratospheric wind field simulation data generation tasks in different regions and seasons.

[0088] The present invention uses a latent space diffusion model to generate stratospheric wind field simulation data. The latent space diffusion model can learn the intrinsic representation of stratospheric wind field sample data by gradually adding noise and denoising processes in the latent space, thereby better capturing the complex distribution of stratospheric wind field sample data. Compared with VAE and GAN, the latent space diffusion model has higher flexibility in modeling high-dimensional, nonlinear and complex distributed data. The latent space diffusion model is generated through a step-by-step denoising process, which helps to retain more original data features. In addition, the latent space diffusion model can maintain a high diversity when generating samples, avoiding the mode collapse problem that may occur in GAN. The training process of GAN is usually unstable because it involves adversarial training of two networks. In contrast, the training process of the latent space diffusion model is more stable because it does not require simultaneous optimization of the generator and discriminator. Although VAE training is relatively stable, it is not as effective as the latent space diffusion model when processing complex data.

[0089] This paper uses a latent space diffusion model to simulate the stratospheric wind field by learning its distribution in three dimensions and its temporal variations from stratospheric wind field data. This method can provide important support for establishing a stratospheric simulation environment and for training intelligent navigation controllers for stratospheric aerostats. It has the following advantages:

[0090] (1) High simulation accuracy: By introducing spatial convolution layers, temporal convolution layers, and spatiotemporal attention layers to process the temporal and spatial information of stratospheric wind field sample data, respectively, the latent space diffusion model can better capture the spatiotemporal variation characteristics of the stratospheric wind field, thereby improving the accuracy and reliability of the latent space diffusion model in generating stratospheric wind field simulation data. The introduction of a super-resolution network enables this method to generate stratospheric wind field simulation data based on the latent space diffusion model, not limited to the resolution of the stratospheric wind field sample data used for training, and can obtain higher-resolution stratospheric wind field simulation data;

[0091] (2) High computational efficiency: This method directly generates simulated wind fields from Gaussian noise, avoiding the use of large-scale numerical models and the calculation of various fluid dynamics equations. This allows the model to complete large-scale simulations in a shorter time, improving computational efficiency and response speed. DDIM accelerates sampling, and compared to typical diffusion models such as DDPM, it can generate the same high-quality stratospheric wind field simulation data with fewer steps.

[0092] (3) Strong scalability: This method is not limited to a specific region or time scale. It has the ability to adapt to different regions and time scales and can be widely used to simulate stratospheric wind fields under various environmental conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] Figure 1 A flowchart of an embodiment of the present invention;

[0094] Figure 2 A diagram showing a conditional control network structure according to an embodiment of the present invention;

[0095] Figure 3 A structural diagram of a conditional encoder according to an embodiment of the present invention;

[0096] Figure 4 This is a diagram of a denoising neural network structure of a latent space diffusion model according to an embodiment of the present invention;

[0097] Figure 5 Schematic diagram of the pre-training process of the latent space diffusion model according to one embodiment of the present invention;

[0098] Figure 6 This is a structural diagram of a super-resolution network model according to an embodiment of the present invention;

[0099] Figure 7 A comparative analysis diagram of stratospheric wind field simulation data and actual wind field generated using a latent space diffusion model in one embodiment of the present invention;

[0100] Figure 8 This is a visualization diagram of stratospheric wind field sample data for training a spatiotemporal super-resolution network model in an embodiment of the present invention; wherein, Figure 8 (a) shows the visualization of the low-resolution stratospheric wind field sample data obtained after artificial coarsening. Figure 8 (b) is a visualization of high-resolution stratospheric wind field sample data before coarsening;

[0101] Figure 9 This is an effect diagram of high-resolution stratospheric wind field simulation data generated using a spatiotemporal super-resolution network model in one embodiment of the present invention. DETAILED DESCRIPTION

[0102] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and implementation examples. It should be understood that the specific implementation examples described herein are intended to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0103] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present application described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0104] Furthermore, the terms "installed," "disposed," "provided with," "connected," "connected," and "socketed" should be interpreted broadly. For example, "connected" can mean a fixed connection, a removable connection, or an integral structure; it can be a mechanical connection or an electrical connection; it can be a direct connection, an indirect connection through an intermediary, or an internal communication between two devices, elements, or components. Those skilled in the art will understand the specific meanings of these terms in this application based on the specific circumstances.

[0105] Reference Figures 1 to 8 The present invention provides a method for generating stratospheric wind field simulation data based on a latent space diffusion model, the flow chart of which is as follows: Figure 1 shown.

[0106] Using the method provided by the present invention to simulate the stratospheric wind field requires collecting stratospheric wind field sample data for training the neural network. Optionally, this embodiment takes the stratosphere in Hunan Province (longitude range: 108°E~114°E, latitude range: 24°N~30°N) as an example, and collects the stratospheric wind field sample data part of the fifth generation atmospheric reanalysis data set (ECMWFReanalysis v5, ERA5) from the European Centre for Medium-Range Weather Forecasts (ECMWF) in this region (maximum pressure is 150hPa, minimum pressure is 50hPa, and there are 5 different pressure levels).

[0107] The collected stratospheric wind field sample data is processed. In this embodiment, the stratospheric wind field sample data is divided into 29 parts in the longitude and latitude directions, and into 5 different levels in the altitude direction. The time period is 24 hours. The wind field data is expressed in the form of longitudinal wind speed u and latitudinal wind speed v, and the dimension of the data is 2. Finally, a four-dimensional spatiotemporal stratospheric dynamic wind field dataset for training the stratospheric wind field simulation data generation model is constructed. The dataset is data with a dimension of (29, 29, 5, 24, 2), that is, the simulated wind field is divided into a three-dimensional grid of (29, 29, 5), and is divided into 24 different time points in time. The wind speed is divided into two components: longitudinal wind speed and latitudinal wind speed. Optionally, this embodiment stores the processed four-dimensional stratospheric dynamic wind field in the format of a pickle file.

[0108] In addition, in order to train the super-resolution network model, the collected data are first interpolated to obtain a second stratospheric wind field sample data set with a data dimension of (32, 32, 5, 24, 2) to make downsampling and upsampling operations more convenient. Then, the collected stratospheric wind field sample data are coarsened to obtain a first stratospheric wind field sample data set with a data dimension of (16, 16, 5, 24, 2). The low-resolution stratospheric wind field sample data and its corresponding second stratospheric wind field sample data are used as training data sets to train the super-resolution network model.

[0109] Specifically, an embodiment of the present invention provides a stratospheric wind field simulation method, comprising the following steps:

[0110] Step S1: construct a conditional control network, input the conditional information for controlling the stratospheric wind field into the conditional control network, and obtain a conditional control vector.

[0111] In some embodiments, in order to better encode the conditions for controlling the stratospheric wind field, a conditional control network is designed. The network structure is as follows: Figure 2 As shown, the system includes two encoding layers and three linear layers. A numerical encoding layer encodes numerical conditions such as time, and a non-numerical encoding layer encodes non-numerical conditions such as the initial state of the wind farm. The outputs of the two encoding layers are concatenated and input into the three linear layers to produce the conditional control vector. The first linear layer serves as the input layer, receiving the outputs of the numerical encoding layer and the non-numerical encoding layer. The second linear layer serves as the hidden layer, and the third linear layer serves as the output layer, outputting the conditional control vector.

[0112] Step S2: Taking the stratospheric wind field sample data and the conditional control vector as input, extracting features from the stratospheric wind field sample data through the conditional encoder in the pre-trained encoder-decoder model and mapping them to a low-dimensional latent space vector; wherein the encoder-decoder model includes two parts: the conditional encoder and the conditional decoder.

[0113] In some embodiments, the encoder-decoder model network structure diagram is as follows Figure 3 As shown, by using a three-dimensional spatial convolution layer to process the spatial dimension information of the original stratospheric wind field sample data, and using a one-dimensional convolution layer to process the time dimension information of the original stratospheric wind field sample data, the encoder model first processes the spatial dimension information of the input original stratospheric wind field sample data, and then processes the result obtained by the three-dimensional spatial convolution layer as the input of the time convolution layer, thereby obtaining an encoder model that can separate the time and space information, and can extract the spatial dimension features and the time dimension features of the original stratospheric wind field sample data respectively, and can learn the distribution characteristics of the stratospheric wind field while learning its dynamic change characteristics.

[0114] The pre-training process of the encoder model includes the following steps:

[0115] Step S21: constructing an encoder-decoder model and initializing parameters of the encoder-decoder model;

[0116] Step S22: inputting the condition information for controlling the stratospheric wind field into the conditional control network to obtain a conditional control vector; inputting the stratospheric wind field sample data and the conditional control vector into the encoder-decoder model to obtain stratospheric wind field simulation data generated by the encoder-decoder model;

[0117] Step S23: calculating a first loss function value based on the stratospheric wind field simulation data and the stratospheric wind field sample data;

[0118] Step S24: updating the parameters of the encoder-decoder model based on back propagation according to the first loss function value;

[0119] Step S22: Iterate steps S22 to S24 until the iteration end condition is reached to obtain a pre-trained encoder-decoder model.

[0120] In some embodiments, the first loss function used to train the encoder-decoder model is calculated using the following formula:

[0121]

[0122] Wherein, Los(r,x) is the first loss function, r represents the generated stratospheric wind field simulation data, x represents the stratospheric wind field sample data, mse(r,x) represents the mean square error between r and x, σ represents the variance of the generated stratospheric wind field simulation data, KL(r,x) represents the KL divergence between r and x, and γ represents the weight of the KL divergence.

[0123] The mean square error is calculated by the following formula:

[0124]

[0125] Where N represents the number of grid points in a stratospheric wind field, r i represents the wind speed value of the i-th grid point in the generated stratospheric wind field simulation data, x i Represents the wind speed value of the i-th grid point in the stratospheric wind field sample data.

[0126] The KL divergence is calculated by the following formula:

[0127]

[0128] Among them, μ r and σ r represent the mean and variance of the generated stratospheric wind field simulation data, μ x and σ x They represent the mean and variance of the corresponding stratospheric wind field sample data used for training.

[0129] Step S3: gradually add random noise to the latent space vector (gradually diffuse the latent space vector) to obtain a noisy latent space vector; use the conditional control vector and the noisy latent space vector as input, predict the noise through a denoising neural network model of a pre-trained latent space diffusion model, perform a reverse denoising process, and obtain a denoised latent space vector; input the denoised latent space vector and the conditional control vector into the conditional decoder of the pre-trained encoder-decoder model for decoding to obtain the first stratospheric wind field simulation data.

[0130] In some embodiments, the denoising neural network model of the latent space diffusion model in step S3 includes a downsampling module, an intermediate module and an upsampling module. The specific network structure is as follows: Figure 4 As shown, the pre-training process includes the following steps:

[0131] Step S31: constructing a denoising neural network model of the latent space diffusion model and initializing the parameters of the denoising neural network model;

[0132] Step S32: inputting the conditional information for controlling the stratospheric wind field into the conditional control network to obtain a conditional control vector, inputting the stratospheric wind field sample data and the conditional control vector into the encoder part of the encoder-decoder model to obtain a latent space vector;

[0133] Step S33: gradually adding random noise to the latent space vector to obtain a noisy latent space vector; inputting the conditional control vector and the noisy latent space vector into a denoising neural network of a latent space diffusion model to obtain predicted noise, and performing a reverse denoising process based on the predicted noise to obtain a denoised latent space vector; inputting the denoised latent space vector and the conditional control vector into the conditional decoder for decoding to obtain stratospheric wind field simulation data;

[0134] Step S34: Calculating a second loss function value based on the stratospheric wind field simulation data obtained through the denoising neural network and decoding;

[0135] Step S35: updating the parameters of the denoising neural network model based on back propagation according to the second loss function value;

[0136] Step S36: Iterate steps S33 to S35 until the iteration end condition is reached to obtain a pre-trained encoder-decoder model.

[0137] The second loss function value in S34 is calculated using the following formula:

[0138]

[0139] Among them, Loss represents the second loss function, ε t represents the random noise added in step t in step S33, T represents the maximum number of steps, ε θ (X t , t) represents the predicted noise obtained by the denoising neural network model in step S33, Represents ε t -ε θ (X t ,t), γ represents the proportion of the difference between the stratospheric wind field simulation data and the stratospheric wind field sample data in the loss function, O represents the stratospheric wind field simulation data decoded by the conditional decoder, and Y represents the corresponding stratospheric wind field sample data used for training.

[0140] The pre-training process of the latent space diffusion model is as follows Figure 5As shown, the original stratospheric wind field sample data is mapped into a low-dimensional latent space for diffusion process, which greatly reduces the data dimension that the diffusion model needs to learn. The forward diffusion process of the latent space diffusion model is to add noise to the latent variables in the latent space, and the reverse diffusion process is to denoise the latent variables. Considering that the diffusion model requires a very long Markov process to generate high-quality stratospheric wind field simulation data, this is a very slow reverse diffusion process. Therefore, the latent space diffusion model described in step S3 adopts DDIM to perform the reverse diffusion process. Through this method, the number of sampling steps in the process of generating stratospheric wind field simulation data can be significantly reduced.

[0141] Step S4: using the first stratospheric wind field simulation data as input, and obtaining second stratospheric wind field simulation data through a pre-trained spatiotemporal super-resolution network model; wherein the resolution of the second stratospheric wind field simulation data is higher than the resolution of the first stratospheric wind field simulation data.

[0142] In some embodiments, the spatiotemporal super-resolution network model in step S4 includes an initial feature extraction module, a feature fusion module and a reconstruction module. The specific network structure is as follows: Figure 6 As shown, the pre-training process of the spatiotemporal super-resolution network model includes the following steps:

[0143] Step S41: constructing a spatiotemporal super-resolution network model and initializing the parameters of the spatiotemporal super-resolution network model;

[0144] Step S42: inputting the three-dimensional wind field data of the first time point of the first stratospheric wind field sample data into the initial feature extraction module of the spatiotemporal super-resolution network model to obtain an initial feature vector I0;

[0145] Step S43: inputting the first stratospheric wind field sample data into the first feature fusion processing unit of the feature fusion module of the spatiotemporal super-resolution network model to obtain a second feature vector I1, concatenating the initial feature vector I0 and the second feature vector I1, and inputting the concatenated data into the second feature fusion processing unit of the feature fusion module to obtain a fused feature vector I2;

[0146] Step S44: inputting the fused feature vector I2 into the reconstruction module of the spatiotemporal super-resolution network model to obtain reconstructed third stratospheric wind field simulation data;

[0147] Step S45: calculating the L2 loss based on the second stratospheric wind field sample data corresponding to the first stratospheric wind field sample data and the third stratospheric wind field simulation data reconstructed by the spatiotemporal super-resolution network model; wherein the resolution of the third stratospheric wind field sample data is higher than the resolution of the first stratospheric wind field sample data; and the second stratospheric wind field sample data is the high-resolution stratospheric wind field sample data that actually corresponds to the first stratospheric wind field sample data;

[0148] Step S46: updating the parameters of the spatiotemporal super-resolution network model based on the L2 loss value through back propagation;

[0149] Step S47: Iterate steps S42 to S46 until the iteration end condition is reached to obtain a pre-trained spatiotemporal super-resolution network model.

[0150] The embodiment of the present invention evaluates the effect of the simulated wind field by comparing the stratospheric wind field simulation data with the actual average wind speed profile of the stratospheric wind field. The difference between the generated stratospheric wind field simulation data and the stratospheric wind field sample data should be as small as possible.

[0151] The comparison between the average wind speed profile of the simulated wind field generated by the latent space diffusion model in this embodiment and the real wind field is as follows: Figure 7 As shown, in this embodiment, the overall average wind speed of the stratospheric wind field sample data (real stratospheric wind field data) used for training is 31.29 m / s, while the overall average wind speed of the simulated wind field generated using the latent space diffusion model is 34.17 m / s.

[0152] Figure 8 (a) shows the visualization of the low-resolution stratospheric wind field sample data obtained after artificial coarsening. Figure 8 (b) is a visualization of high-resolution stratospheric wind field sample data before coarsening. Figure 9 This is a visualization effect diagram of the low-resolution stratospheric wind field sample data obtained after coarsening based on this embodiment after using the spatiotemporal super-resolution network to enhance the resolution. The structural similarity index (SSIM) of the high-resolution stratospheric wind field simulation data obtained by this embodiment using the spatiotemporal super-resolution network and the high-resolution stratospheric wind field sample data can reach up to 0.99, and the peak signal-to-noise ratio (PSNR) can reach up to 60.

[0153] An embodiment of the present application further provides an electronic device, comprising: a memory and a processor;

[0154] The memory is used to store computer programs;

[0155] The processor is configured to call the computer program to execute the method described above.

[0156] An embodiment of the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program. When the computer program is executed on an electronic device, the electronic device implements the method described above.

[0157] An embodiment of the present application further provides a computer program product, including a computer program. When the computer program is run on an electronic device, the electronic device implements the method described above.

[0158] The embodiments of the present application also provide a system, an electronic device, a computer-readable storage medium, and a computer program product. The specific implementation methods can refer to the specific embodiments of the above methods and will not be repeated here.

[0159] Obviously, those skilled in the art should understand that the above-mentioned units or steps of the present application can be implemented using a general-purpose computing device. They can be concentrated on a single computing device or distributed across a network composed of multiple computing devices. Alternatively, they can be implemented using program codes executable by the computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps can be made into a single integrated circuit module for implementation. Thus, the present application is not limited to any specific combination of hardware and software.

[0160] The above description of the embodiments of the present application is only a partial embodiment of the present application, which is used to enable professionals in this field to implement or use the contents of the present application, and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A stratospheric wind field simulation method, characterized in that: The method includes: Step S1: constructing a conditional control network, inputting conditional information for controlling the stratospheric wind field into the conditional control network, and obtaining a conditional control vector; Step S2: Taking the stratospheric wind field sample data and the conditional control vector as input, extracting features from the stratospheric wind field sample data through the conditional encoder in the pre-trained encoder-decoder model and mapping them to a low-dimensional latent space vector; wherein the encoder-decoder model includes two parts: the conditional encoder and the conditional decoder; Step S3: gradually adding random noise to the latent space vector to obtain a noisy latent space vector; using the conditional control vector and the noisy latent space vector as input, predicting the noise through a denoising neural network model of a pre-trained latent space diffusion model, performing a reverse denoising process to obtain a denoised latent space vector; inputting the denoised latent space vector and the conditional control vector into a conditional decoder in a pre-trained encoder-decoder model for decoding to obtain first stratospheric wind field simulation data; Step S4: using the first stratospheric wind field simulation data as input, and obtaining second stratospheric wind field simulation data through a pre-trained spatiotemporal super-resolution network model; wherein the resolution of the second stratospheric wind field simulation data is higher than the resolution of the first stratospheric wind field simulation data.

2. The method according to claim 1, characterized in that In step S1, the conditional control network includes a numerical coding layer, a non-numerical coding layer, and three linear layers, wherein the numerical coding layer is used to encode numerical conditions, and the non-numerical coding layer is used to encode non-numerical conditions. The outputs of the numerical coding layer and the non-numerical coding layer are concatenated and input into the three linear layers to obtain the conditional control vector. Among the three linear layers, the first linear layer serves as an input layer to receive outputs of the numerical coding layer and the non-numerical coding layer, the second linear layer serves as a hidden layer, and the third linear layer serves as an output layer to output the conditional control vector.

3. The method according to claim 1, characterized in that The conditional encoder includes a plurality of groups of spatiotemporal downsampling modules and spatiotemporal attention modules connected in sequence; each group of spatiotemporal downsampling modules and spatiotemporal attention modules includes a spatiotemporal downsampling module and a spatiotemporal attention module connected in sequence; Each spatiotemporal downsampling module includes a spatial downsampling unit and a temporal downsampling unit connected in sequence; The spatial downsampling unit includes a plurality of three-dimensional spatial convolution layers connected in sequence, which are used to process the spatial dimension information of the stratospheric wind field sample data; the temporal downsampling unit includes a one-dimensional temporal convolution layer, which is used to process the temporal dimension information of the stratospheric wind field sample data; each convolution layer is followed by a residual layer; Each spatiotemporal attention module includes a spatial attention layer and a temporal attention layer connected in sequence, which are used to process the spatial dimension information and temporal dimension information of the stratospheric wind field sample data respectively; Sequential connection means that the output of the previous module, layer or unit is used as the input of the next module or layer; The conditional encoder receives the conditional control vector and the stratospheric wind field sample data as input, and the stratospheric wind field sample data is input into the first three-dimensional spatial convolution layer in the conditional encoder; the other three-dimensional spatial convolution layers except the first three-dimensional spatial convolution layer in the conditional encoder, and the one-dimensional temporal convolution layer, take the output of the previous module, layer or unit and the conditional control vector as input; The conditional decoder includes multiple groups of sequentially connected spatiotemporal upsampling modules and spatiotemporal attention modules; each group of spatiotemporal upsampling modules and spatiotemporal attention modules includes a sequentially connected spatiotemporal upsampling module and a spatiotemporal attention module; The spatiotemporal upsampling module includes a spatial upsampling unit and a temporal upsampling unit; the spatial upsampling unit includes a plurality of sequentially connected three-dimensional spatial convolutional layers for upsampling the spatial information of the stratospheric wind field sample data; the temporal upsampling unit includes a one-dimensional temporal convolutional layer for upsampling the temporal information of the stratospheric wind field sample data; each convolutional layer is followed by a residual layer; four-dimensional spatiotemporal stratospheric wind field simulation data is obtained through the spatiotemporal upsampling module.

4. The method according to claim 1, wherein The pre-training process of the encoder-decoder model includes: Step S21: constructing an encoder-decoder model and initializing parameters of the encoder-decoder model; Step S22: inputting the condition information for controlling the stratospheric wind field into the conditional control network to obtain a conditional control vector; inputting the stratospheric wind field sample data and the conditional control vector into the encoder-decoder model to obtain stratospheric wind field simulation data generated by the encoder-decoder model; Step S23: calculating a first loss function value based on the stratospheric wind field simulation data and the stratospheric wind field sample data; Step S24: updating the parameters of the encoder-decoder model based on back propagation according to the first loss function value; Step S22: Iterate steps S22 to S24 until the iteration end condition is reached to obtain a pre-trained encoder-decoder model; Preferably, the first loss function in step S23 is calculated using the following formula: Wherein, Los(r,x) is the first loss function, r represents the generated stratospheric wind field simulation data, x represents the stratospheric wind field sample data, mse(r,x) represents the mean square error between r and x, σ represents the variance of the generated stratospheric wind field simulation data, KL(r,x) represents the KL divergence between r and x, and γ represents the weight of the KL divergence; The mean square error is calculated by the following formula: Where N represents the number of grid points in a stratospheric wind field, r i represents the wind speed value of the i-th grid point in the generated stratospheric wind field simulation data, x i Represents the wind speed value of the i-th grid point in the stratospheric wind field sample data; The KL divergence is calculated by the following formula: Among them, μ r and σ r represent the mean and variance of the generated stratospheric wind field simulation data, μ x and σ x They represent the mean and variance of the corresponding stratospheric wind field sample data used for training.

5. The method according to claim 1, wherein The denoising neural network model of the latent space diffusion model in step S3 includes a downsampling module, an intermediate module and an upsampling module connected in sequence; The downsampling module includes a plurality of downsampling units connected in sequence; The downsampling unit includes a three-dimensional spatial convolution layer, a one-dimensional temporal convolution layer, a residual layer, and a spatiotemporal attention module connected in sequence; The intermediate module includes a spatiotemporal attention module, and the spatiotemporal attention module is provided with residual layers before and after; The upsampling module includes a plurality of upsampling units connected in sequence, each upsampling unit includes a four-dimensional spatiotemporal convolution module and a spatiotemporal attention module connected in sequence; the four-dimensional spatiotemporal convolution module includes a three-dimensional spatial convolution layer, a one-dimensional temporal convolution layer and a residual layer connected in sequence.

6. The method according to claim 1, characterized in that The pre-training process of the denoising neural network model of the latent space diffusion model includes: Step S31: constructing a denoising neural network model of the latent space diffusion model and initializing the parameters of the denoising neural network model; Step S32: inputting the conditional information for controlling the stratospheric wind field into the conditional control network to obtain a conditional control vector, inputting the stratospheric wind field sample data and the conditional control vector into the encoder part of the encoder-decoder model to obtain a latent space vector; Step S33: gradually adding random noise to the latent space vector to obtain a noisy latent space vector; inputting the conditional control vector and the noisy latent space vector into a denoising neural network of a latent space diffusion model to obtain predicted noise, and performing a reverse denoising process based on the predicted noise to obtain a denoised latent space vector; inputting the denoised latent space vector and the conditional control vector into the conditional decoder for decoding to obtain stratospheric wind field simulation data; Step S34: Calculating a second loss function value based on the stratospheric wind field simulation data obtained through the denoising neural network and decoding; Step S35: updating the parameters of the denoising neural network model based on back propagation according to the second loss function value; Step S36: Iterate steps S33 to S35 until the iteration end condition is reached to obtain a pre-trained encoder-decoder model; Preferably, the second loss function value in step S34 is calculated using the following formula: Among them, Loss represents the second loss function, ε t represents the random noise added in step t in step S33, T represents the maximum number of steps, ε θ (X t , t) represents the predicted noise obtained by the denoising neural network model in step S33, Represents ε t -ε θ (X t ,t), γ represents the proportion of the difference between the stratospheric wind field simulation data and the stratospheric wind field sample data in the loss function, O represents the stratospheric wind field simulation data decoded by the conditional decoder, and Y represents the corresponding stratospheric wind field sample data used for training.

7. The method according to claim 1, characterized in that The spatiotemporal super-resolution network model in step S4 includes an initial feature extraction module, a feature fusion module and a reconstruction module; the initial feature extraction module includes a first feature extraction unit, a downsampling layer and a second feature extraction unit connected in sequence; the first feature extraction unit and the second feature extraction unit both include a three-dimensional spatial convolution layer, a spatial attention layer and a residual layer connected in sequence; the feature fusion module includes a first feature fusion processing unit and a second feature fusion processing unit connected in sequence; the first feature fusion processing unit and the second feature fusion processing unit include two sub-processing units, each of which includes a first feature fusion processing unit and a second feature fusion processing unit connected in sequence. The output end of the initial feature extraction module is connected to the input end of the second feature fusion processing unit in the feature fusion module; the four-dimensional spatiotemporal convolution module includes a three-dimensional spatial convolution layer, a one-dimensional temporal convolution layer and a residual layer connected in sequence; the output end of the feature fusion module is connected to the input end of the reconstruction module; the reconstruction module includes a four-dimensional spatiotemporal convolution module, a spatiotemporal attention module, an upsampling layer, a residual layer and a spatiotemporal attention module connected in sequence.

8. The method according to claim 1, characterized in that The pre-training process of the spatiotemporal super-resolution network model includes the following steps: Step S41: constructing a spatiotemporal super-resolution network model and initializing the parameters of the spatiotemporal super-resolution network model; Step S42: inputting the three-dimensional wind field data of the first time point of the first stratospheric wind field sample data into the initial feature extraction module of the spatiotemporal super-resolution network model to obtain an initial feature vector I0; Step S43: inputting the first stratospheric wind field sample data into the first feature fusion processing unit of the feature fusion module of the spatiotemporal super-resolution network model to obtain a second feature vector I1, concatenating the initial feature vector I0 and the second feature vector I1, and inputting the concatenated data into the second feature fusion processing unit of the feature fusion module to obtain a fused feature vector I2; Step S44: inputting the fused feature vector I2 into the reconstruction module of the spatiotemporal super-resolution network model to obtain reconstructed third stratospheric wind field simulation data; Step S45: calculating the L2 loss based on the second stratospheric wind field sample data corresponding to the first stratospheric wind field sample data and the third stratospheric wind field simulation data reconstructed by the spatiotemporal super-resolution network model; wherein the resolution of the third stratospheric wind field sample data is higher than the resolution of the first stratospheric wind field sample data; and the second stratospheric wind field sample data is the high-resolution stratospheric wind field sample data that actually corresponds to the first stratospheric wind field sample data; Step S46: updating the parameters of the spatiotemporal super-resolution network model based on the back propagation according to the L2 loss value; Step S47: Iterate steps S42 to S46 until the iteration end condition is reached to obtain a pre-trained spatiotemporal super-resolution network model.

9. An electronic device, characterized in that: include: memory and processor; The memory is used to store computer programs; The processor is configured to call the computer program to execute the method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed on an electronic device, the electronic device implements the method according to any one of claims 1 to 8.

11. A computer program product comprising a computer program, characterized in that When the computer program is executed on an electronic device, the electronic device is enabled to implement the method according to any one of claims 1 to 8.

Images