A High-Resolution Physical Constraint Downscaling Method for Wind Speed ​​Based on Deep Attention Residual Networks

Abstract

This invention discloses a high-resolution physical constraint downscaling method for wind speed based on a deep attention residual network, involving the interdisciplinary fields of meteorological data downscaling and artificial intelligence. The method includes the following steps: constructing a physically anchored residual learning framework, designing and training a deep attention residual network, integrating physical constraints and generating a high-resolution wind speed field, and optimizing the model and training strategies. This invention utilizes a physically anchored residual learning framework, an encoder-decoder architecture integrating convolutional block attention modules, a loss function that fuses wind field divergence penalty terms, and non-negative physical constraints to solve the problems of high computational cost, insufficient physical consistency, weak generalization ability in complex terrain areas, and reliance on manual feature engineering in traditional downscaling methods. It achieves accurate and efficient mapping from coarse-resolution meteorological data to high-resolution, physically reliable wind speed fields, improving the model's representational ability and operational application value in multiple geographical scenarios.

Description

Technical Field

[0001] This invention relates to the field of meteorological data downscaling and artificial intelligence, specifically a high-resolution physical constraint downscaling method for wind speed based on a deep attention residual network. Background Technology

[0002] High spatiotemporal resolution wind speed data is indispensable in atmospheric environment simulation, wind energy resource assessment, local climate research, and extreme weather event analysis. However, while current mainstream global climate models and meteorological reanalysis data can provide physically consistent and long-term global coverage, their spatial resolution is only between 0.25° and 1°. This makes it difficult to accurately capture the local circulation and abrupt wind speed changes caused by complex underlying surfaces such as mountains, canyons, coastlines, and urban clusters, which severely restricts their application in refined regional management and decision support.

[0003] To bridge the scale gap between global climate models and local applications, dynamic downscaling and statistical downscaling have become two mainstream technical approaches. Dynamic downscaling methods can accurately simulate local physical processes, but they consume enormous computational resources, requiring hundreds or even thousands of CPU hours to simulate climate scenarios spanning several to decades, making it difficult to meet the large-scale application needs of multi-member, long-term, and multi-scenario climate predictions. Traditional statistical downscaling methods are more computationally efficient, but rely on simple linear assumptions and cumbersome manual feature engineering, making it difficult to capture the complex nonlinear relationships between topography, surface roughness, and wind speed. Furthermore, they are prone to generating spatially discontinuous prediction fields, and their generalization performance significantly decreases in insufficiently covered terrain types. In recent years, deep learning technologies, represented by Convolutional Neural Networks (CNNs), have provided a new direction for statistical downscaling of meteorological data due to their powerful capabilities in automatically extracting spatial features and learning nonlinear relationships. Related research has explored the application of models such as U-Net and Generative Adversarial Networks (GANs) in temperature and precipitation downscaling. However, purely end-to-end data-driven models are criticized as black boxes, lacking clear physical consistency guarantees. The prediction results may deviate too much from the background field or violate basic physical intuition, reducing their credibility in scientific research and practical applications.

[0004] Existing patents still have significant shortcomings in incorporating physical constraints for wind speed downscaling. Taking patent application CN118366046A as an example, it adopts a dual-path independent processing structure and a three-path terrain-aided information processing network, resulting in a large number of model parameters, complex structure, high training and inference costs, and is not conducive to business deployment. Although terrain data is introduced, it lacks explicit constraints on physical laws such as wind field mass conservation and momentum conservation, which may lead to physical inconsistencies in the prediction results. The feature interaction mechanism is simplistic and does not achieve adaptive attention calibration for channel and spatial dimensions, making it difficult to dynamically focus on key features. The loss function uses simple linear weighting, which does not fully balance the constraint strength of wind field vector magnitude and direction, and the training stability needs to be optimized.

[0005] In summary, effectively embedding physical knowledge into deep learning frameworks to construct downscaling models that are both efficient, accurate, and physically reliable has become a core challenge in the interdisciplinary field of meteorological artificial intelligence. Existing technologies generally suffer from high computational costs, limited model expressive power, lack of physical constraints, insufficient generalization ability, and low automation of feature engineering. They also have significant shortcomings in model lightweighting, embedding physical mechanisms, and adaptive fusion of multimodal features. Therefore, this paper proposes a high-resolution, physically constrained downscaling method for wind speed based on deep attention residual networks to overcome these problems. Summary of the Invention

[0006] The purpose of this invention is to provide a high-resolution physical constraint downscaling method for wind speed based on deep attention residual networks, so as to solve the problems mentioned in the background art.

[0007] To address the aforementioned technical problems, this invention provides a high-resolution physical constraint downscaling method for wind speed based on deep attention residual networks, comprising the following steps: Step (1) Construct a physical anchoring residual learning framework: Decompose the target high-resolution wind speed field into the sum of the physical background field after upsampling of the coarse-resolution background field and the residual correction term learned by the deep neural network. The physical background field is obtained by bilinear interpolation and used as the physical anchor point. Step (2) Design and train deep attention residual network: The network adopts an encoder-decoder architecture. The encoder extracts multi-scale features through three levels of progressive downsampling. After each level, a convolutional block attention module (CBAM) is embedded to perform adaptive feature calibration. The decoder reconstructs the high-resolution residual field through upsampling and skip connections, and introduces regularization techniques to prevent overfitting. Step (3) Physical constraint integration and high-resolution wind speed field generation: The predicted residual field is added to the physical background field, and non-negative constraints are applied through nonlinear functions to ensure that the output wind speed field is physically reasonable; Step (4) Model optimization and training strategy: The mean squared error loss function combining L2 regularization and wind field divergence penalty term is adopted, and the adaptive moment estimation optimizer and learning rate decay strategy are used to train the model.

[0008] Furthermore, the encoder's three-stage downsampling structure is configured as follows: EncoderBlock1 maps the 8-channel input to 64 channels, with a 3×3 convolutional kernel and a 5×5 receptive field; EncoderBlock2 maps the 64 channels to 128 channels, with a 3×3 convolutional kernel and a 9×9 receptive field, achieving a 2x downsampling; EncoderBlock3 maps the 128 channels to 256 channels, with a 3×3 convolutional kernel and a 13×13 receptive field, achieving a 4x downsampling. Each encoder stage contains two convolutional layers, followed by a batch normalization layer and a ReLU activation function.

[0009] Furthermore, the implementation process of the Convolutional Block Attention (CBAM) module is as follows: Channel attention: The input feature map is simultaneously subjected to global average pooling and global max pooling to generate two channel descriptors, which are then fed into a shared multilayer perceptron to generate channel attention weights, which are multiplied with the original feature map channel by channel. Spatial attention: The channel-calibrated feature maps are subjected to average pooling and max pooling along the channel dimension, and then concatenated and generated through a 7×7 convolutional layer to generate spatial attention weights. These weights are multiplied with the channel-calibrated feature maps at each spatial position to obtain a dual-attention calibrated feature map.

[0010] Furthermore, the reduction ratio r of the multilayer perceptron adopts a dynamic adjustment mechanism, specifically: the input region is divided into flat areas, complex terrain areas, or transition zones by a lightweight scene classifier, and r=32, r=8, or r=16 are assigned accordingly; or it is based on the terrain complexity index: , Through mapping functions: , Adaptive calculation of the r value, where The mean and standard deviation of the training set complexity. This is the adjustment coefficient.

[0011] Furthermore, the total loss function is: in: The mean square error between the predicted wind speed and the true value; This is the L2 weight decay term; This is a divergence penalty term; This is a hyperparameter used to balance the strength of various constraints.

[0012] Furthermore, the decoder adopts a three-level upsampling structure symmetrical to the encoder, and achieves upsampling through bilinear interpolation or transposed convolution. Each level receives the attention calibration feature map of the corresponding level of the encoder through skip connections and performs channel concatenation. The output layer of the decoder maps the high-dimensional features into a single-channel residual field through a 1×1 convolution kernel. After the convolutional layer of the decoder, a ReLU activation function and a batch normalization layer are connected in sequence, and a Dropout layer with a dropout rate of 0.1 is introduced.

[0013] Furthermore, the specific implementation of physical constraints is as follows: pass: ; Negative wind speed values ​​are truncated to 0, where R represents the physical background field, and R represents the residual field. The wind field divergence is calculated using the central difference method, and the divergence on the discrete grid is approximately: , and This represents the grid spacing.

[0014] Furthermore, the input features include a coarse-resolution background field, high-resolution static features, and optional dynamic features. The coarse-resolution background field is the wind speed field output by a global or regional numerical weather prediction model. The static features include digital elevation models, topographic relief, surface roughness length, land use type, and distance from the coastline. The dynamic features are time information encoded by sine / cosine transform. All input features are resampled to the target high-resolution grid and Z-score normalized, and then concatenated along the channel dimension as the network input tensor.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. A physically anchored residual learning framework was constructed to ensure the dynamic consistency and physical rationality of the downscaling results: This invention abandons the traditional strategy of directly fitting complex wind fields using an "end-to-end" black box model and innovatively proposes a physically anchored residual learning framework. This framework decomposes the high-resolution wind speed field into the sum of a coarse-resolution background field and residual correction terms learned by the deep network. By using the upsampled background field, which contains large-scale dynamic consistency, as a benchmark, the deep network only needs to learn relatively simple corrections (S1) caused by local topography and non-uniform underlying surface. This design not only significantly reduces the learning difficulty and convergence time of the model, but also fundamentally prevents the generated results from physically deviating from the dynamic framework of the original climate model, ensuring the physical consistency between the downscaled wind field and the large-scale circulation, and greatly improving the scientific value and operational credibility of the results.

[0016] 2. An adaptive hybrid attention mechanism is introduced, significantly improving the model's representation ability and generalization performance in complex terrain areas: This invention creatively integrates the Convolutional Block Attention Module (CBAM) into the encoder-decoder architecture, achieving dual adaptive calibration of multi-source features. The channel attention mechanism can dynamically evaluate the importance weights of different physical feature channels (such as elevation, roughness, coastline distance, etc.), enabling the model to autonomously identify key driving factors in different geographical scenarios (such as mountains, coastlines, and cities) (S2.2.1). The spatial attention mechanism can accurately focus on key areas sensitive to wind speed changes (such as topographic relief zones, land-sea junctions, and obstacle wakes), prioritizing the allocation of limited computing resources to these complex areas (S2.2.2). This ability to "focus on what when" makes the model more accurate in extracting features from complex underlying surfaces, significantly enhancing the model's scene adaptability and generalization ability, and overcoming the shortcomings of traditional methods in areas with abrupt terrain changes.

[0017] 3. This invention achieves a highly efficient end-to-end automated process, avoiding the tedious and expert-dependent manual feature engineering of traditional statistical methods: Through a designed deep convolutional network, this invention can automatically learn the complex nonlinear mapping (S2) of the high-resolution residual field from the original coarse-resolution feature field. Unlike traditional statistical downscaling, this method eliminates the need for tedious manual feature design and extraction (e.g., calculating local means, gradients, curvature, and other statistics), achieving fully automated end-to-end processing from data input to output. This not only significantly improves the method's efficiency and lowers the application threshold but also avoids biases introduced by improper subjective feature selection, ensuring the objectivity and repeatability of the method.

[0018] 4. Deep integration of physical constraints and data-driven approaches achieves an effective balance between scientific credibility and precision and efficiency: This invention is not a purely "data-driven" black-box model, but rather deeply embeds meteorological prior knowledge into the deep learning process through a residual learning framework, non-negative wind speed constraints, and physically consistent network design (S4). The generated downscaling results possess both the high accuracy and nonlinear fitting capabilities of deep learning models, while strictly adhering to fundamental physical principles such as non-negative wind speed and large-scale dynamic coordination. This "physical guidance + data-driven" paradigm effectively solves the physical inconsistency problem often criticized in pure deep learning models, producing results that combine the cutting-edge nature of data science with the rigor of atmospheric science, demonstrating extremely high credibility in both research and operational applications.

[0019] 5. The network structure is scientifically rigorous and incorporates effective regularization strategies to ensure excellent model performance and robust generalization ability: The encoder-decoder architecture used in this invention can effectively extract and reconstruct multi-scale spatial features. Simultaneously, batch normalization and dropout regularization techniques are widely employed in the network (S2.3), effectively suppressing overfitting during training. This allows the model to perform well not only on the training set but also on unseen validation and test sets, as well as under different geographical regions and climatic conditions (S3), maintaining stable and reliable prediction performance, demonstrating good practicality and generalization value.

[0020] 6. Possessing broad applicability and significant application value, this invention provides reliable technical support for refined meteorological services across multiple fields: The downscaling method proposed in this invention has high versatility. Its core framework is not only applicable to wind speed downscaling, but can also be extended to the refined forecasting and reconstruction of various meteorological elements such as temperature, humidity, and precipitation after fine-tuning. Therefore, the outputs of this invention can directly serve multiple high-value fields such as wind energy resource assessment, urban planning, atmospheric environment simulation, and severe weather early warning, providing a powerful core technical tool for national and regional refined meteorological services, disaster prevention and mitigation, and sustainable development, with very broad application prospects. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of a high-resolution physical constraint downscaling method for wind speed based on a deep attention residual network according to the present invention. Figure 2 This is a model architecture diagram of the channel attention mechanism for a high-resolution physical constraint downscaling method for wind speed based on a deep attention residual network, as described in this invention. Figure 3 This is a model architecture diagram of the spatial attention mechanism of a high-resolution physical constraint downscaling method for wind speed based on a deep attention residual network according to the present invention. Figure 4 This is a model architecture diagram of a hybrid attention mechanism for a high-resolution physical constraint downscaling method for wind speed based on a deep attention residual network, as described in this invention. Figure 5 This is a schematic diagram of the unscaled GFS numerical weather prediction product of the high-resolution physical constraint downscaling method for wind speed based on deep attention residual networks according to the present invention. Figure 6 This is a schematic diagram of a downscaled GFS numerical weather prediction product based on a high-resolution physical constraint downscaling method for wind speed according to the present invention. Figure 7This is a schematic diagram of the fitting between the downscaled GFS numerical product and the real observation of a high-resolution physical constraint downscaling method for wind speed based on a deep attention residual network according to the present invention. Figure 8 The present invention provides a high-resolution physical constraint downscaling method for wind speed based on deep attention residual networks, which includes the model training loss function and error variation curve. Detailed Implementation

[0022] The technical solutions of 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 some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please see Figures 1-8 The present invention provides a technical solution: See Figures 1-8 As shown, an embodiment of a high-resolution physical constraint downscaling method for wind speed based on deep attention residual networks is presented: By constructing a deep learning framework that embeds physical information, we achieve accurate and efficient mapping from coarse-resolution numerical weather prediction products to high-resolution, physically consistent wind speed fields. Specifically, this includes the following steps: S1. Construct a physically anchored residual learning framework: To address potential physical inconsistencies in purely data-driven models, this invention proposes a physically anchored residual learning paradigm. This framework does not require deep neural networks to directly learn the complex and dynamically large full-range wind speed field; instead, it allows them to learn a relatively smooth and small-scale residual correction term.

[0024] Specifically, the target high-resolution wind speed field As shown in equation (1): in, This represents the target high-resolution wind speed field. It is a field upsampled from the coarse-resolution wind speed background field to the target high-resolution field using bilinear interpolation. It carries large-scale circulation dynamic coordination information from the original numerical model and serves as a physical anchor point. The deep neural network to be trained is tasked with learning wind speed corrections caused by local fine-scale features (such as complex terrain and surface roughness heterogeneity) that cannot be reflected in coarse-resolution fields.

[0025] This design greatly reduces the learning difficulty of the model and fundamentally ensures that the generated high-resolution field does not physically deviate from the large-scale dynamic background, significantly improving the physical rationality of the results and their value for business applications.

[0026] S2. Design and training of deep attention residual networks: This step is the core of the invention for achieving high-precision downscaling. The deep attention residual network... An encoder-decoder architecture integrating an attention mechanism is adopted, and its mathematical expression is shown in equation (2): in, and These represent the encoder and decoder, respectively. These are all the trainable parameters of the model.

[0027] S2.1 Multi-level feature encoding and adaptive attention calibration: encoder It is responsible for extracting multi-scale spatial context information from the input multi-source features, from coarse to fine. It employs a three-level progressive downsampling structure, with each level containing two convolutional layers, progressively extracting features through nonlinear transformations. The mathematical description of the level encoder is Equation (3): in, This represents the convolution operation. It is the ReLU activation function. Indicates batch normalization. and These are the weights and bias parameters of the convolutional layer, respectively. For input features.

[0028] The configurations for each level are as follows: EncoderBlock1: Input channels 8→64, convolution kernel 3×3, receptive field 5×5, mainly extracts low-level features (edges, textures, and other basic patterns). EncoderBlock2: Input channels 64→128, convolution kernel 3×3, receptive field 9×9, capturing mid-level features (local terrain structure features). EncoderBlock3: Input channels 128→256, convolution kernel 3×3, receptive field 13×13, learns high-level features (global spatial distribution pattern). S2.2, Integrated Convolutional Block Attention Module: A key aspect of this invention is that a convolutional block attention module is embedded after each level of the encoder. This module achieves adaptive recalibration of the feature map through a serial channel attention and spatial attention mechanism.

[0029] S2.2.1 Channel attention mechanism: The role of channel attention is to evaluate the differences in importance of different feature channels (i.e., different input physical variables) to the current task.

[0030] Step 1: Feature aggregation.

[0031] For the input feature map Global average pooling and global max pooling are performed respectively to generate two channel descriptors as shown in equations (4) and (5): in These represent the channel descriptors after average pooling and max pooling, respectively.

[0032] Step 2: Weight generation.

[0033] The two descriptors are fed into a shared multilayer perceptron to generate the final channel attention weights. As shown in equation (6): in, and For the weights of the MLP, To reduce the ratio, For ReLU function, This is the Sigmoid function.

[0034] In the channel attention submodule of CBAM, the reduction ratio The degree of dimensional compression in the intermediate layer is controlled, and its selection is based on the following considerations: Balancing parameter count with computational efficiency: It can reduce the number of MLP parameters to approximately 1 / 3 of the original number of channels with almost no loss of expressive power. This significantly reduces the computational burden.

[0035] Sufficiency of channel information aggregation: Experiments show that when When the value is between 8 and 32, the channel attention mechanism has the best effect on feature recalibration. By centering within this range, we can avoid information loss due to excessive compression and prevent overfitting caused by redundant parameters.

[0036] Compatibility with common network architectures: Widely used in classic attention networks such as ResNet and SENet, it has been proven to be an empirically effective and stable setup.

[0037] Distinguishing characteristics of wind fields: In wind speed downscaling tasks, the input features (such as elevation, roughness, and distance from the coast) have clear physical meanings, and the number of channels is usually no more than 10. This is sufficient to achieve effective inter-channel weight allocation.

[0038] In the wind speed downscaling task, the channel attention mechanism exhibits clear physical interpretability: in areas with complex topography, the elevation (DEM) and topographic relief channels are given higher weights; in the land-sea interface zone, the importance of the distance from the coast and surface roughness channels is significantly increased; during the nighttime period, the weight of time-encoded features (such as the sine and cosine representations of hours) increases significantly; while in flat areas, the weight of topographically related channels is relatively reduced.

[0039] To further improve the model's adaptability to different underlying surface types, this invention designs a dynamic adjustment mechanism, enabling... The value can be automatically adjusted based on the geometric and topographic features of the input area: 1) Scene-based classification guidance choose: Train a lightweight scene classifier (based on DEM, land use, etc.) to divide the input region into: Flat areas (plains, ocean) Complex terrain areas (mountains, canyons) Transitional zone (coastline, urban-rural fringe) Different assignments are given according to different scenarios. Values ​​(e.g., used in complex terrain areas) To enhance sensitivity, use in flat areas To reduce the amount of computation.

[0040] 2) Terrain complexity adaptive calculate: Define terrain complexity index (The ratio of standard deviation to mean), and design a mapping function: in Let the mean and standard deviation of the complexity be the values ​​of the training set. This is the adjustment coefficient.

[0041] This mechanism enables the model to automatically reduce its speed in complex terrain areas. This enhances the ability of the channel to distinguish information.

[0042] Multi-scale attention parallel dynamic selection: Multiple different Attention branches of value (e.g.) =8,16,32), dynamically fusing the outputs of each branch through learnable gating weights to achieve fully data-driven operation. Adaptive.

[0043] S2.2.2 Spatial Attention Mechanism The role of spatial attention is to focus on key spatial areas in the feature map that are sensitive to changes in wind speed.

[0044] Step 1: Feature compression.

[0045] Feature map after channel attention calibration Average pooling and max pooling are performed along the channel dimension, as shown in equations (7) and (8): in , Step 2: Spatial weight calculation.

[0046] The two feature maps are concatenated and then passed through a convolutional layer to generate a spatial attention map. As shown in equation (9): in It is a 7×7 convolution. For channel splicing, This is the Sigmoid function.

[0047] The distribution of spatial attention weights is highly correlated with geographical features: in complex terrain areas, such as mountains and canyons where wind speeds accelerate, the weights increase significantly; in transitional zones, such as areas with strong thermal contrasts, such as the boundaries between land and sea and between urban and rural areas, the weights also increase significantly; downstream of obstacles, such as the leeward wake areas of building complexes or forests, also show high attention weights; while in flat and open areas, the weights are relatively low due to the relatively gentle changes in wind speed.

[0048] S2.2.3 Attention Convergence: In deep neural networks, the core idea of ​​the attention mechanism is to mimic the selective attention mechanism of the human visual system, enabling the model to adaptively focus on the most relevant parts of the input features. For the wind speed downscaling task, the impact of different geographical regions and feature channels on local wind speed varies significantly. To address this, we introduce a Convolutional Block Attention Module (CBAM) after each layer of the encoder. This module achieves adaptive calibration of multi-source features through sequential channel attention and spatial attention mechanisms.

[0049] The mathematical expression of the CBAM module is shown in equations (10) and (11): in, For the input feature map, and These represent the channel and spatial attention maps, respectively. This represents element-wise multiplication. The output features are those after dual attention calibration.

[0050] S2.3 Progressive Feature Decoding and Residual Field Generation: decoder It is responsible for upsampling the abstract high-level features extracted by the encoder and reconstructing them into a high-resolution residual field. The task is to gradually recover the spatial size of the feature map by using a structure that is roughly symmetrical to the encoder, through transposed convolution or nearest neighbor upsampling combined with convolution.

[0051] To compensate for potential loss of detail during downsampling, the decoder introduces skip connections, directly passing feature maps from corresponding encoder levels to the decoder to achieve feature fusion between high and low levels. Ultimately, the decoder outputs a residual correction field with the same size as the target wind speed field. Throughout the network, regularization techniques such as Dropout are also introduced to randomly "shut down" some neurons, effectively preventing overfitting and enhancing the model's generalization ability.

[0052] The decoder uses a symmetrical convolutional structure to gradually restore the high-dimensional feature mapping to the target residual field as shown in equation (12): in This represents the predicted residual field. Dropout regularization was introduced during the decoding process, with a random dropout rate of 0.1, to enhance the model's generalization ability.

[0053] S3, Physical Constraint Integration and High-Resolution Wind Field Generation: In obtaining the residual field predicted by the network Then, it is compared with the physical background field. The sums yield a preliminary predicted field. To ensure the final output conforms to basic meteorological principles, a simple physical constraint—a nonnegativity constraint—is applied. Wind speed, as a scalar, should not be negative. Therefore, the final predicted wind speed field is generated by a nonlinear function, as shown in equation (13): in, The operation truncates any possible negative values ​​to 0, thus ensuring that the generated wind speed field is physically plausible.

[0054] S4. Model Optimization and Training Strategies: The model is trained by minimizing the difference between the predicted field and the true high-resolution observed field. A mean squared error with L2 regularization is used as the loss function, and a wind field divergence penalty term is added to the loss function to balance fitting accuracy and model complexity, ensuring that the predicted wind field conforms to the basic principles of atmospheric dynamics. The optimization process uses an adaptive moment estimator optimizer with a learning rate decay strategy to ensure the stability and efficiency of the training process. By training on large-scale historical data, the model learns a complex and universal mapping relationship from coarse-resolution features to high-resolution residuals.

[0055] S4.1 Loss function strategy that introduces wind field divergence penalty term: S4.1.1 Divergence Calculation Module Design: For predicted high-resolution wind fields (U and V component form), its horizontal divergence field can be calculated: On a real discrete grid, the central difference method can be used for approximate calculations. Let the grid spacing be... and Then the divergence calculation can be expressed as: S4.1.2 Construction of the divergence penalty term: To encourage the generation of near-source-free / sink-free wind fields (suitable for large-scale background compatibility requirements), an L2 norm penalty term for divergence is added to the loss function: S4.1.3, Expansion of the Total Loss Function The total loss function consists of the following three parts: in: The mean square error between the predicted wind speed and the true value; This is the L2 weight decay term; This is a divergence penalty term; This is a hyperparameter used to balance the strength of various constraints.

[0056] It should be noted that: We selected the complex topographical region of southeastern coastal China as our study area. Using ECMWFERA5 reanalysis data (approximately 31 km resolution) as coarse-resolution input, combined with static data such as 30-meter resolution DEM and land use, the goal was to downscale wind speed to 12.5 km resolution.

[0057] After training, the model performed excellently on the independent test set. Compared with direct bilinear interpolation, the wind speed field generated by our method at a resolution of 12.5 km reduced the root mean square error by approximately 40%, and significantly outperformed traditional interpolation methods and benchmark deep learning models that did not incorporate attention mechanisms and physical constraints in depicting the spatial distribution details of wind speed in complex mountainous and coastal areas. The spatial attention map clearly shows the model's focus on areas of abrupt topographic changes and land-sea boundaries, and the channel attention weights also align with meteorological understanding of the impact of different underlying surface features on wind speed, demonstrating the model's good physical interpretability.

[0058] Summarize: A physically anchored residual learning framework was constructed to ensure the dynamic consistency and physical rationality of the downscaling results: This invention abandons the traditional strategy of directly fitting complex wind fields using an "end-to-end" black box model and innovatively proposes a physically anchored residual learning framework. This framework decomposes the high-resolution wind speed field into the sum of a coarse-resolution background field and residual correction terms learned by the deep network. By using the upsampled background field, which contains large-scale dynamic consistency, as a benchmark, the deep network only needs to learn relatively simple corrections (S1) caused by local topography and non-uniform underlying surface. This design not only significantly reduces the learning difficulty and convergence time of the model, but also fundamentally prevents the generated results from physically deviating from the dynamic framework of the original climate model, ensuring the physical consistency between the downscaled wind field and large-scale circulation, and greatly improving the scientific value and operational credibility of the results.

[0059] An adaptive hybrid attention mechanism is introduced, significantly improving the model's representation ability and generalization performance in complex terrain areas: This invention creatively integrates the Convolutional Block Attention Module (CBAM) into the encoder-decoder architecture, achieving dual adaptive calibration for multi-source features. The channel attention mechanism can dynamically evaluate the importance weights of different physical feature channels (such as elevation, roughness, coastline distance, etc.), enabling the model to autonomously identify key driving factors in different geographical scenarios (such as mountains, coastlines, and cities) (S2.2.1). The spatial attention mechanism can accurately focus on key areas sensitive to wind speed changes (such as topographic relief zones, land-sea junctions, and obstacle wakes), prioritizing the allocation of limited computing resources to these complex areas (S2.2.2). This ability to "focus on what when" makes the model more accurate in extracting features from complex underlying surfaces, significantly enhancing the model's scene adaptability and generalization ability, and overcoming the shortcomings of traditional methods in areas with abrupt terrain changes.

[0060] This invention achieves a highly efficient end-to-end automated process, avoiding the tedious and expert-dependent manual feature engineering of traditional statistical methods. Through a designed deep convolutional network, it automatically learns the complex nonlinear mapping (S2) of the high-resolution residual field from the original coarse-resolution feature field. Unlike traditional statistical downscaling, this method eliminates the need for cumbersome manual feature design and extraction (e.g., calculating local means, gradients, curvature, etc.), achieving fully automated end-to-end processing from data input to output. This significantly improves the method's efficiency, lowers the application threshold, and avoids biases introduced by improper subjective feature selection, ensuring the objectivity and repeatability of the method.

[0061] By deeply integrating physical constraints and data-driven approaches, this invention achieves an effective balance between scientific credibility and precision and efficiency. It is not a purely "data-driven" black-box model, but rather deeply embeds meteorological prior knowledge into the deep learning process (S4) through a residual learning framework, non-negative wind speed constraints, and a physically consistent network design. The generated downscaling results possess both the high accuracy and nonlinear fitting capabilities of deep learning models, while strictly adhering to fundamental physical principles such as non-negative wind speed and large-scale dynamic coordination. This "physical guidance + data-driven" paradigm effectively solves the physical inconsistency problem often criticized in pure deep learning models. The resulting products combine the cutting-edge nature of data science with the rigor of atmospheric science, exhibiting extremely high credibility in both research and operational applications.

[0062] The network structure is scientifically rigorous and incorporates effective regularization strategies to ensure excellent model performance and robust generalization ability. The encoder-decoder architecture used in this invention can effectively extract and reconstruct multi-scale spatial features. Simultaneously, batch normalization and dropout regularization techniques are widely employed in the network (S2.3) to effectively suppress overfitting during training. This allows the model to perform well not only on the training set but also on unseen validation and test sets, as well as under different geographical regions and climatic conditions (S3), maintaining stable and reliable prediction performance, thus possessing good practicality and generalization value.

[0063] Possessing broad applicability and significant application value, this invention provides reliable technical support for refined meteorological services across multiple fields. The downscaling method proposed in this invention is highly versatile; its core framework is not only applicable to wind speed downscaling but, with fine-tuning, can be extended to refined forecasting and reconstruction of various meteorological elements such as temperature, humidity, and precipitation. Therefore, the outputs of this invention can directly serve multiple high-value fields such as wind energy resource assessment, urban planning, atmospheric environment simulation, and severe weather early warning, providing a powerful core technical tool for national and regional refined meteorological services, disaster prevention and mitigation, and sustainable development, with a very broad application prospect.

Claims

1. A high-resolution physical constraint downscaling method for wind speed based on a deep attention residual network, characterized in that, Includes the following steps: Step (1) Construct a physical anchoring residual learning framework: Decompose the target high-resolution wind speed field into the sum of the physical background field after upsampling of the coarse-resolution background field and the residual correction term learned by the deep neural network. The physical background field is obtained by bilinear interpolation and used as the physical anchor point. Step (2) Design and train deep attention residual network: The network adopts an encoder-decoder architecture. The encoder extracts multi-scale features through three levels of progressive downsampling. After each level, a convolutional block attention module (CBAM) is embedded to perform adaptive feature calibration. The decoder reconstructs the high-resolution residual field through upsampling and skip connections, and introduces regularization techniques to prevent overfitting. Step (3) Physical constraint integration and high-resolution wind speed field generation: The predicted residual field is added to the physical background field, and non-negative constraints are applied through nonlinear functions to ensure that the output wind speed field is physically reasonable; Step (4) Model optimization and training strategy: The mean squared error loss function combining L2 regularization and wind field divergence penalty term is adopted, and the adaptive moment estimation optimizer and learning rate decay strategy are used to train the model.

2. The high-resolution physical constraint downscaling method for wind speed based on deep attention residual networks as described in claim 1, characterized in that: The encoder's three-stage downsampling structure is configured as follows: EncoderBlock1 maps the 8-channel input to 64 channels, with a 3×3 convolutional kernel and a 5×5 receptive field; EncoderBlock2 maps the 64 channels to 128 channels, with a 3×3 convolutional kernel and a 9×9 receptive field, achieving a 2x downsampling; EncoderBlock3 maps the 128 channels to 256 channels, with a 3×3 convolutional kernel and a 13×13 receptive field, achieving a 4x downsampling. Each encoder stage contains two convolutional layers, followed by a batch normalization layer and a ReLU activation function.

3. The high-resolution physical constraint downscaling method for wind speed based on deep attention residual networks as described in claim 1, characterized in that: The implementation process of the Convolutional Block Attention (CBAM) module is as follows: Channel attention: The input feature map is simultaneously subjected to global average pooling and global max pooling to generate two channel descriptors, which are then fed into a shared multilayer perceptron to generate channel attention weights, which are multiplied with the original feature map channel by channel. Spatial attention: The channel-calibrated feature maps are subjected to average pooling and max pooling along the channel dimension, and then concatenated and generated through a 7×7 convolutional layer to generate spatial attention weights. These weights are multiplied with the channel-calibrated feature maps at each spatial position to obtain a dual-attention calibrated feature map.

4. The high-resolution physical constraint downscaling method for wind speed based on deep attention residual networks as described in claim 3, characterized in that: The reduction ratio r of the multilayer perceptron is dynamically adjusted, specifically: the input region is divided into flat areas, complex terrain areas, or transition zones using a lightweight scene classifier, and r=32, r=8, or r=16 are assigned accordingly; or it is based on the terrain complexity index: ， Through mapping functions: ， Adaptive calculation of the r value, where The mean and standard deviation of the training set complexity. This is the adjustment coefficient.

5. The high-resolution physical constraint downscaling method for wind speed based on deep attention residual networks as described in claim 1, characterized in that: The total loss function is: in: The mean square error between the predicted wind speed and the true value; This is the L2 weight decay term; This is a divergence penalty term; This is a hyperparameter used to balance the strength of various constraints.

6. The high-resolution physical constraint downscaling method for wind speed based on deep attention residual networks as described in claim 1, characterized in that: The decoder adopts a three-level upsampling structure symmetrical to the encoder. Upsampling is achieved through bilinear interpolation or transposed convolution. Each level receives the attention calibration feature map of the corresponding level of the encoder through skip connections and performs channel stitching. The output layer of the decoder maps high-dimensional features to a single-channel residual field through a 1×1 convolutional kernel. After the decoder convolutional layer, a ReLU activation function and a batch normalization layer are connected in sequence, and a Dropout layer with a dropout rate of 0.1 is introduced.

7. The high-resolution physical constraint downscaling method for wind speed based on deep attention residual networks as described in claim 1, characterized in that: The specific implementation of physical constraints is as follows: pass: ； Negative wind speed values ​​are truncated to 0, where R represents the physical background field, and R represents the residual field. The wind field divergence is calculated using the central difference method, and the divergence on the discrete grid is approximately: , and This represents the grid spacing.

8. The high-resolution physical constraint downscaling method for wind speed based on deep attention residual networks as described in claim 1, characterized in that: Input features include coarse-resolution background field, high-resolution static features, and optional dynamic features; The coarse-resolution background field is the wind speed field output by the global or regional numerical weather prediction model. Static features include digital elevation model, topographic relief, surface roughness length, land use type, and distance from the coastline. Dynamic features are time information encoded by sine / cosine transform. All input features are resampled to the target high-resolution grid and Z-score normalized. After being concatenated along the channel dimension, they are used as the network input tensor.

Images