Regional gas downscaling generation method based on diffusion model

Abstract

The invention relates to the technical field of agricultural meteorological prediction, in particular to a regional gas downscaling generation method based on a diffusion model, and the method comprises the steps: extracting spatial-temporal characteristics of different time steps and spatial positions through a multi-scale convolutional network, and then carrying out the noise optimization and denoising of a prediction result through a diffusion model. By combining a physical consistency discriminator and a quantum heuristic optimization technology, physical errors in meteorological prediction are further corrected. And finally, through self-supervised online fine tuning, an accurate agricultural meteorological prediction result is output. Compared with a traditional prediction method, the method has the advantages that the space-time feature capturing capability is remarkably improved, the problems of precision and stability in long-time scale meteorological prediction in the prior art are effectively solved, and the method has high application value.

Description

Technical Field

[0001] The invention relates to the technical field of agricultural meteorological forecasting, and in particular to a regional atmospheric downscaling generation method based on a diffusion model. Background Art

[0002] Accurate prediction of meteorological data plays a vital role in the fields of agricultural meteorology, environmental monitoring and disaster warning. Traditional meteorological downscaling methods usually rely on statistical models or relatively simple machine learning algorithms, such as linear regression and support vector machines. However, these methods face significant challenges when dealing with highly complex spatiotemporal data. For example, existing statistical models such as Kriging interpolation can provide good predictions when the data is sparse or the structure is relatively simple, but when dealing with complex high-dimensional spatiotemporal data, especially when predicting extreme weather events, it is often difficult to accurately capture the nonlinear characteristics and spatiotemporal dependencies in the data. Deep learning-based methods, such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs), have solved this problem to a certain extent, but these methods still face the challenge of how to deal with spatial and temporal dependencies at the same time, especially in long-term meteorological forecasts. The accuracy and stability of existing technologies are still difficult to meet the needs of practical applications.

[0003] Existing diffusion models have also been used to model spatiotemporal data, but many methods fail to effectively combine multi-scale features and physical constraints, resulting in limitations in the model's prediction effect. In particular, in the field of agricultural meteorological forecasting, there is a lack of models that can simultaneously capture local and global spatiotemporal features, which affects the accuracy and robustness of long-term meteorological forecasts. Summary of the invention

[0004] In view of the many problems existing in the above-mentioned prior art, the present invention provides a regional downscaling generation method based on a diffusion model. The present invention improves the prediction ability of meteorological data by combining a multi-scale convolutional neural network (CNN) and a diffusion model. The multi-scale convolutional network is used to extract the features of meteorological images from different spatiotemporal scales, and these features reflect the dynamic changes at different time and space scales. The diffusion model gradually removes the noise in the prediction results through a multi-step inverse diffusion process, optimizes the spatiotemporal feature representation, and enhances the prediction ability of long-term meteorological changes. The local energy, mass conservation and precipitation distribution errors are calculated by a physical consistency discriminator, and corrected in combination with a quantum heuristic optimization algorithm, which further improves the accuracy and stability of the model. The present invention effectively overcomes the shortcomings of the prior art in processing complex spatiotemporal data, and can provide high-precision and stable long-term meteorological forecasts, which is particularly suitable for application fields such as agricultural meteorological forecasts.

[0005] A regional atmospheric downscaling generation method based on a diffusion model, the method comprising the following steps: Using regional model meteorological field data, historical satellite observation meteorological field data, terrain data, and land use data, a fused conditional vector is generated through multi-modal feature encoding and a multi-head self-attention mechanism; Initial high-dimensional noise data is sampled from a standard normal distribution, and the initial high-dimensional noise data and the fused conditional vector are jointly input into a pre-trained diffusion model. Preliminary target resolution meteorological field data is generated through inverse diffusion updates, where variational Bayesian inference, differential evolution optimization, and the Lagrange multiplier method are embedded in the inverse diffusion update process to achieve physical constraint correction; For the preliminary target resolution meteorological field data, a physical consistency discriminator is used to calculate the physical error vectors of local energy, mass conservation, and precipitation distribution. The local and global correction problems are decomposed based on the alternating direction multiplier method, a correction vector is obtained by combining quantum-inspired optimization, and the preliminary target resolution meteorological field data is corrected online using automatic differentiation technology to generate feedback-corrected meteorological field data; For the feedback-corrected meteorological field data, multi-scale convolutional neural networks and wavelet transforms are used to extract low-scale, mid-scale, and high-scale features. A spatial graph structure is constructed using graph signal processing algorithms to achieve signal smoothing between regions, and the final target resolution meteorological field data is output through self-supervised online fine-tuning.

[0006] Preferably, the multi-modal feature encoding module performs three-layer convolution operations on regional model meteorological field data using a convolutional neural network with a convolutional layer, activation layer, and pooling layer structure to extract local spatial features, uses a convolutional neural network with fixed convolutional kernel parameters to extract global reflectivity information from historical satellite observation meteorological field data, uses a graph convolutional network based on neighborhood node distance aggregation information to extract terrain undulation features from terrain data, uses a multi-layer convolutional neural network to extract spectral and texture features from land use data, and linearly transforms through a fully connected layer and then uses a multi-head self-attention mechanism to weight and fuse the foregoing features to generate a fused conditional vector.

[0007] Preferably, the initial high-dimensional noise data is obtained by randomly sampling a standard normal distribution with a mean of zero and a variance of one, and the size of the sampled noise data is consistent with the spatial resolution represented by the regional model meteorological field data.

[0008] Preferably, variational Bayesian inference is embedded in the inverse diffusion update process. The variational Bayesian inference is implemented by constructing a multi-layer neural network based on the likelihood function of the local energy distribution, and the posterior probability of the noise scheduling parameter in the inverse diffusion process is updated using the obtained likelihood function.

[0009] Preferably, the differential evolution optimization algorithm is adopted in the inverse diffusion update process. The differential evolution optimization algorithm quantifies the numerical difference between the preliminary target resolution meteorological field data and the predetermined target energy spectrum by setting a fitness function, and performs mutation, crossover, and selection operations on the candidate correction vectors to iteratively update the candidate solutions, thereby determining the correction vector for physical constraint correction.

[0010] Preferably, the Lagrange multiplier method is applied in the inverse diffusion update process. By calculating the local energy error, mass conservation error, and precipitation distribution error, and generating a physical constraint correction term according to the preset multiplier coefficient, the pre-updated state data is corrected to correct the physical indicators.

[0011] Preferably, the physical consistency discriminator adopts a feedforward neural network structure, performs operations on each pixel in the preliminary target resolution meteorological field data, outputs the local energy error, mass conservation error, and precipitation distribution error by setting an activation function, and forms a physical error vector by combining the error values in a fixed order.

[0012] Preferably, the alternating direction multiplier method decomposes the global correction problem into a local correction sub-problem and a global correction sub-problem, solves the local update amount and the global update amount by setting local penalty parameters and global penalty parameters respectively, and forms an overall update amount by linearly weighting and integrating the update amounts; the quantum-inspired optimization algorithm uses a simulated annealing strategy to perform a global search on the candidate correction vector composed of the local update amount and the global update amount. The global search evaluates the candidate correction vector based on the fitness function and selects the optimal correction vector according to the evaluation result.

[0013] Preferably, the automatic differentiation technology calculates the gradients of the preliminary target resolution meteorological field data and the correction vector through the backpropagation algorithm, and updates the preliminary target resolution meteorological field data using the gradient descent method to achieve online correction and generate the meteorological field data after feedback correction.

[0014] Preferably, for the meteorological field data after feedback correction, a convolutional neural network with low-scale convolutional layers, medium-scale convolutional layers, and high-scale convolutional layers is used to extract spatial features of different scales respectively, the data is decomposed in the frequency domain using wavelet transform, a spatial graph structure is constructed based on geographical coordinate information, and a graph signal processing algorithm is used to smooth the adjacent pixel information. Subsequently, the final target resolution meteorological field data is output through self-supervised online fine-tuning.

[0015] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows: By combining a multi-scale convolutional network with a diffusion system, the present invention solves the defect that the prior art cannot fully capture spatio-temporal dependence relationships. By extracting spatio-temporal features at different scales through the multi-scale convolutional network and combining with the denoising process of the diffusion model, the present invention effectively improves the accuracy of agricultural meteorological data in long-term prediction. Specifically, through the noise addition of the diffusion model and the multi-step inverse diffusion process, the present invention can optimize the representation of spatio-temporal features, ensure the accurate capture of complex spatio-temporal dynamics, and thus enhance the robustness and accuracy of the model for long-term meteorological changes. In addition, the physical consistency discriminator and quantum-inspired optimization method of the present invention provide additional physical constraint corrections for meteorological data, further ensuring the reliability of the prediction results.

[0016] Through this combination of technologies, the present invention not only achieves remarkable results in improving the accuracy of meteorological prediction, but also enables the model to adapt to new data changes in real time through self-supervised fine-tuning and online optimization technologies, providing more flexible and accurate meteorological prediction services. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flowchart of the method of the present invention; Figure 2 is a schematic diagram of physical consistency detection and online correction in the present invention; Figure 3 is a schematic diagram of multi-scale post-processing in the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0018] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present disclosure. In the following detailed description, for the sake of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure. However, it is obvious that one or more embodiments can be implemented without these specific details. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessarily confusing the concepts of the present disclosure.

[0019] The terms used herein are merely for describing specific embodiments and are not intended to limit the present disclosure. The terms "including", "comprising", etc. used herein indicate the presence of the described features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0020] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0021] AsFigure 1 As shown in the figure, a regional precipitation downscaling generation method based on a diffusion model, the method comprising the following steps: Using regional model meteorological field data, historical satellite observation meteorological field data, terrain data, and land use data, generate a fused conditional vector through multi-modal feature encoding and a multi-head self-attention mechanism; The present invention aims to achieve regional precipitation downscaling generation based on a diffusion model. Its core idea is to construct a fused conditional vector using multi-modal data from different sources, providing accurate and controllable prior information for the diffusion model to generate meteorological field data at the target resolution. The data sources include regional model meteorological field data, historical satellite observation meteorological field data, terrain data, and land use data. After preprocessing each data type, its unique physical and statistical features are extracted through a specially designed multi-modal feature encoding module, and then the various features extracted are weighted and fused using a multi-head self-attention mechanism to generate a representative fused conditional vector.

[0022] This fused conditional vector can comprehensively reflect the regional large-scale meteorological background, local climate characteristics, terrain undulations, and land cover conditions, forming a key condition to guide the diffusion model to gradually generate meteorological field data at the target resolution during the inverse diffusion process. During the generation process, strict control of physical constraints is achieved through inverse diffusion updates using variational Bayesian inference, differential evolution optimization, and the Lagrange multiplier method, ensuring that the generated data meets the predetermined requirements in terms of physical indicators such as local energy, mass conservation, and precipitation distribution. Further, a physical consistency discriminator is used to detect errors in the preliminarily generated data, a correction vector is obtained based on the alternating direction multiplier method and quantum-inspired optimization, and online correction is implemented in combination with automatic differentiation technology. Finally, the meteorological field data at the target resolution optimized through multi-scale post-processing and graph signal processing is output. Through the collaborative work of the above modules, the present invention not only takes into account the characteristics of each data but also realizes the deep fusion of multi-modal data, providing a solution with strong physical constraints and statistical consistency for regional precipitation downscaling generation. Its overall principle and implementation steps show a clear technical route and operability in each link of data preprocessing, feature extraction, conditional fusion, inverse diffusion generation, and online correction, providing detailed guidance for subsequent specific implementation.

[0023] Preferably, the multi-modal feature encoding module performs three-layer convolution operations on the regional model meteorological field data using a convolutional neural network with convolutional layers, activation layers, and pooling layers to extract local spatial features, uses a convolutional neural network with fixed convolutional kernel parameters to extract global reflectivity information from historical satellite observation meteorological field data, uses a graph convolutional network based on neighborhood node distance aggregation information to extract terrain undulation features from terrain data, uses a multi-layer convolutional neural network to extract spectral and texture features from land use data, and linearly transforms through a fully connected layer and then uses a multi-head self-attention mechanism to weight and fuse the aforementioned features to generate a fused conditional vector.

[0024] In the process of feature extraction from the regional model meteorological field data, a convolutional neural network with convolutional layers, activation layers, and pooling layers is used to achieve multi-level extraction of local spatial features. Specifically, the convolutional neural network is designed with a three-layer structure, and each layer includes a convolutional operation, a non-linear activation function, and a pooling operation. First, the first-layer convolutional operation is performed on the input regional model meteorological field data. A convolutional kernel with a set size of 3×3 is used to scan the local receptive field of the original data, and a preliminary feature map is obtained through linear filtering. Its mathematical expression is

[0025] where, represents the input data, and represent the first-layer convolutional kernel weights and biases respectively, represents the convolutional operation, and ReLU is the rectified linear unit. Subsequently, after passing through the max-pooling layer to perform downsampling, reducing the spatial dimension while retaining significant features, and then using the pooling result as the input of the second-layer convolution, repeating similar convolution, activation, and pooling steps to obtain the intermediate feature map . In the third-layer convolution, local detailed information is further extracted through fine-grained convolution operations to obtain a deeper feature representation , during the whole process, the convolutional kernel size, stride, and pooling window of each layer are determined according to experimental tuning to ensure that the extracted features can fully express the local meteorological changes and spatial gradient information in the regional model meteorological field data. This process can not only capture the minute changes of meteorological parameters such as temperature, humidity, and wind speed in the local area but also reflect the weather system structure at the regional scale, providing a fine local feature basis for the generation of subsequent fusion condition vectors. In addition, this convolutional neural network adopts batch normalization technology to stabilize the training process and uses a dropout layer to prevent overfitting, further enhancing the robustness and generalization ability of feature extraction. In actual operation, an open-source deep learning framework can be used to build the network in the embodiment, and the network is trained using regional model meteorological field data samples, and hyperparameters are adjusted to achieve the optimal feature extraction effect, so as to ensure that the extracted local spatial features can accurately reflect the meteorological conditions of the target area.

[0026] For historical satellite observation meteorological field data, a convolutional neural network with fixed convolutional kernel parameters is used to extract global reflectivity information, aiming to obtain meteorological features with stability and representativeness in a long time series. In this method, the network structure is set as a group of convolutional layers with fixed parameters. The convolutional kernel size is fixed at 5×5, the stride is set to 1, and the convolutional kernel parameters are not dynamically adjusted, so as to ensure the consistency of the features extracted at different time sections. Mathematically, its feature extraction process can be expressed as:

[0027] where, is the historical satellite observation meteorological field data matrix, is the preset fixed convolutional kernel parameter, is the bias term, and the ReLU function is used to introduce non-linear mapping. After the convolutional operation, a global average pooling layer is used to statistically aggregate the obtained features, and the calculation formula is:

[0028] where represents the total number of pixels in the feature map. This operation can integrate local information into global statistical features, thus reflecting the large-scale climate background and the overall reflectivity level of the meteorological field. Through the fixed-parameter convolutional layer, the instability caused by parameter update can be effectively reduced, so that the extracted global reflectivity features remain highly consistent in time, providing a globally stable prior feature input for regional climate downscaling. This implementation scheme is suitable for processing long-term satellite observation data. After data normalization and standardization, the data is input into the fixed convolutional kernel network to obtain representative global reflectivity features, and these features are used as part of the subsequent feature fusion to effectively capture the large-scale meteorological change trend.

[0029] For the feature extraction of terrain data, the present invention adopts a graph convolutional network based on the aggregated information of neighborhood node distances to capture terrain undulation and local elevation change information. In specific implementation, the terrain data is discretized into a regular grid, each pixel in the grid is regarded as a node in the graph, and an adjacency matrix is constructed according to the geographical location to calculate the distances between nodes. Each element in the adjacency matrix reflects the geometric relationship and elevation difference between adjacent nodes. The core calculation formula of the graph convolutional network is:

[0030] where, is the initial node feature matrix (e.g., containing elevation values, slope values, etc.), is the graph adjacency matrix, is the node degree matrix, and its diagonal elements are the number of connected edges of the corresponding nodes, is the training parameter matrix, is the activation function, such as ReLU. Through the graph convolution operation, the features of each node are updated under the aggregation of the information of adjacent nodes, thus reflecting the local continuity and change trend of the terrain. To further enhance the feature expression ability, multiple layers of graph convolution can be adopted in the network. The first layer aggregates the direct neighborhood information, and the second layer further captures the comprehensive effect of the second-order neighborhood, so that the extracted terrain features not only reflect the elevation information of a single node, but also contain the overall terrain trend of the surrounding area. In the practical application of the present invention, digital elevation model data can be used for training. By preprocessing the elevation data (such as noise filtering and normalization), high-quality graph-structured data is constructed, and then features representing terrain undulation, slope change, and watershed distribution are extracted through the graph convolutional network, providing accurate terrain information input for subsequent feature fusion.

[0031] In the process of feature extraction of land use data, the present invention adopts a multi-layer convolutional neural network to extract the spectral and texture features of the data. First, the land use data is preprocessed, including data format conversion, normalization, and spatial alignment, to ensure that the input data meets the network requirements. Then, the preprocessed land use image data is subjected to convolution operations through a multi-layer convolutional neural network. This network consists of consecutive convolutional layers, activation layers, and pooling layers. Each layer of convolution operation uses a convolutional kernel of a fixed size (such as 3×3 or 5×5), introduces a non-linear mapping through a non-linear activation function (such as ReLU), and uses the pooling layer to reduce the spatial resolution of the feature map, thereby extracting local features reflecting land cover types, texture details, and spectral distributions. After the extracted features are processed by several layers of convolution, a linear transformation is performed through a fully connected layer to map the multi-dimensional features into the same latent space. This process can be expressed as:

[0032] where, represents the feature vector output by the convolutional network, and are the weights and biases of the fully connected layer. Subsequently, to achieve cross-modal feature fusion, a multi-head self-attention mechanism is adopted to weight and fuse the features extracted from each data source. In the specific calculation process, the features output by the fully connected layer are divided into a query matrix , a key matrix , and a value matrix , and their calculation formula is:

[0033] where is the dimension of the key matrix, and the function is used to normalize the similarity scores. The finally output features are concatenated after parallel calculation by multiple attention heads and mapped back to the original dimension, thereby generating a fused conditional vector. This mechanism effectively captures the interdependent relationships between the features of each modality and highlights the key features through attention weights, ensuring that the output fused conditional vector contains both spectral and texture features in information expression and also fuses spatial geometric information, becoming an important prior information for guiding the diffusion model to generate meteorological field data at the target resolution. In the embodiment, a publicly available dataset can be used to train the above-mentioned multi-layer convolutional neural network and self-attention module, and the network structure parameters can be adjusted to obtain the best feature fusion effect.

[0034] As Figure 2 shown, initial high-dimensional noise data is sampled from the standard normal distribution, and the initial high-dimensional noise data and the fused conditional vector are jointly input into a pre-trained diffusion model, and preliminary target resolution meteorological field data is generated through inverse diffusion updates, where variational Bayesian inference, differential evolution optimization, and the Lagrange multiplier method are embedded in the inverse diffusion update process to achieve physical constraint correction; The present invention realizes regional precipitation downscaling generation based on a diffusion model. The core task is to generate meteorological field data at the target resolution from initial random noise data through an inverse diffusion process. The diffusion model adopts initial high-dimensional noise data sampled from the standard normal distribution and inputs it into a pre-trained diffusion network in combination with a fused conditional vector. In the inverse diffusion update process, by embedding variational Bayesian inference, differential evolution optimization, and the Lagrange multiplier method, the noise scheduling parameters are dynamically adjusted and the process of generating target meteorological field data is optimized to ensure that the generated meteorological field data meets the expected requirements in terms of physical constraints (such as local energy, mass conservation, precipitation distribution, etc.).

[0035] Variational Bayesian inference provides posterior probability updates for noise adjustment. Differential evolution optimization updates candidate solutions through an optimization algorithm and obtains a correction vector. The Lagrange multiplier method further controls the physical constraint correction term to ensure that the generated data is consistent with the actual physical phenomena. Finally, through the combination of multi-scale convolutional neural networks, wavelet transforms, and graph signal processing algorithms, the generated meteorological field data is further optimized to improve the accuracy and consistency of its spatial features, ensuring that the final output meteorological data has high resolution, high physical consistency, and reasonable geospatial correlation.

[0036] Preferably, the initial high-dimensional noise data is obtained by randomly sampling from a standard normal distribution with a mean of zero and a variance of one. The size of the sampled noise data is consistent with the spatial resolution represented by the regional model meteorological field data.

[0037] The generation of the initial high-dimensional noise data adopts the method of random sampling from a standard normal distribution. The mean of the noise data is zero and the variance is one, ensuring that the generated noise data satisfies the basic characteristics of the standard normal distribution. The sampled noise data has the same spatial resolution and dimension as the regional model meteorological field data. This processing method ensures that the generated data can match the spatial characteristics of the target meteorological field. For example, if the spatial resolution of the regional model meteorological field data is 1 km, then the size of the sampled noise data will correspond to a resolution of 1 km. Specifically, the sampling of this standard normal distribution can be achieved by using a random number generator (such as numpy.random.normal(0, 1, size=(m, n)) in Python), ensuring that the dimension of the noise data is consistent with the target meteorological field data, where m and n are the number of rows and columns of the generated noise data, corresponding to the latitude and longitude lengths of the target meteorological field data. The initial noise data generated in this way provides a necessary randomness basis for the subsequent diffusion model generation process, and also ensures the consistency of the initial data and the target data in spatial resolution, ensuring physical consistency during the generation process.

[0038] Preferably, variational Bayesian inference is embedded in the inverse diffusion update process. The variational Bayesian inference is implemented by constructing a multi-layer neural network based on the likelihood function of the local energy distribution, and the obtained likelihood function is used to update the posterior probability of the noise scheduling parameters in the inverse diffusion process.

[0039] During the inverse diffusion update process, variational Bayesian inference updates the posterior probability of the noise scheduling parameters by constructing a multi-layer neural network based on the likelihood function of the local energy distribution. Specifically, the goal of variational Bayesian inference is to infer the most likely parameter values based on the input noise data and the fusion condition vector, thereby guiding the diffusion model update process. First, the model uses the local energy distribution as the likelihood function to construct a multi-layer neural network, where each layer of the network is responsible for processing features at different scales, and gradually extracts the global features of the data through layer-by-layer transmission between layers. The training objective of this neural network is to minimize the loss function, which consists of the difference between the noise data and the actual meteorological field data, as well as the physical consistency error. The neural network adjusts the network parameters through the backpropagation algorithm to make the output parameter values more consistent with the true distribution of the target meteorological field data.

[0040] In this process, variational Bayesian inference can not only update the prior knowledge of the model according to the observed data, but also continuously optimize the model parameters through the iterative process, thereby providing accurate noise scheduling information for generating the target meteorological field data. For example, when processing precipitation data in the meteorological field, variational Bayesian inference can update the noise distribution in the precipitation generation process according to the spatial distribution pattern and local change characteristics of the precipitation data, so as to ensure that the finally generated data is more in line with the actual situation in terms of local precipitation values and spatial distribution.

[0041] Preferably, a differential evolution optimization algorithm is adopted in the inverse diffusion update process. The differential evolution optimization algorithm quantifies the numerical difference between the preliminary target resolution meteorological field data and the predetermined target energy spectrum by setting a fitness function, and performs mutation, crossover, and selection operations on the candidate correction vectors to iteratively update the candidate solutions, thereby determining the correction vector for physical constraint correction.

[0042] In the inverse diffusion update process, the differential evolution optimization algorithm is used to adjust the correction vector to achieve physical constraint correction. Differential evolution optimization is a population-based global optimization algorithm that can effectively search for the global optimal solution by performing mutation, crossover, and selection operations among candidate solutions. In the present invention, the differential evolution algorithm quantifies the numerical difference between the generated preliminary target resolution meteorological field data and the predetermined target energy spectrum by setting a fitness function. The calculation method of the fitness function is:

[0043] where is the model prediction value of the th sample point, is the corresponding target energy spectrum value, is the correction vector to be optimized, is the total number of sample points. The differential evolution algorithm randomly introduces mutations among candidate solutions to form new solutions, then generates offspring solutions through crossover operations, and selects the most suitable solutions from the parents and offspring through selection operations. The iterative process continuously adjusts the correction vector until the fitness function is minimized, thereby obtaining the optimal physical constraint correction vector. This process can ensure that the finally generated meteorological field data is as close as possible to the predetermined target data in terms of global energy distribution, thus ensuring physical consistency and the accuracy of practical applications.

[0044] Preferably, the Lagrange multiplier method is applied in the inverse diffusion update process. By calculating the local energy error, mass conservation error, and precipitation distribution error, and generating a physical constraint correction term according to the preset multiplier coefficient, the pre-updated state data is corrected to calibrate the physical indicators.

[0045] The Lagrange multiplier method is applied in the inverse diffusion update process to calculate the local energy error, mass conservation error, and precipitation distribution error, and generate a physical constraint correction term to correct the pre-updated state data. The Lagrange multiplier method is a classic method for solving optimization problems with constraints. It adds the constraint conditions to the optimization objective by introducing Lagrange multipliers. In the specific operation process, first, the local energy error, mass conservation error, and precipitation distribution error between the initially generated target resolution meteorological field data and the actual observation data are calculated to obtain the physical error vector. These error vectors represent the deviations of the generated data in physical properties and need to be corrected through physical constraint correction. On this basis, the Lagrange multiplier method introduces Lagrange multiplier terms and minimizes these error terms together with the constraint conditions. The final correction term is:

[0046] where is the error function, , and are the physical constraint terms for local energy, mass conservation, and precipitation distribution respectively, , and are the Lagrange multipliers used to adjust the weights of each physical constraint. By optimizing this Lagrange function, the finally generated correction vector can effectively correct the initially generated meteorological field data to meet the actual physical constraint requirements. The present invention effectively solves the contradiction between physical consistency and the quality of generated data, ensuring that the generated data not only conforms to statistical laws but also conforms to actual physical phenomena.

[0047] For the preliminary target resolution meteorological field data, a physical consistency discriminator is used to calculate the physical error vectors of local energy, mass conservation, and precipitation distribution. Based on the alternating direction multiplier method, the local and global correction problems are decomposed. A correction vector is obtained by combining quantum-inspired optimization, and the preliminary target resolution meteorological field data is corrected online using automatic differentiation technology to generate the meteorological field data after feedback correction. The present invention addresses the problem of regional downscaling of precipitation. A generation method based on a diffusion model is adopted, and through the inverse diffusion process, the initial random high-dimensional noise is gradually transformed into the meteorological field data with the target resolution. The core lies in using a physical consistency discriminator to quantitatively detect the physical indicators of the preliminarily generated data, decomposing the correction problem based on the alternating direction multiplier method, obtaining the optimal correction vector through quantum-inspired optimization, and finally implementing online correction of the preliminary target resolution meteorological field data by combining automatic differentiation technology to generate the meteorological field data after feedback correction. This method comprehensively considers key physical indicators such as local energy, mass conservation, and precipitation distribution to ensure that the generated data meets the predetermined constraints in terms of statistical characteristics and physical laws.

[0048] Specifically, first, initial high-dimensional noise data is sampled from the standard normal distribution. This noise data has statistical characteristics of a mean of zero and a variance of one, and its size is consistent with the spatial resolution described by the regional model meteorological field data. Then, the sampled noise data and the fusion conditional vector generated through multi-modal feature encoding and multi-head self-attention mechanism are jointly input into the pre-trained diffusion model, and the noise is gradually removed using the inverse diffusion update formula. During the inverse diffusion update process, through embedded variational Bayesian inference, differential evolution optimization, and the Lagrange multiplier method, the noise scheduling parameters and physical constraint correction terms are updated in real time to ensure that the preliminarily generated target resolution meteorological field data can meet the requirements of local energy, mass conservation, and precipitation distribution in terms of physical indicators. Subsequently, a physical consistency discriminator is used to calculate the physical errors of each pixel, and then the alternating direction multiplier method is used to decompose the global correction problem into local and global sub-problems. Then, the quantum-inspired optimization algorithm is combined to globally search for the candidate correction vector, and the gradient is calculated using the backpropagation algorithm through automatic differentiation technology. The gradient descent method is used to perform online correction on the preliminary target resolution meteorological field data, thereby generating the final meteorological field data after feedback correction.

[0049] This method not only ensures that the data gradually evolves from random noise to data with higher physical consistency during the generation process but also, through a strict physical constraint correction mechanism, ensures that the final output data has both a high spatial resolution and meets the requirements of actual physical phenomena, providing a complete and operable technical solution for regional downscaling of precipitation.

[0050] Preferably, the physical consistency discriminator adopts a feedforward neural network structure to perform operations on each pixel in the preliminary target resolution meteorological field data, and outputs the local energy error, mass conservation error, and precipitation distribution error by setting an activation function, and combines the error values in a fixed order to form a physical error vector.

[0051] In the present invention, the physical consistency discriminator adopts a feedforward neural network structure to perform point-by-point operations on each pixel in the preliminary target resolution meteorological field data. Its basic principle is to map the input data to a high-dimensional feature space and extract complex features through multi-layer linear transformation and non-linear activation functions. Specifically, when implemented, each pixel data input to the feedforward neural network includes temperature, humidity, wind speed, and other relevant meteorological quantities, which are used as the input vector of the neural network after normalization processing. The neural network consists of an input layer, several hidden layers, and an output layer. Each layer of the hidden layer uses linear transformation calculations, that is:

[0052] And through an activation function, such as ReLU (Rectified Linear Unit), a non-linear mapping is performed to obtain:

[0053] Until the output layer, its output obtains an output with a fixed numerical range by setting a specific activation function (such as Sigmoid or Tanh). This output represents the numerical values of the pixel in terms of local energy error, mass conservation error, and precipitation distribution error. In actual operation, to ensure the consistency of the calculation of each physical index, first, the network input needs to be strictly standardized to convert data with different dimensions into feature vectors with the same dimension. Subsequently, the weight parameters of the neural network are obtained through training on a large amount of historical observation data and numerical simulation data, enabling the network to accurately predict physical errors. For the local energy error, its calculation principle can be based on the difference between the measured value and the model prediction value, and the formula can be expressed as:

[0054] Where, represents the actually measured temperature or energy value, Denote the predicted value of the neural network output; similarly, the mass conservation error and precipitation distribution error are calculated according to their respective physical constraints. Finally, these error values are combined into a physical error vector in a fixed order through a fully connected layer in a predetermined order. This vector contains multiple scalars, each scalar corresponding to the error of a key physical index. In practical applications, the batch processing method can be used to calculate the entire meteorological field data to obtain a global physical error distribution map. This distribution map not only reflects local errors but also shows the error transfer trend between regions, providing an intuitive basis for subsequent correction. In the embodiment, by setting a certain error threshold, such as the local energy error not exceeding 0.6 (in standardized values), to ensure that subsequent correction operations are carried out under strict physical constraints, thereby improving the physical consistency of the generated data. The training of this feedforward neural network can use the backpropagation algorithm and use the mean square error as the loss function to ensure that the deviation between the predicted output and the actual error is minimized, thereby providing a high-quality physical error vector output for online correction.

[0055] Preferably, the alternating direction multiplier method decomposes the global correction problem into a local correction sub-problem and a global correction sub-problem, solves the local update amount and the global update amount by setting local penalty parameters and global penalty parameters respectively, and forms an overall update amount through linear weighted integration of the update amounts.

[0056] In the present invention, the alternating direction multiplier method (ADMM) is used to decompose the global correction problem into a local correction sub-problem and a global correction sub-problem. Its core idea is to transform the original problem into a series of sub-problems to be solved respectively by introducing auxiliary variables and Lagrange multipliers, and then construct the overall update amount. Specifically, when implementing, first define a global objective function, which is used to describe the difference between the preliminary target resolution meteorological field data and the predetermined physical target. This difference includes local energy error, mass conservation error, and precipitation distribution error. Let the original objective function be:

[0057] where, denotes the meteorological field data to be corrected, is the physical constraint matrix, is the target physical index. Using the alternating direction multiplier method, first introduce an auxiliary variable so that the original problem is decomposed into:

[0058] where the function is used to represent the physical constraint term of global correction. The update steps of the ADMM algorithm include the alternating optimization of the variables and , and the Lagrange multiplier The updates, and their update formulas are respectively:

[0059]

[0060]

[0061] Among them, is the regularization parameter, is the number of iterations. By setting appropriate local penalty parameters and global penalty parameters, the local update amount and the global update amount can be solved respectively. The local update amount reflects the error between the preliminary target resolution meteorological field data and the target physical indicators in the local area, while the global update amount reflects the physical inconsistency of the entire area data as a whole. After multiple iterations, the local update amount and the global update amount are linearly weighted and integrated according to a fixed weight to form an overall update amount, which is used to guide the online correction process. Specifically, the overall update amount can be expressed as:

[0062] Among them, represents the local update amount, represents the global update amount, and are the linear weighting coefficients. The present invention ensures that local and global corrections are carried out simultaneously, which can not only make fine adjustments for local physical errors, but also ensure that the overall meteorological field data meets the physical constraint requirements on the global scale. In actual operation, in the embodiments, appropriate local and global penalty parameters are preset using historical data and numerical simulation results, and the alternating direction method of multipliers is used for iterative solution until the overall update amount converges to the optimal solution, thereby providing a stable update basis for the subsequent quantum-inspired optimization stage.

[0063] The quantum-inspired optimization algorithm uses a simulated annealing strategy to globally search the candidate correction vector composed of the local update amount and the global update amount. The global search evaluates the candidate correction vector based on a fitness function and selects the optimal correction vector according to the evaluation result.

[0064] In the present invention, the quantum-inspired optimization algorithm uses a simulated annealing strategy for global search of candidate correction vectors, and its core lies in searching for the global optimal solution through random perturbation and temperature parameter control. Specifically, based on the candidate correction vector composed of the local update amount and the global update amount decomposed by the alternating direction method of multipliers, the simulated annealing algorithm is used for global optimization. The simulated annealing algorithm is based on the physical annealing process, and an initial temperature , and gradually reduce the temperature in an exponential decay manner, randomly generate candidate solutions at each temperature, and evaluate them according to the fitness function. The fitness function is used to quantify the difference between the candidate correction vector and the target physical index, and the mathematical expression is:

[0065] in, Indicates candidate correction vectors, Indicates A preset target correction vector, is the vector dimension. In each iteration, the algorithm calculates the acceptance probability based on the fitness difference between the current temperature and the candidate solution. The acceptance probability is usually expressed using the Boltzmann distribution function as follows:

[0066] in, Represents the difference between the candidate solution fitness function value and the current solution, is the current temperature. If the new candidate solution reduces the value of the fitness function, the candidate solution is accepted, otherwise the poor candidate solution is accepted with a certain probability, thereby avoiding falling into the local optimum. By continuously lowering the temperature and iteratively updating, the optimal correction vector is finally obtained. In actual implementation, this process requires determining the initial temperature, cooling rate, and number of iterations, which can usually be determined through experimental data debugging. For example, in practical applications, the initial temperature can be set to 1.0, the cooling rate to 0.95, and the number of iterations to 500 steps to ensure that a stable optimal solution is obtained in the global search. The simulated annealing algorithm has strong global search capabilities and robustness, and can find a better solution in complex optimization problems, so that the correction vector plays a key role in the correction of physical constraints, ensuring that the meteorological field data obtained by subsequent online correction is close to the predetermined target in terms of physical indicators.

[0067] Preferably, the automatic differentiation technology calculates the gradient of the preliminary target resolution meteorological field data and the correction vector through a back-propagation algorithm, and uses a gradient descent method to update the preliminary target resolution meteorological field data, thereby achieving online correction to generate feedback-corrected meteorological field data.

[0068] The automatic differentiation technology is used in the present invention to calculate the gradient of the preliminary target resolution meteorological field data and the obtained correction vector, and the gradient descent method is used to perform online correction on the data, thereby generating feedback-corrected meteorological field data. Automatic differentiation is based on the back propagation algorithm, and its core principle is to derive the complex composite function layer by layer and multiply the gradients of each layer to achieve efficient calculation of the overall gradient. In the specific implementation, the preliminary target resolution meteorological field data is represented as a variable , the correction vector is expressed as , the objective function is defined as:

[0069] Among them, represents the Euclidean norm. The backpropagation algorithm calculates the gradient of the loss function with respect to by the chain rule, that is:

[0070] Using this gradient information, the value of is updated by the gradient descent method, and its update formula is:

[0071] Among them, is the meteorological field data after the th update, is the learning rate parameter, and usually a suitable value is determined through experiments, such as 0.01 or 0.001. This online correction process adjusts the data according to the latest calculated gradient in each iteration, making the generated meteorological field data gradually approach the ideal target. In actual operation, the automatic differentiation technology can be implemented through deep learning frameworks (such as TensorFlow or PyTorch). This technology automatically records the computational graph and performs gradient propagation during the update process to ensure the efficiency and accuracy of the computational process. Through online correction, not only can the physical inconsistencies in the preliminary data be corrected, but also the data changes can be responded to in real time to achieve continuous correction, so that the finally output meteorological field data meets the predetermined standards in terms of local features and overall physical indicators, thus meeting the strict requirements for data accuracy and physical consistency in regional precipitation scale generation.

[0072] Such as Figure 3 shown, the multi-scale convolutional neural network and wavelet transform are used to extract low-scale, medium-scale, and high-scale features from the meteorological field data after the feedback correction, the graph signal processing algorithm is used to construct a spatial graph structure to realize signal smoothing between regions, and the final target resolution meteorological field data is output through self-supervised online fine-tuning.

[0073] In the present invention, the ultimate goal is to ensure that the generated meteorological field data meets the requirements in terms of physical consistency, spatial feature accuracy and regional smoothness by combining a variety of advanced technologies. First, the feedback-corrected meteorological field data is subjected to multi-scale convolutional neural network (CNN) and wavelet transform to extract low-scale, medium-scale and high-scale spatial features. The convolutional neural network captures local spatial features of different scales through multi-level convolution operations and retains spatial structural information. The wavelet transform further decomposes the data in the frequency domain and extracts features in each frequency band, which is crucial for improving spatial resolution and accuracy. Next, based on the geographic coordinate information, a spatial graph structure is constructed, and the smoothing of inter-regional signals is achieved through the graph signal processing algorithm, thereby eliminating high-frequency noise and inconsistency in the data and maintaining a balance between large scale and local details. Finally, the output target resolution meteorological field data is further optimized by self-supervised online fine-tuning to ensure its stability and accuracy in practical applications. This series of steps ensures the consistency and efficiency of meteorological field data in different spatial scales, frequency domains and physical characteristics, and provides reliable data support and processing paths for regional downscaling generation.

[0074] Preferably, a convolutional neural network with low-scale convolutional layers, medium-scale convolutional layers and high-scale convolutional layers is used to extract spatial features of different scales from the meteorological field data after feedback correction, wavelet transform is used to decompose the data in the frequency domain, a spatial graph structure is constructed based on geographic coordinate information, and a graph signal processing algorithm is used to smooth the information of adjacent pixels, and then self-supervised online fine-tuning is used to achieve the output of meteorological field data with the final target resolution.

[0075] In the present invention, the meteorological field data after feedback correction are further processed by a multi-scale convolutional neural network. Specifically, the convolutional neural network extracts features of different spatial scales through low-scale, medium-scale, and high-scale convolutional layers respectively, so as to ensure that the meteorological features from large scales to local micro-scales can be fully captured during the processing. First, the low-scale convolutional layer uses smaller convolutional kernels (such as 3×3 or 5×5) to scan the input meteorological data and extract local spatial information at a fine granularity, such as the temperature, humidity, and wind speed changes of individual pixels. These features can capture the local dynamic changes in the meteorological data, such as the changes in local climate patterns. Next, the convolutional kernel size of the medium-scale convolutional layer increases (such as 7×7 or 9×9), and feature extraction is carried out on a larger receptive field, focusing on the larger-scale patterns and trends in the meteorological field, such as the changes in regional airflows, the formation of systematic meteorological phenomena (such as monsoons, fronts), etc. Finally, the high-scale convolutional layer performs wider-range spatial feature extraction through larger-sized convolutional kernels (such as 15×15 or 20×20). This layer can capture the global features of the meteorological field, such as large-scale climate fluctuations, abnormal weather events, etc. The convolutional operation of each layer uses the ReLU activation function for non-linear mapping, enabling the network to capture more complex meteorological field data features.

[0076] The output of the convolutional operation of each layer is downsampled through a pooling layer, reducing the computational amount while retaining key information. The pooling operation usually uses the max-pooling or average-pooling method to select the most significant features from the output of the convolutional layer respectively. The pooled data enters the fully connected layer for further processing to ensure that the features extracted from each scale can be effectively combined together. This process can not only improve the efficiency of spatial feature extraction but also improve the spatial resolution of the meteorological field data through the information fusion of different scales, thereby improving the accuracy and reliability of the downscaled meteorological data.

[0077] After the multi-scale convolutional neural network, the data is decomposed in the frequency domain through wavelet transform. Wavelet transform is an effective signal processing technology. By converting the signal from the time domain to the frequency domain, it can finely analyze the features of the signal at different scales. In the processing of meteorological field data, the main purpose of using wavelet transform is to separate information of different frequencies and extract low-frequency features (such as large-scale meteorological patterns) and high-frequency features (such as local changes, noise, etc.). Specifically, the meteorological field data is first decomposed through wavelet transform to obtain a set of low-frequency and high-frequency coefficients. The low-frequency coefficients contain the overall trend and large-scale patterns of the data, while the high-frequency coefficients capture the detailed information in the meteorological data, such as local weather phenomena, sudden meteorological events, etc.

[0078] The mathematical formula of wavelet transform is usually expressed as:

[0079] Among them, is the transformed signal, is the original signal, is the wavelet basis function, is the scale factor, which controls the expansion and compression of the wavelet basis function. By extracting and decomposing the frequency domain characteristics at different scales, the downscaling results of meteorological data can be optimized from both global and local levels, and noise can be eliminated while retaining the true meteorological patterns.

[0080] Based on geographical coordinate information, a spatial graph structure is constructed to further optimize the smoothing of meteorological field data. The core of the spatial graph structure is to regard each pixel as a node in the graph, and the edges between nodes are constructed through geographical coordinate information. These edges reflect the spatial adjacency relationship, that is, the distance and relationship between adjacent pixels. The construction method of the graph usually uses a similarity measure based on geographical distance to connect adjacent nodes. For example, the Euclidean distance or Manhattan distance is used to define the edges between nodes, and the weights of the edges can be set based on these distances or other relevant geographical features.

[0081] The key of the graph signal processing algorithm lies in processing the signal defined on the graph structure, mainly by smoothing the signals between adjacent nodes to eliminate the inconsistencies in the data. When processing meteorological data, this algorithm can smooth the local errors by weighted averaging of the adjacent nodes in the graph, using the spatial relationship between nodes. For example, if there is an obvious error in a meteorological field data point, the graph signal processing algorithm will smooth and adjust this point based on the values of its neighboring nodes, making the value of this point more conform to the trend of its neighboring area. Specifically, the aggregation of node features can be achieved through graph convolution operations. The general form of graph convolution operations is:

[0082] Among them, is the output feature matrix, is the normalized adjacency matrix of the graph, is the input feature matrix, is the weight matrix, is the activation function. Through multi-layer graph convolution processing, the signals in the spatial graph structure can be smoothed layer by layer, eliminating noise and retaining useful meteorological information, and finally obtaining fine regional meteorological field data.

[0083] Self-supervised online fine-tuning is a key technology in the present invention, aiming to further optimize the accuracy and consistency of the generated meteorological field data. In this process, the model makes online adjustments based on the error between the output generated by itself and the actual observed data to achieve gradual improvement of the generated data. Self-supervised learning is a variant of unsupervised learning, in which the model makes self-adjustments by minimizing the error with the real data. Specifically, when implemented, the model compares the currently output meteorological data with the real observed data, calculates the error, and updates the model parameters through the backpropagation algorithm. The error calculation formula is:

[0084] where, is the meteorological field data predicted by the current model, is the actual observed value, is the number of data samples. By calculating the gradient through backpropagation, the model continuously optimizes its parameters, gradually reducing the difference between the predicted value and the actual observed value, and finally outputs accurate meteorological field data. During the implementation process, the fine-tuning strategy can adopt the batch-based stochastic gradient descent method (SGD) or its variants (such as the Adam optimizer) to ensure the efficiency and convergence during the optimization process. Through self-supervised online fine-tuning, the model can adaptively make adjustments and continuously improve the results during the generation process, ensuring that the finally output meteorological field data has high accuracy and physical consistency.

[0085] Those skilled in the art should understand that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects.

[0086] The above are only the embodiments of the present application and are not used to limit the present application. For those skilled in the art, various changes and modifications can be made to the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included within the scope of the claims of the present application.

Claims

1. A regional atmospheric downscaling generation method based on a diffusion model, characterized in that: The method comprises the following steps: Using regional model meteorological field data, historical satellite observation meteorological field data, terrain data and land use data, a fusion conditional vector is generated through multimodal feature encoding and multi-head self-attention mechanism; Sampling initial high-dimensional noise data from a standard normal distribution, and inputting the initial high-dimensional noise data and the fusion condition vector into a pre-trained diffusion model, generating preliminary target resolution meteorological field data through inverse diffusion updating, wherein variational Bayesian inference, differential evolution optimization, and Lagrange multiplier method are embedded in the inverse diffusion updating process to achieve physical constraint correction; The physical error vectors of local energy, mass conservation and precipitation distribution are calculated for the preliminary target resolution meteorological field data using a physical consistency discriminator, and the local and global correction problems are decomposed based on the alternating direction multiplier method, and the correction vector is obtained by combining quantum heuristic optimization, and the preliminary target resolution meteorological field data is corrected online using automatic differentiation technology to generate feedback-corrected meteorological field data; A multi-scale convolutional neural network and wavelet transform are used to extract low-scale, medium-scale and high-scale features from the feedback-corrected meteorological field data. A graph signal processing algorithm is used to construct a spatial graph structure to achieve inter-regional signal smoothing. The meteorological field data with the final target resolution is output through self-supervised online fine-tuning.

2. The method according to claim 1, characterized in that The multimodal feature encoding module uses a convolutional neural network with a structure of convolutional layer, activation layer and pooling layer to perform three-layer convolution operation on the regional model meteorological field data to extract local spatial features. A convolutional neural network with fixed convolution kernel parameters is used to extract global reflectivity information for historical satellite observation meteorological field data. A graph convolutional network based on neighborhood node distance aggregation information is used to extract terrain undulation features for terrain data. A multi-layer convolutional neural network is used to extract spectral and texture features for land use data. After linear transformation through the fully connected layer, the multi-head self-attention mechanism is used to perform weighted fusion of the above features to generate a fusion condition vector.

3. The method according to claim 1, characterized in that The initial high-dimensional noise data is obtained by randomly sampling a standard normal distribution with a mean of zero and a variance of one. The size of the noise data generated by the sampling is consistent with the spatial resolution represented by the regional model meteorological field data.

4. The method according to claim 1, characterized in that: Variational Bayesian inference is embedded in the inverse diffusion updating process. The variational Bayesian inference is realized by constructing a multi-layer neural network based on the likelihood function of the local energy distribution, and the obtained likelihood function is used to perform a posterior probability update on the noise scheduling parameters in the inverse diffusion process.

5. The method according to claim 1, characterized in that The differential evolution optimization algorithm is used in the inverse diffusion update process. The differential evolution optimization algorithm sets the fitness function to quantify the numerical difference between the preliminary target resolution meteorological field data and the predetermined target energy spectrum, and performs mutation, crossover and selection operations on the candidate correction vectors to iteratively update the candidate solutions, thereby determining the correction vector for physical constraint correction.

6. The method according to claim 1, characterized in that The Lagrange multiplier method is applied in the inverse diffusion update process. By calculating the local energy error, mass conservation error and precipitation distribution error, and generating physical constraint correction terms based on the preset multiplier coefficients, the pre-updated state data are corrected to correct the physical indicators.

7. The method according to claim 1, characterized in that The physical consistency discriminator adopts a feedforward neural network structure to perform operations on each pixel in the preliminary target resolution meteorological field data. It outputs the local energy error, mass conservation error and precipitation distribution error by setting the activation function, and combines the error values ​​in a fixed order to form a physical error vector.

8. The method according to claim 1, characterized in that The alternating direction multiplier method decomposes the global correction problem into a local correction subproblem and a global correction subproblem, solves the local update amount and the global update amount respectively by setting the local penalty parameter and the global penalty parameter, and integrates the update amounts by linear weighting to form the overall update amount; the quantum heuristic optimization algorithm adopts a simulated annealing strategy to perform a global search on the candidate correction vectors composed of the local update amount and the global update amount, the global search evaluates the candidate correction vectors based on the fitness function, and selects the optimal correction vector according to the evaluation result.

9. The method according to claim 1, characterized in that: The automatic differentiation technology calculates the gradient of the preliminary target resolution meteorological field data and the correction vector through the back propagation algorithm, and uses the gradient descent method to update the preliminary target resolution meteorological field data, so as to realize online correction and generate feedback corrected meteorological field data.

10. The method according to claim 1, characterized in that A convolutional neural network with low-scale convolutional layers, medium-scale convolutional layers and high-scale convolutional layers is used to extract spatial features of different scales from the meteorological field data after feedback correction. The wavelet transform is used to decompose the data in the frequency domain, and a spatial graph structure is constructed based on the geographic coordinate information. The graph signal processing algorithm is used to smooth the information of adjacent pixels. Subsequently, self-supervised online fine-tuning is used to achieve the output of meteorological field data with the final target resolution.

Images