Meteorological downscaling method fusing numerical weather forecast and AI weather forecast

Abstract

The invention relates to the technical field of weather forecast, in particular to a weather downscaling method fusing numerical weather forecast and AI weather forecast, which comprises the following steps: acquiring numerical weather forecast data and satellite remote sensing observation data, and performing unified preprocessing, interpolation and mapping to generate refined forecast data and observation reference data. Then, local compensation based on a deep neural network and model-independent element learning are adopted to carry out regional adaptive compensation, a joint objective function is constructed to implement variational assimilation and joint optimization, and meanwhile, a generative adversarial network and a graph neural network are combined with reinforcement learning to realize error compensation and node-level adaptive correction; and finally, continuously updating parameters through closed-loop feedback iteration to generate updated forecast data. According to the method, the problems of insufficient local details and poor model adaptability in the downscaling process in the prior art are effectively solved, and the forecasting precision and robustness are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological forecasting, and particularly to a meteorological downscaling method that integrates numerical weather forecasting and AI weather forecasting. Background Art

[0002] Meteorological downscaling technology is of great significance for improving the spatial details of forecasts and capturing local meteorological phenomena, and can provide more refined meteorological information for practical applications such as power grid dispatching and disaster warning. The existing technology mainly uses traditional linear interpolation methods for downscaling (Chinese invention patent, publication number: CN118981060A, title: A meteorological forecasting method integrating downscaling and multi-model integration), combines Kalman filtering to correct the numerical forecast results, and uses radial basis function neural networks to integrate the forecast results of multiple parties; however, due to the relatively coarse numerical model grid, linear interpolation cannot fully recover local non-linear characteristics, and the insufficient adaptability of Kalman filtering and traditional neural networks to complex meteorological fields, there are obvious defects in forecast accuracy, local detail expression, and model robustness. At the same time, the downscaling and multi-model integration processing in the existing technology is often a one-way process, lacking a dynamic feedback mechanism, and it is difficult to achieve real-time correction of forecast errors and adaptive update of model parameters, thus limiting the forecast performance when facing complex non-linear atmospheric movements. Summary of the Invention

[0003] Aiming at the many problems existing in the above-mentioned existing technology, the present invention provides a meteorological downscaling method that integrates numerical weather forecasting and AI weather forecasting. The present invention uses unified preprocessing, local compensation by deep neural networks, meta-learning, variational assimilation, and generative adversarial networks to achieve data closed-loop feedback and parameter adaptive update. The present invention gradually reduces the forecast error through multiple iterations, so that the downscaled forecast results are significantly improved in both spatial details and physical rationality.

[0004] A meteorological downscaling method that integrates numerical weather forecasting and AI weather forecasting, comprising the following steps: Collect the forecast data output by the numerical model and the satellite observation data within the corresponding forecast area, and perform preprocessing on the data to unify the data format, and use interpolation and mapping techniques to obtain refined forecast data and observation reference data respectively; Perform local compensation processing on the refined forecast data based on a deep neural network, and through meta-learning technology, pre-train on historical data to achieve regional adaptive compensation to generate local compensation data, and add the refined forecast data and the local compensation data point by point to form preliminary refined data. At the same time, construct a joint objective function including mean square error and physical constraints, perform variational assimilation and joint optimization on the preliminary refined data, use a generative adversarial network to perform error compensation under physical constraints, and use a graph neural network combined with reinforcement learning to achieve node-level adaptive correction to obtain corrected forecast data; Format the corrected forecast data to form the final forecast data, calculate the grid-level forecast error using the observation benchmark data, and convert the forecast error into a feedback gradient, which is used to update the parameters of local compensation, variational assimilation, and adaptive correction, and generate updated data through closed-loop feedback iteration.

[0005] Preferably, the preprocessing operation adopts the Z-score normalization method, standardizes each data item by calculating the mean and standard deviation of the forecast data and satellite observation data respectively, and converts the spatial coordinates of the two types of data into a unified standard coordinate system.

[0006] Preferably, the interpolation and mapping technology adopts the bilinear interpolation method to calculate the weighted average of adjacent pixel values in the forecast data and satellite observation data, and combines the non-linear mapping algorithm based on convolution operation to improve the spatial resolution of the forecast data and satellite observation data respectively, so as to obtain refined forecast data and observation benchmark data respectively.

[0007] Preferably, the local compensation process adopts a U-shaped network with an encoding and decoding structure. The U-shaped network uses continuous convolution operations and downsampling to extract features, and then restores details through upsampling and skip connections, and sets a self-attention module in the network to enhance the local feature response.

[0008] Preferably, the meta-learning technology adopts the model-agnostic meta-learning method, pre-trains the parameters of the deep neural network on no less than three independent historical data sets, and fine-tunes the network parameters in the regional adaptation stage to generate local compensation data.

[0009] Preferably, the joint objective function consists of the mean square error loss and the physical constraint loss. The physical constraint loss calculates the deviation of physical quantities in the forecast data based on the principles of momentum conservation and energy conservation, and sets a fixed value as the weight coefficient to constrain the physical characteristics.

[0010] Preferably, the variational assimilation and joint optimization steps adopt the gradient descent algorithm, calculate the gradient of the loss function with respect to the network parameters through automatic differentiation technology, and update the parameters according to the calculated gradient to optimize the network parameters.

[0011] Preferably, the generative adversarial network adopts a convolutional autoencoder structure to form a generator, extracts the error features in the jointly optimized forecast data through encoding and reconstructs the error compensation data through decoding, while the discriminator adopts a convolutional neural network structure to distinguish the generated error compensation data from the real data, so as to achieve error compensation under physical constraints.

[0012] Preferably, the graph neural network combines reinforcement learning to aggregate the grid node features divided in the prediction area by using a graph convolutional network, and sets a function with the reduction of the prediction error as the reward to train the correction weights of each node, so as to achieve node-level adaptive correction.

[0013] Preferably, in the closed-loop feedback iteration step, the prediction error is calculated grid by grid by using the observed reference data and the final prediction data, and the obtained prediction error is converted into a feedback gradient, which is used to continuously update the parameters of each link of local compensation processing, variational assimilation and adaptive correction, so that the grid-level prediction error is gradually reduced in the iteration process.

[0014] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows: By adopting local compensation processing of deep neural network, model-agnostic meta-learning technology, variational assimilation joint optimization, generative adversarial network and graph neural network combined with reinforcement learning, the effect of closed-loop feedback iterative optimization of prediction data is realized; By constructing a deep neural network with an encoder-decoder structure and a self-attention mechanism, local compensation is performed on the downscaled prediction data, and the network parameters are pre-trained on multi-region historical data by using meta-learning, so that the model can quickly adapt to new regions; By constructing a joint objective function containing mean square error and physical constraints, variational assimilation and gradient descent are used to continuously update the network parameters; Using a generative adversarial network to compensate for the error of the jointly optimized data under physical constraints, and then combining a graph convolutional network and reinforcement learning to adaptively correct the node-level features in the prediction area, effectively reducing the global and local prediction errors; Through the above technical means, the present invention not only solves the problems of insufficient detail recovery of downscaled data and poor model adaptability in the prior art, but also constructs a feedback iteration mechanism, so that the prediction data gradually approaches the true observation in continuous update, thereby improving the overall prediction accuracy and robustness. Description of the Drawings

[0015] Figure 1 It is a schematic flow chart of the method of the present invention; Figure 2 It is a schematic diagram of local compensation and meta-learning in the present invention; Figure 3 It is a schematic diagram of the joint objective function and optimization in the present invention; Figure 4 It is a schematic diagram of the adaptive correction of the graph neural network and reinforcement learning in the present invention; Figure 5 It is a schematic diagram of feedback iteration and parameter update in the present invention. Detailed Embodiments

[0016] 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, numerous 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 also 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.

[0017] 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.

[0018] 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.

[0019] As Figure 1 shown, a meteorological downscaling method that combines numerical weather prediction and AI weather prediction includes the following steps: Collect the forecast data output by the numerical model and the satellite observation data within the corresponding forecast area, and perform preprocessing on the data to unify the data format, and use interpolation and mapping techniques to obtain refined forecast data and observation reference data respectively; The present invention aims to achieve the unification of multi-source data and the improvement of spatial resolution in the meteorological downscaling process by integrating numerical weather prediction and artificial intelligence weather prediction technologies, so as to provide high-quality basic data for subsequent meteorological forecast correction and fusion. To this end, first, collect the forecast data output by the numerical model, which are obtained by the numerical weather prediction model through physical process calculations and usually include key meteorological elements such as temperature, humidity, wind speed, and precipitation. Its spatial resolution is limited by the model grid size and it is difficult to reflect local details. At the same time, use satellite remote sensing technology to collect satellite image data within the forecast area. These image data can provide real-time information on the surface and atmospheric states and have high detail resolution in space. Since the two data sources differ in collection methods, output formats, spatial projections, and resolutions, preprocessing operations must be performed on them to achieve consistency in data format, numerical scale, and spatial coordinates.

[0020] The preprocessing operation eliminates the dimension and systematic errors between the two types of data by performing unified normalization and coordinate transformation on the forecast data and satellite observation data respectively, and lays a foundation for subsequent data fusion. Further, by adopting interpolation and mapping techniques, the spatially resolved data is processed to improve the spatial resolution. The interpolation technique can smoothly transition the coarse-resolution data in space, while the mapping technique restores local details through non-linear transformation, thereby obtaining refined forecast data and observation reference data respectively. This data processing flow not only solves the problems of inconsistent data formats and mismatched resolutions between data sources, but also provides a unified and reliable input for local correction, model joint optimization, and closed-loop feedback iteration in subsequent weather forecasting. As a whole, the present invention can ensure the efficient docking and fusion of data in each link of data collection, preprocessing, spatial interpolation, and mapping, ensuring that the numerical model and satellite observation data are processed at the same spatial scale and data format, thereby improving the accuracy and stability of meteorological downscaling forecasting. Generally speaking, the present invention effectively eliminates the differences between data sources by implementing standardized and unified preprocessing on the collected data, and then using interpolation and mapping techniques to improve the resolution, providing refined forecast data and observation reference data that match the actual observation data for the artificial intelligence processing module, ensuring that the subsequent fusion and correction algorithms are carried out at a unified scale, and thus achieving the goal of synergistically improving the meteorological forecasting resolution with multi-source data.

[0021] Preferably, the preprocessing operation adopts the Z-score normalization method to standardize each data item by calculating the mean and standard deviation of the forecast data and satellite observation data respectively, and converts the spatial coordinates of the two types of data into a unified standard coordinate system.

[0022] In terms of the data preprocessing operation, the present invention preferably adopts the Z-score normalization method to perform standardized processing on the collected numerical model forecast data and satellite image data respectively. In the specific operation process, first, statistical analysis is performed on each numerical value in the forecast data to calculate the mean value (denoted as ) and standard deviation (denoted as ) of all data items, and then according to the formula:

[0023] the original numerical value is converted into a standardized value , where Represents the normalized data value. This method ensures that the normalized forecast data presents a standard normal distribution, thereby eliminating the numerical deviation caused by factors such as sensor characteristics, data acquisition time and measurement errors. At the same time, the satellite image data is also normalized using the same statistical indicators, that is, the mean and standard deviation of each pixel in each image are calculated respectively, and the above normalization formula is applied to all pixel data, so that both types of data are converted into standardized data. This operation not only makes the forecast data and observation data consistent in numerical range, but also provides a unified scale basis for subsequent data fusion. In addition, the preprocessing operation also involves uniformly converting the spatial coordinates of the two types of data into a preset standard coordinate system, such as WGS84 or Universal Transverse Mercator projection. In the process of spatial coordinate conversion, the longitude and latitude or projection coordinate value of each data point is recalculated by the projection conversion algorithm based on the projection information of the original data and the parameters of the target coordinate system, ensuring that the forecast data and satellite observation data can be geographically one-to-one corresponding. In practical applications, in order to improve processing efficiency and data consistency, the above-mentioned normalization and coordinate conversion operations can be automated through programming, and batch processing technology is used to uniformly operate large-scale data sets, thereby achieving efficient and accurate data preprocessing. The Z score normalization method used in the present invention has the advantages of simple calculation, intuitive results and wide application range, and is particularly suitable for meteorological downscaling applications that require unified processing of data from different sources. Through preprocessing operations, not only can the systematic errors between different data sources be reduced, but also subsequent interpolation, mapping and artificial intelligence correction steps can be seamlessly connected on the data scale, providing a stable and reliable data basis for the overall downscaling process.

[0024] Preferably, the interpolation and mapping technology uses a bilinear interpolation method to calculate the weighted average of adjacent pixel values in the forecast data and satellite observation data, and combines a nonlinear mapping algorithm based on convolution operations to respectively enhance the spatial resolution of the forecast data and satellite observation data, thereby obtaining refined forecast data and observation benchmark data, respectively.

[0025] In terms of interpolation and mapping technology, the present invention preferably adopts a bilinear interpolation method combined with a nonlinear mapping algorithm based on convolution operation to perform spatial resolution enhancement processing on forecast data and satellite image data respectively, so as to obtain refined forecast data and observation benchmark data. First, bilinear interpolation is a weighted average calculation method based on the four neighboring data points around the target point. Its basic principle is: suppose the point to be found is located inside the rectangle formed by the four known data points, and suppose the upper left, upper right, lower left, and lower right data values of the rectangle are respectively , , and , the horizontal distance of the target point relative to the upper left corner is , the vertical distance is , then the target point value can be calculated by the following formula:

[0026] where and both range from 0 to 1, reflecting the position ratios of the target point in the horizontal and vertical directions. By this method, the low-resolution data can be spatially smoothed and interpolated, thus initially improving the resolution. For the forecast data and satellite image data, after bilinear interpolation processing respectively, preliminary refined data with a higher grid density can be generated. Next, the present invention further uses a non-linear mapping algorithm based on convolution operation to refine the interpolated data to restore the local features and detailed information in the data. This non-linear mapping algorithm usually adopts a deep convolutional network structure, and performs non-linear transformation on the data through multiple convolutional operations and activation functions (such as the ReLU function). In specific operations, first, the image data processed by bilinear interpolation is used as the network input, local spatial features are extracted through consecutive convolutional layers, then the linear combination is mapped to a non-linear output through the activation function, and finally, after several convolutional processes, image data with a further improved resolution and richer details is output.

[0027] For example, the convolutional layer can be set with a convolutional kernel size of 3×3 or 5×5, and the stride and padding parameters are adjusted according to the actual data characteristics; the selection of the activation function and the number of layers are optimized according to the training samples and the expected output. After training, this convolutional network can accurately capture the texture, edges and local change information in the data, thus realizing the refinement process. In addition, to ensure that the forecast data and satellite image data have a consistent spatial structure after interpolation and mapping processing, the present invention stipulates that the same interpolation algorithm and non-linear mapping network structure are adopted for the two types of data, so as to ensure that the refined forecast data and the observed reference data can correspond one by one in grid division, spatial distribution and numerical range. Using the above interpolation and mapping techniques can not only effectively improve the spatial resolution of the data, but also restore the local features lost due to physical model approximation in the low-resolution data, providing refined input data for subsequent artificial intelligence correction and data fusion. In short, in the present invention, through the method of combining bilinear interpolation and non-linear mapping of the convolutional network, the resolution improvement and detail restoration of the forecast data and satellite image data are realized, ensuring that the two types of data have consistent and high-quality spatial information in subsequent multi-source data fusion and artificial intelligence processing links, thereby improving the overall accuracy and reliability of the meteorological downscaling forecast system.

[0028] such as Figure 2As shown in the figure, local compensation processing based on a deep neural network is performed on the refined forecast data. After pre-training on historical data through meta-learning technology, regional adaptive compensation is achieved to generate local compensation data. The refined forecast data and the local compensation data are superimposed point by point to form preliminary refined data. At the same time, a joint objective function containing mean square error and physical constraints is constructed to perform variational assimilation and joint optimization on the preliminary refined data. An error compensation is carried out using a generative adversarial network under physical constraints, and a graph neural network combined with reinforcement learning is used to achieve node-level adaptive correction to obtain corrected forecast data; The present invention aims to achieve the deep integration of numerical weather forecasting and artificial intelligence weather forecasting in meteorological downscaling. The core objective is to improve the spatial resolution and detail restoration ability of forecast data, so as to provide high-quality input data for subsequent meteorological forecast correction and data fusion. To achieve the above objectives, the present invention first uniformly collects and preprocesses the forecast data output by the numerical model and the satellite remote sensing image data collected within the forecast area. The collected forecast data is from a numerical weather forecasting model, and its output data reflects various key meteorological elements such as temperature, humidity, wind speed, and precipitation. However, due to the use of a relatively large grid in the numerical model calculation, its spatial resolution is relatively low, making it difficult to meet the requirements of capturing local meteorological characteristics; while satellite image data directly reflects the state of the atmosphere and the surface, with a high resolution and can provide local detail information.

[0029] To effectively integrate these two types of data, they need to be first converted into a unified data format. Specifically, the present invention performs preprocessing operations on the two types of collected data respectively, and uses a unified normalization method and spatial coordinate transformation to convert the original data into a data format with the same numerical range, projection method, and grid structure. The purpose of this preprocessing step is to eliminate systematic errors caused by differences in sensor characteristics, acquisition times, and projection parameters between different data sources, and ensure that the data in subsequent processing can be effectively docked at the same scale. Then, the present invention uses interpolation and mapping techniques to improve the spatial resolution of the preprocessed data respectively.

[0030] Interpolation techniques use mathematical methods to calculate the numerical transition between adjacent grids or pixels, thereby constructing data on a finer grid based on the original data. Mapping techniques, on the other hand, restore local details through non-linear transformations to ensure high-precision spatial matching between numerical model data and satellite image data. This step ultimately obtains refined forecast data and observational reference data, providing high-resolution and detail-rich basic data for the subsequent artificial intelligence correction process. Generally speaking, the overall design of the present invention is to construct refined meteorological forecast data from multi-source data through preprocessing, interpolation, and mapping steps, and based on this, carry out subsequent artificial intelligence processing processes such as local compensation, joint optimization, and closed-loop feedback to achieve efficient fusion between numerical forecasts and observational data and high-quality output of downscaling forecasts. From data collection, preprocessing to data refinement, each step of the present invention takes into account data formats, spatial resolutions, and numerical consistency issues, ensuring that the overall process is repeatable and stable in operation, providing an innovative and operable solution path for meteorological downscaling forecasts.

[0031] Preferably, the local compensation process uses a U-shaped network with an encoding and decoding structure. The U-shaped network uses continuous convolutional operations and downsampling to achieve feature extraction, and then restores details through upsampling and skip connections. A self-attention module is set in the network to enhance local feature responses.

[0032] In the process of processing refined forecast data, the local compensation process is a key link to achieve regional detail restoration. The present invention preferably uses a U-shaped network with an encoding and decoding structure as the core algorithm for local compensation. This network structure has symmetric encoding and decoding paths, which can extract multi-scale features during downsampling and restore image details during upsampling.

[0033] Specifically, the encoding path maps the input refined forecast data into a low-dimensional feature representation through continuous convolutional layers and downsampling operations, retaining the global information and low-frequency features of the data during this process; subsequently, the decoding path uses upsampling operations and convolutional layers to gradually restore the low-dimensional features to a feature map of the same size as the original data, and at the same time uses skip connections to directly transfer the high-resolution features in the encoding path to the decoding layer to make up for the possible loss of local details during the sampling process. To further enhance the network's response ability to local features, the present invention introduces a self-attention module into the U-shaped network. This module calculates the correlation between each position in the input feature map, assigns different weights to each local area, enabling the network to automatically focus on key areas when processing meteorological data with complex local changes, thereby achieving more refined local compensation. For example, for areas such as boundary layer development or local convective activities, the network can capture and amplify the subtle changes in these areas, and then generate local compensation data that conforms to actual observations.

[0034] In practical applications, the adaptive compensation for different meteorological elements can be achieved by adjusting the parameters of the convolution kernel size, number of layers, activation function, and attention module, ensuring that the generated local compensation data can accurately reflect the local physical process. This local compensation processing not only improves the detail expression ability of the forecast data but also provides accurate regional compensation information for subsequent data joint optimization and error correction during the process of integrating numerical model and satellite observation data. By combining the U-shaped network with the self-attention module, the present invention fully utilizes the multi-scale information of meteorological data in local compensation processing and uses a large amount of historical data for supervised learning during the training process to ensure the stability and applicability of the network parameters, thus providing a solid technical support for meteorological downscaling forecasting. For example, in actual downscaling forecasting, when the input data reflects the large-scale temperature field and local temperature anomaly regions, the network can extract the global temperature field information through encoding and restore the local temperature anomaly details during the decoding process, making the output local compensation data finely present the change of temperature gradient in space.

[0035] In addition, the U-shaped network structure adopted in the present invention has high generality and scalability and is applicable to local compensation processing of different regions and different meteorological elements. Its design principle and parameter settings can be adjusted according to specific application requirements. Generally speaking, this local compensation processing method effectively solves the problems of difficult capture of detail changes and insufficient regional adaptability in traditional local compensation methods by introducing a deep convolutional network, an encoding-decoding structure, and a self-attention mechanism, provides accurate and detailed local compensation data for subsequent data superposition and joint optimization, and significantly improves the performance and reliability of the entire downscaling forecasting system.

[0036] Preferably, the meta-learning technology adopts a model-agnostic meta-learning method to pre-train the parameters of the deep neural network on no less than three independent historical data sets and fine-tune the network parameters during the regional adaptation stage to generate local compensation data.

[0037] The present invention preferably uses a model-agnostic meta-learning method to pre-train a deep neural network to achieve the goal of regional adaptability compensation. The core idea of meta-learning technology is to enable the model to quickly adapt and update parameters through a small number of samples when facing new tasks, thus solving the problem of insufficient generalization ability of traditional methods under different regions and conditions. In the specific implementation process, first, multiple independent historical data sets are selected as training samples, and these data sets cover meteorological data under different regions and different climate conditions to ensure the diversity of training samples. Use these data sets to pre-train the deep neural network, adopt the Model-Agnostic Meta-Learning (abbreviated as MAML) algorithm, and in each training process, calculate the loss function of each task and obtain the gradient of each task parameter with respect to the overall loss to optimize the parameter initialization. Its basic process is as follows: for a given task, first calculate the task loss and update the network parameters using the gradient descent method; then, test the update effect on other tasks and globally adjust the initial parameters through backpropagation to obtain new parameters that can adapt to multiple tasks. Mathematically, let the initial parameter be , for task calculate the loss , then the parameter update formula for each task is

[0038] Next, calculate the overall loss on all tasks and update the initial parameters. The update formula is:

[0039] where and are the learning rates for the inner update and the outer update, respectively.

[0040] After several iterations, the obtained model parameters only need to make a small amount of gradient adjustment when encountering new regional data, and can quickly adapt to the new meteorological environment. By pre-training the network using the present invention, the network can capture general meteorological features and local compensation laws, so that in practical applications, by quickly fine-tuning the new regional data, accurate local compensation data can be generated. In terms of embodiments, it is assumed that pre-training is performed on the historical data sets of regions A, B, and C, and the data of each region respectively includes meteorological elements such as temperature, humidity, and wind speed. After updating the model parameters through the MAML algorithm, when performing local compensation processing on the forecast data of the new region D, the number of fine-tuning iterations required is significantly reduced, and the output local compensation data is more accurate in details. The effect of this technology is to significantly improve the adaptation speed and accuracy of the network to new tasks, and at the same time reduce the risk of model failure caused by differences in training samples under different climate conditions. By adopting the model-agnostic meta-learning technology, the present invention ensures that the local compensation processing can achieve consistent performance under different regional conditions, thus providing reliable local compensation data for subsequent data superposition and joint optimization, and ultimately improving the accuracy and robustness of the entire downscaling forecasting system.

[0041] In addition, the application of the meta-learning technology makes the initialization of the model parameters have stronger universality, and thus greatly reduces the time and computational resource consumption for large-scale training for each new region. In the data preparation, pre-training, regional adaptation, and fine-tuning and other links of the present invention, strict supervised learning and cross-validation strategies are adopted to ensure that the generated local compensation data has significant stability and accuracy statistically and meets the requirements of actual meteorological forecasting. In summary, the present invention effectively realizes the rapid generation of local compensation data by adopting the model-agnostic meta-learning method, fully pre-training the deep neural network, and fine-tuning the network parameters in the regional adaptation stage, provides dynamic and highly adaptable data support for subsequent downscaling processing, and greatly enhances the application ability and promotion potential of the system under multi-region and multi-climate conditions.

[0042] Preferably, as Figure 3 shown, the joint objective function is composed of the mean square error loss and the physical constraint loss. Among them, the physical constraint loss calculates the physical quantity deviation in the forecast data based on the principles of momentum conservation and energy conservation, and sets a fixed value as the weight coefficient to constrain the physical characteristics.

[0043] The joint objective function constructed by the present invention in the joint optimization process is composed of the mean square error loss and the physical constraint loss. Its design idea is to simultaneously consider the accuracy optimization driven by data and the rationality constraint of the physical model. The mean square error loss is used to measure the numerical difference between the preliminarily refined data and the real observed data, and its mathematical expression is:

[0044] where represents the true value of the th grid point, represents the model prediction value, is the total number of grid points. This loss function is often used in numerical weather prediction to quantify the prediction error. The physical constraint loss calculates the deviation of physical quantities in the forecast data based on the principles of momentum conservation and energy conservation. Its core lies in comparing the change trends of key physical quantities in the forecast data with the theoretical expectations, thereby introducing additional constraints into the objective function. The physical constraint loss is set as:

[0045] where represents the forecast value of the th physical quantity, represents the theoretical value derived based on the conservation principle, is the number of physical quantities. The combined objective function is:

[0046] where is the weight coefficient, and its value is determined through experiments within a specific range (e.g., 0.5 to 2) to balance the impacts of the mean squared error and the physical constraint loss. This design ensures that during the optimization process, the model not only pursues the minimization of data fitting but also satisfies the constraints of physical laws, thereby enabling the output forecast data to have both low numerical errors and conform to the theoretical physical characteristics.

[0047] In specific implementation, to calculate the physical constraint loss, first, it is necessary to determine the theoretical expectations of key physical quantities in the forecast data, such as temperature, humidity, wind speed, etc., which can be obtained through existing physical models or empirical formulas; subsequently, the difference between the actual forecast value and the theoretical value is taken, and the absolute values are summed to obtain the total physical constraint loss. For example, in temperature forecasting, the theoretical temperature distribution can be determined based on the principle of heat balance, and then the deviation between the actual temperature forecast and the theoretical value is calculated. The combined objective function calculates the gradients of each layer's parameters through automatic differentiation techniques during the training process, and then updates the model parameters to achieve the minimization of the overall objective.

[0048] The advantage of this method is that it can not only ensure high-precision fitting of the data-driven part but also guarantee that the output results conform to the basic physical laws, thereby improving the credibility and interpretability of the forecast results. In the embodiment, by adjusting the weight coefficient The value can flexibly balance the roles of the mean square error and physical constraints under different meteorological scenarios, enabling the model to neither deviate from physical laws due to excessive pursuit of numerical optimality nor affect the data fitting effect due to excessive reliance on physical constraints during local compensation and joint optimization. Generally speaking, the joint objective function constructed in the present invention provides a dual constraint mechanism for downscaling forecasting, which not only realizes the fine control of numerical errors but also ensures the reasonable expression of physical characteristics, thus significantly improving the overall performance and stability of the forecasting system in practical applications.

[0049] Preferably, the variational assimilation and joint optimization steps adopt the gradient descent algorithm. The gradient of the loss function with respect to the network parameters is calculated through automatic differentiation technology, and the parameters are updated according to the calculated gradient to achieve the optimization of the network parameters.

[0050] In the present invention, the variational assimilation and joint optimization steps are the key links to further refine the forecasting data and optimize the parameters. Preferably, the gradient descent algorithm is used as the optimization means. The gradients of the parameters in the joint objective function are calculated through automatic differentiation technology, and the parameters are updated according to the gradient direction. The basic idea of variational assimilation is to combine the observed data with the forecasting data. By constructing an objective function (i.e., the aforementioned joint objective function), the model achieves a balance between data fitting and physical constraints, so that the finally output forecasting data is both numerically and physically close to the real situation. In actual operation, variational assimilation usually adopts an iterative update strategy, that is, in each iteration, the gradient of the joint objective function is calculated according to the current model parameters, and after updating the parameters, the objective function is recalculated until the preset convergence criterion is reached. Mathematically, let the model parameters be , then the parameter update formula is:

[0051] where represents the learning rate, represents the gradient of the joint objective function with respect to the parameter at the -th iteration. This process uses an automatic differentiation framework to calculate the gradient, which can accurately obtain the sensitivity of each parameter in the complex network to the objective function, thereby guiding the update of the parameters. The joint optimization part, on the basis of variational assimilation, aims at the local errors and systematic biases existing in the forecasting data, and gradually reduces the error between the model output and the real observed data through continuous iteration.

[0052] Specifically, in each iteration, the forecast output is first calculated using the current parameters, then the current output is evaluated through a joint objective function composed of the forecast error and the physical constraint loss, and finally the model parameters are adjusted based on the gradient information obtained by automatic differentiation. Through multiple iterations, the present invention can gradually make the model parameters tend to the optimal state and finally obtain the jointly optimized forecast data. During the implementation process, the variational assimilation and joint optimization steps have strict settings for the learning rate, the number of iterations, and the convergence conditions, and the selection of these parameters is determined based on a large number of historical data experiments. For example, by performing cross-validation on the historical data set, the learning rate range most suitable for the current forecast scenario can be determined, so as to ensure that the optimization process will neither fall into a local optimal solution nor have a problem of too slow convergence speed. In addition, to ensure data consistency, the same numerical precision and time step are used for both the gradient calculation and parameter update in the joint optimization process to avoid model instability caused by calculation errors. In summary, the variational assimilation and joint optimization steps adopted in the present invention achieve the minimization of the joint objective function through gradient descent and automatic differentiation techniques. Under the dual effects of numerical errors and physical constraints, the fineness and accuracy of the forecast data are significantly improved, and a solid parameter update basis is provided for the subsequent closed-loop feedback iteration. Through this step, not only can the errors be further corrected on the basis of the preliminary refined data, but also the model parameters can be dynamically adjusted to ensure that the forecast data gradually approaches the true observation in each iteration, thereby continuously improving the performance of the overall downscaling forecast system.

[0053] Preferably, the generative adversarial network uses a convolutional autoencoder structure to form a generator, which extracts the error features in the jointly optimized forecast data through encoding and reconstructs the error compensation data through decoding, while the discriminator uses a convolutional neural network structure to distinguish between the generated error compensation data and the real data, so as to achieve error compensation under physical constraints.

[0054] In the present invention, to achieve error compensation and fine correction of forecast data, a generative adversarial network structure is preferably adopted. This network includes a generator composed of a convolutional autoencoder and a discriminator composed of a convolutional neural network. The design goal of the generator is to extract error features from the jointly optimized forecast data and reconstruct the error compensation data through the encoding and decoding processes, thereby locally correcting the forecast data. The generator first uses multiple convolutional layers in the encoder part to extract features from the input data. The encoder adopts continuous convolutional operations and downsampling operations to convert the original data into a low-dimensional feature representation, which can capture the global information and local variations in the forecast data. Subsequently, the decoder part gradually restores the data size through upsampling operations and convolutional layers, and uses skip connections to fuse the high-resolution information saved during the encoding process into the decoding process to ensure that the output data has rich detailed information. The output of the generator is the error compensation data, which reflects the local difference between the jointly optimized forecast data and the actual forecast. At the same time, the discriminator adopts a standard convolutional neural network structure to discriminate between the error compensation data output by the generator and the real data. The discriminator extracts features through multiple convolutional operations and pooling operations, and finally outputs a binary classification result through a fully connected layer, indicating whether the input data conforms to the physical constraint conditions. During the training process, the generator and the discriminator are alternately optimized. The generator tries to deceive the discriminator so that it cannot distinguish between the generated data and the real data, while the discriminator continuously improves the recognition accuracy. This adversarial training process is achieved by optimizing the objective function, where the goal of the generator is to minimize the gap between the generated data and the real data, and the goal of the discriminator is to maximize the discrimination ability between the two. In the specific mathematical expression, the loss function of the generator can be expressed as:

[0055] And the loss function of the discriminator is:

[0056] Where represents the output generated by the generator for the input , and represents the probability value judged by the discriminator for the input . Through the above adversarial process, the generator can gradually adjust the output data under physical constraints, so that the error compensation data is both numerically close to the real data and conforms to the physical characteristics of the forecast field. In practical applications, to ensure that the generative adversarial network works effectively under physical constraint conditions, in the present invention, the convolutional kernel size, number of layers, activation function, and optimization algorithm of the generator and the discriminator are carefully debugged and verified to ensure stable training under different meteorological scenarios.

[0057] Adopting the convolutional autoencoder structure not only has high - efficient feature compression ability, but also can effectively reconstruct details during the decoding process, enabling the error - compensation data generated by the generator to finely capture local deviations. The design of the discriminator ensures strict screening of the physical rationality of the generated data, thereby providing accurate feedback to the generator. Through this generative adversarial mechanism, the present invention realizes an effective transformation from data - driven to physical - constraint in the error - correction link of forecast data, significantly improving the overall accuracy and robustness of meteorological downscaling forecasts. For example, in a certain meteorological field forecast, the error - compensation data generated by the generative adversarial network successfully corrected the deviation in the local temperature anomaly area, making the matching degree between the final forecast result and the observed data significantly improved. Generally speaking, this generative adversarial network structure provides an innovative means for the present invention to achieve error compensation under physical constraints. Its operable details include network structure design, parameter setting, loss - function definition, and optimization of the training process, all of which provide strong support for achieving high - quality forecast data output.

[0058] Preferably, as Figure 4 shown, the graph neural network combined with reinforcement learning uses a graph convolutional network to aggregate the grid - node features of the forecast area division, and sets a function with the reduction of forecast error as the reward to train the correction weights of each node to achieve node - level adaptive correction.

[0059] In the present invention, to achieve node - level adaptive correction within the forecast area, a method of combining graph neural network and reinforcement learning is preferably adopted. Specifically, the present invention divides the forecast area into several grid nodes, and each node contains a multi - dimensional feature vector composed of jointly optimized forecast data and error features compensated by the generative adversarial network. The graph neural network aggregates the features between each node through graph convolutional operations, realizing the sharing and transmission of global information and local information. This graph convolutional network adopts the standard graph convolutional formula:

[0060] where represents the node - feature matrix of the th layer, is the adjacency matrix including self - connection, is the corresponding degree matrix, is the weight matrix of the th layer, As an activation function, this formula realizes the normalized weighted aggregation of node features. On this basis, a reward function is introduced in the reinforcement learning part to measure the degree of reduction in the prediction error of each node after correction. Specifically, the reward function is set as the reduction amplitude of the node prediction error, and the prediction error is obtained by comparing grid by grid with the observed reference data; the reinforcement learning algorithm adopts the policy gradient method to adjust the correction weight according to the node features, with the goal of minimizing the overall prediction error.

[0061] During the training process, each node obtains the corresponding reward according to the current state and correction weight, and the network continuously updates the policy parameters to maximize the expected reward. The core of this method is to use the graph neural network to capture the complex spatial correlation between nodes, and dynamically adjust the correction weights of each node through the reinforcement learning mechanism, so that the final node-level adaptive correction result can effectively reduce the overall prediction error. During the implementation process, the present invention details the number of layers of the graph convolutional network, the node feature dimension of each layer, and the construction method of the adjacency matrix. At the same time, for the reinforcement learning part, key parameters such as the form of the reward function, the learning rate, and the discount factor are set.

[0062] For example, in a certain prediction area, if the local error is large at some grid nodes, the system transmits the information of adjacent nodes through the graph neural network, and automatically adjusts the correction weights of these nodes in combination with the feedback of the reward function, so that the predicted data after correction is closer to the real observed data. The present invention not only improves the accuracy of node-level correction, but also realizes the dynamic optimization of the overall prediction data through the adaptive mechanism. Generally speaking, the combination of the graph neural network and reinforcement learning provides a new type of node-level adaptive correction method for the present invention. The principle is to use graph convolution to extract spatial features, and then optimize the correction strategy through reinforcement learning to achieve fine adjustment of the prediction error. Through the present invention, the system can automatically adjust the correction weights of each grid node in various meteorological scenarios, thereby significantly improving the accuracy and stability of the overall meteorological downscaling prediction. This technical solution has high operability. Its key steps include grid division, graph convolution operation, reward function setting, and policy update. Each step can be realized through existing deep learning and reinforcement learning platforms, providing reliable theoretical and technical support for practical applications.

[0063] As Figure 5 shown, the format of the corrected prediction data is sorted to form the final prediction data, the grid-level prediction error is calculated using the observed reference data, and the prediction error is converted into a feedback gradient, which is used to update the parameters of local compensation, variational assimilation, and adaptive correction, and update data is generated through closed-loop feedback iteration.

[0064] The closed-loop feedback iteration part of the present invention aims to achieve the dynamic update and optimization of the parameters of local compensation, variational assimilation, and adaptive correction by sorting out the format of the corrected and predicted data, calculating errors, and generating feedback gradients, and finally generate continuously improved updated prediction data.

[0065] Specifically, first, the corrected and predicted data obtained after various levels of processing is sorted out in format, and the data is unified into a standardized data format, which includes a unified spatial grid, time stamps, and standardized meteorological physical quantity values, providing a consistent data basis for subsequent error calculation; next, using the pre-collected observation benchmark data, which has high spatial and numerical accuracy after strict preprocessing, a grid-by-grid calculation method is used to compare the final prediction data with the observation benchmark data to obtain the prediction error of each grid cell; then, the calculated prediction error is converted into a feedback gradient through a specific algorithm. The feedback gradient is the key signal guiding the update of model parameters, which reflects the change law of the error distribution on the spatial grid and its sensitivity to the parameters of local compensation, variational assimilation, and adaptive correction modules; finally, using the closed-loop feedback mechanism, the parameters of each module are iteratively updated according to the feedback gradient, and the grid-level prediction error is gradually reduced through multiple feedback iterations, so that the entire system can continuously approach the real observation during the downscaling prediction process, and the finally output updated prediction data reaches an ideal state in terms of spatial resolution and detail restoration.

[0066] Generally speaking, this closed-loop feedback iteration process not only ensures error correction during the data fusion process, but also realizes system adaptive optimization through continuous parameter updates, thereby gradually improving the accuracy and stability of meteorological downscaling prediction in multiple iterations. The core advantage of the present invention is to use the feedback gradient to achieve error propagation and accurately adjust each processing module according to the error information in each iteration, so that the entire prediction system can achieve dynamic optimization from coarse to fine and from global to local while maintaining physical constraints and data consistency, thereby providing more accurate and reliable downscaling data support for actual meteorological prediction.

[0067] Preferably, the closed-loop feedback iteration step calculates the prediction error grid by grid using the observation benchmark data and the final prediction data, and converts the obtained prediction error into a feedback gradient, and the feedback gradient is used to continuously update the parameters of local compensation processing, variational assimilation, and adaptive correction, so that the grid-level prediction error is gradually reduced during the iteration process.

[0068] In the closed-loop feedback iteration process, calculating the forecast error grid by grid using the observed benchmark data and the final forecast data is a crucial step. The principle lies in making a detailed comparison between the forecast value and the true observed value within each grid cell to obtain the error distribution. When specifically implemented, first define the forecast error of each grid cell as the difference between the forecast value and the observed benchmark value within that grid, using the formula:

[0069] where, represents the forecast error of the grid cell located in row and column , is the value of the observed benchmark data for this grid cell, is the value in the final forecast data. To ensure the accuracy of error calculation, it is required to calculate each grid cell one by one to obtain a complete error matrix, which reflects the error distribution of the forecast data in each region of space. Next, convert the error matrix into a feedback gradient. This process uses the principle of the backpropagation algorithm to map each error value into gradient information, which describes the sensitivity of the current error to the local model parameters. During the specific calculation process, through the gradient function convert the error into the feedback gradient , and the feedback gradient can be expressed as:

[0070] where represents the loss function, They are the model parameters that affect the forecast output in the corresponding grid cells. When using this formula, the error of each grid cell will be conducted through the corresponding gradient function to generate a feedback gradient matrix, which will be used to guide the update of subsequent model parameters. In operation, a high-precision computing platform is used to batch process the error data, and the gradient is accurately calculated through automatic differentiation technology, making the feedback gradient have high numerical stability. For example, when the forecast error of the grid cells in a certain area is large, the corresponding feedback gradient will also be large, indicating that the model parameters in this area need to be significantly adjusted; conversely, the feedback gradient in the area with smaller errors is also correspondingly smaller, indicating that the model has converged relatively well in this area. This process ensures that the error information can be accurately transmitted to each parameter update module, providing a specific adjustment direction and quantitative basis for subsequent local compensation, variational assimilation, and adaptive correction modules. Through this error conversion mechanism, not only the fine measurement of the global forecast error is realized, but also an operable gradient signal is provided for local parameter adjustment, thereby promoting the effective progress of the entire closed-loop feedback iteration process, gradually reducing the forecast error and improving the overall forecast quality. This step requires strict control of data resolution, numerical accuracy, and gradient calculation method in practical applications to ensure the accuracy and efficiency of the error conversion process, thus laying a solid foundation for improving the overall performance of the system.

[0071] In the closed-loop feedback iteration process, the parameter update mechanism is the core link to achieve the adaptive optimization of the model. In the specific operation process, based on the feedback gradient obtained by the aforementioned conversion of the forecast error, each model parameter is updated according to the gradient descent algorithm to ensure that the forecast error can be effectively reduced in the next iteration. The parameter update formula usually adopts the standard gradient descent formula, that is:

[0072] where represents the parameter of the model at the th iteration, is the preset learning rate, is the gradient of the loss function with respect to the parameter calculated through the feedback gradient in the current iteration. The application of this formula ensures that each parameter update is in the direction of reducing the overall forecast error. To implement this update process, the present invention uses automatic differentiation technology to calculate the gradient of the joint objective function, thereby accurately obtaining the feedback gradient matrix, which not only reflects the influence of the errors of each grid cell on the local parameters, but also takes into account the manifestation of physical constraint factors in the forecast data. In operation, by setting a reasonable learning rate parameter the amplitude of parameter update can be balanced, avoiding excessive parameter fluctuations caused by too large a learning rate or too slow a convergence speed due to too small a learning rate.

[0073] In actual implementation, batch update or mini-batch stochastic gradient descent strategy can be adopted for parameter update, so as to improve the iteration efficiency while ensuring the update stability. For example, in practical applications, when the feedback gradient of the forecast error in a certain area is large, the local compensation parameters corresponding to this area will be updated significantly to quickly correct the forecast deviation; while when the error feedback gradient is small, the parameter update is relatively small to ensure the best balance between local and global forecast data. This update mechanism not only ensures that each closed-loop feedback can effectively reduce the forecast error, but also continuously optimizes the entire downscaling forecast system through continuous adjustment of model parameters, gradually achieving fine capture of complex meteorological field data. The present invention can also be combined with the adaptive learning rate adjustment technology to dynamically adjust the learning rate parameter according to the change trend of the feedback gradient, thereby further improving the stability and convergence speed of the update process. Generally speaking, the parameter update mechanism realizes the continuous optimization of the parameters of the local compensation, variational assimilation and adaptive correction modules through the strict gradient descent algorithm and automatic differentiation technology, provides strong technical support for the improvement of the overall forecast performance of the system, and ensures the continuous reduction of the forecast error during multiple iterations.

[0074] The core of the closed-loop feedback iteration process is to continuously adjust the model parameters through multiple iterations, so that the error between the forecast data and the observed reference data gradually shrinks. For this reason, the present invention stipulates strict iteration strategies and convergence criteria. Specifically, the closed-loop feedback iteration process starts from the initial feedback gradient calculation, regenerates the updated forecast data after each parameter update, then calculates the new forecast error for the updated forecast data grid by grid using the observed reference data, and converts the new error into a new feedback gradient, and so on in a cycle. Mathematically, this process can be described as: Let the initial forecast error be , and the error corresponding to the -th forecast data generated by parameter update is , and the feedback gradient is . When the following conditions are met: or , the iteration process is considered to converge, where is the preset convergence threshold. This iteration process realizes continuous error gradient calculation and parameter update through automatic differentiation technology, and records the error change trend in each iteration to ensure that the system parameters converge near the global optimal solution. In practical applications, to improve the iteration efficiency, an early stopping strategy can be adopted, that is, stop the iteration in advance when the error change is lower than the preset threshold in several consecutive iterations to save computing resources.

[0075] In an embodiment, for example, in a meteorological downscaling forecast, after 10 parameter updates through closed-loop feedback iteration, the grid-level error is reduced from an initial average of 5 degrees Celsius to below 0.5 degrees Celsius, meeting the preset convergence condition. This process not only reflects the trend of gradually decreasing error, but also ensures the effective correction of both local areas and the overall forecast through the dynamic propagation of feedback gradients across the entire forecast area. During the iteration process, the parameter updates at each link are interrelated, and the parameter adjustment of the local compensation module directly affects the effects of variational assimilation and adaptive correction, thus forming a self-adaptive optimization closed-loop system. The entire process needs to be implemented in combination with a high-performance computing platform, and strict time step and numerical precision control are adopted to ensure data consistency and calculation accuracy during multiple iterations. Through this closed-loop feedback iteration mechanism, the forecast system can continuously self-adjust and optimize, so as to maintain a high forecast accuracy and stability when facing a complex and changing meteorological field, and achieve an organic balance between data-driven and physical constraints.

[0076] In the actual implementation of the present invention, the continuous update of forecast data and the continuous improvement of system performance are achieved through closed-loop feedback iteration, and its effects have been verified in multiple meteorological field forecast cases. Specifically, during implementation, the system first formats the initially generated corrected forecast data to ensure that the data has a unified spatial resolution, time identifier, and numerical unit, and then calculates the error for each grid cell using the preprocessed observation benchmark data to obtain a refined forecast error matrix. Through automatic differentiation technology, the error matrix is converted into a feedback gradient matrix, and each value in this matrix directly reflects the sensitivity of the error of the corresponding grid cell to the model parameters.

[0077] In one embodiment, the average value of the initial forecast error in a certain area is 4 degrees Celsius. Through closed-loop feedback iteration, the average error value gradually decreases after each iteration; after the 5th iteration, the average error value drops to 1.2 degrees Celsius, and after the 10th iteration, the average error is reduced to 0.6 degrees Celsius, reaching the preset convergence standard. The parameter updates during this iteration process not only improve the forecast accuracy globally, but also specifically correct the forecast anomalies in local areas, resulting in a significant reduction in local errors. To achieve this goal, the system adopts a batch parameter update and adaptive learning rate adjustment strategy during the iteration process, records the parameter update situation and the error reduction amplitude for each iteration, and adjusts the learning rate according to the error change to ensure that the parameter update is neither too aggressive nor too slow. In addition, to ensure the accuracy of feedback gradient calculation, the system uses high-precision floating-point calculations and introduces an error smoothing algorithm during the feedback process to avoid parameter fluctuations caused by numerical instability. This closed-loop feedback iteration mechanism shows high robustness and adaptability in practical applications. Even in the face of sudden meteorological events or large data noise, the system can still effectively correct the forecast data through continuous iteration and improve the reliability of the forecast results.

[0078] For example, in a weather forecast involving enhanced local convective activities, due to the drastic changes in local temperature and humidity data, the initial forecast error was large. However, through closed-loop feedback iteration, the system completed parameter adjustment in a short time and achieved timely correction of the local anomaly forecast. Generally speaking, this embodiment proves the effectiveness of the closed-loop feedback iteration mechanism in improving the accuracy of meteorological downscaling forecasts. Its operable details cover error calculation, feedback gradient generation, parameter update, and iteration convergence strategies, providing a detailed and replicable optimization process for the actual forecast system and significantly enhancing the adaptability and forecast stability of the system in complex meteorological environments.

[0079] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. 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.

[0080] The above are only the embodiments of the present application and are not intended 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 meteorological downscaling method that integrates numerical weather prediction and AI weather prediction, characterized in that, It includes the following steps: Collect the forecast data output by the numerical model and the satellite observation data within the corresponding forecast area, and perform preprocessing on the data to unify the data format. Use interpolation and mapping techniques to obtain refined forecast data and observation reference data respectively; Perform local compensation processing on the refined forecast data based on a deep neural network. Through meta-learning technology, pre-train on historical data and then achieve regional adaptive compensation to generate local compensation data. Add the refined forecast data and local compensation data point by point to form preliminary refined data. At the same time, construct a joint objective function that includes mean square error and physical constraints, perform variational assimilation and joint optimization on the preliminary refined data, use a generative adversarial network to perform error compensation under physical constraints, and use a graph neural network combined with reinforcement learning to achieve node-level adaptive correction to obtain corrected forecast data; Organize the format of the corrected forecast data to form the final forecast data. Calculate the grid-level forecast error using the observation reference data, and convert the forecast error into a feedback gradient. The feedback gradient is used to update the parameters of local compensation, variational assimilation, and adaptive correction, and generate updated data through closed-loop feedback iteration.

2. The method according to claim 1, characterized in that The preprocessing operation uses the Z-score normalization method. Standardize each data item by calculating the mean and standard deviation of the forecast data and satellite observation data respectively, and convert the spatial coordinates of the two types of data into a unified standard coordinate system.

3. The method according to claim 1, characterized in that, The interpolation and mapping techniques use bilinear interpolation to calculate the weighted average of adjacent pixel values in the forecast data and satellite observation data, and combine a non-linear mapping algorithm based on convolution operation to improve the spatial resolution of the forecast data and satellite observation data respectively, so as to obtain refined forecast data and observation reference data respectively.

4. The method according to claim 1, characterized in that The local compensation processing uses a U-shaped network with an encoding and decoding structure. The U-shaped network uses continuous convolution operations and downsampling to extract features, and then restores details through upsampling and skip connections. A self-attention module is set in the network to enhance local feature responses.

5. The method according to claim 1, characterized in that, The meta-learning technology uses the model-agnostic meta-learning method to pre-train the parameters of the deep neural network on no less than three independent historical data sets, and fine-tune the network parameters in the regional adaptation stage to generate local compensation data.

6. The method according to claim 1, wherein The joint objective function consists of mean square error loss and physical constraint loss. Among them, the physical constraint loss calculates the deviation of physical quantities in the forecast data based on the principles of conservation of momentum and conservation of energy, and sets a fixed value as the weight coefficient to constrain physical characteristics.

7. The method according to claim 1, wherein The variational assimilation and joint optimization steps use the gradient descent algorithm. Calculate the gradient of the loss function with respect to the network parameters through automatic differentiation technology, and update the parameters according to the calculated gradient to achieve the optimization of the network parameters.

8. The method according to claim 1, characterized in that, The generative adversarial network uses a convolutional autoencoder structure to form a generator. Extract the error features in the jointly optimized forecast data through encoding and reconstruct the error compensation data through decoding. The discriminator uses a convolutional neural network structure to distinguish between the generated error compensation data and the real data, so as to achieve error compensation under physical constraints.

9. The method according to claim 1, wherein The graph neural network combines reinforcement learning to aggregate the grid node features of the forecast area division using a graph convolutional network, and sets a function with the reduction of the forecast error as the reward to train the correction weights of each node to achieve node-level adaptive correction.

10. The method according to claim 1, characterized in that The closed-loop feedback iteration step calculates the forecast error grid by grid using the observed benchmark data and the final forecast data, and converts the obtained forecast error into a feedback gradient, which is used to continuously update the parameters of local compensation processing, variational assimilation, and adaptive correction, so that the grid-level forecast error is gradually reduced during the iteration process.

Images