Method and system for meteorological data spatial downscaling

Abstract

The invention discloses a method and a system for spatial downscaling of meteorological data, and belongs to the technical field of spatial downscaling. The method comprises the following steps: acquiring a historical meteorological data set, and carrying out standardization processing on the historical meteorological data set to obtain a training set; using the training set to train a deep learning spatial downscaling model fused with physical mechanism constraints to obtain a spatial downscaling model; and obtaining target meteorological data, and performing spatial downscaling on the target meteorological data through the spatial downscaling degree model. According to the invention, spatial downscaling of current low-resolution meteorological data can be realized, so that a data basis is provided for realization of a regional refined meteorological model and construction of a large meteorological model.

Description

Technical Field

[0001] The present invention relates to the technical field of spatial downscaling, and more specifically, to a method and system for spatial downscaling of meteorological data. Background Art

[0002] Spatial downscaling of meteorological data is of great significance in improving the spatial resolution of meteorological forecasts, supporting local decision-making, improving environmental impact assessments, studying local climate change, and improving the quality of meteorological services. Large-scale climate models usually have low spatial resolution and are difficult to capture local meteorological phenomena (such as urban climate effects, mountain climate effects, etc.).

[0003] Through spatial downscaling technology, the data output by these large-scale models can be converted into higher-resolution data, thereby improving the prediction accuracy of local meteorological events. At the same time, local governments and management agencies need high-resolution meteorological data for disaster prevention and mitigation planning, infrastructure design, and environmental protection.

[0004] At present, there are mainly traditional methods and deep learning-based methods for meteorological spatial downscaling. Traditional downscaling methods are mainly divided into two categories, namely statistical downscaling and dynamical downscaling. Statistical downscaling establishes a statistical relationship between large-scale variables and local-scale variables, and after verification, it is applied to the data generated by the global climate model to obtain local-scale climate information.

[0005] Dynamical downscaling is based on dynamic equations and uses the initial and boundary conditions given by global climate models to simulate climate processes in specific regions, providing higher resolution and more realistic local-scale climate information, thereby being able to more carefully simulate the impact of various factors such as topography and land use on climate.

[0006] In the spatial downscaling task, traditional interpolation methods have several problems. First, traditional interpolation methods usually lose image details and textures because they mainly rely on the information of neighboring pixels for prediction. Second, traditional interpolation methods often lead to blurring and reduced sharpness of image edges, which cannot meet the requirements of applications with high image clarity. Finally, if there is noise in the original image, traditional interpolation methods may amplify the noise, thereby affecting the quality of the image. Summary of the invention

[0007] In view of the above problems, the present invention proposes a method for spatial downscaling of meteorological data, comprising:

[0008] Acquire a historical meteorological data set, and perform standardization processing on the historical meteorological data set to obtain a training set;

[0009] Using the training set to train a deep learning spatial downscaling model that integrates physical mechanism constraints, to obtain a spatial downscaling model;

[0010] Target meteorological data is acquired, and the target meteorological data is spatially downscaled using the spatial downscaling model.

[0011] Optionally, a historical meteorological data set is obtained, and the historical meteorological data set is standardized to obtain a training set, including:

[0012] According to meteorological elements, historical meteorological data is obtained, the size of high-resolution snapshots in the historical meteorological data is randomly cropped, and patches are set for snapshots with a resolution of 720×1440 in the historical meteorological data to obtain preprocessed historical meteorological data, and the high-resolution data in the preprocessed historical meteorological data is downsampled using average pooling to obtain low-resolution data, and the low-resolution data is standardized to obtain a training set.

[0013] Optionally, meteorological elements include: wind speed at 10 meters above the ground, temperature at 2 meters above the ground, and total water vapor in the atmospheric column.

[0014] Optionally, the training set satisfies a distribution with mean 0 and standard deviation 1.

[0015] Optional network structure of deep learning spatial downscaling model integrating physical mechanism constraints, including:

[0016] Shallow feature extraction module, deep feature extraction module and image reconstruction module;

[0017] The image reconstruction module is used to aggregate the shallow features of the data extracted by the shallow feature extraction module and the deep features of the data extracted by the deep feature extraction module to reconstruct a high-resolution image.

[0018] Optionally, when training the deep learning spatial downscaling model integrating physical mechanism constraints using the training set, the training set is input into the deep learning spatial downscaling model integrating physical mechanism constraints with a preset data dimension;

[0019] The preset data dimension is: C×H×W;

[0020] Among them, H and W are the number of longitude and latitude grid points, and C is the number of meteorological elements.

[0021] In another aspect, the present invention further provides a system for spatial downscaling of meteorological data, comprising:

[0022] A data acquisition unit, used for acquiring a historical meteorological data set and performing standardization processing on the historical meteorological data set to obtain a training set;

[0023] A training unit, used to train the deep learning spatial downscaling model integrating physical mechanism constraints using the training set to obtain a spatial downscaling model;

[0024] The output unit is used to obtain target meteorological data and spatially downscale the target meteorological data using the spatial downscaling model.

[0025] Optionally, the data acquisition unit obtains a historical meteorological data set and performs standardization processing on the historical meteorological data set to obtain a training set, including:

[0026] According to meteorological elements, historical meteorological data is obtained, the size of high-resolution snapshots in the historical meteorological data is randomly cropped, and patches are set for the snapshots of the historical meteorological data to obtain preprocessed historical meteorological data, and the high-resolution data in the preprocessed historical meteorological data is downsampled using average pooling to obtain low-resolution data, and the low-resolution data is standardized to obtain a training set.

[0027] Optionally, meteorological elements include: wind speed at 10 meters above the ground, temperature at 2 meters above the ground, and total water vapor in the atmospheric column.

[0028] Optionally, the training set satisfies a distribution with mean 0 and standard deviation 1.

[0029] Optional network structure of deep learning spatial downscaling model integrating physical mechanism constraints, including:

[0030] Shallow feature extraction module, deep feature extraction module and image reconstruction module;

[0031] The image reconstruction module is used to aggregate the shallow features of the data extracted by the shallow feature extraction module and the deep features of the data extracted by the deep feature extraction module to reconstruct a high-resolution image.

[0032] Optionally, when the training unit uses the training set to train the deep learning spatial downscaling model integrating physical mechanism constraints, the training set is input into the deep learning spatial downscaling model integrating physical mechanism constraints with a preset data dimension;

[0033] The preset data dimension is: C×H×W;

[0034] Among them, H and W are the number of longitude and latitude grid points, and C is the number of meteorological elements.

[0035] In yet another aspect, the present invention further provides a computing device, comprising: one or more processors;

[0036] a processor for executing one or more programs;

[0037] When the one or more programs are executed by the one or more processors, the above-described method is implemented.

[0038] In yet another aspect, the present invention further provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed, the method described above is implemented.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] The present invention provides a method for spatial downscaling of meteorological data, comprising: obtaining a historical meteorological data set, and performing standardization processing on the historical meteorological data set to obtain a training set; using the training set to train a deep learning spatial downscaling model that integrates physical mechanism constraints to obtain a spatial downscaling model; obtaining target meteorological data, and spatially downscaling the target meteorological data through the spatial downscaling model. The present invention can achieve spatial downscaling of current low-resolution meteorological data, thereby providing a data basis for the realization of regional refined meteorological models and the construction of large meteorological models. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 is a flow chart of the method of the present invention;

[0042] Figure 2 It is a structural schematic diagram of the spatial downscaling model of the method of the present invention;

[0043] Figure 3 This is a schematic diagram of the operation principle of the physical mechanism constraint layer of the method of the present invention;

[0044] FIG4( a ) is a visualization result diagram of the spatial downscaling of the wind speed at 10 meters above the ground by 8 times using the method of the present invention;

[0045] FIG4( b ) is a visualization result of the spatial downscaling of the wind speed at 10 meters above the ground by 16 times using the method of the present invention;

[0046] FIG5( a ) is a visualization result of the spatial downscaling of the temperature at 2 meters above the ground surface by 8 times using the method of the present invention;

[0047] FIG5( b ) is a visualization result of the spatial downscaling of the temperature at 2 meters above the surface by 16 times using the method of the present invention;

[0048] FIG6( a ) is a visualization result diagram of the spatial downscaling of the total amount of water vapor in the atmospheric column by 8 times using the method of the present invention;

[0049] FIG6( b ) is a visualization result of the spatial downscaling of the total amount of water vapor in the atmospheric column by 16 times using the method of the present invention;

[0050] Figure 7 It is a structural diagram of the system of the present invention. DETAILED DESCRIPTION

[0051] Now, exemplary embodiments of the present invention are described with reference to the accompanying drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. These embodiments are provided to disclose the present invention in detail and completely and to fully convey the scope of the present invention to those skilled in the art. The terms used in the exemplary embodiments shown in the accompanying drawings are not intended to limit the present invention. In the accompanying drawings, the same units / elements are marked with the same reference numerals.

[0052] Unless otherwise specified, the terms (including technical terms) used herein have the commonly understood meanings to those skilled in the art. In addition, it is understood that the terms defined in commonly used dictionaries should be understood to have the same meanings as those in the context of the relevant fields, and should not be understood as idealized or overly formal meanings.

[0053] Embodiment 1:

[0054] The present invention proposes a method for spatial downscaling of meteorological data, such as Figure 1 As shown, including:

[0055] Step 1: Obtain a historical meteorological data set, and perform standardization on the historical meteorological data set to obtain a training set;

[0056] Step 2: using the training set to train a deep learning spatial downscaling model that integrates physical mechanism constraints to obtain a spatial downscaling model;

[0057] Step 3: Obtain target meteorological data, and spatially downscale the target meteorological data using the spatial downscaling model.

[0058] The process of obtaining a historical meteorological data set and standardizing the historical meteorological data set to obtain a training set includes:

[0059] According to meteorological elements, historical meteorological data is obtained, the size of high-resolution snapshots in the historical meteorological data is randomly cropped, and patches are set for the snapshots of the historical meteorological data to obtain preprocessed historical meteorological data, and the high-resolution data in the preprocessed historical meteorological data is downsampled using average pooling to obtain low-resolution data, and the low-resolution data is standardized to obtain a training set.

[0060] Among them, meteorological elements include: wind speed 10 meters above the ground, temperature 2 meters above the ground, and the total amount of water vapor in the atmospheric column.

[0061] The training set has a distribution with a mean of 0 and a standard deviation of 1.

[0062] Among them, the network structure of the deep learning spatial downscaling model integrating physical mechanism constraints includes:

[0063] Shallow feature extraction module, deep feature extraction module and image reconstruction module;

[0064] The image reconstruction module is used to aggregate the shallow features of the data extracted by the shallow feature extraction module and the deep features of the data extracted by the deep feature extraction module to reconstruct a high-resolution image.

[0065] Wherein, when the deep learning spatial downscaling model integrating physical mechanism constraints is trained using the training set, the training set is input into the deep learning spatial downscaling model integrating physical mechanism constraints with a preset data dimension;

[0066] The preset data dimension is: C×H×W;

[0067] Among them, H and W are the number of longitude and latitude grid points, and C is the number of meteorological elements.

[0068] The present invention is further described below with reference to specific cases:

[0069] S1. Collect the ECMWF fifth generation reanalysis data set (ERA5). The selected meteorological elements include wind speed at 10 meters above the surface, temperature at 2 meters above the surface, and total water vapor in the atmospheric column. The wind speed at 10 meters above the surface is defined as Where u and v are the wind speeds at 10 meters above the ground in the longitudinal and latitudinal directions, respectively. These three meteorological elements represent different key predictor variables. Accurate prediction of wind speed is crucial for wind energy and energy resource planning, and higher resolution is usually required. The temperature at 2 meters above the ground plays an important role in monitoring extreme events such as heat waves and climate change. The total amount of water vapor in the atmospheric column can diagnose precipitation with small-scale characteristics. After acquiring the data, each high-resolution snapshot is randomly cropped to a smaller size for training. For the ERA5 dataset with a spatial resolution of 720×1440, the patch size is defined as 128×128. The number of patches for each snapshot is selected as 8. The high-resolution data is then downsampled using average pooling to obtain low-resolution data. Finally, the data is preprocessed by normalizing the ERA5 dataset, that is, subtracting the mean and dividing by the standard deviation. Normalization transforms the original data into a distribution that satisfies a mean of 0 and a standard deviation of 1. The formula for normalization is:

[0070]

[0071] Where μ and σ are the mean and standard deviation of X. The time range of the sample data set is from 2008 to 2013, the time resolution is 1 hour, the training data set is selected from 2008, 2010, 2011 and 2013, the validation data set is selected from 2012, and the test data set is selected from 2009.

[0072] S2. Establish a deep learning spatial downscaling model that integrates physical mechanism constraints and train the deep learning model.

[0073] The network structure of the deep learning spatial downscaling model integrating physical mechanism constraints is as follows: Figure 1 As shown in Figure 1, it mainly consists of three parts: a shallow feature extraction module, a deep feature extraction module, and an image reconstruction module. The shallow feature extraction module uses a convolutional layer to capture shallow low-level features in the early stage of the network. It directly sends these features to the image reconstruction module to retain low-frequency information. The deep feature extraction module consists of a residual Swin Transformer block (RSTB). RSTB calculates self-attention and realizes information interaction between windows by utilizing the Swin Transformer v2 layer (S2TL). Each RSTB has a residual connection to aggregate features at different levels. Finally, the image reconstruction module aggregates the previously extracted shallow and deep features to reconstruct a high-resolution image.

[0074] Given a low-resolution meteorological grid data X LR ∈R C×H×W , where C, H and W represent the number, height and width of meteorological elements in the meteorological image respectively. A 3×3 convolutional layer Conv(·) is used to extract shallow features F0∈Conv(·) C×H×W , where C represents the number of feature channels, the specific formula is:

[0075] F0=Conv(X LR ), (3)

[0076] The convolutional layer is good at processing shallow visual information and has good local information extraction capabilities. It can effectively extract shallow features of the image, and can perform more stable optimization and generate better results. At the same time, it can also map the input image space to a higher-dimensional feature space, paving the way for subsequent deep feature extraction.

[0077] The deep feature extraction module contains M RSTB modules and a convolution layer with a convolution kernel size of 3×3. The intermediate features F1, F2, ... F M And the final output deep feature F D The extraction is performed block by block, and the specific implementation process is as follows:

[0078]

[0079] in It is the kth RSTB module, in which the convolutional layer is used to introduce the inductive bias into the network of the Transformer architecture to prepare for the subsequent fusion of shallow and deep features.

[0080] The image reconstruction module reconstructs a high-resolution meteorological image by fusing the shallow and deep features extracted previously. The specific process is as follows:

[0081] X SR =f REC (F0+ F D ), (5)

[0082] where f REC Represents the image reconstruction module. The shallow feature F0 is mainly composed of low-frequency information, while the deep feature F D The shallow feature F0 is directly input into the image reconstruction module through long-distance jump connection, so that the deep feature extraction module can focus more on extracting high-frequency information, which makes the network training more stable. At the same time, in order to improve the spatial resolution of the feature map, a sub-pixel convolution layer is used when sampling the feature map.

[0083] When modeling meteorological elements, a principled relationship such as the law of conservation of mass needs to be established between low-resolution and high-resolution samples. For this reason, the present invention introduces a hard constraint that satisfies the law of conservation of mass and embeds it into the neural network as part of the network architecture. The advantage of this hard constraint is that it can also ensure that the generated meteorological image satisfies the constraint during reasoning. The soft constraint of using additional loss terms encourages the network output to be close to the value that satisfies the constraint, which can minimize the loss during training and speed up the convergence.

[0084] Consider downscaling by a factor of N in each linear dimension, let n = N 2 , the value y in the high-resolution block i , i=1,...n corresponds to low-resolution pixel value x, {y i |i=1,...,n} This set is called a superpixel. The mass conservation law is used as a constraint expression:

[0085]

[0086] According to the above formula, the algorithm of the present invention designs an additive constraint layer to be added at the end of the network structure, such as Figure 1 The following operations are used to obtain the intermediate output and reset it. The calculation formula is as follows:

[0087]

[0088] in represents the intermediate output of the network before the constraint layer, y i ,i=1,...n represents the final output after the constraint layer. Figure 3 An example is shown to further illustrate the working principle of the constraint layer. It is worth noting that the designed additive constraint layer can handle any input.

[0089] The present invention uses the MAE loss function to train the network model. The MAE loss function optimizes the overall model by calculating the pixel difference between the generated image and the original image. The loss function calculation formula is as follows:

[0090]

[0091] Where X SR and X HR They represent the SR meteorological image and HR meteorological image generated by the network respectively.

[0092] The method for training a deep learning spatial downscaling model that integrates physical mechanism constraints is:

[0093] (1) The multi-meteorological element forecast data output by the numerical weather forecast model are downsampled through average pooling to obtain low-resolution two-dimensional meteorological data, which are used as the input data of the neural network model. The input data dimension is C×H×W, where H and W are the number of longitude and latitude grid points, respectively, and C is the number of meteorological elements.

[0094] (2) The ECMW fifth generation reanalysis (ERA5) data are used as output.

[0095] (3) Training the neural network model: The C×H×W input data is input into the feature extraction layer once, and after being processed by the feature extraction layer, it is immediately input into the first residual attention module. The obtained feature map is then passed through the RSTB module in sequence. This process is repeated 6 times. The obtained output data enters a full convolution layer, and finally outputs a 720×1440 global data space downscaled prediction value. The hyperparameters of the neural network model are adjusted according to the decrease of the loss function, mainly including the learning rate, the number of convolution kernels, the activation function, etc. When the loss function no longer decreases, the neural network model training is completed.

[0096] The present invention can make full use of data and adopt a data-driven approach. Based on its excellent super-resolution capability, it demonstrates superior performance in spatial downscaling tasks, and realizes spatial downscaling of current low-resolution meteorological data, thereby providing a data basis for the realization of regional refined meteorological models and the construction of large meteorological models. At the same time, it is superior to the current traditional interpolation-based spatial downscaling method in terms of the performance of various indicators, and greatly improves the timeliness and accuracy of downscaling.

[0097] The deep learning spatial downscaling model method of the present invention, which integrates physical mechanism constraints, can make full use of data and achieve superior performance in spatial downscaling tasks through its excellent super-resolution capability. Root mean square error (RMSE), peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) are selected as evaluation indicators.

[0098] Root mean square error;

[0099] Root mean square error (RMSE) is an evaluation index often used to measure the difference between the predicted value and the true value. It can be used to evaluate the difference between SR images and HR images. At the same time, RMSE has the same unit as the original data, which is easy to explain and understand. The lower the RMSE value, the closer the SR image is to the HR image, indicating that the generated image quality is better. The calculation formula is as follows:

[0100]

[0101] Among them, X and Y represent HR image and SR image respectively, X i represents the i-th pixel value of the HR image, Y i represents the i-th pixel value of the SR image, and N represents the total number of pixels.

[0102] Peak signal-to-noise ratio;

[0103] For image super-resolution related tasks, peak signal-to-noise ratio (PSNR) is a commonly used evaluation indicator, which represents the ratio of high-frequency information in an image and is used to measure the similarity between SR images and HR images. The higher the PSNR value, the closer the SR image is to the HR image, indicating that the generated image quality is better. The calculation formula is as follows:

[0104]

[0105] Among them MAX X represents the maximum value of pixel value, and MSE(·) represents the mean square error between SR image and HR image.

[0106] Structural similarity;

[0107] Structural similarity SSIM is an indicator used to evaluate the structural perception of an image. It makes a comprehensive comparison between SR images and HR images in three dimensions: brightness, structure, and contrast. The SSIM value range is between 0 and 1. The closer the SSIM value is to 1, the more similar the SR image and the HR image are, indicating that the generated image quality is better. The calculation formula is as follows:

[0108]

[0109] where μ X and μ Y represents the mean of X and Y, σ X and σ Y represents the standard deviation of X and Y, σ XY Represents the covariance of X and Y. c1 and c2 are constants used to avoid the denominator being zero.

[0110] Quality deviation;

[0111] In order to quantify the quality difference between SR images and HR images, the present invention calculates the quality deviation between the two by the following formula. When the quality deviation value is zero, it means that the SR image generated by the model completely follows the law of mass conservation. The calculation formula of the quality deviation is as follows:

[0112]

[0113] Where x represents the value of a pixel in the HR image, y i ,i=1,...n represents all values ​​in the high-resolution block of the SR image corresponding to the low-resolution pixel.

[0114] Comparative experiment: The algorithm model proposed in this invention is implemented based on PyTorch, and the algorithm is trained and tested on Nvidia GeForceRTX 4090GPU (24GB video memory). The network model weights of the algorithm are randomly initialized and trained from scratch, and no pre-trained weights are used. The hyperparameters of the training phase, the number of RSTBs, the number of S2TLs, the window size, the number of channels, and the number of attention heads are set to 6, 6, 8, 180, and 6 respectively, the batch size is set to 32, the number of iterations is set to 200, and the initial learning rate is set to 10 -4 For all experiments in this paper, the AdamW optimizer was used for training, with momentum set to 0.9 and weight decay set to 10. -5 , and select the MAE loss function to train the network model.

[0115] In order to verify the effectiveness of the algorithm of the present invention, comparative experiments were carried out with other mainstream methods in the field of image super-resolution, including Bicubic, SRCNN, ESPCN, EDSR, WDSR, Swin IR and Swin2SR.

[0116] The overall prediction effect is shown in Table 1. The experimental results on the ERA5 dataset show that the network model proposed in this invention has significant improvements in evaluation indicators such as RMSE, PSNR, SSIM and quality deviation compared with the current mainstream methods in the field of image super-resolution.

[0117] Table 1

[0118]

[0119]

[0120] The analysis of experimental results shows that in the 8-fold magnification experiment, the present invention surpasses other mainstream methods in the three evaluation indicators of RMSE, PSNR and SSIM, achieving excellent performance of 0.081, 33.759 and 0.927, and is better than Swin2SR. This result proves that physical mechanism constraints are an effective way to improve the accuracy of spatial downscaling tasks. In addition, as the scale factor increases, even in the 16-fold magnification experiment, the present invention maintains its advantage, achieving performance of 0.145, 28.35 and 0.843 in RMSE, PSNR and SSIM, respectively. It is worth noting that compared with SRCNN, ESPCN, EDSR and WDSR based on CNN architecture, SwinIR, Swin2SR and the present invention based on Transformer architecture show better performance. This is partly due to the fact that the large-scale parameter quantity of Transformer can better fit complex meteorological grid data to a large extent.

[0121] Figures 4(a)(b), 5(a)(b), and 6(a)(b) show the visual comparison of wind speed at 10 meters above the ground, temperature at 2 meters above the ground, and total water vapor in the atmospheric column. Figures (a) and (b) of each figure show the visualization results with scale factors of 8 and 16, respectively. Since the quantitative comparison results of EDSR and WDSR in Table 1 are similar, the visualization results of WDSR will not be shown here to optimize the layout of the images.

[0122] Embodiment 2:

[0123] The present invention also proposes a system 200 for spatial downscaling of meteorological data, such as Figure 7 As shown, including:

[0124] The data collection unit 201 is used to obtain a historical meteorological data set and perform standardization processing on the historical meteorological data set to obtain a training set;

[0125] A training unit 202 is used to train the deep learning spatial downscaling model integrating physical mechanism constraints using the training set to obtain a spatial downscaling model;

[0126] The output unit 203 is used to obtain target meteorological data and spatially downscale the target meteorological data using the spatial downscaling model.

[0127] The data collection unit 201 obtains a historical meteorological data set and performs standardization processing on the historical meteorological data set to obtain a training set, including:

[0128] According to meteorological elements, historical meteorological data is obtained, the size of high-resolution snapshots in the historical meteorological data is randomly cropped, and patches are set for the snapshots of the historical meteorological data to obtain preprocessed historical meteorological data, and the high-resolution data in the preprocessed historical meteorological data is downsampled using average pooling to obtain low-resolution data, and the low-resolution data is standardized to obtain a training set.

[0129] Among them, meteorological elements include: wind speed 10 meters above the ground, temperature 2 meters above the ground, and the total amount of water vapor in the atmospheric column.

[0130] The training set has a distribution with a mean of 0 and a standard deviation of 1.

[0131] Among them, the network structure of the deep learning spatial downscaling model integrating physical mechanism constraints includes:

[0132] Shallow feature extraction module, deep feature extraction module and image reconstruction module;

[0133] The image reconstruction module is used to aggregate the shallow features of the data extracted by the shallow feature extraction module and the deep features of the data extracted by the deep feature extraction module to reconstruct a high-resolution image.

[0134] The training unit 202 uses the training set to train the deep learning spatial downscaling model integrating physical mechanism constraints, and inputs the training set into the deep learning spatial downscaling model integrating physical mechanism constraints with a preset data dimension;

[0135] The preset data dimension is: C×H×W;

[0136] Among them, H and W are the number of longitude and latitude grid points, and C is the number of meteorological elements.

[0137] The present invention can achieve spatial downscaling of current low-resolution meteorological data, thereby providing a data basis for the realization of regional refined meteorological models and the construction of large meteorological models.

[0138] Embodiment 3:

[0139] Based on the same inventive concept, the present invention also provides a computer device, which includes a processor and a memory, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is used to execute the program instructions stored in the computer storage medium. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is suitable for implementing one or more instructions, and is specifically suitable for loading and executing one or more instructions in the computer storage medium to implement the corresponding method flow or corresponding functions, so as to implement the steps of the method in the above embodiment.

[0140] Embodiment 4:

[0141] Based on the same inventive concept, the present invention also provides a storage medium, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device for storing programs and data. It is understandable that the computer-readable storage medium here can include both a built-in storage medium in a computer device and an extended storage medium supported by the computer device. The computer-readable storage medium provides a storage space, which stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by a processor are also stored in the storage space, and these instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the steps of the method in the above embodiment.

[0142] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes. The schemes in the embodiments of the present invention may be implemented in various computer languages, for example, object-oriented programming language Java and literal scripting language JavaScript, etc.

[0143] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0144] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0145] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0146] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0147] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for spatial downscaling of meteorological data, characterized in that: include: Acquire a historical meteorological data set, and perform standardization processing on the historical meteorological data set to obtain a training set; Using the training set to train a deep learning spatial downscaling model that integrates physical mechanism constraints, to obtain a spatial downscaling model; Target meteorological data is acquired, and the target meteorological data is spatially downscaled using the spatial downscaling model.

2. The method according to claim 1, characterized in that The step of obtaining a historical meteorological data set and performing standardization processing on the historical meteorological data set to obtain a training set includes: According to meteorological elements, historical meteorological data is obtained, the size of high-resolution snapshots in the historical meteorological data is randomly cropped, and patches are set for the snapshots of the historical meteorological data to obtain preprocessed historical meteorological data, and the high-resolution data in the preprocessed historical meteorological data is downsampled using average pooling to obtain low-resolution data, and the low-resolution data is standardized to obtain a training set.

3. The method according to claim 2, characterized in that The meteorological elements include: wind speed 10 meters above the ground, temperature 2 meters above the ground and the total amount of water vapor in the atmospheric column.

4. The method according to claim 2, characterized in that: The training set satisfies a distribution with a mean of 0 and a standard deviation of 1.

5. The method according to claim 1, characterized in that The network structure of the deep learning spatial downscaling model integrating physical mechanism constraints includes: Shallow feature extraction module, deep feature extraction module and image reconstruction module; The image reconstruction module is used to aggregate the shallow features of the data extracted by the shallow feature extraction module and the deep features of the data extracted by the deep feature extraction module to reconstruct a high-resolution image.

6. The method according to claim 1, characterized in that When training the deep learning spatial downscaling model integrating physical mechanism constraints using the training set, the training set is input into the deep learning spatial downscaling model integrating physical mechanism constraints with a preset data dimension; The preset data dimension is: C×H×W; Among them, H and W are the number of longitude and latitude grid points, and C is the number of meteorological elements.

7. A system for spatial downscaling of meteorological data, characterized in that: include: A data acquisition unit, used for acquiring a historical meteorological data set and performing standardization processing on the historical meteorological data set to obtain a training set; A training unit, used to train the deep learning spatial downscaling model integrating physical mechanism constraints using the training set to obtain a spatial downscaling model; The output unit is used to obtain target meteorological data and spatially downscale the target meteorological data using the spatial downscaling model.

8. The system according to claim 7, characterized in that The data acquisition unit acquires a historical meteorological data set and performs standardization processing on the historical meteorological data set to obtain a training set, including: According to meteorological elements, historical meteorological data is obtained, the size of high-resolution snapshots in the historical meteorological data is randomly cropped, and patches are set for the snapshots of the historical meteorological data to obtain preprocessed historical meteorological data, and the high-resolution data in the preprocessed historical meteorological data is downsampled using average pooling to obtain low-resolution data, and the low-resolution data is standardized to obtain a training set.

9. The system according to claim 8, characterized in that The meteorological elements include: wind speed 10 meters above the ground, temperature 2 meters above the ground and the total amount of water vapor in the atmospheric column.

10. The system according to claim 8, characterized in that The training set satisfies a distribution with a mean of 0 and a standard deviation of 1.

11. The system according to claim 7, characterized in that The network structure of the deep learning spatial downscaling model integrating physical mechanism constraints includes: Shallow feature extraction module, deep feature extraction module and image reconstruction module; The image reconstruction module is used to aggregate the shallow features of the data extracted by the shallow feature extraction module and the deep features of the data extracted by the deep feature extraction module to reconstruct a high-resolution image.

12. The system according to claim 7, characterized in that When the training unit uses the training set to train the deep learning spatial downscaling model integrated with physical mechanism constraints, the training set is input into the deep learning spatial downscaling model integrated with physical mechanism constraints with a preset data dimension; The preset data dimension is: C×H×W; Among them, H and W are the number of longitude and latitude grid points, and C is the number of meteorological elements.

13. A computer device, characterized in that: include: one or more processors; a processor for executing one or more programs; When the one or more programs are executed by the one or more processors, the method according to any one of claims 1 to 6 is implemented.

14. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed, the method according to any one of claims 1 to 6 is implemented.

Images