A typhoon near-surface wind field downscaling method based on a hybrid attention transformer framework

Abstract

The application discloses a typhoon near-ground wind field downscaling method based on a mixed attention Transformer framework, and is a high-resolution typhoon near-ground wind field downscaling system TCHAT established based on an HAT model. The system combines high spatial resolution information of observation data and long-time sequence characteristics of reanalysis products, directly reads reanalysis data to perform a downscaling process from a low-resolution wind field to a high-resolution wind field, and generates high-resolution typhoon wind fields with high precision and long time sequences. The system can support high-quality and long-time sequence typhoon near-ground two-dimensional wind field data sets, provide a solid data foundation for systematically depicting typhoon structure, typhoon evolution, intensity and scale characteristics, and is helpful to typhoon climate and typhoon disaster assessment research.

Description

Technical Field

[0001] This invention belongs to the fields of computer vision and meteorological technology, and in particular relates to a typhoon near-surface wind field downscaling system, which will be directly applied to high-resolution near-surface wind field reconstruction, intensity and scale estimation of typhoons. Background Technology

[0002] High spatiotemporal resolution near-surface two-dimensional wind fields of typhoons can be used for analysis of asymmetry, inner and outer core structure, and typhoon evolution processes. This provides crucial data for typhoon climate research and initial fields for models and patterns, while also supporting direct estimation of intensity and scale. Changes in intensity and scale are important indicators for understanding typhoon climate response and disaster changes, and are of great significance for typhoon disaster assessment. Existing high-resolution wind field observation data, such as synthetic aperture radar (SAR) and near-surface wind field data from the Hurricane Wind Analysis System (HWIND), can accurately reflect the intensity and structural characteristics of typhoons. However, such data have limited temporal coverage and a small sample size, making it difficult to support systematic analysis and long-term trend research on typhoon climate change. Reanalysis products and model data have long time series, but low spatial resolution, making it impossible to accurately resolve the structural, intensity, and scale information of typhoons.

[0003] Currently, methods for obtaining high-resolution typhoon wind field information mainly include traditional dynamic and statistical downscaling methods, as well as the rapidly developing deep learning methods. Dynamic downscaling methods often simulate single typhoons, requiring high computational costs and long processing times, making them unsuitable for long-term climate studies. Statistical downscaling typically relies on simple methods such as linear regression, which struggle to characterize the complex nonlinear mappings present in meteorological data. Furthermore, statistical downscaling is heavily dependent on long-term, high-quality historical observation data, while the limited number of typhoon event samples further restricts the accuracy and applicability of these methods. Traditional two-stage generative super-resolution methods based on deep learning are prone to error accumulation and dependency, affecting wind field accuracy. Diffusion probability models typically require multiple iterative denoising steps to generate the final high-resolution sample, resulting in slow inference speeds and high computational resource requirements during training and inference. Additionally, these models lack estimation capabilities for two-dimensional high-resolution wind fields.

[0004] Therefore, existing technologies have significant limitations in reconstructing high-resolution near-surface two-dimensional wind fields of typhoons, making it difficult to support systematic analysis and long-term trend research on typhoon climate change. Summary of the Invention

[0005] Purpose of the invention: To address the problems in high-resolution near-surface two-dimensional wind field reconstruction of typhoons using existing technologies, this invention provides a typhoon near-surface wind field downscaling method based on the hybrid attention Transformer framework. This method is more efficient, accurate, and reliable, and can fully leverage the complementary advantages of observational and reanalysis data to construct a uniform high-resolution near-surface two-dimensional wind field dataset for typhoons over long time series. This provides a solid foundation for systematically revealing the structural characteristics, evolution process, and quantitative estimation of intensity and scale of typhoons, and offers important support for typhoon climate research and disaster assessment.

[0006] Technical Solution: To achieve the above-mentioned objectives, this invention adopts the following technical solution: a method for downscaling near-surface wind fields of typhoons based on a hybrid attention Transformer framework, comprising the following steps: S1, Data Preprocessing: Obtain high-resolution typhoon near-surface wind field observation dataset as target data, and after screening, cropping and normalization, pair it with ERA5 in the analysis data to complete the construction of the data sample set, and divide the data sample set into training set, test set and validation set according to the proportion; S2, Deep Learning Model Construction: Based on the HAT model of Transformer, a generator is built that includes a shallow feature extraction module, a deep feature extraction module, and an image reconstruction module. At the same time, a discriminator is built that includes three parts: downsampling, upsampling, and discriminative output. The values ​​of the adversarial loss function and the feature matching loss function are calculated through the output of the discriminator. S3, Deep learning model training: The total loss is a weighted loss of pixel-level loss, adversarial loss, and feature matching loss, and the model parameters are updated, optimized, and saved using the total loss; the mean squared error (MSE) is used as the evaluation metric to determine the optimized model.

[0007] Furthermore, the data preprocessing in step S1 includes the following steps: S11. Obtain the high-resolution typhoon near-surface wind field observation dataset HWIND since 1998. Select samples with an intensity of 34kt or higher from the typhoon best path dataset, and transform samples with a spatial resolution of less than 0.05° to the target resolution of 0.05° using bilinear interpolation. S12, based on the location of the typhoon center, extract 80 pixels above, below, left, and right of the typhoon center and crop them to obtain a 160×160 two-dimensional array; S13, the typhoon near-surface wind field observation dataset HWIND uses U-wind speed field, V-wind speed field and full wind speed field, and splices them with two-dimensional array to form a target data sample of 160×160×3 three-dimensional array; and generates high-resolution ocean-land mask according to the latitude and longitude range of HWIND for discriminator input. S14. Based on ERA5 reanalysis data, the lowest sea level pressure position within a 1°×1° range around the center of the typhoon's best path dataset is found as the new typhoon center. The new typhoon center is then cropped by 16 pixels above, below, left, and right, and stitched together with the normalized sea level pressure field, ocean-land mask, U-wind field, V-wind field, and full wind speed to obtain a 32×32×5 three-dimensional array of ERA5 input data samples. S15, match the ERA5 input data samples with the target data samples one by one to complete the construction of the data sample set, and then divide the data sample set into training set, test set and validation set in a ratio of 6:3:1.

[0008] Furthermore, in step S15, when the data sample set is proportionally divided into training set, test set, and validation set, data samples from the same typhoon event are grouped into the same dataset.

[0009] Furthermore, the generator mentioned in step S2 includes: The shallow feature extraction module includes a 3×3 convolutional layer, which is used to map the input low-resolution wind field to a high-dimensional feature space. The deep feature extraction module includes several residual hybrid attention groups (RHAGs) and a convolutional layer with a 3×3 kernel. The RHAG consists of several hybrid attention blocks (HAT Blocks), an overlapping cross attention block (OCAB), and a 3×3 convolutional layer. Each HAT Block contains two layer normalization blocks, a channel attention block (CAB), a local window multi-head self-attention (W-MSA), and a multilayer perceptron to extract local and global wind field features. The OCAB module contains two layer normalization blocks, a multilayer perceptron, and a cross-window attention block (OCA) to fuse wind field information from neighboring grid regions. The image reconstruction module contains two convolutional layers and pixel rearrangement upsampling. After the first 3×3 convolution, pixel rearrangement upsampling is performed, and after the last convolution, the image is mapped to the output.

[0010] Furthermore, the discriminator in step S2 includes: Downsampling is performed by using a multi-layered spectral normalized convolutional layer with a kernel size of 4×4 to progressively compress the spatial resolution of the wind speed field and extract multi-scale wind field features. Upsampling is used to gradually restore spatial resolution through bilinear interpolation and convolution operations, while skip connections are used to fuse local details in shallow layers with global structures in deep layers. The output, consisting of several convolutional layers and a single-channel output layer, is used to generate grid-by-grid true / false discrimination results.

[0011] Furthermore, the pixel-level loss mentioned in step S3 is based on the mean absolute error loss L1 Loss. For the typhoon wind field downscaling task, a new Gaussian Decay Loss is designed as the pixel-level loss and calculated using the following formula: , In the formula, and These represent the distances from grid point i to the grid center in the horizontal and vertical directions, respectively. Let i be the truth value of grid point i. σ is the model prediction value for grid point i, and the Gaussian parameter σ is the optimal value determined through experimental tuning, with a value of 40.

[0012] Furthermore, the total loss described in step S3 The calculation formula is as follows: , In the formula, It's a pixel-level loss. It is feature matching loss. It is about combating losses.

[0013] Furthermore, in step S3, during the deep learning model training process, the generator TCHAT and the discriminator U-Net adopt an asymmetric training strategy: the generator is trained 10 times, while the discriminator is trained once. The generator TCHAT has 500,000 training iterations, a batch size of 4, and uses the Adam optimizer with an initial learning rate of 0.0001. The discriminator U-Net has a batch size of 4, uses the Adam optimizer, and has an initial learning rate of 0.0001.

[0014] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0015] (1) The high-resolution typhoon near-surface two-dimensional wind field downscaling system of the present invention has good precision and accuracy.

[0016] (2) This invention significantly reduces computational costs and improves computational efficiency while ensuring the accuracy of scaling down.

[0017] (3) This invention effectively integrates the high spatial resolution information of observation data with the long time series characteristics of reanalysis products, realizes the fine reconstruction of low-resolution wind field to high resolution, and constructs a high-resolution, high-quality long time series typhoon near-surface two-dimensional wind field dataset. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the generator TCHAT model and the discriminator U-Net model architecture used in the training process of the present invention.

[0019] Figure 2 This invention provides a comparative analysis of the model performance under two functions: Gaussian Decay Loss and Mean Absolute Error Loss. Specifically, (a) the prediction error performance of Gaussian Decay Loss and MAE L1 Loss is compared at different radial distances; and (b) the MSE difference between Gaussian Weighted L1 Loss and Ordinary L1 Loss at different radial distances.

[0020] Figure 3 The graph shows the changes in the loss function and evaluation index of the generator TCHAT model during training, with the number of iterations. (a) The mean square error of the model on the validation set is saved during training; (b) The structural similarity index of the model on the validation set is saved during training; (c) The total loss of the generator during training changes with the number of iterations; (d) The discriminator loss during training changes with the number of iterations.

[0021] Figure 4 The following is a comparison of the average radial wind field profiles of the actual observations, TCHAT model results, and bilinear interpolation results of this invention, including: (a) a comparison of all typhoon time samples; (b) a comparison of typhoon time samples with wind speeds between 33 and 49 m / s; and (c) a comparison of typhoon time samples with wind speeds above 49 m / s.

[0022] Figure 5 This is a schematic diagram of the application process of the present invention.

[0023] Figure 6 To illustrate the application of this invention during three periods of the 2010 hurricane "Earl" that did not make landfall, and to compare it with actual observations and bilinear interpolation results, the color is filled in to represent the full wind speed field, and the arrows represent the U and V two-dimensional wind fields. Among them: (a) the actual observed HWIND wind field at 06:00:00 on August 30, 2010; (b) the wind field obtained from the downscaling result of the TCHAT model at 06:00:00 on August 30, 2010; (c) the wind field obtained from bilinear interpolation of ERA5 data at 06:00:00 on August 30, 2010; (d) the actual observed HWIND wind field at 00:00:00 on September 2, 2010; (e) the wind field obtained from the downscaling result of the TCHAT model at 00:00:00 on September 2, 2010; (f) the wind field obtained from bilinear interpolation of ERA5 data at 00:00:00 on September 2, 2010; (g) the actual observed HWIND wind field at 21:00:00 on September 3, 2010; (h) the wind field obtained from 21:00:00 on September 3, 2010. Wind field obtained by downscaling TCHAT model; (i) Wind field obtained by bilinear interpolation of ERA5 data on 2010 09.03 21:00:00.

[0024] Figure 7 To illustrate the application of this invention at three time points during the 2004 hurricane "Jenny" landfall and to compare it with actual observations and bilinear interpolation results, the following data is presented: (a) Actual observed HWIND wind field at 03:00:00 on September 26, 2004; (b) Wind field obtained from the downscaling result of the TCHAT model at 03:00:00 on September 26, 2004; (c) Wind field obtained from bilinear interpolation of ERA5 data at 03:00:00 on September 26, 2004; (d) Actual observed HWIND wind field at 09:00:00 on September 26, 2004; (e) Wind field obtained from the downscaling result of the TCHAT model at 09:00:00 on September 26, 2004; (f) Wind field obtained from bilinear interpolation of ERA5 data at 09:00:00 on September 26, 2004; (g) Wind field obtained from 15:00:00 on September 26, 2004. Actual observation of HWIND wind field; (h) Wind field obtained from downscaling results of TCHAT model on 2004 09.26 15:00:00; (i) Wind field obtained from bilinear interpolation of ERA5 data on 2004 09.26 15:00:00. Detailed Implementation

[0025] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0026] This invention establishes a high-resolution typhoon near-surface wind field downscaling system (TCHAT) based on the HAT model. This system combines the high spatial resolution information of observational data with the long time-series characteristics of reanalysis products. By directly reading reanalysis data, it performs a downscaling process from low-resolution wind fields to high-resolution wind fields. The generated high-resolution typhoon wind fields have high accuracy and long time series, which can support high-quality, long-time-series typhoon near-surface two-dimensional wind field datasets. This provides a solid data foundation for systematically characterizing typhoon structure, typhoon evolution, intensity, and scale characteristics, and is helpful for typhoon climate and typhoon disaster assessment research.

[0027] This invention utilizes ERA5 reanalysis data and HWIND high-resolution near-surface wind field observation data, and trains a hybrid attention HAT model based on the Transformer architecture to ultimately transform the input low-resolution wind field into a high-resolution wind field. Based on this, a uniform, long-term, climatologically significant high-resolution typhoon near-surface two-dimensional wind field dataset is constructed, and climatological estimation of typhoon intensity and scale is performed.

[0028] like Figure 5As shown, the overall steps of the typhoon near-surface wind field downscaling method based on the hybrid attention Transformer framework of this invention are as follows:

[0029] (1) Perform data preprocessing on the input ERA5 element fields (sea level pressure field, ocean-land mask, U wind field, V wind field, and full wind speed field), normalize each element using different standards, and splice them into a three-dimensional array X of size 32×32×5.

[0030] (2) The trained typhoon two-dimensional wind field downscaling TCHAT model is used to downscale X to obtain a three-dimensional array of 160×160×3. The array is then split and dimensionalized to obtain the target individual element fields of 160×160: U wind field, V wind field, and full wind speed field.

[0031] (3) The split element fields are reverse normalized according to the standard to restore their original physical quantity range, thereby obtaining the original U wind field, V wind field and full wind speed field.

[0032] To achieve the above functions, it is first necessary to write preprocessing code for different feature fields of ERA5 reanalysis data; secondly, it is necessary to select a suitable deep learning model architecture; and finally, based on the selected model architecture, the model parameters are trained, and the parameters of the best model are saved. The design of these three steps will be explained in detail below.

[0033] Step S1, Data Preprocessing:

[0034] Target data processing: The U-wind speed field, V-wind speed field, and total wind speed field from the high-resolution typhoon near-surface wind field observation dataset HWIND (1998-2013) were selected as the target data, totaling 1697 time-series samples. The spatial resolution of this target data is 0.05°. The main processing steps are as follows:

[0035] Step S11: Based on the typhoon optimal path dataset, sample screening is performed, selecting only samples with an intensity of 34kt (tropical storm, TS) or higher, and samples with a spatial resolution of less than 0.05° are bilinearly interpolated to the target resolution.

[0036] Step S12: Based on the typhoon center location, extract 80 pixels above, below, left, and right of the typhoon center, resulting in a 160×160 dimension. After cropping, delete target data with dimensions less than 155×155 and process each sample into an .npy file for later use. The land and ocean masks are generated from the corresponding latitude and longitude ranges (1 for ocean, 0 for land).

[0037] Step S13: During the training process, for each typhoon time in the training and validation sets, the U-wind speed field, V-wind speed field, and total wind speed field are processed using the maximum-minimum normalization method, as shown in the following formula: , in, The normalized value. The original data, and Typically, these are the minimum and maximum values ​​of the data. In this invention, for different elements, [the following values ​​are taken]. and The normalized parameters obtained from the model tests for different fixed values ​​are shown in Table 1.

[0038] Table 1. Normalization range of different elements

[0039] If the array dimension is less than 160×160, use edge value padding to expand the array to 160×160, and then stitch the three elements together to form a 160×160×3 three-dimensional array.

[0040] Step S14, Input Data Processing: ERA5 reanalysis data is used as input data, with samples corresponding one-to-one with the target data samples. The spatial resolution of this input data is 0.25°. Based on the center of the typhoon's optimal path dataset, the location of the lowest sea-level pressure within a 1°×1° radius is found as the new typhoon center. The typhoon center is then cropped by 16 pixels vertically and horizontally (32×32 dimensions) and processed into an .npy file for later use. Similar to the target data processing, the sea-level pressure field, ocean-land mask, U-wind field, V-wind field, and full wind speed field are normalized (see Table 1) and then concatenated into a 32×32×5 three-dimensional array.

[0041] Step S15, Construction of training dataset: A total of 1697 time samples and 145 typhoon events were obtained. All typhoon events were divided into training set, test set and validation set in a ratio of 6:3:1. At the same time, since the actual typhoon intensity distribution is different, in order to improve the model's learning ability for typhoons of different intensities, the wind speed intensity probability distribution of the training set and the validation set is kept similar, while the intensity distribution of the test set is deliberately designed to be different, so as to simulate the evaluation of the model's generalization ability to unknown typhoons in actual applications. The specific division is shown in Table 2.

[0042] Table 2 Number of samples in the training dataset

[0043] S2, Deep Learning Model Construction

[0044] A schematic diagram of the overall model structure is shown below. Figure 1 As shown, (1) Basic architecture of high-resolution typhoon two-dimensional wind field model: This invention uses the Transformer-based HAT model as its basic architecture, consisting of three parts: shallow feature extraction (extracting preliminary low-level features such as local texture and edges from the input low-resolution data), deep feature extraction (deep feature extraction uses an attention mechanism to simultaneously capture local details and overall structure of the input data, thereby obtaining a higher-level feature representation), and image reconstruction (converting features into high-resolution output). Shallow feature extraction consists of one convolutional layer with a kernel size of 3×3, which maps the input low-resolution wind field to a high-dimensional feature space.

[0045] Deep feature extraction comprises several Residual Hybrid Attention Groups (RHAGs) and a 3×3 convolutional layer. Each RHAG consists of several Hybrid Attention Blocks (HAT Blocks), an Overlapping Cross Attention Block (OCAB), and a 3×3 convolutional layer. Each HAT Block contains two layer normalization modules, a Channel Attention Block (CAB), a Local Window Multi-Head Self-Attention (W-MSA), and a multilayer perceptron to extract local and global wind field features. The OCAB module contains two layer normalization modules, a multilayer perceptron, and an OCA (Overlapping Window Attention) module to fuse wind field information from neighboring grid regions.

[0046] The image reconstruction part includes two convolutional layers. After the first 3×3 convolution, the pixels are upsampled through rearrangement. After the last convolution, the image is mapped to the output.

[0047] Since the original model only supports single-to-single (1-to-1) and three-to-three (3-to-3) image input and output, the model input and output were modified to npy format files to adapt to meteorological data, and the model was adjusted to be compatible with input and output of any channel. In addition, the original model only supports 3x and 2nx downscaling, so a 5x downscaling module was added to meet the task requirements.

[0048] (2) Discriminator U-Net structure

[0049] To improve the generator model's ability to learn asymmetric wind field structures, this invention introduces an additional U-Net-based discriminator to calculate adversarial loss and feature matching loss.

[0050] The architecture of the discriminator is as follows Figure 1As shown, the model consists of three parts: downsampling, upsampling, and discriminative output. The downsampling part uses multiple spectral normalization (SN) convolutional layers with a kernel size of 4×4 to progressively compress the spatial resolution of the wind speed field and extract multi-scale wind field features. The upsampling part uses bilinear interpolation and convolution operations to progressively restore the spatial resolution, while simultaneously fusing shallow local details with deep global structures through skip connections, thereby enhancing the ability to discriminate typhoon wind speed spatial distribution patterns. The discriminative output part contains several convolutional layers and a single-channel output layer, generating grid-by-grid true / false discrimination results. This allows the model to not only distinguish the authenticity of the overall wind speed field but also to discriminate high-wind-speed areas, typhoon eyes, and eyewall structures at local scales, thus providing more meteorologically relevant adversarial constraints for the generator's downscaling wind field reconstruction.

[0051] S3, Deep Learning Model Training

[0052] The loss function and evaluation metrics of the model in this invention are important settings during the training of deep learning models, and will be described in detail below: (1) Loss function settings: The model training process relies on changes in the loss function to update the model parameters. In this invention, during training, when the generator model outputs a reconstructed image, a weighted combination of three loss functions is used to update the model: a) Pixel-wise Reconstruction Loss For pixel-level loss, this invention, based on the Mean Absolute Error Loss (L1 Loss), newly designs Gaussian Decay Loss for the task of downscaling typhoon wind fields. This loss function aims to assign higher weight to the difference between the true value and the model prediction in the typhoon center region, thereby strengthening the model's learning of the high-wind-speed region in the typhoon center. The formula is as follows: , in, and These represent the distances of grid point i from the grid center in the horizontal and vertical directions, respectively. The closer the point is to the center, the greater its weight. Let i be the truth value of grid point i. This is the model prediction value for grid point i. The Gaussian parameter σ is the optimal value determined through experimental tuning, with a value of 40. This parameter effectively emphasizes the important features of the typhoon's central region while avoiding excessively rapid weight decay that could lead to the loss of surrounding information, thus achieving the best reconstruction effect. Compared to the traditional mean absolute error loss, this loss function significantly improves the reconstruction of the typhoon's central region; see [link to details]. Figure 2 .

[0053] b) Adversarial loss (GAN Loss) and Feature Matching Loss

[0054] FeatureMatchingLoss and GANLoss are calculated from the discriminator's discrimination results. Specifically, L1Loss is used for FeatureMatchingLoss, and WGAN-GP is used for GANLoss. The feature matching loss and adversarial loss are calculated from the discriminator's feature layer output and discrimination result output, respectively, and the model is optimized.

[0055] Traditional discriminators typically use the generator's output directly for discrimination, while this invention employs a different approach: First, output the generator results. Additional land and ocean masking channels were added to form m, the purpose of this processing is to mask artifacts caused by terrain differences; Then m rotated 180° to obtain rot; The final input data to the discriminator is m- rot; The ground truth values ​​are also processed in the same way and input into the discriminator. The purpose of this design is to enable the model to more sensitively capture the asymmetric features of the wind field and improve the accuracy of the wind field structure.

[0056] In summary, the total loss for model parameter updates is: .

[0057] (2) Evaluation indicators:

[0058] During model training, the current model is saved every certain number of iterations. After training is completed, the mean squared error (MSE) is used to evaluate the numerical deviation between the generated wind field and the observed wind field, and the model is selected accordingly.

[0059] (3) Deep learning model training and model selection

[0060] a) Hyperparameter settings:

[0061] In this invention, the generator and discriminator employ an asymmetric training strategy (the generator is trained multiple times for each discriminator training iteration). This strategy, along with the number of training iterations, batch size, and learning rate, are hyperparameters that significantly impact the training performance of the deep learning model. After adjusting the controlled variables, the optimal hyperparameter combination for the model on the test set was obtained, as shown in Table 3. The changes in the loss function and evaluation metrics with the number of training iterations are shown in [Table 3]. Figure 3 .

[0062] Table 3 Model Hyperparameter Settings

[0063] (Where Adam(1e-4) indicates that Adam is used as the optimization method, with an initial learning rate of 0.0001. The Adam optimizer combines momentum and adaptive learning rate techniques, and can automatically adjust the learning rate of each parameter based on the estimated first moment (i.e., mean) and second moment (i.e., uncentered variance).)

[0064] b) Model training performance and optimal model selection: During training, the generator saves the model parameters every 2000 iterations. Since a typhoon is a complex nonlinear system and the training sample size is limited, to avoid model overfitting, this study did not select the model with the smallest training loss function, but instead used the model with the smallest MSE on the validation set as the final model.

[0065] like Figure 3 As shown, the generator loss and discriminator loss steadily decrease with increasing iteration count, gradually stabilizing after 200,000 iterations and reaching a plateau. The validation set MSE gradually decreases with model training, stabilizing after 400,000 iterations, then plateauing with increasing iteration count, reaching its minimum at 444,000 iterations. Ultimately, this invention selects the model trained at the lowest validation MSE value (7.5837), i.e., the model trained at the 444,000th iteration, as TCHAT.

[0066] After training, this invention can automatically process large amounts of reanalysis data to achieve rapid, high-quality high-resolution downscaling of typhoon two-dimensional wind fields.

[0067] Application Examples: The specific implementation process of the present invention is as follows: The present invention is applied to three periods of Hurricane Earl (which did not make landfall) in 2010 and three periods of Hurricane Jenny (which made landfall) in 2004.

[0068] Figure 5This is a schematic diagram of the typhoon near-surface wind field downscaling system described in this invention, specifically including: (1) Data preprocessing for reanalysis:

[0069] For the U-shaped wind field, V-shaped wind field, and sea level pressure field, based on the typhoon center location in the typhoon optimal path dataset, 10°×10° corresponding data are cropped with the typhoon center as the center. Then, the lowest sea level pressure within this range is found as the new typhoon center, and corresponding data of 16 pixels above, below, left, and right of the typhoon (i.e., 32×32) are extracted. Based on the land and sea information within this range, a land-ocean mask of the same size is generated (ocean is 1, land is 0). The full wind speed field is calculated from the corresponding U-shaped and V-shaped wind fields, and each element is processed into a separate npy format file.

[0070] (2) Application of automated programs:

[0071] The pre-processed 5 element npy files and the normalized parameters obtained from the model training process are input into the typhoon near-surface two-dimensional wind field downscaling model TCHAT. The TCHAT model outputs the U, V and full wind speed fields after a 5-fold downscaling. The model automatically splits different elements and restores them according to the corresponding inverse normalization parameters, outputting different elements as npy files.

[0072] Figure 6 and Figure 7 The results of applying this invention to Hurricane Earl and Hurricane Jane are shown respectively (and compared with the real high-resolution field and bilinear interpolation baseline).

[0073] The high-resolution near-surface two-dimensional wind field downscaling system for typhoons of this invention, by employing deep learning technology, brings several significant benefits, especially in improving efficiency, wind field accuracy, and reducing costs. The main advantages are as follows:

[0074] (1) The high-resolution typhoon near-surface two-dimensional wind field downscaling system of the present invention has good precision and accuracy. Specifically, the MSE of the downscaling results of the TCHAT model is 7.5837 compared with the actual HWIND high-resolution observations. Compared with the results of traditional statistical downscaling methods, which use linear correction methods that are difficult to capture the nonlinear relationship of typhoon wind fields, the model proposed in this invention shows higher accuracy. Compared with the prior art, the radial wind field profile of typhoons obtained by the present invention is basically consistent with the actual observations, and there is a significant improvement compared with the commonly used downscaling reference baseline ( Figure 4 ).

[0075] (2) This invention significantly reduces computational costs and improves computational efficiency while ensuring downscaling accuracy. During the training phase, three NVIDIA 4090 graphics cards were used, taking a total of 45 hours and 31 minutes to complete model training. After the model training was completed, when only inference was performed, downscaling was performed on 42 typhoons (502 typhoon time samples) using one NVIDIA 4090 graphics card, taking a total of 41 seconds, which can meet the requirements for efficient generation of large-scale typhoon wind field data. For dynamic downscaling, high-resolution simulation of a typhoon usually takes 2 to 3 days.

[0076] Therefore, this invention exhibits high computational efficiency in both the training and inference phases: the model learns within a reasonable timeframe during training, and the inference phase can rapidly generate high-resolution wind fields from a large number of typhoon samples, far faster than traditional dynamic downscaling methods, and also shows certain advantages over current deep learning methods. Simultaneously, this invention maintains the accuracy and physical plausibility of the downscaled wind fields, making it suitable for downscaling and analyzing large-scale typhoon wind fields.

[0077] (3) This invention effectively integrates the high spatial resolution information of observational data with the long-term characteristics of reanalysis products, achieving refined reconstruction of low-resolution wind fields to high resolution, and constructing a high-resolution, high-quality long-term typhoon near-surface two-dimensional wind field dataset. Furthermore, the two-dimensional wind field dataset obtained by this invention provides good estimation of parameters such as maximum typhoon wind speed and 17m wind radius, and also shows significant advantages in characterizing the evolution of typhoon intensity over time. It can effectively capture the rapid changes in tropical cyclone intensity, which is of great significance for improving the accuracy of disaster risk assessment. This indicates that this invention can provide a solid data foundation for systematically characterizing typhoon structure, evolution, intensity, and scale characteristics, and support research on long-term typhoon climate change and typhoon disaster assessment.

[0078] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for downscaling near-surface wind fields of typhoons based on a hybrid attention Transformer framework, characterized in that... Includes the following steps: S1, Data Preprocessing: Obtain high-resolution typhoon near-surface wind field observation dataset as target data, and after screening, cropping and normalization, pair it with ERA5 reanalysis data to complete the construction of data sample set, and divide the data sample set into training set, test set and validation set according to proportion; S2, Deep Learning Model Construction: Based on the HAT model of Transformer, a generator is built that includes a shallow feature extraction module, a deep feature extraction module, and an image reconstruction module. At the same time, a discriminator is built that includes three parts: downsampling, upsampling, and discriminative output. The values ​​of the adversarial loss function and the feature matching loss function are calculated through the output of the discriminator. The discriminator includes: Downsampling is performed by using a multi-layer spectral normalization convolutional layer with a kernel size of 4×4 to progressively compress the spatial resolution of the wind speed field and extract multi-scale wind field features. Upsampling is used to gradually restore spatial resolution through bilinear interpolation and convolution operations, while skip connections are used to fuse local details in shallow layers with global structures in deep layers. The discriminant output, comprising several convolutional layers and a single-channel output layer, generates grid-by-grid true / false discrimination results. The feature matching loss and adversarial loss are calculated from the discriminator's feature layer output and discrimination result output, respectively. First, the generator output results are... Add an additional ocean-to-land mask channel to form m, to mask artifacts caused by terrain differences; then... m rotated 180° to obtain rot; The final input data to the discriminator is The true values ​​are also processed in the same way and input into the discriminator so that the model can capture the asymmetric features of the wind field. S3, Deep Learning Model Training: The total loss is a weighted average of pixel-level loss, adversarial loss, and feature matching loss, and the model parameters are updated, optimized, and saved using this total loss. The mean squared error (MSE) is used as the evaluation metric to determine the optimized model. The pixel-level loss is based on the mean absolute error loss (L1 Loss), but for the typhoon wind field downscaling task, a new Gaussian Decay Loss is designed as the pixel-level loss and calculated using the following formula. ， In the formula, and These represent the distances from grid point i to the grid center in the horizontal and vertical directions, respectively. Let i be the truth value of grid point i. This is the model prediction value for grid point i. The Gaussian parameter σ is the optimal value determined through experimental tuning, and its value is 40. The formula for calculating the total loss (All Losses) is as follows: ， In the formula, It's a pixel-level loss. It is feature matching loss. It is about combating losses.

2. The typhoon near-surface wind field downscaling method based on the hybrid attention Transformer framework according to claim 1, characterized in that: The data preprocessing in step S1 includes the following steps: S11. Obtain the high-resolution typhoon near-surface wind field observation dataset HWIND since 1998. Select samples with an intensity of 34kt or higher from the typhoon best path dataset, and transform samples with a spatial resolution of less than 0.05° to the target resolution of 0.05° using bilinear interpolation. S12, based on the location of the typhoon center, extract 80 pixels above, below, left, and right of the typhoon center and crop them to obtain a 160×160 two-dimensional array; S13, the typhoon near-surface wind field observation dataset HWIND uses U-wind speed field, V-wind speed field and full wind speed field, and splices them with two-dimensional array to form a target data sample of 160×160×3 three-dimensional array; and generates high-resolution ocean-land mask according to the latitude and longitude range of HWIND for discriminator input. S14. Based on ERA5 reanalysis data, the lowest sea level pressure position within a 1°×1° range around the center of the typhoon's optimal path dataset is found as the new typhoon center. The new typhoon center is then cropped by 16 pixels above, below, left, and right. It is then stitched together with the normalized sea level pressure field, ocean-land mask, U-wind speed field, V-wind speed field, and full wind speed field to obtain a 32×32×5 three-dimensional array of ERA5 input data samples. S15, match the ERA5 input data samples with the target data samples one by one to complete the construction of the data sample set, and then divide the data sample set into training set, test set and validation set in a ratio of 6:3:

1.

3. The typhoon near-surface wind field downscaling method based on the hybrid attention Transformer framework according to claim 2, characterized in that: In step S15, when the data sample set is divided into training set, test set and validation set according to the proportion, data samples of the same typhoon event are divided into the same dataset.

4. The typhoon near-surface wind field downscaling method based on the hybrid attention Transformer framework according to claim 1, characterized in that: The generator mentioned in step S2 includes: The shallow feature extraction module includes a 3×3 convolutional layer, which is used to map the input low-resolution wind field to a high-dimensional feature space. The deep feature extraction module includes several residual hybrid attention groups (RHAGs) and a convolutional layer with a 3×3 kernel. The RHAG consists of several hybrid attention blocks (HAT Blocks), an overlapping cross attention block (OCAB), and a 3×3 convolutional layer. Each HAT Block contains two layer normalization blocks, a channel attention block (CAB), a local window multi-head self-attention (W-MSA), and a multilayer perceptron for extracting local and global wind field features. The overlapping cross attention block (OCAB) contains two layer normalization blocks, a multilayer perceptron, and an overlapping cross attention block (OCA) for fusing wind field information from neighboring grid regions. The image reconstruction module contains two convolutional layers and pixel rearrangement upsampling. After the first 3×3 convolution, pixel rearrangement upsampling is performed, and after the last convolution, the image is mapped to the output.

5. The typhoon near-surface wind field downscaling method based on the hybrid attention Transformer framework according to claim 1, characterized in that: In step S3, during the deep learning model training process, the generator TCHAT and the discriminator U-Net adopt an asymmetric training strategy. The generator is trained for 10 iterations, while the discriminator is trained once. The generator TCHAT has 500,000 training iterations, a batch size of 4, and uses the Adam optimizer with an initial learning rate of 0.0001. The discriminator U-Net has a batch size of 4, uses the Adam optimizer, and has an initial learning rate of 0.0001.

Images