Wind field downscaling method based on physical constraint and artificial intelligence

Abstract

The invention discloses a wind field downscaling method based on physical constraint and artificial intelligence, and relates to the technical field of meteorological numerical forecasting, and the method comprises the steps: obtaining an input data set which comprises topographic data, low-resolution forecasting data and observation data; feature extraction and data processing are carried out on the input data set to obtain a training data set, the training data set is utilized to train a neural network model fused with the attention mechanism, and a loss function of the neural network model comprises physical constraint terms. According to the method, feature information of the terrain and the air pressure field is brought into model input, feature extraction is carried out on the air pressure field through principal component analysis, and the adaptability of the model to different terrain and weather backgrounds is improved.

Description

Technical Field

[0001] This invention relates to the field of meteorological numerical forecasting technology, specifically a wind field downscaling method based on physical constraints and artificial intelligence. Background Technology

[0002] Accurate wind field forecasting is a core issue in atmospheric science research and operational applications. Over the past two decades, numerical weather prediction has achieved significant breakthroughs in dynamic framework construction, physical process parameterization, and data assimilation, resulting in a substantial improvement in forecast accuracy. However, constrained by computational resource limitations and the inherent nonlinear characteristics of atmospheric systems, conventional numerical models struggle to directly output high-resolution wind field products. To overcome this limitation, researchers have developed various downscaling techniques, including statistical downscaling methods (such as linear regression and artificial neural networks) and dynamic downscaling methods (such as nested regional climate models). In recent years, deep learning technology has demonstrated significant advantages in meteorological downscaling, with downscaling models based on convolutional neural networks (CNNs) and generative adversarial networks (GANs) outperforming traditional statistical methods in multiple validation cases.

[0003] Current mainstream wind field downscaling methods suffer from two key drawbacks: First, existing deep learning models often treat meridional wind, zonal wind, and total wind speed as independent variables, neglecting the intrinsic relationships between these physical quantities, leading to a lack of physical consistency in forecast results. Second, the models are insufficiently adaptable to complex terrain and multi-scale weather systems, especially in areas with dramatic topographic relief and during severe weather events, where forecast errors increase significantly. These problems severely limit the practical application value of deep learning downscaling techniques in operational forecasting. Furthermore, existing methods have technical shortcomings when processing multi-source heterogeneous data, such as insufficient feature extraction and a single attention mechanism, making it difficult to effectively integrate auxiliary information such as topography and pressure fields. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the present invention provides a wind field downscaling method based on physical constraints and artificial intelligence, which can solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, the present invention provides the following technical solution: a wind field downscaling method based on physical constraints and artificial intelligence, comprising: acquiring an input dataset, wherein the input dataset includes terrain data, low-resolution forecast data, and observation data;

[0007] The input dataset is subjected to feature extraction and data processing to obtain a training dataset. The training dataset is then used to train a neural network model that incorporates an attention mechanism. The loss function of the neural network model includes a physical constraint term.

[0008] The real-time input data is fed into the trained neural network model to generate high-resolution wind field forecast data.

[0009] As a preferred embodiment of the wind field downscaling method based on physical constraints and artificial intelligence described in this invention, the method involves: collecting the terrain data and the low-resolution forecast data, wherein the low-resolution forecast data includes wind field forecast data and pressure field forecast data.

[0010] The observation data is collected, and the observation data is high-resolution wind field observation data;

[0011] The terrain data is subjected to linear interpolation to match the resolution of the terrain data with that of the low-resolution forecast data.

[0012] As a preferred embodiment of the wind field downscaling method based on physical constraints and artificial intelligence described in this invention, the following steps are taken: principal component analysis is performed on the pressure field forecast data to extract feature vectors, and the training dataset is constructed based on the feature vectors and the high-resolution wind field observation data.

[0013] A UNET neural network model integrating dual cross-attention modules is constructed. The loss function of the neural network model includes a wind field physical constraint term, wherein the physical constraint term is used to maintain physical consistency among meridional wind, zonal wind and all wind speeds.

[0014] If the training error of the neural network model is greater than a preset threshold, the model hyperparameters are adjusted and training continues; otherwise, the trained neural network model is output.

[0015] As a preferred embodiment of the wind field downscaling method based on physical constraints and artificial intelligence described in this invention, the following is provided: an encoder and decoder structure of the UNET neural network is constructed, and a dual cross-attention module is set between the encoder and decoder, wherein the dual cross-attention module includes a channel attention submodule and a spatial attention submodule;

[0016] The loss function of the neural network model is constructed. The loss function includes a reconstruction loss term and a physical constraint term. The physical constraint term calculates the full wind speed based on the meridional wind component and the zonal wind component. If the deviation between the calculated full wind speed and the predicted full wind speed exceeds a preset threshold, the weight coefficient of the physical constraint term is increased.

[0017] If a scale change in the input feature map is detected, the channel attention submodule in the dual cross attention module is activated to perform feature recalibration, and the spatial attention submodule is activated to perform position-adaptive weighting.

[0018] As a preferred embodiment of the wind field downscaling method based on physical constraints and artificial intelligence described in this invention, wherein: the loss value of the neural network model during the training process is calculated, and the loss value includes reconstruction loss value and physical constraint loss value;

[0019] If the rate of decrease of the reconstruction loss value is less than a first preset threshold, the learning rate and batch size of the neural network model are adjusted; if the physical constraint loss value is greater than a second preset threshold, the weight of the physical constraint term is increased.

[0020] If the loss value is still greater than the third preset threshold after a preset number of iterations, the parameter update magnitude of the neural network model is adjusted by gradient pruning; otherwise, the currently trained neural network model is output.

[0021] As a preferred embodiment of the wind field downscaling method based on physical constraints and artificial intelligence described in this invention, the method includes: acquiring real-time input data, which includes topographic data, low-resolution forecast data, and pressure field forecast data of the area to be predicted;

[0022] Principal component analysis was performed on the pressure field forecast data in the real-time input data to extract feature vectors;

[0023] The feature vector is input into the trained neural network model to generate the high-resolution wind field forecast data, which includes meridional wind components, zonal wind components, and total wind speed.

[0024] As a preferred embodiment of the wind field downscaling method based on physical constraints and artificial intelligence described in this invention, wherein: the feature vector extracted from the pressure field forecast data by the principal component analysis is adapted to the dual cross-attention module, and the dual cross-attention module performs feature recalibration and position adaptive weighting according to the scale change of the input feature map;

[0025] The physical constraints of the neural network model are combined with the loss value monitoring during the training process, and the physical consistency of the model training process is ensured by adjusting the weight coefficients of the physical constraints.

[0026] The trained neural network model, after being standardized and saved, is used for the high-resolution wind field forecasting of the real-time input data.

[0027] To further address the aforementioned technical problems, the present invention provides the following technical solution: a system for a wind field downscaling method based on physical constraints and artificial intelligence, comprising: a data acquisition module for acquiring the input dataset, wherein the input dataset includes the terrain data, the low-resolution forecast data, and the observation data;

[0028] The feature extraction module is used to perform feature extraction and data processing on the input dataset to obtain the training dataset;

[0029] A model building module is used to build the neural network model that incorporates the attention mechanism;

[0030] The parameter optimization module is used to determine the loss function of the neural network model and adjust the weight coefficients of the physical constraint terms.

[0031] A model training module is used to train the neural network model using the training dataset;

[0032] The feature analysis module is used to perform principal component analysis on the pressure field forecast data in the real-time input data and extract the feature vector;

[0033] The forecast generation module is used to input the feature vector into the trained neural network model to generate the high-resolution wind field forecast data, which includes the meridional wind component, the zonal wind component, and the total wind speed.

[0034] A computer device includes a memory and a processor, the memory storing a computer program, characterized in that the processor executes the computer program to implement the steps of the wind field downscaling method based on physical constraints and artificial intelligence as described above.

[0035] A computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the wind field downscaling method based on physical constraints and artificial intelligence as described above.

[0036] The beneficial effects of this invention are as follows: First, by incorporating topographic and pressure field features into the model input and employing principal component analysis to extract features from the pressure field, the model's adaptability to different terrain and weather backgrounds is improved. Second, by integrating a dual cross-attention module into the UNET architecture, the model's ability to capture multi-scale features is enhanced, allowing the downscaling results to better preserve the spatial structure characteristics of the wind field. Finally, by designing a loss function that includes physical constraints, the model can simultaneously generate high-resolution forecasts of meridional wind, zonal wind, and full wind speed in a single calculation, ensuring consistency among these physical quantities while reducing computational resource consumption. These improvements enable this invention to maintain forecast accuracy while enhancing the model's practicality in real-world applications. Attached Figure Description

[0037] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a flowchart of the method in this invention;

[0039] Figure 2 This is a system architecture diagram of the present invention;

[0040] Figure 3 This is the overall flowchart of the present invention;

[0041] Figure 4 This is a diagram of the model structure in this invention;

[0042] Figure 5 This is a diagram of spatial cross-attention and channel cross-attention in this invention;

[0043] Figure 6 This is a diagram of the computer device used in this invention. Detailed Implementation

[0044] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0045] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0046] Example 1, referring to Figures 1-5 As an embodiment of the present invention, a wind field downscaling method based on physical constraints and artificial intelligence is provided.

[0047] Figure 1 A flowchart illustrating an overall method for wind field downscaling based on physical constraints and artificial intelligence is shown, including:

[0048] S1: Obtain the input dataset, which includes terrain data, low-resolution forecast data, and observation data;

[0049] S11: Collect topographic data and low-resolution forecast data, including wind field forecast data and pressure field forecast data;

[0050] S12: Collect observation data, which is high-resolution wind field observation data;

[0051] S13: Perform linear interpolation on the terrain data to match the resolution of the terrain data with that of the low-resolution forecast data.

[0052] The above constitutes a complete solution for multi-source data acquisition and processing.

[0053] Specifically, the dataset integrates three key types of information: topographic data, low-resolution forecast data, and observational data. Among them, the low-resolution forecast data includes two parts: wind field forecast data and pressure field forecast data. These data together constitute the basic data source for model training and forecast generation.

[0054] For pressure field forecast data, principal component analysis (PCA) is used for feature extraction. First, the 3D pressure field is transformed into a 2D array p (m samples, n features), and then centered.

[0055]

[0056] Where p i Let i be the air pressure field of the i-th sample. p represents the average pressure field. c This represents the centered pressure field. Then, the covariance matrix C is calculated based on the centered pressure field.

[0057]

[0058] Perform eigenvalue decomposition on the covariance matrix, select the eigenvectors corresponding to the first k eigenvalues ​​as principal components, construct the projection matrix W, and then... c Projected onto matrix W:

[0059] p r =p c W

[0060] Where p r The extracted feature vectors are then processed. Next, the feature vectors of the pressure field, the low-resolution wind field forecast data from the model, the interpolated topographic data, and the high-resolution wind field observation data are standardized.

[0061]

[0062] Where v is a given variable (eigenvector of the pressure field, low-resolution wind field forecast data from the model, interpolated topographic data, and high-resolution wind field observations), vmean It is the average value of v, v std It is the standard deviation of v.

[0063] To match the topographic data with the low-resolution forecast data at the grid scale, a linear interpolation method was used to process the topographic data, ensuring its resolution remained consistent with the low-resolution forecast data. High-resolution wind field observations were collected and, after quality control, were used for model training.

[0064] Through the data acquisition and processing steps described above, this method successfully constructs a multi-source input dataset with coordinated spatiotemporal resolution and continuous physical features. Compared to the use of a single data source in existing technologies, this method extracts pressure field features through principal component analysis and combines it with standardization processing to effectively extract large-scale weather background information, providing more comprehensive and reliable input information for subsequent model training. Furthermore, the standardization operations during data preprocessing lay the foundation for the fusion and application of data from different sources, effectively improving the model's adaptability to different weather systems and terrain conditions.

[0065] S2: Perform feature extraction and data processing on the input dataset to obtain the training dataset. Use the training dataset to train the neural network model that incorporates the attention mechanism. The loss function of the neural network model includes a physical constraint term.

[0066] S21: Perform principal component analysis on the pressure field forecast data, extract eigenvectors, and construct a training dataset based on the eigenvectors and high-resolution wind field observation data;

[0067] S22: Construct a UNET neural network model that integrates dual cross-attention modules. The loss function of the neural network model includes a wind field physical constraint term, which is used to maintain physical consistency between meridional wind, zonal wind and all wind speeds.

[0068] S221: Construct the encoder and decoder structure of the UNET neural network, and set up a dual cross-attention module between the encoder and decoder. The dual cross-attention module includes a channel attention submodule and a spatial attention submodule.

[0069] S222: The loss function for constructing the neural network model. The loss function includes a reconstruction loss term and a physical constraint term. The physical constraint term calculates the total wind speed based on the meridional wind component and the zonal wind component. If the deviation between the calculated total wind speed and the predicted total wind speed exceeds a preset threshold, the weight coefficient of the physical constraint term is increased.

[0070] S223: If a scale change in the input feature map is detected, the channel attention submodule in the dual cross attention module is activated to perform feature recalibration, and the spatial attention submodule is activated to perform position adaptive weighting.

[0071] After extracting the pressure field feature vector from S21, the key is how to effectively integrate these physically meaningful features into a deep learning model. This invention designs a feature fusion strategy that uses a dual-cross attention module to dynamically couple the pressure field feature vector with the wind field data. Specifically, the extracted pressure field feature vector is denoted as p_r, and the low-resolution wind field data is denoted as w_l. The feature fusion process is represented as follows:

[0072] F in =φ(W p p r +W w w l )

[0073] Among them W p and W w The weight matrix is ​​a learnable matrix. The activation function is nonlinear. This fusion method ensures that the large-scale circulation information carried by the pressure field can guide the extraction of wind field features. The fused feature map F in The data is then fed into a neural network based on the UNET architecture for processing.

[0074] The model is based on the classic UNET architecture and innovatively introduces a dual cross-attention module. The model's input includes pressure field feature vectors after principal component analysis, standardized low-resolution wind field forecast data, and interpolated terrain data. These multi-source input data are first processed by an encoder, which gradually reduces the feature map size through convolution and pooling operations to extract multi-scale features; the decoder then gradually restores the feature map size through upsampling operations to reconstruct the high-resolution output.

[0075] A specially designed dual cross-attention module is placed between the encoder and decoder, such as... Figure 5 As shown. This module is designed to better handle the multi-source heterogeneous data acquired in S1, including the channel attention submodule and the spatial attention submodule. These two submodules work together, respectively responsible for feature recalibration and location adaptive weighting, ensuring that the model can fully utilize the large-scale weather background information carried by the pressure field feature vectors.

[0076] To ensure the physical plausibility of the prediction results, the model employs a combined loss function that includes a reconstruction loss term and a physical constraint term:

[0077]

[0078] Among them, u f v f w f These represent the model-predicted meridional wind, zonal wind, and total wind speed, uo v o w o These are the corresponding observed values. Physical constraint terms. This is used to maintain physical consistency between meridional wind, zonal wind, and total wind speed. When the deviation between the total wind speed calculated based on the meridional and zonal wind components and the predicted total wind speed exceeds a preset threshold, the model will automatically increase the weight coefficient w4 of the physical constraint term to strengthen the physical constraints.

[0079] The core innovation of the dual-cross attention module lies in achieving bidirectional interaction between feature maps in the channel and spatial dimensions. Let the input feature map be... Where C is the number of channels, and H and W are the height and width of the feature map, respectively. The module first calculates the channel attention weights:

[0080] M c =σ(avgpool(F)W1W2)

[0081] Where W1 and W2 are learnable transformation matrices, and σ is the sigmoid activation function. Channel attention weights M c This reflects the importance relationship between different meteorological element channels.

[0082] Simultaneously, the module calculates the spatial attention weights:

[0083] M s =σ(conv([maxpool(F);avgpool(F)]))

[0084] The square brackets indicate feature concatenation operations, and `conv` represents a 1×1 convolutional layer. Spatial attention weights M s It captures the importance of different locations in the feature map, especially the feature responses in areas with complex terrain.

[0085] The final feature enhancement is achieved through double weighting:

[0086]

[0087] in This indicates element-wise multiplication. This dual attention mechanism enables the model to simultaneously focus on the physical relationships and spatial differences in importance between meteorological elements, improving its ability to capture wind field characteristics in complex terrain areas.

[0088] The handling of feature map scale changes employs an adaptive mechanism. When a scale change in the input feature map is detected, the system simultaneously activates two sub-modules within the dual cross-attention module: the channel attention sub-module performs feature recalibration, adjusting the importance weights of features from different channels; the spatial attention sub-module performs position-adaptive weighting, enhancing the feature representation of key regions. This dual attention mechanism enables the model to adaptively adjust its feature extraction strategy based on the specific characteristics of the input data.

[0089] This model design ensures effective capture of multi-scale wind field features while guaranteeing the physical validity of the prediction results through physical constraints. Experiments show that compared to the traditional UNET structure, this improved model exhibits better prediction performance in wind field downscaling tasks, especially demonstrating a significant advantage in wind field prediction under the influence of complex weather systems. By introducing a dual cross-attention module and incorporating physical constraints, the model successfully achieves a balance between feature extraction capability and physical consistency, providing reliable technical support for high-precision wind field downscaling forecasts.

[0090] S23: If the training error of the neural network model is greater than the preset threshold, adjust the model hyperparameters and continue training; otherwise, output the trained neural network model.

[0091] S231: Calculate the loss value of the neural network model during the training process. The loss value includes the reconstruction loss value and the physical constraint loss value.

[0092] S232: If the rate of decrease of the reconstruction loss value is less than the first preset threshold, adjust the learning rate and batch size of the neural network model; if the physical constraint loss value is greater than the second preset threshold, increase the weight of the physical constraint term.

[0093] S233: If the loss value is still greater than the third preset threshold after a preset number of iterations, then the parameter update magnitude of the neural network model is adjusted by gradient pruning; otherwise, the currently trained neural network model is output.

[0094] After model construction is completed, the stability of the training process and the physical reasonableness of the results become key considerations. This invention designs an adaptive training strategy that guides model optimization by monitoring changes in the loss value in real time. Based on the loss function designed above:

[0095]

[0096] Among them, u f v f w f These represent the model-predicted meridional wind, zonal wind, and total wind speed values, respectively. o v o w oFor each observation, w1, w2, w3, and w4 are the weighting coefficients for each loss.

[0097] During training, the model continuously calculates and monitors reconstruction loss and physical constraint loss. When the rate of decline in reconstruction loss slows, the system automatically multiplies the learning rate (lr) by a decay factor β (β is set to 0.1) and doubles the batch size to help the model escape local optima. Simultaneously, if the physical constraint loss exceeds a preset threshold ε (ε is set to 0.1), the weight coefficients are dynamically adjusted to strengthen the physical constraints.

[0098] w4 = w4 × (1 + γ)

[0099] Where γ is the weight growth factor, with a value of 0.2.

[0100] During model training, the weight coefficient w4 of the physical constraint term adopts an adaptive adjustment mechanism. Let the physical constraint loss at the t-th iteration be:

[0101]

[0102] Where u f v f These are the meridional and zonal wind components predicted by the model, respectively; w is the predicted total wind speed; and the superscript t indicates the current iteration number.

[0103] Define the index of the degree of satisfaction of physical constraints:

[0104]

[0105] Where η t This represents the ratio of the physical constraint loss in the current iteration to that in the previous iteration, reflecting the changing trend of the degree to which the physical constraints are satisfied. When η t When the value exceeds the preset threshold δ (δ is set to 1.2), it indicates that the degree of physical constraint satisfaction has decreased. At this time, the weight coefficient update follows the following:

[0106]

[0107] in These are the weight coefficients for the current iteration. The updated weight coefficients are represented by γ, which is the weight growth factor (value 0.2) used to control the magnitude of weight adjustments.

[0108] This dynamic adjustment mechanism can be effective when physical constraints are well satisfied (η) t Approaching 1), the weight coefficient remains stable; when the deviation of physical constraints increases, the weight coefficient increases adaptively to strengthen the constraint on physical laws; the max function ensures that the weight coefficient will not experience negative growth.

[0109] This adjustment ensures that the physical relationships between meridional winds, zonal winds, and all wind speeds remain consistent. During training, to avoid the impact of gradient explosion on model stability, the system initiates a gradient pruning mechanism when the loss value remains high after K consecutive iterations (K set to 5).

[0110]

[0111] Where g represents the original gradient, g clipped Let θ be the gradient after clipping, and let θ be the clipping threshold, set to 1.0.

[0112] Through this multi-layered adaptive adjustment, the model training exhibits excellent convergence characteristics. In the early stages of training, the balanced weight allocation allows the model to fully learn the data features; as training progresses, the dynamic strengthening of physical constraints ensures that the prediction results always conform to the basic physical laws of the wind field. This training strategy, together with the data preprocessing scheme and model structure constructed above, forms an organic whole, fully leveraging the advantages of pressure field feature extraction and the dual cross-attention mechanism.

[0113] This training scheme, based on dynamic monitoring and adaptive adjustment, not only improves the training efficiency of the model, but more importantly, significantly enhances the physical plausibility of the prediction results. When the model's performance metrics on the validation set stabilize and the physical constraints are well satisfied, the entire training process is complete, laying a solid foundation for subsequent practical applications.

[0114] S3: Input the real-time input data into the trained neural network model to generate high-resolution wind field forecast data.

[0115] S31: Acquire real-time input data, which includes topographic data, low-resolution forecast data, and pressure field forecast data of the area to be predicted.

[0116] S32: Perform principal component analysis on the pressure field forecast data in the real-time input data and extract the feature vector;

[0117] S33: Input the feature vector into the trained neural network model to generate high-resolution wind field forecast data, which includes meridional wind components, zonal wind components, and total wind speed.

[0118] After model training is completed, this invention can be used to generate high-resolution wind field forecast data in real time. The process mainly includes three key steps: real-time data acquisition, forecast generation, and result verification.

[0119] First, the system acquires real-time input data, including topographic data, low-resolution forecast data, and pressure field forecast data for the area to be predicted. The acquired real-time data is then standardized.

[0120]

[0121] Where x represents the input variable, μ x and σ x These are the mean and standard deviation saved during the training phase, respectively. This processing ensures that the data distribution in the forecast phase remains consistent with that in the training phase. In the forecast generation phase, the model first calculates the confidence index for the real-time forecast:

[0122]

[0123] Where F p For the current forecast field, F h For historical similarity fields, the closer the CI value is to 1, the higher the forecast reliability. When the CI value is lower than the preset threshold τ (τ is set to 0.8), the system will automatically adjust the forecast weights.

[0124] w adj =w0×(1+α(1-CI))

[0125] Where w0 is the initial weight, α is the adjustment coefficient (set to 0.5), and w adj These are the adjusted weights. This adaptive weight adjustment mechanism can effectively improve the stability of forecasts.

[0126] The system also includes a dedicated physical consistency verification module for the generated forecast results. When a physical inconsistency is detected in the forecast wind field, it is corrected using the least squares method.

[0127]

[0128] Where u i v i w i These represent the meridional wind, zonal wind, and total wind speed forecasts for the i-th grid point, respectively, where n is the total number of grid points. This verification mechanism ensures the physical validity of the final forecast product.

[0129] This forecasting scheme not only accurately depicts the spatial distribution characteristics of wind fields, but also demonstrates excellent forecasting capabilities under complex terrain and special weather system influences. In particular, under the influence of severe convective weather systems, the forecast results still maintain high spatiotemporal continuity and physical rationality, thanks to real-time monitoring of confidence indicators and rigorous verification of physical consistency.

[0130] In this way, the present invention realizes the end-to-end conversion from multi-source input data to high-resolution wind field forecast products, providing a practical intelligent downscaling solution for meteorological forecasting services.

[0131] Furthermore, the feature vectors extracted from the pressure field forecast data by principal component analysis are adapted to the dual cross-attention module. The dual cross-attention module performs feature recalibration and position adaptive weighting based on the scale changes of the input feature map.

[0132] The physical constraints of the neural network model are combined with the monitoring of the loss value during training. The physical consistency of the model during training is ensured by adjusting the weight coefficients of the physical constraints.

[0133] After the trained neural network model is standardized and saved, it is used for high-resolution wind field forecasting based on real-time input data.

[0134] In summary, this method innovatively integrates a dual-cross attention mechanism and a physical constraint loss function, achieving effective preservation of wind field physical consistency and adaptive extraction of multi-scale features. The main technical advantages of this invention are: first, the introduction of physical constraint terms ensures the physical consistency of meridional wind, zonal wind, and overall wind speed forecasts; second, the designed dual-cross attention module significantly improves the model's ability to capture features at different scales; and third, a complete model training and parameter optimization mechanism is established, guaranteeing the stability and reliability of the forecast results.

[0135] Example 2, refer to Figure 2 As an embodiment of the present invention, a wind field downscaling method system based on physical constraints and artificial intelligence is provided, comprising:

[0136] The system comprises the following modules: a data acquisition module for acquiring the input dataset, which includes topographic data, low-resolution forecast data, and observation data; a feature extraction module for extracting features from the input dataset and processing the data to obtain the training dataset; a model building module for constructing a neural network model incorporating an attention mechanism; a parameter optimization module for determining the loss function of the neural network model and adjusting the weight coefficients of the physical constraint terms; a model training module for training the neural network model using the training dataset; a feature analysis module for performing principal component analysis on the pressure field forecast data in the real-time input data to extract feature vectors; and a forecast generation module for inputting the feature vectors into the trained neural network model to generate high-resolution wind field forecast data, which includes meridional wind components, zonal wind components, and total wind speed.

[0137] Example 3, referring to Figure 6This is one embodiment of the present invention, which differs from the previous embodiment in that: if the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0138] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0139] More specific examples (a non-exhaustive list) of computer-readable media include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0140] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0141] Example 4 is an embodiment of the present invention, which provides a wind field downscaling method based on physical constraints and artificial intelligence. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0142] To verify the effectiveness of this invention, a complex terrain region in Southwest China was selected as the experimental area, spanning from January to December 2023. The experiment compared the proposed method with three existing downscaling methods: traditional bilinear interpolation, the Single-Attention UNET model, and a Physics-Constrained Neural Network (PHN). Experimental data included pressure field data (0.25° resolution) from ERA5 reanalysis data and measured wind field data from automatic weather stations (approximately 3 km interval). To ensure the scientific rigor and comparability of the experiment, a five-fold cross-validation method was used, dividing the dataset into training, validation, and test sets in an 8:1:1 ratio. During model training, the Adam optimizer was used with an initial learning rate of 0.001 and a batch size of 64, trained for 500 epochs on an NVIDIA A100 GPU until convergence.

[0143] In wind field forecasting in complex terrain regions, the blocking, lifting, and thermal effects of topography lead to significant local circulation characteristics. This invention effectively captures these local characteristics by combining a dual-cross attention mechanism with dynamic physical constraints. Particularly during periods influenced by severe weather systems, this invention demonstrates a clear advantage. Experimental results show that this invention outperforms comparative methods in terms of root mean square error (RMSE), mean absolute error (MAE), and correlation coefficient (R) for wind speed forecasts. The improvement in forecast performance is particularly significant in mountainous areas with topographic relief greater than 500 meters. Analysis of the spatial distribution characteristics of the forecast results reveals that this invention better characterizes local circulation features such as valley winds and slope winds, thanks to the effective extraction of topographic features by the dual-cross attention module and the dynamic adjustment mechanism of physical constraints.

[0144] Table 1 Comparison of Method Evaluation Indicators

[0145]

[0146]

[0147] In-depth analysis of the experimental data reveals that this invention significantly outperforms existing methods in several key indicators. Regarding forecast accuracy, the wind speed RMSE of this invention is only 1.24 m / s, a reduction of 56.6% compared to the traditional bilinear interpolation method, 34.4% compared to the single-attention UNET model, and 24.8% compared to the physically constrained neural network. Particularly noteworthy is the invention's capture rate of 87.5% for extreme weather events, far exceeding other methods, indicating stronger forecasting capabilities under the influence of strong weather systems. The terrain sensitivity index reaches 0.89, demonstrating the model's effective identification and response to the impact of complex terrain. The physical consistency score is 0.95, close to the ideal value of 1.0, proving the effectiveness of the dynamic physical constraint mechanism. Although the computation time of this invention is slightly longer than traditional methods, the slight increase in computational cost is acceptable considering the significant improvement in forecast accuracy. The forecast lead time improvement rate reaches 35.6%, indicating that this invention not only improves forecast accuracy but also extends the effective forecast lead time, which is of great significance for operational applications.

[0148] Importantly, the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A wind field downscaling method based on physical constraints and artificial intelligence, characterized in that, include: Obtain the input dataset, which includes terrain data, low-resolution forecast data, and observation data; The input dataset is subjected to feature extraction and data processing to obtain a training dataset. The training dataset is then used to train a neural network model that incorporates an attention mechanism. The loss function of the neural network model includes a physical constraint term. Real-time input data is fed into the trained neural network model to generate high-resolution wind field forecast data.

2. The wind field downscaling method based on physical constraints and artificial intelligence as described in claim 1, characterized in that: The terrain data and the low-resolution forecast data are collected, wherein the low-resolution forecast data includes wind field forecast data and pressure field forecast data; The observation data is collected, and the observation data is high-resolution wind field observation data; The terrain data is subjected to linear interpolation to match the resolution of the terrain data with that of the low-resolution forecast data.

3. The wind field downscaling method based on physical constraints and artificial intelligence as described in claim 2, characterized in that: Principal component analysis was performed on the pressure field forecast data to extract feature vectors, and the training dataset was constructed based on the feature vectors and the high-resolution wind field observation data. A UNET neural network model integrating dual cross-attention modules is constructed. The loss function of the neural network model includes a wind field physical constraint term, wherein the physical constraint term is used to maintain physical consistency among meridional wind, zonal wind and all wind speeds. If the training error of the neural network model is greater than a preset threshold, the model hyperparameters are adjusted and training continues; otherwise, the trained neural network model is output.

4. The wind field downscaling method based on physical constraints and artificial intelligence as described in claim 3, characterized in that: Construct the encoder and decoder structure of the UNET neural network, and set a dual cross-attention module between the encoder and decoder. The dual cross-attention module includes a channel attention submodule and a spatial attention submodule. The loss function of the neural network model is constructed. The loss function includes a reconstruction loss term and a physical constraint term. The physical constraint term calculates the full wind speed based on the meridional wind component and the zonal wind component. If the deviation between the calculated full wind speed and the predicted full wind speed exceeds a preset threshold, the weight coefficient of the physical constraint term is increased. If a scale change in the input feature map is detected, the channel attention submodule in the dual cross attention module is activated to perform feature recalibration, and the spatial attention submodule is activated to perform position-adaptive weighting.

5. The wind field downscaling method based on physical constraints and artificial intelligence as described in claim 4, characterized in that: Calculate the loss value of the neural network model during the training process, the loss value including reconstruction loss value and physical constraint loss value; If the rate of decrease of the reconstruction loss value is less than a first preset threshold, the learning rate and batch size of the neural network model are adjusted; if the physical constraint loss value is greater than a second preset threshold, the weight of the physical constraint term is increased. If the loss value is still greater than the third preset threshold after a preset number of iterations, the parameter update magnitude of the neural network model is adjusted by gradient pruning; otherwise, the currently trained neural network model is output.

6. The wind field downscaling method based on physical constraints and artificial intelligence as described in claim 5, characterized in that: Acquire real-time input data, which includes topographic data, low-resolution forecast data, and pressure field forecast data of the area to be predicted. Principal component analysis was performed on the pressure field forecast data in the real-time input data to extract feature vectors; The feature vector is input into the trained neural network model to generate the high-resolution wind field forecast data, which includes meridional wind components, zonal wind components, and total wind speed.

7. The wind field downscaling method based on physical constraints and artificial intelligence as described in claim 6, characterized in that: The feature vector extracted from the pressure field forecast data by the principal component analysis is adapted to the dual cross-attention module. The dual cross-attention module performs feature recalibration and position adaptive weighting based on the scale change of the input feature map. The physical constraints of the neural network model are combined with the loss value monitoring during the training process, and the physical consistency of the model training process is ensured by adjusting the weight coefficients of the physical constraints. The trained neural network model, after being standardized and saved, is used for the high-resolution wind field forecasting of the real-time input data.

8. A system employing the wind field downscaling method based on physical constraints and artificial intelligence as described in any one of claims 1 to 7, characterized in that, include: The data acquisition module is used to acquire the input dataset, which includes the terrain data, the low-resolution forecast data, and the observation data; The feature extraction module is used to perform feature extraction and data processing on the input dataset to obtain the training dataset; A model building module is used to build the neural network model that incorporates the attention mechanism; The parameter optimization module is used to determine the loss function of the neural network model and adjust the weight coefficients of the physical constraint terms. A model training module is used to train the neural network model using the training dataset; The feature analysis module is used to perform principal component analysis on the pressure field forecast data in the real-time input data and extract the feature vector; The forecast generation module is used to input the feature vector into the trained neural network model to generate the high-resolution wind field forecast data, which includes the meridional wind component, the zonal wind component, and the total wind speed.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the wind field downscaling method based on physical constraints and artificial intelligence as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the wind field downscaling method based on physical constraints and artificial intelligence as described in any one of claims 1 to 7.

Images