Global surface wind field downscaling method based on generative adversarial network

Abstract

The present application relates to a kind of global surface wind field downscaling method based on generative adversarial network, belong to artificial intelligence and atmospheric science cross technical field.The global surface wind field downscaling method based on generative adversarial network includes: obtaining low-resolution wind field data, high-resolution wind field label data and terrain data, with low-resolution wind field grid unit as center pixel interpolation expansion and with terrain data channel splicing construction training data pair, the generative adversarial network of construction containing generator and discriminator, with asymmetric update proportion alternately training generator and discriminator, input low-resolution wind field data to be downscaling in the generator after training output high-resolution wind field prediction result.The method is established accurate spatial correspondence in full map scale by center pixel alignment strategy, and simultaneously optimizes the training efficiency of generative adversarial network.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and atmospheric science, specifically relating to a global surface wind field downscaling method based on generative adversarial networks. Background Technology

[0002] High-quality, high-resolution marine environmental data are crucial for ocean dynamics research, air-sea interaction analysis, and marine disaster forecasting. However, limited by computational resources and the physical parameterization schemes of numerical models, the output data resolution of current operational global or regional forecasting systems is generally coarse, failing to meet the needs for detailed characterization of small- and medium-scale ocean processes. Spatial downscaling techniques serve as a key bridge between coarse-resolution data and the demand for high-precision local data. Their goal is to transform large-scale, low-resolution data into high-resolution feature fields with realistic textures and local extrema. In recent years, deep learning-based super-resolution methods have been introduced into downscaling tasks, with generative adversarial networks (GANs) attracting attention due to their potential to recover high-frequency details.

[0003] Existing deep learning-based wind field downscaling methods typically employ generative adversarial networks (GANs) to construct a mapping relationship between low-resolution and high-resolution data. While such approaches can improve resolution to some extent, they still suffer from the following drawbacks: First, the training process often employs a strategy of randomly cropping local image patches, which disrupts the global spatial correspondence between low-resolution grids and high-resolution targets, making it difficult to guarantee spatial consistency in large-scale wind field downscaling. Second, the training imbalance between the generator and discriminator is prominent, and model stability relies on complex regularization methods (such as gradient penalties), resulting in high computational costs. Third, the lack of data preprocessing design tailored to the downscaling factor and network pooling / upsampling operations leads to deviations in the spatial mapping relationship.

[0004] Therefore, how to establish an accurate spatial correspondence between low-resolution grids and high-resolution targets within a large-scale full-map area, and optimize the training efficiency of generative adversarial networks, is a technical problem that urgently needs to be solved. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a global surface wind field downscaling method based on generative adversarial networks (GANs). By interpolating and expanding the dimensionality using low-resolution grid cells as center pixels, a precise spatial correspondence between low-resolution and high-resolution grids is established across a large-scale full map. By setting the ratio of the high-to-low resolution spatial grid distance to an integer power of 2, coupled with an even-number input dimension constraint, the output size is adapted to the network sampling rate, avoiding feature map size misalignment and improving spatial consistency of features. The use of an asymmetric training strategy with high-frequency generator updates suppresses excessively rapid discriminator convergence, maintains adversarial balance, reduces the reliance on complex regularization methods such as gradient penalties, and decreases computational overhead. Introducing terrain data as an additional input channel provides spatial prior information to the model, enhancing its adaptability to different underlying surface types.

[0006] This invention provides a global surface wind field downscaling method based on generative adversarial networks, comprising the following steps: Data acquisition: Acquire low-resolution wind field data, high-resolution wind field label data, and terrain data. The ratio of the spatial grid spacing of the low-resolution wind field data to that of the high-resolution wind field label data is an integer power of 2. Acquire terrain data and interpolate the terrain data to the same spatial resolution as the high-resolution wind field data. Data alignment: Using grid cells in the low-resolution wind field data as center pixels, interpolation alignment and dimensional expansion are performed on the low-resolution wind field data so that the spatial position of each low-resolution grid cell corresponds to an integer multiple of the grid region in the high-resolution space. The spatial resolution of the expanded data is consistent with that of the high-resolution wind field label data. The terrain data is interpolated to the same spatial grid as the high-resolution wind field label data, and channel concatenation is performed with the interpolated and expanded low-resolution wind field data to obtain the model input data. The high-resolution wind field label data is used as training supervision labels. Generative Adversarial Network Construction: A generative adversarial network is constructed, which includes a generator and a discriminator. The generator adopts a U-Net encoder-decoder structure to generate predicted high-resolution data based on the model input data. The discriminator adopts a convolutional neural network structure to distinguish between high-resolution wind field label data and predicted high-resolution data. Model training: Input the model input data into the generator to generate high-resolution prediction data; input the high-resolution prediction data and high-resolution wind field label data into the discriminator for discrimination; update the generator multiple times for each discriminator update; adjust the parameters of the generator and discriminator alternately based on the discrimination results of the discriminator and the high-resolution wind field label data until the preset training conditions are met; Wind field downscaling inference: Input the low-resolution wind field data to be downscaled into the trained generator and output high-resolution wind field prediction results.

[0007] This technical solution establishes a precise spatial correspondence between low-resolution and high-resolution grids across a large-scale full-map by interpolating and expanding the dimensionality using low-resolution grid cells as center pixels. By setting the ratio of the high-to-low resolution spatial grid distance to an integer power of 2 and constraining the even-numbered dimensions of the input features, the scaling of the output size adapts to the overall scaling factor constraints of network pooling and upsampling. Under the condition of equal-size convolutional padding, it avoids the problem of feature map size misalignment and improves the spatial consistency of features. By employing an asymmetric training strategy with high-frequency generator updates, it suppresses the discriminator's excessively fast convergence, maintains adversarial balance, reduces the dependence of training on complex regularization methods such as gradient penalties, and reduces computational overhead. By introducing terrain data as an additional input channel, it provides spatial prior information to the model, enhancing its adaptability to different underlying surface types.

[0008] In some embodiments, during the data acquisition step, the ratio of the high-resolution and low-resolution spatial grid spacing is a power of 2.

[0009] This technical solution sets the ratio of spatial grid spacing to an integer power of 2 (8, or 2³), ensuring a mathematically high match between the upsampling ratio at the generator front-end and the pooling / upsampling operations within the network. Specifically, by setting the input feature map to an even dimension, the height and width of the output feature map after 8x bilinear upsampling are integer multiples of 16. This adapts to the pooling requirements of the U-Net encoder and the upsampling ratio of the decoder (4 times 2x), ensuring dimensional alignment of features at each level and skip connections under the condition of equal-sized convolution padding. This design avoids feature misalignment and boundary space offset problems caused by dimensions not meeting the downsampling ratio constraints, enabling skip connections to accurately transmit spatial information and reducing the difficulty for the network to learn spatial mappings.

[0010] In some embodiments, during the data alignment step, the interpolation alignment and dimension expansion are performed using a bilinear interpolation configuration with corner alignment mode, so that the corners of the low-resolution grid are aligned with the corners of the dimension-expanded high-resolution grid, and the center point of each low-resolution grid cell is mapped to the center position of the corresponding high-resolution grid region.

[0011] This technical solution aligns the corner points of the low-resolution grid with the corner points of the expanded high-resolution grid, making the physical grid center points of the original low-resolution grid and the grid center points of the upsampled high-resolution grid one-to-one. This avoids the inherent half-pixel geographic coordinate offset problem in the non-corner alignment mode, thereby eliminating the wind field spatial phase drift caused by coordinate offset.

[0012] In some embodiments, during the model training step, the generator is updated five times for every update of the discriminator.

[0013] This technical solution sets the update ratio of the discriminator to the generator to be 1:5, which allows the generator to have more frequent parameter update opportunities. This effectively suppresses the discriminator from dominating the adversarial process in the early stages of training due to its excessively fast convergence speed, maintains the dynamic balance between the generator and the discriminator, and avoids the problem of vanishing gradients in the generator caused by an overly strong discriminator.

[0014] In some embodiments, the generator in the generative adversarial network construction step includes a preprocessing module, an encoder path, and a decoder path. The preprocessing module upsamples the low-resolution wind field data by 8 times and concatenates it with the terrain data, outputting an initial feature map to the encoder path. The encoder path includes four downsampling modules, each of which reduces the feature map space size by half through a 2×2 max pooling operation, extracting deep features step by step. The decoder path includes four upsampling modules, each of which receives the feature map output by the previous upsampling module and concatenates it with the feature map passed from the corresponding layer of the encoder through skip connections, and doubles the feature map space size through transposed convolution.

[0015] This technical solution uses a preprocessing module to upsample low-resolution wind field data by 8 times and stitches it with topographic data, enabling spatial registration of wind field data and topographic prior fields before input to the encoder, thus avoiding the difficulty of resolution alignment within the network for multi-source inputs. Deep semantic features are extracted step by step through four downsampling modules in the encoder path, and spatial resolution is restored step by step through four upsampling modules in the decoder path with skip connections, achieving effective fusion of cross-scale features and accurate reconstruction of spatial information.

[0016] In some embodiments, during the generative adversarial network (GAN) construction step, the discriminator includes a primary feature extraction layer, a deep feature extraction layer, and an output layer connected in sequence. The primary feature extraction layer uses a 4×4 convolutional layer with a stride of 2, combined with a LeakyReLU activation function, to extract local spatial features, which are then output to the deep feature extraction layer. The deep feature extraction layer consists of multiple stacked convolutional layers, each followed by a batch normalization layer and a LeakyReLU activation function, progressively reducing the spatial resolution of the feature map and increasing the number of channels. The output layer uses a 1×1 convolutional layer to reduce the channel dimensionality and outputs probability values ​​after passing through a Sigmoid activation function.

[0017] This technical solution uses a large convolutional kernel (4×4) with a stride of 2 for downsampling in the initial feature extraction layer to quickly extract local spatial features of the wind field. By stacking multiple convolutional layers in the deep feature extraction layer, introducing batch normalization layers, and gradually reducing spatial resolution while increasing the number of channels, the gradient vanishing problem in deep network training is effectively prevented, transforming spatial information into high-dimensional abstract semantic features. By using 1×1 convolutional output layer to reduce dimensionality and then activating the output probability value with a Sigmoid function, accurate discrimination between real high-resolution data and generated data is achieved.

[0018] In some embodiments, during the model training step, a loss function is used to alternately optimize the generator and discriminator, and the loss function includes a generator loss function and a discriminator loss function.

[0019] This technical solution provides a clear optimization objective for the training of generative adversarial networks by alternately optimizing the generator and discriminator using a loss function. This enables the generator and discriminator to continuously improve their respective performance in adversarial games, laying the foundation for the subsequent introduction of specific loss function forms.

[0020] In some embodiments, the discriminator loss function is the binary cross-entropy loss, expressed as: (1); In equation (1), This represents the output probability of the discriminator. The input is low-resolution and the prior conditions include terrain. This represents the high-resolution predicted data output by the generator. This represents high-resolution wind field label data. Expressing expectations, This indicates the data distribution.

[0021] This technical solution uses binary cross-entropy loss as the loss function of the discriminator, enabling the discriminator to output a probability value close to 1 for real high-resolution data and a probability value close to 0 for the predicted data output by the generator. This accurately distinguishes high-resolution wind field label data from predicted high-resolution data and provides an effective gradient feedback signal for the generator.

[0022] In some embodiments, the generator loss function is an adversarial loss and The weighted sum of norm losses is expressed as: (2); In equation (2), express Norm, used to measure the absolute deviation between predicted wind field and actual wind field; The norm loss weighting coefficient.

[0023] This technical solution employs counter-loss and The weighted sum of norm losses serves as the generator's loss function, driving the generator to continuously improve the realism of the generated data to deceive the discriminator, while simultaneously... The norm loss-driven constraint ensures that the generated data approximates the true label at the pixel level, guaranteeing the accuracy of the downscaling results in macroscopic numerical values; Weighting coefficients of norm loss Setting the weight to 50 effectively prevents the generator from overfitting to a single deterministic structure, avoiding excessive smoothing of high-frequency information in the subgrid, thus achieving a balance between the reconstruction accuracy of macroscopic physical quantities and the high-frequency spatial heterogeneity of local small-scale wind field disturbances.

[0024] In some embodiments, a model evaluation step is also included, which includes: comparing the high-resolution wind field prediction results with high-resolution wind field label data and calculating evaluation metrics; the evaluation metrics include at least one of Pearson correlation coefficient PCC, root mean square error RMSE, mean absolute error MAE, peak signal-to-noise ratio PSNR and structural similarity index SSIM.

[0025] This technical solution uses evaluation indicators of different dimensions to quantitatively evaluate the downscaling results, which can comprehensively measure the multidimensional differences between the downscaling results and the real high-resolution data.

[0026] Based on the above scheme, the global surface wind field downscaling method based on generative adversarial networks in this embodiment of the invention establishes an accurate spatial correspondence between low-resolution and high-resolution grids by interpolating and expanding the dimensionality using low-resolution grid cells as center pixels across a large-scale map. By setting the ratio of spatial grid distances to integer powers of 2, combined with even-numbered input dimension constraints, the scaling factor is highly compatible with the number of network pooling and upsampling operations, avoiding the feature map size misalignment problem and improving spatial consistency. By employing an asymmetric training strategy with high-frequency generator updates, the discriminator's convergence is suppressed, adversarial balance is maintained, and the dependence on complex regularization methods such as gradient penalties is reduced, thus reducing computational overhead. By introducing terrain data as an additional input channel, spatial prior information is provided to the model, enhancing its adaptability to different underlying surface types. Attached Figure Description

[0027] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1This is a flowchart of the global surface wind field downscaling method based on generative adversarial networks of the present invention; Figure 2 This is an overall framework diagram of the generative adversarial network in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the upsampling process of the feature map in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram illustrating the improvement of geospatial resolution in the single pixel step dimension in Embodiment 1 of the present invention; Figure 5 This is an architecture diagram of the generator in Embodiment 1 of the present invention; Figure 6 This is an architecture diagram of the discriminator in Embodiment 1 of the present invention. Detailed Implementation

[0028] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0029] In the description of this invention, it should be understood that the terms "center", "lateral", "longitudinal", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0030] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0031] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0032] Example 1 like Figure 1As shown, in one embodiment of the global surface wind field downscaling method based on generative adversarial networks of the present invention, the global surface wind field downscaling method based on generative adversarial networks includes the following steps: S1, Data Acquisition In this step, low-resolution wind field data, high-resolution wind field label data, and terrain data are acquired. The ratio of the spatial grid spacing of the low-resolution wind field data to that of the high-resolution wind field label data is an integer power of 2, which is 8 in this embodiment, i.e., 2³.

[0033] like Figure 2 As shown, in this embodiment, the wind speed variable is selected from the fifth-generation global climate and weather reanalysis data ERA5 from the European Centre for Medium-Range Weather Forecasts (ECMWF). and The ERA5 dataset was used as the source of wind field data. The global 0.25° high-resolution data was used as the training labels, and the corresponding 2° coarse-grid data was used as the model input. The 0.25° high-resolution data corresponds to a 720×1440 grid, and the 2° coarse-grid data corresponds to a 90×180 grid. The ratio of the spatial grid spacing between the two is 8 (2³).

[0034] The terrain data used was ETOPO digital elevation map data, with an original spatial resolution of 60 arcseconds (approximately 0.0167°). The ETOPO data was resampled to the same 0.25° spatial grid as the high-resolution wind field label data through bilinear interpolation, ensuring accurate spatial registration between the two data.

[0035] The wind field data is subjected to min-max normalization, utilizing the global maximum and minimum values ​​of the ERA5 high-resolution dataset. and The components are scaled to the range [-1, 1]. The terrain data is also normalized to a uniform numerical range.

[0036] This step introduces ETOPO elevation data as terrain constraints, providing the model with physically based spatial prior information. This enables a single model to cover downscaling tasks across the entire land and sea region, thereby effectively enhancing the model's adaptability to different underlying surface types (such as land-sea boundaries, islands, and complex mountains) and improving the physical realism and spatial heterogeneity of downscaling results in complex terrain areas.

[0037] S2, Data Alignment In this step, during the data preprocessing stage, the grid cells in the low-resolution wind field data are used as the center pixels, and interpolated and aligned with the center of the high-resolution data grid. This ensures that the spatial range of each low-resolution grid cell corresponds to a power of 2 (8 in this embodiment, i.e., 2³) of the high-resolution grid spatial region. The high- and low-resolution data image pairs at each time step are used as input data for model training, and the high-resolution wind field label data is used as training supervision labels. In the early forward propagation stage of the generator network (the generator's preprocessing module), the wind field data and terrain data corresponding to the high- and low-resolution data are input to the generator network's initial front-end module. First, the low-resolution wind field data is interpolated and expanded to increase its spatial resolution to match the high-resolution wind field label data. Then, the expanded wind field data and aligned terrain data are concatenated along the channel dimension to construct an initial fusion feature map containing multi-source physical constraints. This fusion feature map then enters the generator's subsequent encoding and decoding networks, participating in the deep pooling downsampling and upsampling reconstruction process.

[0038] Specifically, a bilinear interpolation alignment configuration with `align_corners=True` is used to perform pre-upsampling with a scale factor of 8, thereby improving the resolution of the low-resolution wind field to a high-resolution grid. Taking a geographic coverage area that needs to be downscaled to improve resolution as an example, such as... Figure 3 As shown, the spatial dimensions of the feature map satisfy... , The number of pixels (matrix size) changes during this process. Figure 4 This is a schematic diagram illustrating the geospatial resolution improvement along a single pixel step size dimension. The change in this process is the geospatial step size (grid precision / resolution), meaning the corresponding geospatial step size changes from... Refine to ,satisfy The transformation relationship remains unchanged, while the geographical coverage (total latitude and longitude span) of the entire region remains constant. In the formula, This represents the height (i.e., the number of pixel rows) of the low-resolution wind field feature map. This represents the width (i.e., the number of pixel columns) of the low-resolution wind field feature map. This indicates the height of the high-resolution wind field feature map. Indicates the width of the high-resolution wind field feature map; This indicates the physical step size (resolution) of the geographic space corresponding to the low-resolution grid, which is 2° in this embodiment; This indicates the physical step size (resolution) of the geospatial area corresponding to the high-resolution grid, which is 0.25° in this embodiment.

[0039] The above process demonstrates the mathematical fit between feature maps and geographic spatial extent: the linear magnification factor of the feature map ( 8) The reduction factor of the geospatial step size ( 8) They are strictly reciprocals of each other. At the same time, the alignment configuration of align_corners=True ensures that the physical grid center points of the original low-resolution grid are mapped one-to-one with the grid center points of the high-resolution grid after upsampling. This avoids the half-pixel geographic coordinate offset problem of align_corners=False from the source and ensures strict consistency of geographic coverage before and after interpolation.

[0040] Unlike the conventional U-Net-GAN design that leaves all upsampling operations to the decoder via transposed convolution, this step places an 8x resolution boost before the encoder. The upsampled forced field can be directly stitched together with a high-resolution terrain prior field of the same size, ensuring strict spatial physical registration between multiple input sources.

[0041] In this embodiment, the ERA5 data is divided into a training set and a validation set according to time, with the training set accounting for 80% (2005-2012) and the validation set accounting for 20% (2013-2014). During the construction of the training data, the high- and low-resolution data pairs corresponding to the year and time are interpolated and aligned, and then the aligned high- and low-resolution image pairs are normalized. Similarly, the 2° low-resolution data of the validation set undergoes the same interpolation, alignment, and normalization process.

[0042] The benefits of this step are as follows: First, because the low-resolution grid is strictly set to diffuse outwards from the center pixel of the high-resolution grid, the spatial coordinate origin remains locked during multiple pooling and upsampling operations. The generator does not need to consume additional parameters to "guess" or "translate" spatial positions when reconstructing details, eliminating the spatial phase drift problem common in downscaling and ensuring spatial topology consistency. Second, unlike conventional U-Net-GAN designs that delegate all upsampling to the decoder, this step places the 8x resolution upsampling before the encoder, allowing the upsampled wind field data and the terrain prior field to be directly stitched together, ensuring strict registration of the spatial physical positions between multiple input sources. Third, by setting the input features to an even number of dimensions, the scaling factor (8 times) and the number of pooling / upsampling operations in the generator network (4 times, 2 times each) are mathematically well matched. That is, after 8 times upsampling, the feature map height and width are integer multiples of 16, which can meet the generator's requirement of 4 times 2 times pooling and upsampling. Each pooling and upsampling operation can achieve integer dimension scaling, avoiding the problem of feature map size misalignment.

[0043] S3, Generative Adversarial Network In this step, a Generative Adversarial Network (GAN) is constructed, consisting of a generator and a discriminator. The generator employs an encoder-decoder structure (i.e., the U-Net architecture) to generate high-resolution prediction data based on the model input data. The discriminator uses a convolutional neural network structure to distinguish between high-resolution wind field label data and high-resolution prediction data.

[0044] like Figure 5 As shown, the generator architecture employs the U-Net encoder-decoder framework, aiming to fuse low-resolution meteorological data with high-resolution terrain features and reconstruct a high-resolution wind field. Specifically, the shrinking path (encoder) progressively reduces the spatial size through pooling operations; the expanding path (decoder) progressively restores resolution and reconstructs details through upsampling. Skip connections are used to connect the encoder and decoder layers to optimize information flow.

[0045] In terms of specific architecture, the generator's input includes low-resolution meteorological variables. (Space dimensions are) and high-resolution terrain data (Space dimensions are) The data first enters the preprocessing module, where bilinear interpolation is used to transform the low-resolution input data. Upsample 8 times Subsequently, with terrain data The features are concatenated along the channel dimension to form a unified initial feature map. Then, the encoder path is entered, consisting of one initial double convolutional module and four downsampling modules; each downsampling module has a kernel size of [missing information]. The max pooling operation halves the spatial resolution, followed by two consecutive " The process involves convolution, batch normalization (BN), and ReLU activation, progressively expanding the number of channels to 64, 128, 256, and 512, ultimately reaching 1024 in the deepest layer. The code then enters the decoder path, containing four upsampling modules. Each module first doubles the spatial resolution of the feature map through transposed convolution (or bilinear upsampling) and performs skip connections with the feature maps of the corresponding encoder layers to preserve spatial details. Next, a double convolution module progressively reduces the number of channels to 512, 256, 128, and 64. Finally, a kernel with a size of [missing information] is used. The output convolutional layer maps feature channels to the number of target variables (e.g., or (channel), outputting the final high-resolution wind field prediction results. .

[0046] Regarding the functional relationship and forward propagation process, let the low-resolution input be... Terrain input is First, an 8x upsampling operator is used. Channel splicing operator Generate initial features The formula is After entering the encoder, set... It is a composite function containing two convolutions, BN, and ReLU. for Max pooling, feature extraction at each level is defined as follows: ,as well as ,in , This is the deepest feature layer (1024 channels). After entering the decoder, let... For a 2x upsampling operation, feature reconstruction at each level is achieved through skip connections, defined as follows: ,as well as ,in Finally, the output layer passes through... Convolution Operator Generate high-resolution wind field prediction results The formula is In the formula, For the initial fused feature map The first layer encoded feature map extracted by double convolution; For the first The feature map output by the layer encoder (as the first layer encoder) (input of the layer) For the first The feature map output by the layer encoder (where (This refers to the deepest feature map of the encoder). The index of the decoder level (value range is...) to ); deepest feature After upsampling and The first layer of decoded feature map is output through skip connections and double convolution processing; For the first The feature map output by the layer decoder (as the first layer decoder) (Input of layer upsampling); For the first The feature map output by the layer decoder. This U-Net architecture achieves efficient fusion of cross-scale features and accurate generation of spatial information through a symmetrical encoding / decoding structure and skip connections.

[0047] like Figure 6As shown, the discriminator architecture employs convolutional layers to progressively reduce the dimension of the feature maps, ultimately outputting a discrimination result to distinguish between real and generated images. The discriminator is essentially a binary classification convolutional neural network, designed to extract spatial features to determine whether the input wind field data comes from real ERA5 high-resolution data. ), or the high-resolution wind field prediction results output by the model ( ).

[0048] In terms of specific architecture, the input layer of the discriminator receives a size of... Wind field characteristic map (of which) The number of channels corresponds to the zonal wind direction. and the wind (Variables, etc.) The data first enters the initial feature extraction layer, using a 4×4 convolutional layer with a stride of 2 for downsampling, and introducing nonlinearity with the LeakyReLU activation function to initially extract the local spatial features of the wind field. Then, it enters the deep feature extraction layer, stacking multiple modules consisting of "4×4 convolutional layers + batch normalization layers + LeakyReLU activation function". In this stage, the spatial resolution of the feature map is progressively reduced in dimensionality, while the number of channels increases, thus transforming spatial information into high-dimensional abstract semantic features; the introduction of the BN layer (batch normalization layer) effectively prevents gradient vanishing problems during deep network training. Finally, in the output layer, a 1×1 convolutional layer is used to perform channel dimensionality reduction on the final feature map, mapping it to a scalar. This scalar is processed by the Sigmoid activation function, outputting a probability value between [0,1]. , which represents the confidence level of the discriminator in considering the input to be real high-resolution wind field data (1 represents real, 0 represents generated).

[0049] Regarding the functional relationship and forward propagation process, let the input variable of the discriminator be... (represent or Its forward computation process can be represented as a function mapping. First, it undergoes a first layer of convolution and activation, the formula is as follows: ; then after The formulas for convolution, normalization, and activation in the intermediate layers are as follows: The final output layer formula is: Wherein, in the formula, Input wind field data to the discriminator; This represents the convolution operation. This represents the total number of convolutional layers. The index of the intermediate layer of the network (the range of values ​​is...) to ); This is the weight matrix of the first convolutional kernel. This is the bias vector of the first convolutional kernel; For input data The feature map output after the first convolution and activation layer; For the first The weight matrix of the convolutional kernel. For the first The bias vector of the convolution kernel; For the first The feature map output by the first layer (as the first layer) (input of the layer) For the first Feature maps output by the layer; and The final output layer (the first) The weight matrix and bias vector of the convolution kernel (layer) (Note that here) For the bias vector, as described below (Different meanings) For the first Feature maps output by the layer (used as input to the output layer); For batch normalization operations, ,in The mathematical independent variable of the activation function (representing the feature tensor or scalar value passed to the function). It is a negative slope coefficient; This is the Sigmoid activation function.

[0050] The beneficial effect of this step is that the spatial alignment strategy used in the previous steps has good mathematical compatibility with the network architecture of this step. During the data preprocessing stage, the center pixel alignment strategy strictly locks the coordinate reference between the 2° low-resolution grid and the 0.25° high-resolution grid; the 8x upsampling magnification of the generator front-end is strictly consistent with the 8x resolution improvement ratio of the data, ensuring that the network learns features on the high-resolution grid throughout. Simultaneously, by setting the input features to an even-number dimension, the height and width of the output feature map after 8x bilinear upsampling are integer multiples of 16, adapting to the total magnification constraint of the generator U-Net encoder pooling and decoder upsampling (4 times 2x). Under the condition of equal-size convolution padding, the feature map can maintain an integer spatial size after each pooling and upsampling operation, ensuring dimensional alignment of features at each level and skip connections. This dual mathematical adaptation of data and network ensures, on the one hand, that the dimensions of feature maps and skip connections at each layer are strictly consistent, avoiding feature misalignment problems; on the other hand, in conjunction with the center pixel alignment at the data level, it suppresses possible spatial phase drift in meteorological downscaling from the source, so that the network does not need to consume additional computing power to learn coordinate offsets, but can focus parameter optimization on the mapping of physical features, reducing the learning difficulty and improving convergence stability.

[0051] S4, Model Training In this step, the model input data is fed into the generator to generate predicted high-resolution data. The predicted high-resolution data and the high-resolution wind field label data, which serve as training supervision labels, are then fed into the discriminator for discrimination. The generator is updated multiple times (5 times) for each discriminator update. Based on the discriminator's discrimination results and the high-resolution wind field label data, the parameters of the generator and discriminator are adjusted alternately until the preset training conditions are met.

[0052] In this embodiment, the generator and discriminator are trained using the RMSprop optimizer. The generator has an initial learning rate of 0.0002, the discriminator has an initial learning rate of 0.002, and the batch size is 64. All max-pooling layers in the network use pooling operations with a kernel size of 2, and the activation function is uniformly LeakyReLU.

[0053] In terms of loss function design, the discriminator uses binary cross-entropy loss (BCELoss) to distinguish between real high-resolution data and generated data; the generator uses adversarial loss and... A weighted combination of norm losses. The specific loss function is defined as follows: The goal of the discriminator is to distinguish real high-resolution data. Downscaled data from generator output Its loss function uses binary cross-entropy loss (BCELoss): (1); In equation (1), This represents the output probability of the discriminator. The input is low-resolution and the prior conditions include terrain. This represents the high-resolution predicted data output by the generator. This represents high-resolution wind field label data. Expressing expectations, This indicates the data distribution.

[0054] The generator's goal consists of two parts: first, to deceive the discriminator (to make...) The first is to approximate 1), and the second is to make the generated data approximate the real data. Therefore, the total loss function of the generator is the adversarial loss and The weighted sum of norm losses is expressed as: (2); In equation (2), express Norm, used to measure the absolute deviation between predicted wind field and actual wind field; These are the norm loss weight coefficients. In model tuning, they will be... Set it to 50. This is compared to a weight that is too high (such as 100). This approach effectively prevents the generator from overfitting to a single deterministic structure and avoids excessive smoothing of high-frequency information in the subgrid, thus achieving an optimal balance between the accuracy of macroscopic physical quantity reconstruction and local small-scale wind field perturbations (high-frequency spatial heterogeneity). Addressing the issue of gradient vanishing in GAN training due to the discriminator's simple task (only binary classification) leading to rapid convergence, while the generator (which needs to simultaneously fit vector wind fields and complex terrain constraints, resulting in complex optimization objectives), this embodiment designs an asymmetric high-frequency generator adversarial training strategy: for every one update of the discriminator, the generator is updated five times (update ratio 5:1). This strategy is specifically designed for the particular combination of "BCE adversarial loss + strong L1 norm loss": by tilting computational power and iteration count towards the generator, it provides the generator with more gradient descent attempts to cross complex local optima, thereby maintaining a dynamic adversarial balance between the two. It should be noted that this ratio has a clear theoretical basis; if the loss function type is changed (e.g., to a Wasserstein GAN requiring high-frequency approximation by the discriminator), this 5:1 ratio will no longer be applicable.

[0055] The model was trained for 100 rounds. To avoid any single network becoming overly dominant, this embodiment does not rely solely on the final round output. Instead, it dynamically monitors the adversarial game process, selects the adversarial critical state with the lowest generator loss and the highest discriminator loss during training, and saves its parameters as optimal hyperparameters. This ensures that the model has good robustness and generalization ability.

[0056] After training, the generator output is denormalized to convert the value range from [-1,1] to the original physical dimension of wind speed (m / s) to restore the actual wind speed value.

[0057] The beneficial effects of this step are as follows: First, the asymmetric high-frequency generator update strategy (5:1) effectively maintains the adversarial balance between the generator and the discriminator, preventing the discriminator from becoming too strong and causing the generator gradient to vanish, thus ensuring that the model converges stably to the optimal state. Second, the explicit mathematical constraints in the aforementioned data alignment step (center pixel alignment determines the spatial mapping relationship) significantly reduce the hypothesis space of the generator, making the gradient descent path of the loss function in this step smoother and more direct. Combined with the 5:1 high-frequency generator update strategy, the model can converge quickly with fewer iterations, significantly reducing training difficulty and computational cost.

[0058] S5, Downscaling Reasoning In this step, the trained generator is used to downscale the low-resolution wind field data to be downscaled, and output high-resolution wind field prediction results.

[0059] Specifically, in practical applications, low-resolution wind field data to be downscaled (such as 2° coarse grid data from ERA5 or coarse resolution outputs from other numerical weather prediction products) is acquired, normalized, and then input into a trained generator. The generator forward propagates and outputs high-resolution wind field prediction results, which are then inversely normalized to obtain the final high-resolution wind field data.

[0060] The beneficial effect of this step is that the resolution improvement from 2° to 0.25° can be completed with only one forward propagation in the inference stage. No iterative optimization or numerical solution is required. The computational efficiency is significantly better than the traditional dynamic downscaling method, which can meet the real-time operation requirements of business.

[0061] S6, Model Evaluation In this step, the high-resolution wind field prediction results are compared with the actual high-resolution wind field reference data, and the evaluation index is calculated.

[0062] To quantitatively assess the accuracy difference between the model output and the reference benchmark, this embodiment uses multiple statistical indicators for evaluation, including at least one of the following: Pearson correlation coefficient, root mean square error, mean absolute error, peak signal-to-noise ratio, and structural similarity index.

[0063] Pearson correlation coefficient This is used to measure the linear correlation between the model output value and the true value, with a value range of [-1, 1]. The closer the value is to 1, the smaller the difference between the two. The expression is: (3); In equation (3), For spatial grid location index, The total number of spatial grid points participating in the calculation. For high-resolution wind field label data (such as or Wind speed variable) in the first The specific value at each spatial location. The wind field prediction results output after downscaling are on the [number]th [year]. The corresponding values ​​at each grid point For high-resolution wind field label data in all The spatial average value at each grid point The wind field prediction results output after downscaling are in all The spatial average value across grid points.

[0064] Root mean square error The expression used to quantify the absolute deviation between the model simulation results and the reference true value is: (4); Mean Absolute Error It is used to measure the average absolute deviation between the model simulation results and the reference true value, and is expressed as: (5); Peak signal-to-noise ratio It is a core metric for evaluating the quality of reconstructed images in image super-resolution or downscaling tasks. It measures the distortion level of the reconstructed image relative to the baseline ground truth based on pixel error, and is measured in dB. A higher value indicates better image restoration and less distortion. At a distortion level greater than 30 dB, the human eye can hardly distinguish obvious distortion. The expression is: (6); In equation (6), , Indicates the first The actual data and model output data values ​​at each spatial grid point; This represents the dataset consisting of both model output data and real data. Indicates in all Traverse the model output within each spatial grid point. and real data Given all possible values, find the global maximum value and its index. This indicates that the extremum search operation is performed across the entire grid point sequence. arrive The above was carried out; Indicates in all Traverse the model output within each spatial grid point. and real data Find the global minimum value for all possible values ​​of . The difference between the two yields the dynamic range (range) after combining the model output and the real data. This range replaces the fixed peak value (e.g., 255) in traditional PSNR to accommodate the uneven data distribution in image downscaling tasks. MSE stands for Mean Squared Error. It measures the overall average error between corresponding points.

[0065] Structural similarity index Measuring the visual similarity between the reconstructed image and the original image using three dimensions—brightness, contrast, and structural information—can compensate for... Relying solely on pixel numerical errors has limitations; the closer the value is to 1, the more consistent the structures of the two images. The expression is: (7); In equation (7), The mean of the high-resolution wind field label data. The mean of the high-resolution wind field prediction results (i.e. , ), The variance of the high-resolution wind field label data. The variance of the high-resolution wind field prediction results. The covariance between high-resolution wind field label data and high-resolution wind field prediction results. and It is a small constant introduced to ensure numerical stability.

[0066] To verify the effectiveness and superiority of the present invention, this embodiment systematically evaluated the downscaling results on the ERA5 validation set (2013-2014).

[0067] Regarding spatial topology consistency, since the low-resolution grid is strictly set to diffuse outwards from the center pixel of the high-resolution grid, the origin of the spatial coordinates remains locked during multiple pooling and upsampling operations. The spatial distribution of correlation coefficients on the validation set shows that the average correlation coefficient within the study area reaches 0.96, with an average correlation coefficient of 0.98 in the ocean region and 0.92 in the land region. This indicates a significant positive correlation between the downscaled data and the true baseline, and eliminates the spatial phase drift problem commonly encountered in downscaling.

[0068] In terms of global land and sea coverage, this invention uses ETOPO elevation data as an additional input channel to provide the model with physically based terrain constraint information. The spatial distribution shows that its values ​​range from 30.00 to 71.17 dB, with an average of 52.22 dB, and only a slight difference of 1.4% between land and sea. The spatial distribution shows an average value of 0.99, with a difference of only 0.22% between marine and land regions. These results indicate that the model performs consistently well on both types of underlying surfaces, without exhibiting the performance bias of "higher accuracy for marine areas and lower accuracy for land areas" commonly seen in traditional methods.

[0069] In terms of reconstructing the spatial structure of extreme events, the spatial pattern of global wind speed evolution during the typhoon events in August-September 2013 within the validation set shows that the results of the generative adversarial network model not only downscaled the coarse-resolution wind field to 0.25°, but also highly matched the real data in terms of extreme value location, intensity and spatial structure (including typhoon structure), fully demonstrating its significant advantage in maintaining fidelity.

[0070] Example 2 The difference between this embodiment and Embodiment 1 is that the data sources are different and the research areas are different.

[0071] The training data used in this embodiment includes the ECMWF ERA-Interim reanalysis dataset, the NCEP CFSR climate prediction system reanalysis dataset, WW3 wave model output data, or simulation results from other regional climate models, replacing the ERA5 dataset in Embodiment 1. Users only need to ensure that the selected dataset contains pairable low-resolution input data and corresponding high-resolution label data, and that the ratio of their spatial grid distances is an integer power of 2 (e.g., 2, 4, 8, 16 times), with precise spatial grid registration and strict temporal alignment, to perform model training and downscaling inference according to the same process as in Embodiment 1.

[0072] The study area of ​​this embodiment is not limited to global sea areas, but can be extended to any sea area, nearshore area, or even the coverage area of ​​land meteorological stations worldwide. It is only necessary to adjust the grid range in data preprocessing according to the regional characteristics.

[0073] Through the description of several embodiments of the global surface wind field downscaling method based on generative adversarial networks of the present invention, it can be seen that the embodiments of the global surface wind field downscaling method based on generative adversarial networks of the present invention have at least one or more of the following advantages.

[0074] 1. The global surface wind field downscaling method based on generative adversarial networks provided by this invention sets the grid cells in the low-resolution wind field data as the center pixels and aligns them with the center of the high-resolution data grid through interpolation. This ensures that the spatial range of each low-resolution grid cell corresponds to an integer power of 2 in the high-resolution grid spatial region. In addition, bilinear interpolation with corner alignment mode is used to expand the dimension of the low-resolution wind field data, aligning the corners of the low-resolution grid with the corners of the expanded high-resolution grid. The center point of each low-resolution grid cell is mapped to the center position of the corresponding high-resolution grid region. This eliminates the need for the generator to consume additional parameters to "guess" or "translate" the spatial position when reconstructing details, thus eliminating spatial phase drift caused by coordinate offset, ensuring the consistency of the downscaling results in spatial topology, and improving computational efficiency.

[0075] 2. The global surface wind field downscaling method based on generative adversarial networks provided by this invention places an 8x interpolation upscaling before the encoder of the generator, so that the upscaled low-resolution wind field data and the interpolated terrain data are directly concatenated before input to the encoder. Unlike the conventional U-Net-GAN design that leaves all upsampling to the decoder for transposed convolution, this avoids the problem of spatial registration within the network due to inconsistent resolution of multiple input sources, and ensures strict registration of the spatial physical location between the wind field data and the terrain prior field.

[0076] 3. The global surface wind field downscaling method based on generative adversarial networks provided in this invention sets the resolution scaling factor of the input features to 8 (i.e., 2³), and combines this with the even-number dimension constraint of the input features, making it highly compatible with the total scaling factor constraint of encoder pooling and decoder upsampling. Under the condition of equal-size convolution padding, the feature map can maintain an integer spatial size after each pooling and upsampling operation of the generator, ensuring the alignment of hierarchical features with skip connection dimensions, avoiding the problem of inconsistent feature map sizes, and reducing the training and optimization difficulty of the network learning spatial mapping.

[0077] 4. The global surface wind field downscaling method based on generative adversarial networks provided by this invention adopts an asymmetric training strategy with high-frequency generator updates. By combining the loss function with setting the update ratio of the discriminator and the generator to 1:5, the generator obtains more frequent parameter update opportunities. This effectively suppresses the discriminator from dominating the adversarial process due to its excessively fast convergence speed in the early stage of training, maintains the dynamic balance between the generator and the discriminator, avoids the problem of generator gradient vanishing caused by an overly strong discriminator, and reduces computational overhead.

[0078] Finally, it should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0079] The above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. A global surface wind field downscaling method based on generative adversarial networks, characterized in that, Includes the following steps: Data acquisition: Acquire low-resolution wind field data, high-resolution wind field label data, and terrain data. The ratio of the high-resolution and low-resolution spatial grid spacing of the low-resolution wind field data to that of the high-resolution wind field label data is an integer power of 2. Acquire terrain data and reduce it to the same spatial resolution as the high-resolution wind field data; Data alignment: Using grid cells in the low-resolution wind field data as center pixels, interpolation alignment and dimensional expansion are performed on the low-resolution wind field data so that the spatial position of each low-resolution grid cell corresponds to an integer multiple of the grid region in the high-resolution space. The spatial resolution of the expanded data is consistent with that of the high-resolution wind field label data. The terrain data is interpolated to the same spatial grid as the high-resolution wind field label data, and channel concatenation is performed with the interpolated and expanded low-resolution wind field data to obtain the model input data. The high-resolution wind field label data is used as training supervision labels. Generative Adversarial Network Construction: A generative adversarial network is constructed, which includes a generator and a discriminator. The generator adopts a U-Net encoder-decoder structure to generate predicted high-resolution data based on the model input data. The discriminator adopts a convolutional neural network structure to distinguish between high-resolution wind field label data and predicted high-resolution data. Model training: Input the model input data into the generator to generate high-resolution prediction data; input the high-resolution prediction data and high-resolution wind field label data into the discriminator for discrimination; update the generator multiple times for each discriminator update; adjust the parameters of the generator and discriminator alternately based on the discrimination results of the discriminator and the high-resolution wind field label data until the preset training conditions are met; Wind field downscaling inference: Input the low-resolution wind field data to be downscaled into the trained generator and output high-resolution wind field prediction results.

2. The global surface wind field downscaling method based on generative adversarial networks according to claim 1, characterized in that, In the data acquisition step, the ratio of the spatial grid spacing between high-resolution and low-resolution data is 2 to the power of 3.

3. The global surface wind field downscaling method based on generative adversarial networks according to claim 1, characterized in that, In the data alignment step, interpolation alignment and dimension expansion are performed using bilinear interpolation configuration in corner alignment mode, which aligns the corners of the low-resolution grid with the corners of the expanded high-resolution grid, and maps the center point of each low-resolution grid cell to the center position of the corresponding high-resolution grid region.

4. The global surface wind field downscaling method based on generative adversarial networks according to claim 1, characterized in that, During the model training process, the discriminator is updated five times for each update of the generator.

5. The global surface wind field downscaling method based on generative adversarial networks according to claim 1, characterized in that, In the generative adversarial network construction steps, the generator includes a preprocessing module, an encoder path, and a decoder path. The preprocessing module upsamples the low-resolution wind field data by 8 times and concatenates it with the terrain data, outputting the initial feature map to the encoder path. The encoder path includes 4 downsampling modules. Each downsampling module reduces the feature map space size by half through a 2×2 max pooling operation, extracting deep features step by step. The decoder path contains four upsampling modules. Each upsampling module receives the feature map output by the previous upsampling module and concatenates it with the feature map passed from the corresponding layer of the encoder through skip connections. The feature map space size is then doubled through transposed convolution.

6. The global surface wind field downscaling method based on generative adversarial networks according to claim 1, characterized in that, In the generative adversarial network construction step, the discriminator includes a primary feature extraction layer, a deep feature extraction layer and an output layer connected in sequence; the primary feature extraction layer uses a convolutional layer with a kernel size of 4×4 and a stride of 2 in conjunction with the LeakyReLU activation function to extract local spatial features, and outputs them to the deep feature extraction layer; The deep feature extraction layer consists of multiple stacked convolutional layers. Each convolutional layer is followed by a batch normalization layer and a LeakyReLU activation function to progressively reduce the spatial resolution of the feature map and increase the number of channels. The output layer uses a 1×1 convolutional layer to reduce the channel dimensionality and outputs probability values ​​after passing through a Sigmoid activation function.

7. The global surface wind field downscaling method based on generative adversarial networks according to claim 5 or 6, characterized in that, In the model training step, loss functions are used to alternately optimize the generator and discriminator. The loss functions include the generator loss function and the discriminator loss function.

8. The global surface wind field downscaling method based on generative adversarial networks according to claim 7, characterized in that, The discriminator loss function is the binary cross-entropy loss, expressed as: （1）； In equation (1), This represents the output probability of the discriminator. The input is low-resolution and the prior conditions include terrain. This represents the high-resolution predicted data output by the generator. This represents high-resolution wind field label data. Expressing expectations, This indicates the data distribution.

9. The global surface wind field downscaling method based on generative adversarial networks according to claim 7, characterized in that, The generator loss function is adversarial loss and The weighted sum of norm losses is expressed as: （2）； In equation (2), express Norm, used to measure the absolute deviation between predicted wind field and actual wind field; The norm loss weighting coefficient.

10. The global surface wind field downscaling method based on generative adversarial networks according to claim 1, characterized in that, It also includes a model evaluation step, which includes: comparing the high-resolution wind field prediction results with the high-resolution wind field label data and calculating the evaluation index; the evaluation index includes at least one of Pearson correlation coefficient, root mean square error, mean absolute error, peak signal-to-noise ratio and structural similarity index.

Images