Physical information guided wind speed field data downscaling method fusing terrain and time perception

Abstract

The invention discloses a physical information guided wind speed field data downscaling method fusing terrain and time perception. The method comprises the steps of 1, multi-source physical constraint modeling and data construction; step 2, constructing a generative downscaling framework based on a conditional diffusion probability model; step 3, designing a noise prediction network structure fusing terrain and time prior; and 4, model training and rapid sampling reasoning. The method solves the problems that the texture of a wind speed field generated by an existing deep learning downscaling method is excessively smooth, high-frequency turbulence details are lost, and the terrain forcing effect and the wind speed seasonal / daily periodic physical law are ignored, and provides a framework combining explicit physical feature engineering and a conditional diffusion model. Through dual deep fusion of terrain and time information, high fidelity, physical consistency and reasoning efficiency of a generated result are considered.

Description

Technical Field

[0001] This invention relates to the fields of wind energy resource assessment and meteorological data processing, specifically to a method for downscaling wind speed field data guided by physical information that integrates topographic and time-awareness. Background Technology

[0002] With the global energy transition, wind energy is playing an increasingly important role as a clean and renewable energy source. High spatiotemporal resolution wind resource datasets are fundamental for macro-level wind farm site selection, micro-level layout optimization, and grid dispatching. Currently, meteorological data mainly comes from numerical weather prediction models or reanalysis data (such as ERA5). However, limited by computational resources and the grid spacing of physical models, the spatial resolution of these data is usually low (e.g., 30km for ERA5), which cannot accurately characterize the local wind field features under complex terrain (such as ridge acceleration effects and canyon winds), making it difficult to directly meet the needs of refined wind energy development.

[0003] To address the issue of insufficient resolution, existing methods are mainly divided into two categories: dynamic downscaling and statistical downscaling. Dynamic downscaling utilizes nested regional climate models, offering clear physical meaning but incurring extremely high computational costs and time consumption. Statistical downscaling, on the other hand, establishes a mapping relationship from low-resolution to high-resolution models, resulting in high computational efficiency. In recent years, deep learning-based super-resolution techniques (such as convolutional neural networks and Transformers) have been applied to wind field downscaling.

[0004] However, existing deep learning-based downscaling methods suffer from the following drawbacks: First, over-smoothing occurs. Traditional convolutional neural networks or Transformer models typically aim to minimize pixel-level errors (such as mean squared error), tending to generate averaged results, leading to the loss of crucial high-frequency details in the wind field (such as gusts and turbulent textures). Second, the physical mechanisms are lacking. Existing methods often treat wind speed fields as ordinary natural image processing, ignoring the physical properties of the wind field. In reality, local terrain is the main driver of wind speed changes, but most methods simply stitch together terrain data as input, failing to effectively capture the multi-scale terrain-wind field interaction. Third, temporal priors are not utilized. Wind speed exhibits significant diurnal variations (such as sea breezes) and seasonal patterns. Existing models often ignore this strong prior knowledge, resulting in poor physical consistency of the generated wind field over time.

[0005] Therefore, how to generate high-resolution wind fields with rich details and conforming to the physical laws of terrain and time while performing high-efficiency calculations is a technical problem that urgently needs to be solved. Summary of the Invention

[0006] The purpose of this invention is to provide a method for downscaling wind speed field data guided by physical information that integrates terrain and time perception, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for downscaling wind speed field data guided by physical information integrating terrain and time awareness includes: Step 1, perform multi-source physical constraint modeling and data construction, including: Step 1.1: Construct wind field data pairs; Step 1.2: Construct multi-channel terrain feature data; Step 1.3: Perform time-periodic feature encoding; Step 2, constructing a generative downscaling framework based on a conditional diffusion probability model, including; Step 2.1: Add noise to the real wind field to construct a forward diffusion process; Step 2.2: Construct an inverse denoising model to predict the added noise; Step 3: Design a noise prediction network structure that integrates terrain and temporal priors. The noise prediction network structure includes residual blocks with integrated channel-spatial attention, a terrain information fusion module, and a temporal information fusion module. Step 4 involves model training and fast sampling inference, including: Step 4.1: Design the loss function and train the model; Step 4.2: Construct a fast sampling inference strategy.

[0008] Further, step 1.1 includes: Define low-resolution wind field data as High-resolution wind field data is , It is the goal of model training. It is one of the input conditions of the model; in practice, it is achieved through bicubic interpolation. Upsampling to Enter the model after ensuring the same dimensions.

[0009] Further, step 1.2 includes: Constructing high-resolution static terrain data It contains the following four channel information: (1) digital elevation model data; (2) horizontal terrain gradient, using the Sobel operator. The data from the digital elevation model are obtained by convolution calculation, reflecting the steepness of the terrain in the east-west direction; (3) Vertical terrain gradient, using the Sobel operator The data of the digital elevation model is obtained by convolution calculation, reflecting the steepness of the terrain in the north-south direction; (4) the land-sea mask is a binary matrix that identifies the boundary between the ocean and the land. These four channel information items are processed by a shallow convolutional network to obtain the final terrain feature map; Step 1.3 includes: The time feature vector τ is constructed using sine / cosine coding: , Here, "doy" indicates the day of the year. This represents the annual cycle constant in days; hour indicates which hour of the day. This represents the daily periodicity constant expressed in hours.

[0010] Further, step 2.1 includes: Using a denoising diffusion probability model as the backbone, the downscaling task is modeled as a process under physical conditions and information. The guided conditional probability distribution generation process, in its forward process, is a fixed Markov chain that feeds data into real high-resolution data. Gaussian noise is gradually added until it becomes pure Gaussian noise. T is a preset parameter representing the total number of iterations required to completely transform the real data into pure Gaussian noise. At any diffusion time step t, the noisy data... Direct sampling is performed using the following formula: , in, Represents the original signal. The introduced random Gaussian noise term is represented by the symbol. express It follows a standard normal distribution with a mean vector of 0 and a covariance matrix that is the identity matrix. t represents the diffusion time step, which is a discrete integer variable ranging from 1 to T. It is a predefined noise variance scheduling parameter corresponding to the diffusion time step t, which monotonically decreases with the diffusion time step t.

[0011] Further, step 2.2 includes: The reverse process aims to transform pure Gaussian noise. The original signal was recovered. Train a noise prediction network Its input is the current noisy data. The output is a prediction of the noise added at that diffusion time step, including the diffusion time step t and all physical condition information c.

[0012] Furthermore, the design of the residual block integrating channel-spatial attention in step 3 includes: Introduce cascaded channel attention modules and spatial attention modules into the residual block; The channel attention module is implemented as follows: First, the input feature maps are processed separately. Global average pooling and global max pooling are performed along the spatial dimension to aggregate spatial information. Then, the two generated feature vectors are input into a multilayer perceptron with shared weights. Finally, the outputs of the multilayer perceptrons are summed and a channel weight map is generated by passing the sigmoid activation function. ; Applying the channel weight map to the input features yields channel-refined features. ; The spatial attention module is implemented as follows: First, the features after channel refinement... Based on this, global average pooling and global max pooling are performed along the channel dimension to obtain two two-dimensional feature maps. These two feature maps are then concatenated and processed through a... The convolutional layers are fused together, and finally a spatial weight map is generated using the Sigmoid function. ; Finally, the spatial weight graph is applied to This yields the final features enhanced by the dual attention mechanism. ; this final feature The residual connection is added to the original input and used as the input for the next layer.

[0013] Furthermore, the design of the terrain information fusion module in step 3.2 includes: Design a dual fusion strategy; Shallow direct fusion: This involves fusing the terrain features generated in step 1.2. Noisy data input from the network and upsampling Direct splicing along the channel dimension; Global Multi-Scale Terrain Attention: A parallel multi-scale terrain feature extraction module is designed, consisting of three parallel branches. Each branch employs a dilated convolution with different dilation rates to extract microscopic terrain details, mesoscale terrain undulations, and macroscopic terrain orientation, respectively. The three feature extraction branches are then fused to generate a global terrain attention map. The global terrain attention map was used to modulate the encoder's output features. : , in, This represents element-wise multiplication, with modulated characteristics. It is passed directly to the decoder via a long hop connection.

[0014] Furthermore, the design of the time information fusion module in step 3.3 includes: First, the position encoding vector at diffusion time step t is generated. Embedded vector with the physical time feature vector τ in step 1.3 Adding them together yields the combined time embedding. : , Within each residual block in the network, a fully connected layer is used to... The mapping is performed using a scaling factor γ and a translation factor β on the feature map. Perform affine transformation: , in τ represents element-wise multiplication of matrices, and Mod represents affine transformation. Through this mechanism, the noise prediction network can dynamically adjust the response of the convolution kernel based on the dual conditions of "diffused noise level" determined by t and "physical time" determined by τ, thereby accurately recovering the wind field characteristics that conform to seasonal and diurnal variation patterns while denoising.

[0015] Further, step 4.1 includes: The loss function is constructed using the L1 norm, and the specific formula is as follows: , in, Represents the loss function. Represents the mathematical expectation. This represents the introduced random Gaussian noise term. This indicates that the noise prediction network is based on noisy data. The noise predicted by the diffusion time step t and physical condition information c Represents the L1 norm; During training, the stochastic gradient descent algorithm is used, based on the loss function. Calculate gradient And update the network parameters θ through backpropagation until the loss function is reached. convergence.

[0016] Further, step 4.2 includes: A denoising diffusion implicit model sampling algorithm is used during the inference phase; Let the diffusion time step of the sampling subsequence be... At any diffusion time step According to the corresponding diffusion time step Noisy samples and noise Calculate the noisy sample from the previous diffusion time step. The formula is as follows: , in, Corresponding to diffusion time steps and Predefined noise variance scheduling parameters, This is for the final high-resolution wind field.

[0017] Compared with the prior art, the beneficial effects of the present invention are:

[0018] 1. Overcoming the problem of overly smoothed wind field textures generated by traditional methods. This invention utilizes a conditional diffusion probability model combined with an L1 norm loss function, which, compared to previous methods based on convolutional neural networks, can effectively restore high-frequency details and edge sharpness of the wind speed field, significantly improving the physical realism and detail richness of the generated data.

[0019] 2. Significantly enhances the physical consistency of downscaling results. By explicitly calculating the terrain gradient and introducing time-periodic encoding, the model can accurately capture the periodic patterns of terrain forcing and wind speed variations with seasons and days, overcoming the deficiency of existing methods in lacking physical mechanism constraints.

[0020] 3. Enhance the model's adaptive extraction capability for complex spatiotemporal features. Design a noise prediction network that integrates channel-spatial attention mechanism and temporal modulation to achieve deep fusion of multi-scale terrain features and dynamic temporal information, effectively improving the model's prediction accuracy under complex terrain and variable climate conditions.

[0021] 4. Balancing generation quality and inference efficiency to meet engineering application needs. A fast sampling strategy based on a denoising diffusion implicit model is designed, compressing the thousands of iterations required by traditional diffusion models to dozens of steps. While ensuring the generation quality of high-resolution wind fields, it significantly reduces computational costs and time overhead, making it valuable for practical engineering applications. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall process of the method of the present invention.

[0023] Figure 2 This is a schematic diagram of the forward diffusion and reverse denoising process of the conditional diffusion model.

[0024] Figure 3 This is a schematic diagram of the residual block structure for integrated channel-space attention. Detailed Implementation

[0025] 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.

[0026] Please see Figure 1 A method for downscaling wind speed field data guided by physical information that integrates terrain and time perception, comprising:

[0027] Step 1, perform multi-source physical constraint modeling and data construction, including:

[0028] Step 1.1, construct wind field data pairs

[0029] Define low-resolution wind field data as , Represents the set of real numbers, symbol The data represents real tensors with corresponding dimensions; high-resolution wind field data is... Where "2" represents the zonal and meridional wind components of the wind speed. This indicates the size of the low-resolution wind field. This indicates the size of a high-resolution wind field. It is the goal of model training. It is one of the input conditions of the model. In practice, it needs to be converted using bicubic interpolation. Upsampling to Enter the model after ensuring the same dimensions.

[0030] Step 1.2: Construct multi-channel terrain feature data

[0031] Topography is a key factor influencing near-surface wind flow. This invention not only uses raw elevation data but also explicitly calculates topographic gradients to reinforce physical constraints. High-resolution static topographic data is constructed. It contains the following four channel information: (1) Digital elevation model data, i.e., normalized elevation data; (2) Horizontal terrain gradient, using the Sobel operator. The data from the digital elevation model are obtained by convolution calculation, reflecting the steepness of the terrain in the east-west direction; (3) Vertical terrain gradient, using the Sobel operator The data from the digital elevation model are obtained by convolution calculation, reflecting the steepness of the terrain in the north-south direction; (4) the land-sea mask, which is a binary matrix that identifies the boundary between the ocean and the land. These four channel information items are processed by a shallow convolutional network to obtain the final terrain feature map. and The format is as follows:

[0032]

[0033] Step 1.3: Perform time-periodic feature encoding.

[0034] To account for the diurnal and seasonal variations in wind speed, this invention introduces a time prior. Since time is periodic (e.g., 11 PM and midnight are adjacent), directly using numerical normalization would disrupt this continuity. Therefore, this invention employs sine / cosine coding to construct the time feature vector. :

[0035]

[0036] Here, "doy" indicates the day of the year. The term represents the annual cycle constant in days, typically 365 or 366; hour indicates which hour of the day. .

[0037] Step 2: Construct a generative downscaling framework based on a conditional diffusion probability model, such as... Figure 2 As shown, it includes:

[0038] Step 2.1: Add noise to the real wind field to construct a forward diffusion process.

[0039] Using a denoising diffusion probability model as the backbone, the downscaling task is modeled as a process under physical conditions and information. The process of generating conditional probability distributions under guidance. The forward process is a fixed Markov chain, generated by feeding data into real high-resolution data. Gaussian noise is gradually added until it becomes pure Gaussian noise. T is a preset parameter representing the total number of iterations required to completely transform the real data into pure Gaussian noise. In this embodiment of the invention, T is set to 1000. At any diffusion time step t, the noisy data... Sampling can be done directly using the following formula:

[0040]

[0041] in, Represents the original signal. The introduced random Gaussian noise term is represented by the symbol. express It follows a standard normal distribution with a mean vector of 0 and a covariance matrix that is the identity matrix. t represents the diffusion time step, which is an integer discrete variable ranging from 1 to T, and its value corresponds to the noise intensity during the diffusion process. It is a predefined noise variance scheduling parameter that monotonically decreases with time step t, meaning that as t increases, the amount of the original signal retained decreases. The less noise, the less noise added. More and more.

[0042] Step 2.2: Construct an inverse denoising model to predict the added noise.

[0043] The reverse process aims to transform pure Gaussian noise. The original signal was recovered. This invention trains a deep neural network (noise prediction network). Its input is the current noisy data. The output is a prediction of the noise added at that diffusion time step, including the diffusion time step t and all physical condition information c.

[0044] Step 3: Design a noise prediction network structure that integrates terrain and temporal priors. The noise prediction network structure includes residual blocks with integrated channel-spatial attention, a terrain information fusion module, and a temporal information fusion module.

[0045] a. Design of residual blocks for integrated channel-spatial attention:

[0046] To enable the network to adaptively focus on key feature channels (such as specific wind speed components) and key spatial regions (such as abrupt topographic changes) in the wind field, such as... Figure 3 As shown, this invention introduces a cascaded channel attention module and a spatial attention module into the residual block. Let the input feature map be... Where C is the number of channels, This represents the size of the feature map.

[0047] The channel attention module is implemented as follows: First, the input feature maps are processed separately. Global average pooling (AvgPool) and global max pooling (MaxPool) are performed along the spatial dimension to aggregate spatial information. Then, the two generated feature vectors are input into a multilayer perceptron (MLP) with shared weights. Finally, the outputs of the MLPs are summed and a channel weight map is generated by applying a sigmoid activation function. The specific calculation formula is as follows:

[0048]

[0049] in, AvgPool represents the sigmoid activation function, MaxPool represents the global average pooling operation, and MLP represents the multilayer perceptron computation operation with shared weights. The channel weight map is applied to the input features to obtain the channel-refined features. :

[0050]

[0051] in, This represents element-wise multiplication.

[0052] The spatial attention module is implemented as follows: First, the features after channel refinement... Based on this, global average pooling and global max pooling are performed along the channel dimension to obtain two two-dimensional feature maps. These two feature maps are then concatenated and processed through a... The convolutional layers are fused together, and finally a spatial weight map is generated using the Sigmoid function. The specific calculation formula is as follows:

[0053]

[0054] in, This represents a convolution operation with a kernel size of 7. Finally, the spatial weight map is applied to... This yields the final features enhanced by the dual attention mechanism. :

[0055]

[0056] This final feature The residual connection is added to the original input and used as the input for the next layer.

[0057] b. Design of the terrain information fusion module:

[0058] Considering that the influence of terrain on wind field is multi-scale, this invention designs a dual fusion strategy.

[0059] Shallow direct fusion: This involves fusing the terrain features generated in step 1.2. Noisy data input from the network and upsampling The layers are directly concatenated along the channel dimension. This allows the network's first convolutional layer to perceive pixel-level local terrain features.

[0060] Global Multi-Scale Terrain Attention: A parallel multi-scale terrain feature extraction module is designed, consisting of three parallel branches. Each branch employs a dilated convolution with different dilation rates (1, 2, and 3) to extract microscopic terrain details, mesoscale terrain undulations, and macroscopic terrain orientation, respectively. The three feature extraction branches are then fused to generate a global terrain attention map. The global terrain attention map was used to modulate the encoder's output features. .

[0061]

[0062] in, This represents element-wise multiplication. Modulated characteristics. The wind field is directly transmitted to the decoder via a long-hop connection, ensuring that the generated wind field strictly conforms to the terrain trend on a large scale.

[0063] c. Design of the time information fusion module:

[0064] To endow the model with "time-aware" capability, this invention utilizes feature-level linear modulation technology. First, the position vector of the diffusion time step t of the diffusion model is encoded. Embedded vector with the physical time feature vector τ in step 1.3 Add, where It is the temporal feature embedding dimension, resulting in combined temporal embedding. :

[0065]

[0066] Within each residual block in the network, a fully connected layer is used to... Mapped to scaling factor Translation factor Where C is the number of channels. For the feature map Perform an affine transformation (Mod):

[0067]

[0068] in 't' represents element-wise multiplication of matrices, and 'mod' represents affine transformation. Through this mechanism, the noise prediction network can dynamically adjust the response of the convolution kernel based on the dual conditions of "the level of diffused noise (determined by t)" and "the time in the physical world (determined by τ)", thereby accurately recovering wind field characteristics that conform to seasonal and diurnal variation patterns while denoising, such as physical phenomena like higher wind speeds in winter and enhanced thermal turbulence in the afternoon.

[0069] Step 4 involves model training and fast sampling inference, including:

[0070] Step 4.1: Design the loss function and train the model.

[0071] To train the noise prediction network To enable accurate estimation of added noise, this invention uses the L1 norm to construct a loss function, the specific formula of which is as follows:

[0072]

[0073] in, Represents the loss function. This represents the mathematical expectation, which is the expectation of the real high-resolution wind field samples over the entire training dataset with a diffusion step size of t. The corresponding conditional information c and random Gaussian noise Take an average. This represents the introduced random Gaussian noise term. This indicates that the noise prediction network is based on noisy data. The noise is predicted by the diffusion time step t and physical condition information c. The L1 norm is used to indicate the L1 loss function, which is less sensitive to outliers and produces more constant gradients. This encourages the model to not only focus on the overall structure but also better preserve high-frequency components in the wind field, such as local gusts and texture details like wind shear caused by terrain. Furthermore, the L1 norm tends to produce sparse solutions, which helps the model remove background noise and generate wind field images with clearer boundaries and higher sharpness.

[0074] During training, the stochastic gradient descent algorithm is used, based on the loss function described above. Calculate gradient And update the network parameters θ through backpropagation until the loss function is reached. Convergence means that the model has the ability to accurately predict noise (i.e., recover the wind field structure) at any noise level.

[0075] Step 4.2, Construct a fast sampling inference strategy

[0076] Standard denoised diffusion probability models require simulating the complete inverse Markov chain process during sampling, resulting in low computational efficiency. To meet the speed requirements of practical engineering applications, this invention employs a denoised diffusion implicit model sampling algorithm in the inference stage. This algorithm generalizes the diffusion process to a non-Markov process, allowing intermediate time steps to be skipped during sampling, thereby generating high-quality wind fields in a very small number of steps.

[0077] Let the diffusion time step of the sampled subsequence be... At any diffusion time step According to the corresponding diffusion time step Noisy samples and network prediction noise Calculate the noisy sample from the previous diffusion time step. The formula is as follows:

[0078]

[0079] in, Corresponding to diffusion time steps and Predefined noise variance scheduling parameters, This is for the final high-resolution wind field.

[0080] Using the above formula, the model can utilize the trained physical constraints and, guided by condition c, quickly iterate to generate high-resolution, high-fidelity wind field data.

[0081] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method of fusing terrain and time-aware physical information guided wind speed field data downscaling, characterized in that, The application relates to a method for generating wind field data, comprising the following steps: Step 1, performing multi-source physical constraint modeling and data construction, comprising: Step 1.1, constructing a wind field data pair; Step 1.2, constructing multi-channel terrain feature data; Step 1.3, performing time periodicity feature coding; Step 2, constructing a generative downscaling framework based on a conditional diffusion probability model, comprising: Step 2.1, adding noise to a real wind field to construct a forward diffusion process; Step 2.2, constructing a reverse denoising model to predict the added noise; Step 3, designing a noise prediction network structure integrating terrain and time priors, wherein the noise prediction network structure comprises a residual block integrating channel-spatial attention, a terrain information fusion module and a time information fusion module; Step 4, performing model training and fast sampling inference, comprising: Step 4.1, performing loss function design and model training; Step 4.2, constructing a fast sampling inference strategy.

2. The method according to claim 1, wherein, The step 1.1 comprises: The low-resolution wind field data is defined as , and the high-resolution wind field data is , is the target of model training, is one of the input conditions of the model; in actual operation, after bicubic interpolation is used to upsample to the same size as , the model is input.

3. The method according to claim 2, wherein, The step 1.2 comprises: Constructing high resolution static terrain data , including the following four channel information: (1) digital elevation model data; (2) horizontal terrain gradient, using Sobel operator Convolution calculation is carried out on the data of digital elevation model, which reflects the steepness of the terrain in the east-west direction; (3) vertical terrain gradient, using Sobel operator Convolution calculation is carried out on the data of digital elevation model, which reflects the steepness of the terrain in the north-south direction; (4) sea-land mask, which is a binary matrix, marking the boundary between ocean and land; The four channel information is processed through a shallow convolutional network to obtain a final terrain feature map; The step 1.3 comprises: A time feature vector tau is constructed by using sine / cosine coding: ， where doy represents the day of the year, represents the annual period constant in days; hour represents the hour of the day, represents the daily period constant in hours.

4. The method according to claim 1, wherein, The step 2.1 comprises: A denoising diffusion probabilistic model is adopted as the backbone framework, and the downscaling task is modeled as a process of generating a conditional probability distribution under the guidance of a physical condition information The forward process is a fixed Markov chain, which gradually adds Gaussian noise to the real high-resolution data T is a preset parameter representing the total number of iterations required to completely convert the real data into pure Gaussian noise, and at any diffusion time step t, the noisy data is directly sampled by the following formula:​ ， wherein denotes the original signal, denotes an introduced random Gaussian noise term, the symbol denotes obeys a standard normal distribution with mean vector 0 and covariance matrix the identity matrix , t denotes a diffusion time step, which is an integer discrete variable taking values in the range from 1 to T, is a predefined noise variance schedule parameter corresponding to diffusion time step t, which is monotonically decreasing with diffusion time step t.

5. The method according to claim 4, wherein, The step 2.2 comprises: The reverse process aims to transform pure Gaussian noise. The original signal was recovered. Train a noise prediction network Its input is the current noisy data. The output is a prediction of the noise added at that diffusion time step, including the diffusion time step t and all physical condition information c.

6. The method according to claim 3, wherein, The design of the residual block integrating channel-spatial attention in the step 3 comprises: A channel attention module and a spatial attention module are introduced in series in the residual block; The implementation manner of the channel attention module is as follows: first, input feature maps are respectively processed Global average pooling and global maximum pooling are performed along the spatial dimension to aggregate spatial information, then the two generated feature vectors are respectively input into a multi-layer perceptron with shared weights, and finally the outputs of the multi-layer perceptron are added and a channel weight map is generated through a Sigmoid activation function ; applying the channel weight map to the input feature to obtain a channel-refined feature ; The implementation manner of the spatial attention module is as follows: firstly, on the basis of the channel-refined feature , global average pooling and global maximum pooling are respectively performed along the channel dimension to obtain two two-dimensional feature maps, then the two feature maps are spliced and fused through a convolution layer, and finally a spatial weight map is generated through a Sigmoid function ; Finally, the spatial weight map is applied to , to obtain the final feature enhanced by the dual attention mechanism ; the final feature will be added to the original input through a residual connection as the input of the next layer.

7. The method according to claim 6, wherein, The design of the terrain information fusion module in the step 3.2 comprises: A double fusion strategy is designed; Shallow direct fusion: concatenating the features of the topographic information generated in step 1.2 with the noisy data input to the network and up-sampled directly in the channel dimension; Global multi-scale terrain attention: a parallel multi-scale terrain feature extraction module is designed, which is composed of three parallel branches, which respectively adopt three different dilated rate hole convolutions. The three kinds of convolutions extract microscopic terrain details, mesoscale terrain undulations and macroscopic terrain trends respectively; a global terrain attention map is generated after the fusion of the three features ; the global terrain attention map is used to modulate the output features of the encoder : ， wherein, denotes element-wise multiplication, modulated features passed directly to the decoder through a long skip connection.

8. The method according to claim 7, wherein, The design of the time information fusion module in the step 3.3 comprises: First the position encoding vector for diffusion time step t is added to the embedding vector for the physical time feature vector τ in step 1.3, resulting in the combined time embedding :​​ ， Within each residual block in the network, the full connection layer is utilized to map the feature maps to an affine transformation with a scaling factor γ and a translation factor β: ， wherein denotes element-wise multiplication of matrices, and Mod denotes an affine transformation. Through this mechanism, the noise prediction network is able to dynamically adjust the response of the convolution kernel according to the dual conditions of the "diffused noise level" determined by t and the "time of the physical world" determined by τ, so as to accurately restore the wind field characteristics conforming to the seasonal and diurnal variation rules while denoising.

9. The method according to claim 1, wherein, The step 4.1 comprises: An L1 norm is used to construct a loss function, and the specific formula is as follows: ， wherein, denotes a loss function, denotes a mathematical expectation, denotes an introduced random Gaussian noise term, denotes that the noise prediction network predicts the noise based on the noisy data , the diffusion time step t, and the physical condition information c, denotes an L1 norm; During training, a stochastic gradient descent algorithm is used to update the network parameters θ according to the loss function Computing the gradient and updating the network parameters θ by backpropagation until the loss function converges.

10. The method of claim 1, wherein The step 4.2 comprises: A denoising diffusion implicit model sampling algorithm is used in the inference stage; Let the diffusion time step of the sampled subsequence be At any diffusion time step The noisy sample at the corresponding diffusion time step and the noise are used to compute the noisy sample at the previous diffusion time step according to the following formula: ， wherein, corresponding to diffusion time steps and a predefined noise variance schedule parameter, is the final generated high-resolution wind field.

Images