Atmospheric pollutant and meteorological data downscaling method, device, equipment and medium

Abstract

The application discloses an atmospheric pollutant and meteorological data downscaling method and device, equipment and medium, relates to the meteorological analysis technical field, and includes: obtaining low-resolution dynamic field and dynamic downscaling simulation data of a target region including pollutant and meteorological data; training a residual correction model using labeled data to obtain a residual correction amount, and obtaining high-resolution reference true values by compensating the simulation data; performing feature dimensionality increase and spatial rearrangement on the low-resolution dynamic field, and combining static geographical priori to construct a multi-source condition constraint vector; adding Gaussian noise to the reference true values, and combining the constraint vector to generate noisy training data; training a Diffusion model with U-Net as the backbone, and using the trained model to downscale the low-resolution dynamic field to obtain high-resolution pollutant and meteorological data results, thereby improving the downscaling efficiency.

Description

Technical Field

[0001] This invention relates to the field of meteorological analysis technology, and in particular to a method, apparatus, equipment and medium for downscaling atmospheric pollutants and meteorological data. Background Technology

[0002] Currently, refined urban environmental governance, near-surface meteorological disaster early warning, and air pollution source tracing place rigid demands on continuous, high-precision, and physically consistent pollutant and meteorological data at the 100-meter scale. Current mainstream 1km resolution dynamic downscaling methods, such as the WRF (Weather Research and Forecasting Model), while adhering to atmospheric physical laws, are limited by computational resources and cannot characterize 100-meter scale forcings such as urban building clusters, complex terrain, and local emissions. While on-site measurements are accurate at individual points, they are spatially sparse, unable to form a continuous field across the entire region, and exhibit significant systematic deviations between dynamic simulations and measurements, hindering the construction of high-resolution true values.

[0003] In recent years, diffusion models have become the cutting-edge mainstream for super-resolution downscaling of meteorological / pollution fields due to their probabilistic generation and detail restoration capabilities. Domestic and foreign teams have achieved the reconstruction of precipitation, temperature, aerosol and other fields at the scale of 1km to hundreds of meters, breaking through the smoothing and deterministic limitations of traditional interpolation and CNN (Convolutional Neural Network) / GAN (Generative Adversarial Network). However, existing cutting-edge solutions still have core shortcomings: First, high-resolution training ground values ​​often directly use uncorrected numerical simulations or sparse interpolation, without integrating dynamic laws and measured residual corrections, resulting in inherent biases in the 1km→100m mapping of model learning; Second, the fusion of multi-source low-resolution dynamic fields and static terrain / underlying surface is merely a simple stitching, lacking attention-enhanced dimensional alignment and information recalibration, and failing to fully utilize geographical priors to guide 10x super-resolution generation; Third, most diffusion downscaling is purely data-driven, without embedding hard physical constraints such as mass, momentum, energy conservation, terrain, and large-scale fields, often resulting in non-physical wind shear, non-conservation of pollutant mass, abrupt temperature changes, and mismatch with terrain, leading to poor physical consistency; Fourth, loss functions only focus on pixel errors or structural similarities, failing to achieve synergistic optimization of numerical accuracy, spatial detail, and physical rationality, making it difficult to balance 100-meter-level precision with atmospheric scientific rationality.

[0004] As can be seen from the above, how to improve the efficiency of downscaling atmospheric pollutants and meteorological data is an urgent problem to be solved. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a method, apparatus, device, and medium for downscaling atmospheric pollutants and meteorological data, which can improve the efficiency of downscaling atmospheric pollutants and meteorological data. The specific solution is as follows: Firstly, this application provides a method for downscaling atmospheric pollutants and meteorological data, including: Acquire low-resolution dynamic field and dynamic downscaling simulation data of the target area that meet the preset low-resolution conditions; the low-resolution dynamic field includes atmospheric pollutant data and meteorological data. A labeled dataset is constructed based on the preset training labels and the dynamic downscaling simulation data, and the initial residual correction model is trained using the labeled dataset to obtain the target residual correction model. The residual correction amount is determined by using the target residual correction model and the dynamic downscaling simulation value corresponding to each grid in the target area. The residual correction amount is then used to perform residual compensation on the dynamic downscaling simulation data to obtain a reference true value that meets the preset high resolution conditions. The low-resolution dynamic field is subjected to feature dimensionality upscaling and spatial rearrangement to obtain dynamic features, and multi-source conditional constraint vectors are determined based on the dynamic features and static geographic prior data. Gaussian noise is added to the reference ground value to obtain noisy high-resolution field data that satisfy the preset high-resolution conditions, and noisy training data is determined based on the multi-source condition constraint vector and the noisy high-resolution field data. An initial Diffusion-Unet model with U-Net as the backdiffusion backbone network is trained based on the noisy training data. The obtained target Diffusion-Unet model is then used to downscale the low-resolution dynamic field, resulting in downscaling results for atmospheric pollutants and meteorological data.

[0006] Optionally, acquiring low-resolution dynamic field and dynamic downscaling simulation data of the target region that meet preset low-resolution conditions includes: The geographical range and time span of the target area are determined, and the spatiotemporal boundary parameters of the target area are generated based on the geographical range and the time span, and the resolution threshold parameters are generated based on the preset low resolution conditions. Based on the spatiotemporal boundary parameters and the resolution threshold parameters, the original observation data of atmospheric pollutants and the original meteorological observation data in the target area are obtained, so as to integrate the original observation data and the original meteorological observation data in the spatiotemporal dimension to obtain a low-resolution dynamic field that meets the preset low-resolution conditions. Based on the low-resolution dynamic field and the spatiotemporal boundary parameters, a regional dynamic downscaling numerical simulation is performed to obtain dynamic downscaling simulation data that matches the spatiotemporal dimension of the low-resolution dynamic field.

[0007] Optionally, the step of constructing a labeled dataset based on preset training labels and the dynamic downscaling simulation data, and using the labeled dataset to train an initial residual correction model to obtain a target residual correction model, includes: Training labels are determined based on the deviation between the measured values ​​of the site and the preset coarse resolution simulation values. The training labels are then matched with the dynamic downscaling simulation data grid by grid and time step by time step based on the spatiotemporal boundary parameters of the target region to obtain a labeled dataset including dynamic downscaling simulation data samples and corresponding preset training label samples. The labeled dataset is divided into a training subset and a validation subset according to a preset ratio; An initial residual correction model is constructed based on the dynamic downscaling simulation data, and the initial residual correction model is iteratively trained using the training subset to obtain a residual correction model to be verified; the residual correction model to be verified is a regression tree model. The verification subset is used to perform grid-by-grid accuracy verification on the residual correction model to be verified, and a target residual correction model is generated when the verification accuracy of the residual correction model to be verified is greater than a preset accuracy threshold.

[0008] Optionally, the step of using the target residual correction model and determining the residual correction amount based on the dynamic downscaling simulation values ​​corresponding to each grid in the target region, and then using the residual correction amount to perform residual compensation on the dynamic downscaling simulation data to obtain a reference true value that meets the preset high-resolution conditions, includes: The grid division rules corresponding to the target area are determined, and the target area is divided into grids using the grid division rules to obtain each grid. The dynamic downscaling simulation values ​​corresponding to each grid are extracted based on the dynamic downscaling simulation data. A grid simulation value sequence is generated based on the dynamic downscaling simulation values, and a residual correction value sequence corresponding to each grid is generated based on the target residual correction model and the grid simulation value sequence. Based on the grid partitioning rules, the residual correction sequence is mapped to the dynamic downscaling simulation data grid by grid, and the dynamic downscaling simulation value of the corresponding grid is algebraically summed with the residual correction to obtain the residual compensation calculation result. Based on the residual compensation calculation result, a reference true value that meets the preset high resolution conditions is generated.

[0009] Optionally, the step of performing feature upsizing and spatial rearrangement on the low-resolution dynamic field to obtain dynamic features, and determining the multi-source conditional constraint vector based on the dynamic features and static geographic prior data, includes: Extract the spatial parameters of the reference truth that satisfy the preset high-resolution conditions, and generate high-resolution reference parameters based on each of the spatial parameters; The low-resolution dynamic field is up-dimensionalized according to the channel dimension to obtain the up-dimensional dynamic features after channel expansion. Based on the high-resolution benchmark parameters, the upgraded dynamic features are spatially rearranged to obtain dynamic features to be processed that match the spatial resolution of the reference true value. The static geographic prior data corresponding to the target area is determined, and the static geographic prior data is subjected to dimension matching and normalization based on the dimension of the dynamic feature to be processed to obtain normalized static features; the static geographic prior data includes resolution terrain elevation and underlying surface type. The dynamic features to be processed and the normalized static features are concatenated along the channel dimension to obtain the concatenated result. The channel weights corresponding to each channel dimension are determined using the channel attention mechanism, and the source condition constraint vector is determined based on the channel weights and the concatenated result.

[0010] Optionally, the step of adding Gaussian noise to the reference ground value to obtain noisy high-resolution field data that satisfy the preset high-resolution conditions, and determining noisy training data based on the multi-source conditional constraint vector and the noisy high-resolution field data, includes: Determine the diffusion time step sequence and the noise intensity parameter corresponding to each time step, and add Gaussian noise to the reference true value step by step based on the diffusion time step sequence and the noise intensity parameter to obtain the initial noisy high-resolution field data corresponding to each time step; The multi-source conditional constraint vector is spatiotemporally aligned with the initial noisy high-resolution field data corresponding to each time step to obtain the spatiotemporally aligned multi-source conditional constraint vector and the target noisy high-resolution field data. Based on the channel dimension, the aligned multi-source conditional constraint vector and the target noisy high-resolution field data are fused to obtain noisy training data; the noisy training data includes paired data; each of the paired data includes the target noisy high-resolution field data corresponding to the time step and the aligned multi-source conditional constraint vector.

[0011] Optionally, the initial Diffusion-Unet model trained based on the noisy training data, with U-Net as the backdiffusion backbone network, is used to scale down the low-resolution dynamic field using the obtained target Diffusion-Unet model, resulting in downscaling results for atmospheric pollutants and meteorological data, including: An initial Diffusion-Unet model with U-Net as the backdiffusion backbone network is constructed, and input data is generated based on the noisy training data corresponding to each time step and the corresponding aligned multi-source conditional constraint vector. Using the initial Diffusion-Unet model and based on the input data, a noise value minimization prediction is determined to obtain the predicted noise value, and the error between the predicted noise value and the true Gaussian noise is determined. The initial Diffusion-Unet model is trained using a multi-scale hybrid loss function and a physical constraint loss function, and the back-diffusion iterative training is performed based on the error. The multi-scale hybrid loss function includes a pixel-level L1 loss function and a structural similarity loss function; the physical constraint loss function includes a conservation constraint loss function and a boundary condition constraint loss function; the conservation constraint loss function includes a mass conservation constraint function, a momentum conservation constraint function, and an energy conservation constraint function; and the boundary condition constraint loss function includes a terrain constraint function and a large-scale field constraint function. When the loss function value corresponding to the current iterative model converges to a preset convergence threshold, the current iterative model is set as the target Diffusion-Unet model; The low-resolution dynamic field is processed to obtain the corresponding low-resolution multi-source conditional constraint vector. Then, the target Diffusion-Unet model is used to perform end-to-end reverse diffusion scaling on the low-resolution multi-source conditional constraint vector to obtain atmospheric pollutant downscaling results and meteorological data downscaling results that meet the preset high-resolution conditions.

[0012] Secondly, this application provides an air pollutant and meteorological data downscaling device, comprising: The low-resolution dynamic field determination module is used to acquire low-resolution dynamic field and dynamic downscaling simulation data of the target area that meet preset low-resolution conditions; the low-resolution dynamic field includes atmospheric pollutant data and meteorological data. The labeled dataset construction module is used to construct a labeled dataset based on preset training labels and the dynamic downscaling simulation data, so as to train an initial residual correction model using the labeled dataset and obtain a target residual correction model. The reference truth generation module is used to determine the residual correction amount based on the target residual correction model and the dynamic downscaling simulation value corresponding to each grid in the target area, so as to use the residual correction amount to perform residual compensation on the dynamic downscaling simulation data and obtain the reference truth value that meets the preset high resolution conditions. The multi-source condition constraint vector determination module is used to perform feature dimensionality upscaling and spatial rearrangement on the low-resolution dynamic field to obtain dynamic features, and to determine the multi-source condition constraint vector based on the dynamic features and static geographic prior data. The noisy training data determination module is used to add Gaussian noise to the reference ground value to obtain noisy high-resolution field data that meet the preset high-resolution conditions, and to determine noisy training data based on the multi-source condition constraint vector and the noisy high-resolution field data. The scaling down module is used to train an initial Diffusion-Unet model with U-Net as the backdiffusion backbone network based on the noisy training data, and to use the obtained target Diffusion-Unet model to scale down the low-resolution dynamic field, thereby obtaining the downscaling results of atmospheric pollutants and meteorological data.

[0013] Thirdly, this application provides an electronic device, comprising: Memory, used to store computer programs; A processor is used to execute the computer program to implement the aforementioned method for downscaling atmospheric pollutants and meteorological data.

[0014] Fourthly, this application provides a computer-readable medium for storing a computer program, wherein the computer program, when executed by a processor, implements the aforementioned method for downscaling atmospheric pollutants and meteorological data.

[0015] As can be seen from the above, before downscaling atmospheric pollutant and meteorological data, this application needs to acquire low-resolution dynamic field and dynamic downscaling simulation data of the target area that meet the preset low-resolution conditions; construct a labeled dataset based on preset training labels and dynamic downscaling simulation data, and use the labeled dataset to train an initial residual correction model to obtain a target residual correction model; use the target residual correction model and the dynamic downscaling simulation values ​​corresponding to each grid in the target area to determine the residual correction amount, and use the residual correction amount to perform residual compensation on the dynamic downscaling simulation data to obtain a reference true value that meets the preset high-resolution conditions; and perform feature upscaling on the low-resolution dynamic field. The space is rearranged to obtain dynamic features, and multi-source conditional constraint vectors are determined based on the dynamic features and static geographic prior data. Gaussian noise is added to the reference ground truth to obtain noisy high-resolution field data that meet the preset high-resolution conditions. Noisy training data is determined based on the multi-source conditional constraint vectors and the noisy high-resolution field data. An initial Diffusion-Unet model with U-Net as the backdiffusion backbone network is trained based on the noisy training data. The obtained target Diffusion-Unet model is used to scale down the low-resolution dynamic field to obtain the downscaling results of atmospheric pollutants and meteorological data.

[0016] Therefore, this application first needs to obtain low-resolution dynamic field and dynamic downscaling simulation data of the target region that meet the preset low-resolution conditions; secondly, it constructs a labeled dataset based on preset training labels and dynamic downscaling simulation data, and uses the labeled dataset to train an initial residual correction model to obtain a target residual correction model; then, it uses the target residual correction model and the dynamic downscaling simulation values ​​corresponding to each grid in the target region to determine the residual correction amount, and uses the residual correction amount to perform residual compensation on the dynamic downscaling simulation data to obtain a reference true value that meets the preset high-resolution conditions; furthermore, it performs feature upscaling and spatial rearrangement on the low-resolution dynamic field. Dynamic features are obtained, and multi-source conditional constraint vectors are determined based on these features and static geographic prior data. Subsequently, Gaussian noise is added to the reference ground truth to obtain noisy high-resolution field data that satisfies the preset high-resolution conditions. Noisy training data is then determined based on the multi-source conditional constraint vectors and the noisy high-resolution field data. Finally, an initial Diffusion-Unet model with U-Net as the backdiffusion backbone is trained based on the noisy training data. This target Diffusion-Unet model is then used to downscale the low-resolution dynamic field, yielding downscaled results for atmospheric pollutants and meteorological data. This improves the efficiency of downscaling atmospheric pollutants and meteorological data, thereby enhancing the user experience. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 This is a flowchart of a method for downscaling atmospheric pollutant and meteorological data disclosed in this application; Figure 2 This is a schematic diagram of a specific atmospheric pollutant and meteorological data downscaling learning framework disclosed in this application; Figure 3 This is a schematic diagram of the structure of an atmospheric pollutant and meteorological data downscaling device disclosed in this application; Figure 4 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Currently, refined urban environmental governance, near-surface meteorological disaster early warning, and air pollution source tracing place rigid demands on continuous, high-precision, and physically consistent pollutant and meteorological data at the 100-meter level. While current mainstream 1km resolution dynamic downscaling follows atmospheric physical laws, it is limited by computational resources and cannot characterize 100-meter-scale constraints such as urban building clusters, complex terrain, and localized emissions. Although on-site measurements are accurate at individual points, they are spatially sparse, unable to form a continuous field across the entire region, and exhibit significant systematic deviations between dynamic simulations and measurements, hindering the construction of high-resolution ground truth. Therefore, this application provides a method for downscaling atmospheric pollutant and meteorological data, which can improve the efficiency of downscaling atmospheric pollutant and meteorological data during the downscaling process.

[0021] See Figure 1 As shown in the figure, an embodiment of the present invention discloses a method for downscaling atmospheric pollutants and meteorological data, including: Step S11: Obtain low-resolution dynamic field and dynamic downscaling simulation data of the target area that meet the preset low-resolution conditions; the low-resolution dynamic field includes atmospheric pollutant data and meteorological data.

[0022] In this embodiment, the technical solution corresponding to the present application is to achieve high-precision downscaling by constructing a deep learning framework for downscaling atmospheric pollutant and meteorological data, and the corresponding schematic diagram of the atmospheric pollutant and meteorological data downscaling learning framework is shown below. Figure 2 As shown.

[0023] First, this application embodiment requires the construction of high-resolution training samples (reference ground truth) at the 100-meter level: for dynamic downscaling simulation data To address the systematic deviations from the measured data at the stations, this application employs a random forest regression algorithm to construct a residual correction model. This model achieves accurate fitting and continuous spatiotemporal correction of the deviations across a 100-meter grid across the entire region, ultimately constructing a reference ground truth that closely reflects the characteristics of the actual atmospheric distribution. .

[0024] The specific process is as follows: First, a dataset is constructed: synchronous measured data from high-density automatic weather stations and environmental monitoring micro-stations within the target area are collected. The dynamic downscaling simulation values ​​of the corresponding 100-meter grid, the 100-meter resolution topographic elevation, and the underlying surface type of each measured station are extracted as input features of the model.

[0025] Specifically, acquiring low-resolution dynamic field and dynamic downscaling simulation data of the target area that meet preset low-resolution conditions may include: determining the geographical range and time span of the target area, generating spatiotemporal boundary parameters of the target area based on the geographical range and time span, and generating resolution threshold parameters based on preset low-resolution conditions; acquiring raw observation data of atmospheric pollutants and raw meteorological observation data in the target area based on the spatiotemporal boundary parameters and resolution threshold parameters, integrating the raw observation data and raw meteorological observation data in spatiotemporal dimensions to obtain a low-resolution dynamic field that meets preset low-resolution conditions; and performing regional dynamic downscaling numerical simulation based on the low-resolution dynamic field and spatiotemporal boundary parameters to obtain dynamic downscaling simulation data that matches the spatiotemporal dimensions of the low-resolution dynamic field.

[0026] Step S12: Construct a labeled dataset based on the preset training labels and the dynamic downscaling simulation data, and use the labeled dataset to train the initial residual correction model to obtain the target residual correction model.

[0027] In this embodiment, the present application requires the residual between the measured values ​​and the simulated values ​​of each station. Labels are trained for the model to form a one-to-one labeled dataset.

[0028] Subsequently, the random forest residual correction model was trained: the input features were standardized and preprocessed to eliminate the influence of dimensions, that is, multiple subsets were extracted from the labeled dataset using the bootstrap sampling method, and a classification and regression tree (CART) was constructed for each subset; when splitting each tree node, some features were randomly selected and the splitting threshold was determined based on the minimum squared error, and the maximum depth of the tree was set to avoid overfitting; the prediction results of all classification and regression trees were integrated, and the average value was used as the model output to complete the training of the random forest residual correction model. The model has the ability to predict residuals of the entire domain at the hundred-meter level.

[0029] Specifically, a labeled dataset is constructed based on preset training labels and dynamic downscaling simulation data. This labeled dataset is then used to train an initial residual correction model to obtain a target residual correction model. This process includes: determining training labels based on the deviation between measured values ​​at stations and preset coarse-resolution simulated values; matching the training labels and dynamic downscaling simulation data grid-by-grid and time-step-by-time based on the spatiotemporal boundary parameters of the target region to obtain a labeled dataset containing dynamic downscaling simulation data samples and corresponding preset training label samples; dividing the labeled dataset into a training subset and a validation subset according to a preset ratio; constructing an initial residual correction model based on the dynamic downscaling simulation data; iteratively training the initial residual correction model using the training subset to obtain a residual correction model to be validated; the residual correction model to be validated is a regression tree model; performing grid-by-grid accuracy validation on the residual correction model to be validated using the validation subset; and generating the target residual correction model when the validation accuracy of the residual correction model to be validated exceeds a preset accuracy threshold.

[0030] Step S13: Utilize the target residual correction model and determine the residual correction amount based on the dynamic downscaling simulation values ​​corresponding to each grid in the target area, so as to use the residual correction amount to perform residual compensation on the dynamic downscaling simulation data and obtain the reference true value that meets the preset high resolution conditions.

[0031] In this embodiment, the present application requires the generation of global 100-meter-level reference true values: generating all 100-meter-level grids within the target area. The values, 100m terrain elevation, and underlying surface type are input into the trained residual correction model, and the residual correction amount is obtained by predicting the entire spatiotemporal continuous value grid by grid. Subsequently, the dynamic downscaling simulation data was corrected globally using the residual compensation method to obtain a high-resolution reference true value at the 100-meter level, and the corresponding calculation formula is as follows: ; in, The dynamic downscaling simulation of meteorological or pollutant variables at location 's' in the target area, before correction, represents the location within the 100-meter grid. The target area is defined by a grid of locations within a hundred-meter radius. This is the reference truth value for that position.

[0032] It is worth mentioning that the above-mentioned construction It integrates atmospheric physical laws from dynamic simulation with real data characteristics from on-site measurements, exhibiting good spatiotemporal continuity and high numerical accuracy. As a target distribution in the training phase of diffusion models, it can effectively guide models to learn the 1km→100m super-resolution mapping law that conforms to real atmospheric conditions.

[0033] Specifically, the residual correction amount is determined using the target residual correction model and based on the dynamic downscaling simulation values ​​corresponding to each grid within the target region. This residual correction amount is then used to perform residual compensation on the dynamic downscaling simulation data to obtain a reference true value that meets preset high-resolution conditions. This process may include: determining the grid division rules corresponding to the target region; dividing the target region into grids using these rules; extracting the dynamic downscaling simulation values ​​corresponding to each grid based on the dynamic downscaling simulation data; generating a grid simulation value sequence based on each dynamic downscaling simulation value; generating a residual correction amount sequence for each grid using the target residual correction model and the grid simulation value sequence; matching the residual correction amount sequence with the dynamic downscaling simulation data grid by grid based on the grid division rules; and algebraically summing the dynamic downscaling simulation values ​​and residual correction amounts for the corresponding grids to obtain the residual compensation calculation results. Finally, a reference true value that meets preset high-resolution conditions is generated based on the residual compensation calculation results.

[0034] Step S14: Perform feature dimensionality upscaling and spatial rearrangement on the low-resolution dynamic field to obtain dynamic features, and determine multi-source condition constraint vectors based on the dynamic features and static geographic prior data.

[0035] In this embodiment, the present application requires multi-source heterogeneous data feature encoding: constructing conditional constraint vectors. To reconstruct specific meteorological features from noise using a guided diffusion model, the first step is to input a 1km low-resolution dynamic field containing pollution and meteorological elements. Meanwhile, the embodiments of this application require the introduction of a 100m resolution static geographic prior consisting of terrain and land use types. Feature encoding is completed by relying on a multimodal feature fusion function: ; ; in, This represents low-resolution meteorological or pollutant dynamic field variables with a resolution of 1km. ,Right now Representative The dynamic characteristic variables of low-resolution meteorological or pollutant data after dimensionality upgrade. For feature dimensionality increase and spatial recombination operators, For feature fusion operators.

[0036] It is worth mentioning that, The operator is mainly used to achieve feature dimensionality enhancement and spatial rearrangement to ensure that the conditional constraint vector C accurately matches the generated 100-meter-scale meteorological field in the feature map dimension. Its operation includes two core steps: information enrichment and spatial rearrangement. It first hides the resolution information in the channel dimension using convolutional layers, enabling a single pixel at a 1km scale to generate 100 sets of differentiated feature values, and then... Spatial rearrangement is completed, and the feature information stored in the channel dimension is unfolded in a 10×10 matrix format with a magnification of 10, allowing the original... 1km scale feature map converted to The 100m scale feature map is used to improve spatial resolution and compress channel dimensions; finally, the... Low-resolution features after upsampling transformation Compared with static geographical prior features During the fusion process With dimensional alignment and information complementarity as the core strategies, it effectively adapts to the characteristic dimensional requirements of 100-meter-level meteorological fields.

[0037] Furthermore, this application's embodiments introduce channel attention (SEBlock, i.e., compression and excitation modules) or spatial attention on the basis of stitching and fusion, allowing the model to automatically learn which channel / spatial location information is more important, thereby improving the representational ability of the fused features. This is suitable for high-precision generation of fields at the hundred-meter level (such as refined precipitation and pressure fields), and the corresponding channel attention formula and stitching and fusion formula are as follows: ; ; in, The fused feature variable represents the concatenation of low-resolution dynamic features and static geographic prior features. This represents the operation of feature stitching along the channel dimension; The dimensional space representing the concatenated feature tensor For batch size, This represents the number of channels after concatenation. and These are the space's height and width, respectively. This represents the fused feature variables after recalibration by the channel attention module; This represents the channel attention enhancement module.

[0038] The SEBlock sub-formula for the channel attention core is shown below: ; in, This represents the final generated multi-source condition constraint vector; This represents a 1×1 convolution operation used for channel compression or channel mapping.

[0039] Furthermore, the corresponding global average pooling formula is as follows: ; in, Representative of fusion features Channel statistical characteristic variables obtained by global average pooling; This represents a global average pooling operation, used to pool the spatial dimensions of each channel; represents the dimensional space of the feature vectors after global average pooling, where For batch size, This represents the number of channels after splicing.

[0040] The ReLU activation function formula is shown below: ; in, This represents the channel attention recalibration weight vector obtained after transformation by the fully connected layer. Representative with This is the mapping result of the fully connected layer as input.

[0041] The channel-by-channel multiplication recalibration formula is as follows: ; in, This represents a channel-wise multiplication operation, used to apply attention weights to the fused features. superior.

[0042] Specifically, the process involves feature upscaling and spatial rearrangement of a low-resolution dynamic field to obtain dynamic features. Based on these dynamic features and static geographic prior data, a multi-source condition constraint vector is determined. This can include: extracting spatial parameters from the reference ground truth that satisfy preset high-resolution conditions to generate high-resolution baseline parameters; upscaling the low-resolution dynamic field according to channel dimensions to obtain upscaled dynamic features with expanded channel numbers; spatially rearranging the upscaled dynamic features based on the high-resolution baseline parameters to obtain dynamic features to be processed that match the spatial resolution of the reference ground truth; determining the static geographic prior data corresponding to the target area, and performing dimension matching and normalization on the static geographic prior data based on the dimensions of the dynamic features to be processed to obtain normalized static features; the static geographic prior data includes resolution terrain elevation and underlying surface type; concatenating the dynamic features to be processed with the normalized static features along the channel dimensions to obtain the concatenated result, and using a channel attention mechanism to determine the channel weights corresponding to each channel dimension, thereby determining the source condition constraint vector based on the channel weights and the concatenated result.

[0043] Step S15: Add Gaussian noise to the reference ground truth to obtain noisy high-resolution field data that meet the preset high-resolution conditions, and determine noisy training data based on the multi-source condition constraint vector and the noisy high-resolution field data.

[0044] In this embodiment, the present application requires the construction of a Diffusion-Unet super-resolution downscaling model (i.e., Diffusion U-Net) that integrates physical prior constraints. That is, the diffusion generation model is used as the core generation framework, and U-Net is used as the backbone network for the reverse diffusion process. A multi-source conditional constraint vector C is introduced to guide the generation process. At the same time, atmospheric dynamic physical constraints and multi-scale hybrid loss functions are embedded to achieve high-precision reconstruction from a 1km low-resolution field to a 100m high-resolution pollution / meteorological field, so as to solve the problems of poor physical consistency and insufficient local fine feature restoration of traditional downscaling methods.

[0045] Subsequently, the forward diffusion process (noise addition and essential feature learning) follows as follows: First, the core of the forward diffusion process is the 100-meter-level reference true value used for training. Gaussian noise is gradually added to degrade the data into pure random noise. At the same time, the model learns the essential characteristics of high-resolution pollution / meteorological fields that conform to atmospheric physical rules, providing a prior distribution for the reverse reconstruction process. This corresponds to the step in the flowchart of "adding Gaussian noise and learning the essential characteristics of high-resolution data that conform to physical rules".

[0046] It is worth mentioning that the forward diffusion process follows the following recursive formula: ; ; Where t is the diffusion time step ( In one specific implementation, T can be set to 1000 in this application embodiment. The signal retention coefficient at the t-th diffusion time step is represented by the noise scheduling coefficient. Calculated variables are used to control the high-resolution meteorological or pollutant field variables of the previous moment. In the current noisy data The retention ratio in. The preset noise scheduling coefficient varies with time step from linear growth to ; It is independent and identically distributed standard Gaussian noise; The noisy high-resolution field data after adding noise at step t.

[0047] In this embodiment, when time step When large enough, This will approximate a standard Gaussian distribution, completing the process of adding noise to the original high-resolution data. Simultaneously, during the forward diffusion process, the constructed conditional constraint vectors will... Noisy data at each time step By performing channel-dimensional stitching, the model can embed low-resolution dynamic field features and static geographical prior information during the noise learning stage, providing global condition guidance for accurate reconstruction through backdiffusion.

[0048] Specifically, Gaussian noise is added to the reference ground value to obtain noisy high-resolution field data that meets preset high-resolution conditions. Noisy training data is then determined based on the multi-source conditional constraint vector and the noisy high-resolution field data. This can include: determining the diffusion time step sequence and the noise intensity parameters corresponding to each time step; adding Gaussian noise to the reference ground value step-by-step based on the diffusion time step sequence and noise intensity parameters to obtain the initial noisy high-resolution field data corresponding to each time step; aligning the multi-source conditional constraint vector and the initial noisy high-resolution field data corresponding to each time step in a spatiotemporal dimension to obtain the spatiotemporally aligned multi-source conditional constraint vector and the target noisy high-resolution field data; and fusing the aligned multi-source conditional constraint vector and the target noisy high-resolution field data based on the channel dimension to obtain the noisy training data. The noisy training data includes paired data; each paired data includes the target noisy high-resolution field data corresponding to the time step and the aligned multi-source conditional constraint vector.

[0049] Step S16: Train an initial Diffusion-Unet model with U-Net as the backdiffusion backbone network based on the noisy training data, and use the obtained target Diffusion-Unet model to scale down the low-resolution dynamic field to obtain the downscaling results of atmospheric pollutants and meteorological data.

[0050] In this embodiment, the present application requires a reverse diffusion reconstruction process based on the U-Net backbone. The reverse diffusion process is the inverse of the forward diffusion process, using the pure Gaussian noise ultimately obtained from the forward diffusion. As input, with conditional constraint vector For global guidance, a pre-trained U-Net network is used to progressively predict and remove Gaussian noise at each step, ultimately reconstructing a high-resolution pollution / meteorological field at the 100-meter level that meets the constraints. This corresponds to the "high-precision reconstruction" step in the flowchart.

[0051] Furthermore, this embodiment employs an improved U-Net network as the noise prediction backbone. The network structure includes a symmetrical downsampling encoder and an upsampling decoder, with skip connections set at corresponding layers in the encoder and decoder to fully preserve the local spatial details and edge features of the high-resolution field. Additionally, group normalization and SiLU activation functions are incorporated into each convolutional module of the network to improve the model's training stability and nonlinear feature fitting ability. The backdiffusion noise prediction and iterative reconstruction formulas are as follows: ; ; in, Gaussian noise predicted for the U-Net network. These are the trainable parameters of the network; This is the cumulative noise figure; C represents the noise variance of the backdiffusion process, and C represents the multi-source conditional constraint vector, used to provide low-resolution meteorological or pollutant dynamic fields, static geographical priors, and other conditional information to the backdiffusion reconstruction process. U-Net represents the backbone neural network structure used to predict noise and perform backdiffusion denoising reconstruction. In one specific embodiment, the embodiments of this application take... ; For random Gaussian noise, when hour The iteration process is terminated.

[0052] In this way, through T-step reverse iterations, the final output is the same as the reference truth value. High-resolution pollution / weather field forecast data with completely consistent dimensions at the 100-meter level The core process of super-resolution downscaling from 1km to 100m was completed.

[0053] Then, a multi-scale hybrid loss function is constructed: In order to ensure that the high-resolution field reconstructed by the model achieves the best results in terms of numerical accuracy and spatial structural similarity, the embodiments of this application construct a multi-scale hybrid loss function, which combines pixel-level L1 loss with structural similarity loss SSIM to achieve synergistic optimization of global numerical bias and local structural features.

[0054] Among them, the total loss function Defined as: ; in, and The loss weighting coefficients are set to 0.8 and 0.2 respectively in this invention, and can be adaptively adjusted according to the characteristics of the target field (meteorological field / pollution field).

[0055] In this embodiment, the pixel-level L1 loss is: The mean absolute error between the predicted value and the reference true value is used to constrain the global numerical accuracy and avoid systematic numerical bias. The formula is as follows: ; in, This represents the total number of pixels in a 100-meter grid. For the first The model prediction values ​​for each grid cell. For the first The reference truth value of each grid.

[0056] The Structural Similarity Index Measure (SSIM) is used to constrain the similarity of spatial structure, texture, and edge features between the predicted field and the ground truth. This addresses the problem of traditional L1 loss leading to overly smooth predictions and loss of locally refined features. The formula is shown below: ; in, , These are the local means of the predicted field and the reference true value, respectively. , These are the local standard deviations. For local covariance; , To maintain numerical stability, the constants are taken as follows: , , The numerical dynamic range of the target field data.

[0057] Furthermore, to address the issue of "physical non-conservation" and non-compliance with atmospheric dynamics in deep learning-generated models, this application embodiment chooses to embed hard physical constraints during model training. This integrates conservation laws and boundary condition constraints from atmospheric dynamics into the model optimization objective, ensuring that the generated hundred-meter-level field conforms to the real atmospheric motion laws, corresponding to the physical constraint module in the flowchart.

[0058] Among them, physical constraint terms Constraints are divided into two categories: conservation constraints and boundary condition constraints. The total physical constraint loss is defined as follows: ; in, , , These are the weighting coefficients for each constraint term, which can be adaptively adjusted according to the target field type.

[0059] It is worth mentioning that the conservation constraints include mass conservation, momentum conservation, and energy conservation constraints, which are set separately for the wind field, pressure field, and temperature field of the meteorological field, as well as the pollutant concentration field of the pollution field.

[0060] mass conservation constraint For the atmospheric wind field and pollutant concentration field, their divergence is constrained to conform to the atmospheric continuity equation to avoid non-physical mass sources and sinks, as shown in the following formula: ; in, air density, The wind field vector, For divergence operators, This represents the time-varying term of density.

[0061] Energy conservation constraint For the temperature field and potential height field, constrain their energy transfer to conform to the first law of thermodynamics and avoid non-physical temperature abrupt changes, the formula is as follows: ; in, For isobaric specific heat, For temperature, For volume, For air pressure, This is a non-insulating heating item.

[0062] The momentum conservation constraint is as follows: for the wind field vector, its motion is constrained to conform to the Navier-Stokes equations and incorporated into the regularization term of the model to avoid non-physical wind field shear.

[0063] The core constraint equation (shallow water wave approximation) is: ; ; The momentum constraint loss (regularization term) is: ; in, The eastern component of the 100-meter grid. The northward wind speed component is represented by a grid of 100 meters (model output / reference true value). The material derivative ( ), used to characterize the rate of change of momentum over time of air particles; Coriolis parameters ( , The angular velocity of Earth's rotation. The target grid latitude, in units of ); The density of air (taken as a constant on a scale of hundreds of meters); , The eastward and northward gradients of the pressure field are calculated from the model's output pressure field at the 100-meter level. The atmospheric turbulent viscosity coefficient (taken as a constant near the ground) ); For the Laplace operator ( ), characterizing the viscous diffusion effect of turbulence; , All are near-ground friction forces (calculated from surface roughness z0 and wind speed modulus). The total number of grid cells in the 100-meter level wind field; subscript For the first The corresponding parameters for a 100-meter grid.

[0064] Boundary condition constraints include terrain constraints and large-scale field constraints, which are set for the forced features of the underlying surface in fields of hundreds of meters.

[0065] Terrain constraints The constraint on the correlation between near-surface meteorological elements (such as wind speed and temperature) and 100m topographic elevation, slope, and aspect conforms to the atmospheric boundary layer theory. For example, wind speed increases logarithmically with the increase of topographic elevation, avoiding non-physical high wind speeds in topographically obstructed areas. The formula is shown below: ; in, For the first Near-surface wind speeds for each grid This represents the terrain elevation of the grid. For surface roughness, The geostrophic wind speed at the corresponding grid point. For reference height.

[0066] Large-scale field constraints The constraint ensures that the spatial average value of the generated 100-meter-level field at the 1km scale is consistent with the input original 1km low-resolution forecast data, guaranteeing the consistency of large-scale circulation characteristics and avoiding deviations in the large-scale background field after downscaling. The formula is shown below: ; Where M is the total number of grids in the original 1km low-resolution field. For the first The average value of 100 100-meter-level subgrids corresponding to a 1km grid. For the first The original input values ​​for a 1km grid.

[0067] Ultimately, the overall optimization objective of model training is a weighted sum of the total loss function and the physical constraint loss: ; in, The global weight coefficient for physical constraint loss is set to 0.5 in this invention. During training, a weight warm-up strategy is adopted, gradually increasing the weight to the target value in the first 10 epochs to avoid excessive suppression of the model's feature learning ability by physical constraints in the early stage of training.

[0068] Specifically, an initial Diffusion-Unet model with U-Net as the backdiffusion backbone is trained based on noisy training data. The resulting target Diffusion-Unet model is then used to downscale the low-resolution dynamic field, yielding downscaled atmospheric pollutant and meteorological data results. This process includes: constructing an initial Diffusion-Unet model with U-Net as the backdiffusion backbone, and generating input data based on the noisy training data at each time step and the corresponding aligned multi-source conditional constraint vectors; using the initial Diffusion-Unet model and the input data to determine the noise minimization prediction, obtaining the predicted noise value, and determining the error between the predicted noise value and the true Gaussian noise; and using a multi-scale mixture loss function and a physical constraint loss function, and based on the error, backdiffusion-Unet model is backscaled. The training process involves iterative training; the multi-scale hybrid loss function includes pixel-level L1 loss and structural similarity loss; the physical constraint loss function includes conservation constraint loss and boundary condition constraint loss; the conservation constraint loss function includes mass conservation constraint, momentum conservation constraint, and energy conservation constraint; and the boundary condition constraint loss function includes terrain constraint and large-scale field constraint. When the loss function value corresponding to the current iterative model converges to a preset convergence threshold, the current iterative model is set as the target Diffusion-Unet model. The low-resolution dynamic field is processed to obtain the corresponding low-resolution multi-source condition constraint vector. Then, the target Diffusion-Unet model is used to perform end-to-end reverse diffusion scale reduction processing on the low-resolution multi-source condition constraint vector to obtain the atmospheric pollutant downscaling result and meteorological data downscaling result that meet the preset high-resolution conditions.

[0069] As can be seen from the above, the embodiments of this application first need to obtain low-resolution dynamic field and dynamic downscaling simulation data of the target region that meet the preset low-resolution conditions; secondly, a labeled dataset is constructed based on the preset training labels and dynamic downscaling simulation data, and an initial residual correction model is trained using the labeled dataset to obtain the target residual correction model; then, the residual correction amount is determined using the target residual correction model and based on the dynamic downscaling simulation values ​​corresponding to each grid in the target region, and the residual correction amount is used to perform residual compensation on the dynamic downscaling simulation data to obtain the reference true value that meets the preset high-resolution conditions; furthermore, feature upscaling and spatial reshaping are performed on the low-resolution dynamic field. The process involves several steps: First, dynamic features are obtained by sorting the data. Multi-source conditional constraint vectors are then determined based on these dynamic features and static geographic prior data. Next, Gaussian noise is added to the reference ground truth to obtain noisy high-resolution field data that satisfies the preset high-resolution conditions. Noisy training data is then determined based on the multi-source conditional constraint vectors and the noisy high-resolution field data. Finally, an initial Diffusion-Unet model with U-Net as the backdiffusion backbone is trained based on the noisy training data. This target Diffusion-Unet model is then used to downscale the low-resolution dynamic field, yielding downscaled results for atmospheric pollutants and meteorological data. This improves the efficiency of downscaling atmospheric pollutants and meteorological data, thereby enhancing the user experience.

[0070] Accordingly, see Figure 3 As shown, this application also provides an air pollutant and meteorological data downscaling device, comprising: The low-resolution dynamic field determination module 11 is used to acquire low-resolution dynamic field and dynamic downscaling simulation data of the target area that meet the preset low-resolution conditions; the low-resolution dynamic field includes atmospheric pollutant data and meteorological data. The labeled dataset construction module 12 is used to construct a labeled dataset based on the preset training labels and the dynamic downscaling simulation data, so as to use the labeled dataset to train the initial residual correction model and obtain the target residual correction model. The reference truth generation module 13 is used to determine the residual correction amount by using the target residual correction model and based on the dynamic downscaling simulation value corresponding to each grid in the target area, so as to use the residual correction amount to perform residual compensation on the dynamic downscaling simulation data and obtain a reference truth value that meets the preset high resolution conditions. The multi-source condition constraint vector determination module 14 is used to perform feature dimensionality upscaling and spatial rearrangement on the low-resolution dynamic field to obtain dynamic features, and to determine the multi-source condition constraint vector based on the dynamic features and static geographic prior data. The noisy training data determination module 15 is used to add Gaussian noise to the reference true value to obtain noisy high-resolution field data that meet the preset high-resolution conditions, and to determine noisy training data based on the multi-source condition constraint vector and the noisy high-resolution field data. The scaling down module 16 is used to train an initial Diffusion-Unet model with U-Net as the backdiffusion backbone network based on the noisy training data, and to use the obtained target Diffusion-Unet model to scale down the low-resolution dynamic field to obtain the downscaling results of atmospheric pollutants and meteorological data.

[0071] In some specific embodiments, the low-resolution dynamic field determination module 11 may specifically include: The spatiotemporal boundary parameter determination unit is used to determine the geographical range and time span of the target area, generate spatiotemporal boundary parameters of the target area based on the geographical range and the time span, and generate resolution threshold parameters based on preset low resolution conditions. The low-resolution dynamic field determination subunit is used to obtain the original observation data of atmospheric pollutants and the original meteorological observation data in the target area based on the spatiotemporal boundary parameters and the resolution threshold parameters, so as to integrate the original observation data and the original meteorological observation data in the spatiotemporal dimension to obtain a low-resolution dynamic field that meets the preset low-resolution conditions. The dynamic downscaling simulation data determination unit is used to perform regional dynamic downscaling numerical simulation based on the low-resolution dynamic field and the spatiotemporal boundary parameters to obtain dynamic downscaling simulation data that matches the spatiotemporal dimension of the low-resolution dynamic field.

[0072] In some specific embodiments, the labeled dataset construction module 12 may specifically include: The training label determination unit is used to determine the training label based on the deviation between the measured value of the site and the preset coarse resolution simulation value. The training label is matched with the dynamic downscaling simulation data grid by grid and time step by time step based on the spatiotemporal boundary parameters of the target area to obtain a labeled dataset including dynamic downscaling simulation data samples and corresponding preset training label samples. A dataset partitioning unit is used to divide the labeled dataset into a training subset and a validation subset according to a preset ratio; The model iterative training unit is used to construct an initial residual correction model based on the dynamic downscaling simulation data, and to iteratively train the initial residual correction model using the training subset to obtain a residual correction model to be verified; the residual correction model to be verified is a regression tree model. The model verification unit is used to perform grid-by-grid accuracy verification on the residual correction model to be verified using the verification subset, and to generate a target residual correction model when the verification accuracy of the residual correction model to be verified is greater than a preset accuracy threshold.

[0073] In some specific embodiments, the reference truth generation module 13 may specifically include: A grid division unit is used to determine the grid division rules corresponding to the target area, so as to divide the target area into grids using the grid division rules, obtain each grid, and extract the dynamic downscaling simulation values ​​corresponding to each grid based on the dynamic downscaling simulation data. The grid simulation value sequence generation unit is used to generate a grid simulation value sequence based on each of the dynamic downscaling simulation values, so as to use the target residual correction model and generate a residual correction quantity sequence corresponding to each grid based on the grid simulation value sequence; The reference truth generation subunit is used to map the residual correction sequence to the dynamic downscaling simulation data grid by grid based on the grid partitioning rules, and to perform algebraic summation of the dynamic downscaling simulation value of the corresponding grid with the residual correction to obtain the residual compensation calculation result, so as to generate a reference truth value that meets the preset high resolution conditions based on the residual compensation calculation result.

[0074] In some specific embodiments, the multi-source condition constraint vector determination module 14 may specifically include: A spatial parameter extraction unit is used to extract spatial parameters of the reference truth that satisfy a preset high-resolution condition, so as to generate high-resolution reference parameters based on each of the spatial parameters. The up-dimensional dynamic feature generation unit is used to up-dimensionalize the low-resolution dynamic field according to the channel dimension to obtain the up-dimensional dynamic features after the channel number expansion. The feature space rearrangement unit is used to perform spatial rearrangement processing on the upgraded dynamic features based on the high-resolution benchmark parameters to obtain dynamic features to be processed that match the spatial resolution of the reference true value. A static geographic prior data determination unit is used to determine the static geographic prior data corresponding to the target area, and to perform dimensional matching and normalization processing on the static geographic prior data based on the dimensions of the dynamic features to be processed, to obtain normalized static features; the static geographic prior data includes resolution terrain elevation and underlying surface type. The concatenated result generation unit is used to concatenate the dynamic features to be processed with the normalized static features along the channel dimension to obtain the concatenated result, and to use the channel attention mechanism to determine the channel weights corresponding to each channel dimension, so as to determine the source condition constraint vector based on each channel weight and the concatenated result.

[0075] In some specific embodiments, the noisy training data determination module 15 may specifically include: The noise intensity parameter determination unit is used to determine the diffusion time step sequence and the noise intensity parameter corresponding to each time step, and add Gaussian noise to the reference true value step by step based on the diffusion time step sequence and the noise intensity parameter to obtain the initial noisy high-resolution field data corresponding to each time step. The spatiotemporal alignment unit is used to align the multi-source conditional constraint vector with the initial noisy high-resolution field data corresponding to each time step in the spatiotemporal dimension, so as to obtain the spatiotemporally aligned multi-source conditional constraint vector and the target noisy high-resolution field data. The feature fusion unit is used to perform feature fusion between the aligned multi-source conditional constraint vector and the target noisy high-resolution field data based on the channel dimension to obtain noisy training data; the noisy training data includes paired data; each of the paired data includes the target noisy high-resolution field data and the aligned multi-source conditional constraint vector corresponding to the time step.

[0076] In some specific embodiments, the scale reduction module 16 may specifically include: The input data generation unit is used to construct an initial Diffusion-Unet model with U-Net as the backdiffusion backbone network, and to generate input data based on the noisy training data corresponding to each time step and the corresponding aligned multi-source conditional constraint vector. The noise value prediction unit is used to determine the noise value minimization prediction based on the initial Diffusion-Unet model and the input data to obtain the predicted noise value, and to determine the error between the predicted noise value and the real Gaussian noise. The model iterative training unit is used to perform backdiffusion iterative training on the initial Diffusion-Unet model based on the error using a multi-scale hybrid loss function and a physical constraint loss function; wherein, the multi-scale hybrid loss function includes a pixel-level L1 loss function and a structural similarity loss function; the physical constraint loss function includes a conservation constraint loss function and a boundary condition constraint loss function; the conservation constraint loss function includes a mass conservation constraint function, a momentum conservation constraint function, and an energy conservation constraint function; the boundary condition constraint loss function includes a terrain constraint function and a large-scale field constraint function; The model setting unit is used to set the current iterative model as the target Diffusion-Unet model when the loss function value corresponding to the current iterative model converges to a preset convergence threshold. The reverse diffusion scaling unit is used to process the low-resolution dynamic field to obtain the corresponding low-resolution multi-source condition constraint vector. Then, the target Diffusion-Unet model is used to perform end-to-end reverse diffusion scaling processing on the low-resolution multi-source condition constraint vector to obtain the atmospheric pollutant downscaling result and meteorological data downscaling result that meet the preset high-resolution conditions.

[0077] Furthermore, embodiments of this application also disclose an electronic device, Figure 4 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the atmospheric pollutant and meteorological data downscaling method disclosed in any of the foregoing embodiments. Furthermore, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0078] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0079] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0080] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the atmospheric pollutant and meteorological data downscaling method executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.

[0081] Furthermore, this application also discloses a computer-readable medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned method for downscaling atmospheric pollutants and meteorological data. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0082] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0083] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0084] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of media known in the art.

[0085] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0086] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for downscaling atmospheric pollutant and meteorological data, characterized in that, include: Acquire low-resolution dynamic field and dynamic downscaling simulation data of the target area that meet the preset low-resolution conditions; the low-resolution dynamic field includes atmospheric pollutant data and meteorological data. A labeled dataset is constructed based on preset training labels and the dynamic downscaling simulation data. An initial residual correction model is trained using the labeled dataset to obtain a target residual correction model. Specifically, training labels are determined based on the deviation between the measured values ​​of the stations and the preset coarse resolution simulation values. The training labels and the dynamic downscaling simulation data are matched sequentially grid-by-grid and time-by-time step based on the spatiotemporal boundary parameters of the target region to obtain a labeled dataset including dynamic downscaling simulation data samples and corresponding preset training label samples. The labeled dataset is divided into a training subset and a validation subset according to a preset ratio; An initial residual correction model is constructed based on the dynamic downscaling simulation data, and the initial residual correction model is iteratively trained using the training subset to obtain a residual correction model to be verified; the residual correction model to be verified is a regression tree model; the residual correction model to be verified is verified grid-by-grid accuracy using the verification subset, and a target residual correction model is generated when the verification accuracy of the residual correction model to be verified is greater than a preset accuracy threshold. The residual correction amount is determined by using the target residual correction model and the dynamic downscaling simulation value corresponding to each grid in the target area. The residual correction amount is then used to perform residual compensation on the dynamic downscaling simulation data to obtain a reference true value that meets the preset high resolution conditions. The low-resolution dynamic field is subjected to feature dimensionality upscaling and spatial rearrangement to obtain dynamic features, and multi-source conditional constraint vectors are determined based on the dynamic features and static geographic prior data. Gaussian noise is added to the reference ground value to obtain noisy high-resolution field data that satisfy the preset high-resolution conditions, and noisy training data is determined based on the multi-source condition constraint vector and the noisy high-resolution field data. An initial Diffusion-Unet model with U-Net as the backdiffusion backbone network is trained based on the noisy training data. The resulting target Diffusion-Unet model is then used to downscale the low-resolution dynamic field, yielding downscaled atmospheric pollutant and meteorological data results. Specifically, the initial Diffusion-Unet model with U-Net as the backdiffusion backbone network is constructed, and input data is generated based on the noisy training data corresponding to each time step and the corresponding aligned multi-source conditional constraint vectors. The initial Diffusion-Unet model is used, and noise value minimization prediction is determined based on the input data to obtain the predicted noise value. The error between the predicted noise value and the true Gaussian noise is then determined. A multi-scale hybrid loss function and a physical constraint loss function are used, and the initial Diffusion-Unet model is iteratively trained using backdiffusion based on the error. The multi-scale hybrid loss function includes a pixel-level L1 loss function and a structural similarity loss function; the physical constraint loss function includes a conservation constraint loss function and a boundary condition constraint loss function. The conservation-type constraint loss function includes mass conservation constraint function, momentum conservation constraint function, and energy conservation constraint function; the boundary condition constraint loss function includes terrain constraint function and large-scale field constraint function; when the loss function value corresponding to the current iteration model converges to a preset convergence threshold, the current iteration model is set as the target Diffusion-Unet model; the low-resolution dynamic field is processed to obtain the corresponding low-resolution multi-source condition constraint vector, and then the target Diffusion-Unet model is used to perform end-to-end reverse diffusion scale reduction processing on the low-resolution multi-source condition constraint vector to obtain atmospheric pollutant downscaling results and meteorological data downscaling results that meet the preset high-resolution conditions.

2. The method for downscaling air pollutants and meteorological data according to claim 1, characterized in that, The acquisition of low-resolution dynamic field and dynamic downscaling simulation data of the target region that meet the preset low-resolution conditions includes: The geographical range and time span of the target area are determined, and the spatiotemporal boundary parameters of the target area are generated based on the geographical range and the time span, and the resolution threshold parameters are generated based on the preset low resolution conditions. Based on the spatiotemporal boundary parameters and the resolution threshold parameters, the original observation data of atmospheric pollutants and the original meteorological observation data in the target area are obtained, so as to integrate the original observation data and the original meteorological observation data in the spatiotemporal dimension to obtain a low-resolution dynamic field that meets the preset low-resolution conditions. Based on the low-resolution dynamic field and the spatiotemporal boundary parameters, a regional dynamic downscaling numerical simulation is performed to obtain dynamic downscaling simulation data that matches the spatiotemporal dimension of the low-resolution dynamic field.

3. The method for downscaling air pollutants and meteorological data according to claim 1, characterized in that, The step of determining the residual correction amount using the target residual correction model and based on the dynamic downscaling simulation values ​​corresponding to each grid within the target region, and then using the residual correction amount to perform residual compensation on the dynamic downscaling simulation data to obtain a reference true value that meets the preset high-resolution conditions, includes: The grid division rules corresponding to the target area are determined, and the target area is divided into grids using the grid division rules to obtain each grid. The dynamic downscaling simulation values ​​corresponding to each grid are extracted based on the dynamic downscaling simulation data. A grid simulation value sequence is generated based on the dynamic downscaling simulation values, and a residual correction value sequence corresponding to each grid is generated using the target residual correction model and based on the grid simulation value sequence. Based on the grid partitioning rules, the residual correction sequence is mapped to the dynamic downscaling simulation data grid by grid, and the dynamic downscaling simulation value of the corresponding grid is algebraically summed with the residual correction to obtain the residual compensation calculation result. Based on the residual compensation calculation result, a reference true value that meets the preset high resolution conditions is generated.

4. The method for downscaling air pollutants and meteorological data according to claim 3, characterized in that, The step of performing feature upsizing and spatial rearrangement on the low-resolution dynamic field to obtain dynamic features, and determining multi-source conditional constraint vectors based on the dynamic features and static geographic prior data, includes: Extract the spatial parameters of the reference truth that satisfy the preset high-resolution conditions, and generate high-resolution reference parameters based on each of the spatial parameters; The low-resolution dynamic field is up-dimensionalized according to the channel dimension to obtain the up-dimensional dynamic features after channel expansion. Based on the high-resolution benchmark parameters, the upgraded dynamic features are spatially rearranged to obtain dynamic features to be processed that match the spatial resolution of the reference true value. The static geographic prior data corresponding to the target area is determined, and the static geographic prior data is subjected to dimension matching and normalization based on the dimension of the dynamic feature to be processed to obtain normalized static features; the static geographic prior data includes resolution terrain elevation and underlying surface type. The dynamic features to be processed and the normalized static features are concatenated along the channel dimension to obtain the concatenated result. The channel weights corresponding to each channel dimension are determined using the channel attention mechanism, and the source condition constraint vector is determined based on the channel weights and the concatenated result.

5. The method for downscaling air pollutants and meteorological data according to claim 1, characterized in that, The step of adding Gaussian noise to the reference ground value to obtain noisy high-resolution field data that satisfy the preset high-resolution conditions, and determining noisy training data based on the multi-source conditional constraint vector and the noisy high-resolution field data, includes: Determine the diffusion time step sequence and the noise intensity parameter corresponding to each time step, and add Gaussian noise to the reference true value step by step based on the diffusion time step sequence and the noise intensity parameter to obtain the initial noisy high-resolution field data corresponding to each time step; The multi-source conditional constraint vector is spatiotemporally aligned with the initial noisy high-resolution field data corresponding to each time step to obtain the spatiotemporally aligned multi-source conditional constraint vector and the target noisy high-resolution field data. Based on the channel dimension, the aligned multi-source conditional constraint vector and the target noisy high-resolution field data are fused to obtain noisy training data; the noisy training data includes paired data; each of the paired data includes the target noisy high-resolution field data corresponding to the time step and the aligned multi-source conditional constraint vector.

6. A device for downscaling atmospheric pollutant and meteorological data, characterized in that, include: The low-resolution dynamic field determination module is used to acquire low-resolution dynamic field and dynamic downscaling simulation data of the target area that meet preset low-resolution conditions; the low-resolution dynamic field includes atmospheric pollutant data and meteorological data. The labeled dataset construction module is used to construct a labeled dataset based on preset training labels and the dynamic downscaling simulation data, so as to train an initial residual correction model using the labeled dataset and obtain a target residual correction model; specifically, the training labels are determined based on the deviation between the measured values ​​of the stations and the preset coarse resolution simulation values, and the training labels and the dynamic downscaling simulation data are matched sequentially grid by grid and time step by time based on the spatiotemporal boundary parameters of the target region to obtain a labeled dataset including dynamic downscaling simulation data samples and corresponding preset training label samples; The labeled dataset is divided into a training subset and a validation subset according to a preset ratio; An initial residual correction model is constructed based on the dynamic downscaling simulation data, and the initial residual correction model is iteratively trained using the training subset to obtain a residual correction model to be verified; the residual correction model to be verified is a regression tree model; the residual correction model to be verified is verified grid-by-grid accuracy using the verification subset, and a target residual correction model is generated when the verification accuracy of the residual correction model to be verified is greater than a preset accuracy threshold. The reference truth generation module is used to determine the residual correction amount based on the target residual correction model and the dynamic downscaling simulation value corresponding to each grid in the target area, so as to use the residual correction amount to perform residual compensation on the dynamic downscaling simulation data and obtain the reference truth value that meets the preset high resolution conditions. The multi-source condition constraint vector determination module is used to perform feature dimensionality upscaling and spatial rearrangement on the low-resolution dynamic field to obtain dynamic features, and to determine the multi-source condition constraint vector based on the dynamic features and static geographic prior data. The noisy training data determination module is used to add Gaussian noise to the reference ground value to obtain noisy high-resolution field data that meet the preset high-resolution conditions, and to determine noisy training data based on the multi-source condition constraint vector and the noisy high-resolution field data. The scaling down module is used to train an initial Diffusion-Unet model with U-Net as the backdiffusion backbone network based on the noisy training data, and to use the obtained target Diffusion-Unet model to scale down the low-resolution dynamic field, obtaining the downscaling results of atmospheric pollutants and meteorological data; specifically, it constructs an initial Diffusion-Unet model with U-Net as the backdiffusion backbone network, and generates input data based on the noisy training data corresponding to each time step and the corresponding aligned multi-source conditional constraint vector; it uses the initial Diffusion-Unet model and the input data to determine the noise value minimization prediction, obtains the predicted noise value, and determines the error between the predicted noise value and the real Gaussian noise; it uses a multi-scale hybrid loss function and a physical constraint loss function, and performs backdiffusion iterative training on the initial Diffusion-Unet model based on the error; wherein, the multi-scale hybrid loss function includes a pixel-level L1 loss function and a structural similarity loss function; the physical constraint loss function includes a conservation constraint loss function and a boundary condition constraint loss function; and the... The conservation-type constraint loss function includes mass conservation constraint function, momentum conservation constraint function, and energy conservation constraint function; the boundary condition constraint loss function includes terrain constraint function and large-scale field constraint function; when the loss function value corresponding to the current iteration model converges to a preset convergence threshold, the current iteration model is set as the target Diffusion-Unet model; the low-resolution dynamic field is processed to obtain the corresponding low-resolution multi-source condition constraint vector, and then the target Diffusion-Unet model is used to perform end-to-end reverse diffusion scale reduction processing on the low-resolution multi-source condition constraint vector to obtain atmospheric pollutant downscaling results and meteorological data downscaling results that meet the preset high-resolution conditions.

7. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the atmospheric pollutant and meteorological data downscaling method as described in any one of claims 1 to 5.

8. A computer-readable medium, characterized in that, Used to store a computer program, wherein the computer program, when executed by a processor, implements the atmospheric pollutant and meteorological data downscaling method as described in any one of claims 1 to 5.

Images