Meteorological downscaling method integrating numerical weather forecast and AI weather forecast

Abstract

The present invention relates to the field of meteorological forecasting technology, and in particular to a meteorological downscaling method that integrates numerical weather forecasting and AI weather forecasting. By collecting numerical weather forecast data and satellite remote sensing observation data, refined forecast data and observation benchmark data are generated through unified preprocessing, interpolation and mapping. Subsequently, local compensation based on deep neural networks and model-independent meta-learning are used to perform regional adaptive compensation, and variational assimilation and joint optimization are implemented by constructing a joint objective function. At the same time, generative adversarial networks and graph neural networks are combined with reinforcement learning to achieve error compensation and node-level adaptive correction. Finally, parameters are continuously updated through closed-loop feedback iteration to generate updated forecast data. The present invention effectively solves the problems of insufficient local details and poor model adaptability in the downscaling process of the prior art, and improves the forecast accuracy and robustness.

Description

Technical Field

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

[0002] Meteorological downscaling technology is crucial for improving forecast spatial detail and capturing local meteorological phenomena. It can provide more refined meteorological information for practical applications such as power grid scheduling and disaster warning. Existing technologies primarily employ traditional linear interpolation downscaling methods (Chinese Invention Patent Publication No. CN118981060A, Title: A Meteorological Forecasting Method Integrating Downscaling and Multi-Model Integration), combined with Kalman filtering to correct numerical forecast results and radial basis function neural networks to integrate multiple forecast results. However, due to the coarse numerical model grids, linear interpolation cannot fully recover local nonlinear characteristics. Furthermore, the Kalman filter and traditional neural networks lack adaptability to complex meteorological fields, resulting in significant deficiencies in forecast accuracy, local detail representation, and model robustness. Furthermore, existing downscaling and multi-model integration processes are often one-way processes lacking dynamic feedback mechanisms, making it difficult to achieve real-time correction of forecast errors and adaptive updating of model parameters. This limits forecast performance when faced with complex nonlinear atmospheric motions. Summary of the Invention

[0003] To address the numerous issues with the aforementioned existing technologies, the present invention provides a meteorological downscaling method that integrates numerical weather forecasting and AI-powered weather forecasting. This method utilizes unified preprocessing, local compensation using deep neural networks, meta-learning, variational assimilation, and generative adversarial networks to achieve closed-loop data feedback and adaptive parameter updates. This method progressively reduces forecast error through multiple iterations, significantly improving the spatial detail and physical plausibility of the downscaled forecast results.

[0004] A meteorological downscaling method integrating numerical weather forecast and AI weather forecast includes the following steps:

[0005] Collect forecast data output by the numerical model and satellite observation data within the corresponding forecast area, preprocess the data to unify the data format, and use interpolation and mapping techniques to obtain refined forecast data and observation benchmark data respectively;

[0006] The refined forecast data is processed with local compensation based on a deep neural network. 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 including mean square error and physical constraints is constructed. Variational assimilation and joint optimization are performed on the preliminary refined data. Error compensation is performed under physical constraints using a generative adversarial network. Graph neural networks combined with reinforcement learning are used to implement node-level adaptive correction to obtain corrected forecast data.

[0007] The corrected forecast data are formatted to form final forecast data, the grid-level forecast error is calculated using the observational benchmark data, and the forecast error is converted into a feedback gradient. The feedback gradient is used to update the parameters of local compensation, variational assimilation, and adaptive correction, and updated data is generated through closed-loop feedback iteration.

[0008] Preferably, the preprocessing operation adopts the Z-score normalization method, which standardizes each data item by calculating the mean value and standard deviation of the forecast data and the satellite observation data respectively, and converts the spatial coordinates of the two types of data into a unified standard coordinate system.

[0009] Preferably, the interpolation and mapping technology uses a bilinear interpolation method to calculate the weighted average of the adjacent pixel values ​​in the forecast data and satellite observation data, and combines a nonlinear mapping algorithm based on convolution operation to respectively enhance the spatial resolution of the forecast data and satellite observation data, thereby obtaining refined forecast data and observation benchmark data, respectively.

[0010] Preferably, the local compensation processing adopts a U-type network with an encoding and decoding structure, which uses continuous convolution operations and downsampling to achieve feature extraction, and then restores details through upsampling and jump connections, and sets a self-attention module in the network to enhance local feature response.

[0011] Preferably, the meta-learning technique adopts a model-independent meta-learning method to pre-train the deep neural network parameters 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.

[0012] Preferably, the joint objective function is composed of mean square error loss and physical constraint loss, wherein 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 a weight coefficient to constrain the physical characteristics.

[0013] Preferably, the variational assimilation and joint optimization steps adopt a 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 optimization of the network parameters.

[0014] Preferably, the generative adversarial network adopts a convolutional autoencoder structure to form a generator, extracts the error features in the joint optimization 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, thereby realizing error compensation under physical constraints.

[0015] Preferably, the graph neural network is combined with reinforcement learning to use a graph convolutional network to aggregate the grid node features of the forecast area division, and a function with a reward of reduced forecast error is set to train the correction weight of each node to achieve node-level adaptive correction.

[0016] Preferably, the closed-loop feedback iteration step calculates the forecast error grid by grid by using the observation baseline data and the final forecast data, and converts the obtained forecast error into a feedback gradient. The feedback gradient is used to continuously update the parameters of each link of local compensation processing, variational assimilation and adaptive correction, so that the grid-level forecast error is gradually reduced during the iteration process.

[0017] Compared with the prior art, the advantages and beneficial effects of the present invention are:

[0018] By adopting local compensation processing of deep neural networks, model-independent meta-learning technology, variational assimilation joint optimization, generative adversarial networks, and graph neural networks combined with reinforcement learning, the effect of closed-loop feedback iterative optimization of forecast data is achieved;

[0019] By constructing a deep neural network with an encoder-decoder structure and a self-attention mechanism, local compensation is performed on the downscaled forecast data. Meta-learning is used to pre-train network parameters on historical data from multiple regions, allowing the model to quickly adapt to new regions.

[0020] By constructing a joint objective function that includes mean square error and physical constraints, variational assimilation and gradient descent are used to achieve continuous update of network parameters;

[0021] Generative adversarial networks are used to compensate for errors in the joint optimization data under physical constraints. Graph convolutional networks and reinforcement learning are then combined to adaptively correct node-level features within the forecast area, effectively reducing global and local forecast errors.

[0022] Through the above-mentioned technical means, the present invention not only solves the problems of insufficient detail recovery of downscaled data and poor model adaptability in the existing technology, but also constructs a feedback iteration mechanism, so that the forecast data gradually approaches the real observation during continuous updating, thereby improving the overall forecast accuracy and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 Schematic diagram of the process of the present invention;

[0024] Figure 2 Schematic diagram of local compensation and meta-learning in the present invention;

[0025] Figure 3 Schematic diagram of the combined objective function and optimization in the present invention;

[0026] Figure 4 Schematic diagram of adaptive correction of graph neural network and reinforcement learning in the present invention;

[0027] Figure 5 Schematic diagram of feedback iteration and parameter update in the present invention. DETAILED DESCRIPTION

[0028] 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 ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure. However, it is apparent that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessary confusion of the concepts of the present disclosure.

[0029] The terms used herein are only for describing specific embodiments and are not intended to limit the present disclosure. The terms "comprise," "include," etc. used herein indicate the presence of the features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0030] 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.

[0031] like Figure 1 As shown, a meteorological downscaling method integrating numerical weather forecast and AI weather forecast includes the following steps:

[0032] Collect forecast data output by the numerical model and satellite observation data within the corresponding forecast area, preprocess the data to unify the data format, and use interpolation and mapping techniques to obtain refined forecast data and observation benchmark data respectively;

[0033] The present invention aims to unify multi-source data and improve spatial resolution in the meteorological downscaling process by integrating numerical weather forecasting and artificial intelligence weather forecasting technology, thereby providing high-quality basic data for subsequent meteorological forecast correction and fusion. To this end, the forecast data output by the numerical model is first collected. These data are obtained by the numerical weather forecast model through physical process calculations and usually include key meteorological elements such as temperature, humidity, wind speed, and precipitation. Their spatial resolution is limited by the model grid size and it is difficult to reflect local details. At the same time, satellite remote sensing technology is used to collect satellite image data in the forecast area. These image data can provide real-time information on the surface and atmospheric conditions and have high detail resolution in space. Since the two data sources differ in acquisition method, output format, spatial projection and resolution, they must be preprocessed to achieve consistency in data format, numerical scale and spatial coordinates.

[0034] The preprocessing operation eliminates the dimensional and systematic errors between the forecast data and satellite observation data by performing unified normalization and coordinate transformation on the two types of data, and lays the foundation for subsequent data fusion. Furthermore, by adopting interpolation and mapping techniques, the spatial resolution of the preprocessed data is improved. The interpolation technique can smoothly transition the coarse-resolution data in space, while the mapping technique uses nonlinear transformation to restore local details, thereby obtaining refined forecast data and observation benchmark data respectively. This data processing flow not only solves the problems of inconsistent formats and resolution mismatches between data sources, but also provides a unified and reliable input for local correction, model joint optimization, and closed-loop feedback iteration in subsequent meteorological forecasts. The present invention can generally ensure efficient data docking and fusion in various links such as data acquisition, preprocessing, spatial interpolation, and mapping, ensuring that the numerical model and satellite observation data are subsequently processed at the same spatial scale and data format, thereby improving the accuracy and stability of meteorological downscaling forecasts. Overall, the present invention implements standardized and unified preprocessing of the collected data, and then uses interpolation and mapping techniques to improve the resolution, effectively eliminating the differences between data sources, and providing the artificial intelligence processing module with refined forecast data and observation benchmark data that match the actual observation data, ensuring that subsequent fusion and correction algorithms are carried out at a unified scale, thereby achieving the goal of collaboratively improving the resolution of meteorological forecasts through multi-source data.

[0035] Preferably, the preprocessing operation adopts the Z-score normalization method, which standardizes each data item by calculating the mean value and standard deviation of the forecast data and the satellite observation data respectively, and converts the spatial coordinates of the two types of data into a unified standard coordinate system.

[0036] In terms of data preprocessing, the present invention preferably uses the Z-score normalization method to standardize the collected numerical model forecast data and satellite image data. In the specific operation process, firstly, the numerical values ​​in the forecast data are statistically analyzed and the average value of all data items (denoted as ) and standard deviation (denoted as ), then follow the formula:

[0037]

[0038] The original value Convert to normalized values ,in Represents the normalized data value. This method ensures that the normalized forecast data exhibits a standard normal distribution, eliminating numerical bias caused by factors such as sensor characteristics, data acquisition time, and measurement errors. Satellite imagery data is also normalized using the same statistical metrics: the mean and standard deviation of each pixel in each image are calculated, and the normalization formula is applied to all pixel data, converting both data types into standardized data. This operation not only ensures consistency in the numerical range of forecast and observation data but also provides a unified scale for subsequent data fusion. Furthermore, preprocessing involves converting the spatial coordinates of both data types into a predefined standard coordinate system, such as WGS84 or Universal Transverse Mercator. During this spatial coordinate conversion, a projection transformation algorithm is used to recalculate the latitude and longitude, or projected coordinates, of each data point based on the projection information of the original data and the parameters of the target coordinate system, ensuring a one-to-one correspondence between the forecast data and the satellite observation data. In practical applications, in order to improve processing efficiency and data consistency, the above-mentioned normalization and coordinate conversion operations can be automated by 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 adopted 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.

[0039] Preferably, the interpolation and mapping technology uses a bilinear interpolation method to calculate the weighted average of the adjacent pixel values ​​in the forecast data and satellite observation data, and combines a nonlinear mapping algorithm based on convolution operation to respectively enhance the spatial resolution of the forecast data and satellite observation data, thereby obtaining refined forecast data and observation benchmark data, respectively.

[0040] 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 let the upper left, upper right, lower left, and lower right data values ​​of the rectangle be 、 、 and , the horizontal distance of the target point relative to the upper left corner is , the longitudinal distance is , then the target point value It can be calculated by the following formula:

[0041]

[0042] in, and Both are between 0 and 1, reflecting the position ratio of the target point in the horizontal and vertical directions. Through this method, low-resolution data can be smoothly interpolated in space, thereby preliminarily improving the resolution. For forecast data and satellite image data, after bilinear interpolation processing, preliminary refined data with higher grid density can be generated. Next, the present invention further uses a nonlinear mapping algorithm based on convolution operation to refine the data after interpolation processing to restore local features and detail information in the data. The nonlinear mapping algorithm usually adopts a deep convolutional network structure, and performs nonlinear transformation on the data through multi-layer convolution operations and activation functions (such as ReLU function). In specific operations, the image data processed by bilinear interpolation is first used as the network input, and the local spatial features are extracted through continuous convolution layers. The linear combination is then mapped to a nonlinear output through the activation function. Finally, after several layers of convolution processing, the output image data with further improved resolution and richer details is output.

[0043] For example, the convolution layer can set the convolution kernel size to 3×3 or 5×5, and the stride and padding parameters can be adjusted according to the actual data characteristics; the selection of the activation function and the number of layers will be optimized according to the training samples and the expected output. After training, the convolution network can accurately capture the texture, edge and local change information in the data, thereby achieving refinement processing. 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 nonlinear mapping network structure are used for the two types of data, thereby ensuring that the refined forecast data and the observation benchmark data can correspond one to one in terms of grid division, spatial distribution and numerical range. The use of the above-mentioned interpolation and mapping technology can not only effectively improve the spatial resolution of the data, but also restore the local features lost in the low-resolution data due to the approximation of the physical model, providing refined input data for subsequent artificial intelligence correction and data fusion. In summary, the present invention achieves resolution improvement and detail restoration of forecast data and satellite image data through a method that combines bilinear interpolation with nonlinear mapping of convolutional networks, 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.

[0044] like Figure 2 As shown, the refined forecast data is processed with local compensation based on a deep neural network. 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 including mean square error and physical constraints is constructed. Variational assimilation and joint optimization are performed on the preliminary refined data. Error compensation is performed under physical constraints using a generative adversarial network. Graph neural networks combined with reinforcement learning are used to implement node-level adaptive correction to obtain corrected forecast data.

[0045] The present invention aims to achieve a deep fusion of numerical weather forecasting and artificial intelligence weather forecasting in meteorological downscaling. The core goal is to improve the spatial resolution and detail restoration capability of forecast data, thereby providing high-quality input data for subsequent meteorological forecast correction and data fusion. In order to achieve the above-mentioned goals, the present invention first performs unified collection and preprocessing operations on the forecast data output by the numerical model and the satellite remote sensing image data collected in the forecast area. The collected forecast data comes from the numerical weather forecast model, and its output data reflects a variety of key meteorological elements such as temperature, humidity, wind speed, and precipitation. However, since the numerical model calculation uses a larger grid, its spatial resolution is relatively low, which is difficult to meet the requirements of capturing local meteorological characteristics; while satellite image data directly reflects the state of the atmosphere and the surface, has a higher resolution, and can provide local detail information.

[0046] In order to achieve effective fusion of these two types of data, they need to be converted into a unified data format first. Specifically, the present invention performs preprocessing operations on the two types of collected data respectively, adopts a unified normalization method and spatial coordinate conversion, and converts the original data into a data format with consistent numerical range, projection method and grid structure. The purpose of this preprocessing step is to eliminate the systematic errors caused by differences in sensor characteristics, acquisition time and projection parameters between different data sources, and ensure that the data in subsequent processing can be effectively connected at the same scale. Next, the present invention uses interpolation and mapping techniques to improve the spatial resolution of the preprocessed data.

[0047] Interpolation technology uses mathematical methods to calculate the numerical transition between adjacent grids or pixels, thereby constructing finer grid data based on the original data. Mapping technology restores local details through nonlinear transformations to ensure that numerical model data and satellite image data are spatially matched with high precision. This step ultimately obtains refined forecast data and observation benchmark data, providing high-resolution, detailed basic data for subsequent artificial intelligence correction links. In general, 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 forecast and observation data and high-quality output of downscaled forecast. From data acquisition, preprocessing to data refinement, the present invention considers data format, spatial resolution and numerical consistency issues at each step to ensure that the overall process is repeatable and stable in operation, providing an innovative and operational solution for meteorological downscaling forecasting.

[0048] Preferably, the local compensation processing adopts a U-type network with an encoding and decoding structure, which uses continuous convolution operations and downsampling to achieve feature extraction, and then restores details through upsampling and jump connections, and sets a self-attention module in the network to enhance local feature response.

[0049] During the refinement of forecast data, local compensation is a key step in restoring regional details. The present invention preferably employs a U-shaped network with an encoding and decoding structure as the core algorithm for local compensation. This network structure features symmetrical encoding and decoding paths, enabling the extraction of multi-scale features during downsampling and the restoration of image details during upsampling.

[0050] In specific implementation, the encoding path maps the input refined forecast data into a low-dimensional feature representation through continuous convolutional layers and downsampling operations, which retains the global information and low-frequency features of the data; then, the decoding path uses upsampling operations and convolutional layers to gradually restore the low-dimensional features to feature maps of the same size as the original data, and at the same time uses jump connections to pass the high-resolution features in the encoding path directly to the decoding layer to compensate for the local details that may be lost during the sampling process. In order to further enhance the network's responsiveness 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 and assigns different weights to each local area, so that the network can 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 activity, the network can capture and amplify subtle changes in these areas, thereby generating local compensation data consistent with actual observations.

[0051] In practical applications, adaptive compensation for different meteorological elements can be achieved by adjusting the convolution kernel size, number of layers, activation function and parameters of the attention module, ensuring that the generated local compensation data can accurately reflect local physical processes. This local compensation processing not only improves the ability to express details of the forecast data, but also provides accurate regional compensation information for subsequent data joint optimization and error correction in the process of fusing numerical models and satellite observation data. By combining the U-shaped network with the self-attention module, the present invention realizes the full utilization of 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 network parameters, thereby providing solid technical support for meteorological downscaling forecasting. For example, in actual downscaling forecasting, when the input data reflects a large-scale temperature field and a local temperature anomaly area, the network can extract the global temperature field information through encoding and restore the local temperature anomaly details during the decoding process, so that the output local compensation data can finely present the changes in temperature gradient in space.

[0052] Furthermore, the U-shaped network structure employed in this invention is highly versatile and scalable, making it suitable for local compensation processing in different regions and for different meteorological elements. Its design principles and parameter settings can be adjusted according to specific application requirements. Overall, this local compensation processing method effectively addresses the difficulties of traditional local compensation methods in capturing detailed changes and lacking regional adaptability by introducing a deep convolutional network, an encoding-decoding structure, and a self-attention mechanism. This provides accurate and detailed local compensation data for subsequent data overlay and joint optimization, significantly improving the performance and reliability of the entire downscaling forecasting system.

[0053] Preferably, the meta-learning technique adopts a model-independent meta-learning method to pre-train the deep neural network parameters 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.

[0054] The present invention preferably adopts a model-independent meta-learning method to pre-train the deep neural network to achieve the goal of regional adaptive 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, thereby solving the problem of insufficient generalization ability of traditional methods in different regions and conditions. In the specific implementation process, multiple independent historical data sets are first selected as training samples. These data sets cover meteorological data in different regions and different climatic conditions to ensure the diversity of training samples. These data sets are used to pre-train the deep neural network, and the model-agnostic meta-learning algorithm (Model-Agnostic Meta-Learning, referred to as MAML) is adopted. In each training process, the loss function of each task is calculated, and the gradient of each task parameter to the overall loss is obtained to optimize the parameter initialization. The basic process is: for a given task, first calculate the task loss, and use the gradient descent method to update the network parameters; then, test the update effect on other tasks, and adjust the initial parameters as a whole through back propagation to obtain new parameters that can adapt to multiple tasks. Mathematically, let the initial parameters be , for the task Calculating losses , then the parameter update formula for each task is

[0055]

[0056] Next, the overall loss is calculated on all tasks and the initial parameters are updated. The update formula is:

[0057]

[0058] in, and are the learning rates for internal and external updates, respectively.

[0059] After several iterations, the obtained model parameters only need a small amount of gradient adjustment when encountering new regional data to quickly adapt to the new meteorological environment. The use of the present invention to pre-train the network enables the network to capture common meteorological characteristics and local compensation laws, so that in practical applications, accurate local compensation data can be generated by quickly fine-tuning the new regional data. In terms of embodiments, assuming that pre-training is performed on historical data sets of regions A, B, and C, each of which contains meteorological elements such as temperature, humidity, and wind speed, after updating the model parameters through the MAML algorithm, when local compensation processing is performed 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 that it significantly improves the speed and accuracy of the network's adaptation to new tasks, while reducing the risk of model failure caused by differences in training samples under different climatic conditions. By adopting model-independent meta-learning technology, the present invention ensures that local compensation processing can achieve consistent performance under different regional conditions, thereby providing reliable local compensation data for subsequent data superposition and joint optimization, and ultimately improving the accuracy and robustness of the entire downscaling forecast system.

[0060] In addition, the application of meta-learning technology makes the initialization of model parameters more universal, thereby greatly reducing the time and computing resource consumption for large-scale training for each new area. The present invention adopts strict supervised learning and cross-validation strategies in data preparation, pre-training, regional adaptation and fine-tuning, etc., to ensure that the generated local compensation data has statistically significant stability and accuracy, meeting the requirements of actual weather forecasts. In summary, the present invention adopts a model-independent meta-learning method to fully pre-train the deep neural network and fine-tune the network parameters in the regional adaptation stage, effectively realizing the rapid generation of local compensation data, providing dynamic and adaptable data support for subsequent downscaling processing, and greatly enhancing the application ability and promotion potential of the system in multiple regions and multiple climate conditions.

[0061] Preferably, Figure 3 As shown in the figure, 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 conservation of momentum and energy, and sets a fixed value as the weight coefficient to constrain the physical characteristics.

[0062] The joint objective function constructed in the joint optimization process of this invention consists of a mean square error loss and a physical constraint loss. Its design concept is to simultaneously take into account data-driven accuracy optimization and the rationality constraints of the physical model. The mean square error loss is used to measure the numerical difference between the preliminary refined data and the actual observed data. Its mathematical expression is:

[0063]

[0064] in Indicates the The true value of the grid points, represents the model prediction value, is the total number of grid points. This loss function is often used in numerical weather forecasting to quantify forecast errors. The physical constraint loss calculates the deviation of physical quantities in the forecast data based on the principles of conservation of momentum and energy. Its core is to compare the changing trends of key physical quantities in the forecast data with theoretical expectations, thereby introducing additional constraints in the objective function. The physical constraint loss is set as:

[0065]

[0066] in Indicates the The predicted value of a physical quantity, represents the theoretical value derived from the conservation principle, is the number of physical quantities. The joint objective function is:

[0067]

[0068] in is a weight coefficient, whose value is experimentally determined to be within a specific range (e.g., 0.5 to 2) to balance the impact of mean square error and physical constraint loss. This design ensures that during the optimization process, the model not only minimizes data fit but also satisfies the constraints of physical laws, resulting in output forecast data with low numerical error and consistent with theoretical physical characteristics.

[0069] In practice, to calculate the physical constraint loss, one first needs to determine the theoretical expected values ​​of key physical quantities in the forecast data, such as temperature, humidity, and wind speed. This can be obtained through existing physical models or empirical formulas. Subsequently, the actual forecast value is subtracted from the theoretical value, and the absolute values ​​are taken and 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 thermal equilibrium, and the deviation between the actual temperature forecast and the theoretical value is then calculated. During training, the joint objective function uses automatic differentiation techniques to calculate the gradients of the parameters at each layer, and then updates the model parameters to minimize the overall objective.

[0070] The advantage of this method is that it can ensure high-precision fitting of the data-driven part and ensure that the output results conform to basic physical laws, thereby improving the credibility and interpretability of the forecast results. The numerical value of can flexibly balance the effects of mean square error and physical constraints in different meteorological scenarios. This ensures that the model neither deviates from physical laws due to excessive pursuit of numerical optimality nor affects the data fitting effect due to excessive reliance on physical constraints during local compensation and joint optimization. Overall, the joint objective function constructed in this invention provides a dual constraint mechanism for downscaling forecasting, which not only achieves precise control of numerical errors but also ensures the reasonable expression of physical characteristics, thereby significantly improving the overall performance and stability of the forecast system in practical applications.

[0071] Preferably, the variational assimilation and joint optimization steps adopt a 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 optimization of the network parameters.

[0072] In the present invention, variational assimilation and joint optimization steps are key links to achieve further refinement of forecast data and parameter optimization. It is preferred to use the gradient descent algorithm as the optimization method, calculate the gradient of each parameter in the joint objective function through automatic differentiation technology, and update the parameters according to the gradient direction. The basic idea of ​​variational assimilation is to combine the observation data with the forecast data, and to achieve a balance between data fitting and physical constraints by constructing the objective function (that is, the aforementioned joint objective function), so that the forecast data finally outputted are close to the actual situation both numerically and physically. 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 the objective function is recalculated after the parameters are updated until the preset convergence standard is reached. Mathematically, let the model parameters be , then the parameter update formula is:

[0073]

[0074] in represents the learning rate, Indicates in The joint objective function parameters This process utilizes an automatic differentiation framework to calculate gradients, accurately determining the sensitivity of each parameter in a complex network to the objective function, thereby guiding parameter updates. The joint optimization process, based on variational assimilation, addresses local errors and systematic biases in the forecast data, gradually reducing the error between the model output and the actual observed data through continuous iteration.

[0075] Specifically, in each iteration, the forecast output is first calculated using the current parameters, and then the current output is evaluated through the 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 joint optimization forecast data. During the implementation process, the variational assimilation and joint optimization steps have strict settings for the learning rate, number of iterations and convergence conditions, and the selection of these parameters is determined based on a large amount of historical data experiments. For example, by cross-validating on a historical data set, the learning rate range that is most suitable for the current forecast scenario can be determined, thereby ensuring that the optimization process will neither fall into a local optimal solution nor have the problem of too slow convergence speed. In addition, to ensure data consistency, the gradient calculation and parameter update in the joint optimization process use the same numerical precision and time step to avoid model instability caused by calculation errors. In summary, the variational assimilation and joint optimization steps employed in this invention minimize the joint objective function through gradient descent and automatic differentiation techniques. This, under the dual influence of numerical errors and physical constraints, significantly improves the precision and accuracy of forecast data and provides a solid foundation for parameter updates in subsequent closed-loop feedback iterations. This step not only allows for further error correction based on the initially refined data but also enables dynamic adjustment of model parameters, ensuring that forecast data progressively approach real-world observations with each iteration, thereby continuously improving the performance of the overall downscaling forecast system.

[0076] Preferably, the generative adversarial network adopts a convolutional autoencoder structure to form a generator, extracts the error features in the joint optimization 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, thereby realizing error compensation under physical constraints.

[0077] In the present invention, in order to achieve error compensation and fine correction of forecast data, a generative adversarial network structure is preferably adopted, which 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 process, thereby locally correcting the forecast data. The generator first uses a multi-layer convolutional layer to extract features from the input data through the encoder part. The encoder uses continuous convolution operations and downsampling operations to convert the original data into a low-dimensional feature representation, which can capture the global information and local changes in the forecast data. Subsequently, the decoder part gradually restores the data size through upsampling operations and convolution layers, and uses jump connections to fuse the high-resolution information saved in 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 differences between the jointly optimized forecast data and the actual forecast. At the same time, the discriminator uses a standard convolutional neural network structure to distinguish the error compensation data output by the generator from the real data. The discriminator extracts features through multi-layer convolution operations and pooling operations, and finally outputs a binary classification result through the fully connected layer, indicating whether the input data meets the physical constraints. During the training process, the generator and the discriminator are optimized alternately. 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 generator's goal is to minimize the gap between the generated data and the real data, and the discriminator's goal is to maximize the ability to distinguish between the two. In specific mathematical expressions, the generator's loss function can be expressed as:

[0078]

[0079] The loss function of the discriminator is:

[0080]

[0081] in Represents the generator input The generated output is, Represents the discriminator's response to the input The probability value of the discriminant. Through the above-mentioned adversarial process, the generator can gradually adjust the output data under physical constraints, so that the error-compensated 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 constraints, the convolution kernel size, number of layers, activation function, and optimization algorithm of the generator and discriminator are carefully debugged and verified to ensure stable training under different meteorological scenarios.

[0082] The use of a convolutional autoencoder structure not only has efficient feature compression capabilities, but also can effectively reconstruct details during the decoding process, so that the error compensation data generated by the generator can accurately capture local deviations. The design of the discriminator ensures that the physical rationality of the generated data is strictly identified, thereby providing accurate feedback to the generator. Through this generative adversarial mechanism, the present invention realizes an effective transformation from data-driven to physical constraints 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 of the local temperature anomaly area, so that the final forecast result and the observed data are significantly improved. In general, the generative adversarial network structure provides the present invention with an innovative means of realizing error compensation under physical constraints. Its operational 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.

[0083] Preferably, Figure 4 As shown in the figure, 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 weight of each node to achieve node-level adaptive correction.

[0084] In the present invention, in order to achieve node-level adaptive correction within the forecast area, a graph neural network combined with reinforcement learning is preferably used. Specifically, the present invention divides the forecast area into several grid nodes, each of which contains a multidimensional feature vector consisting of jointly optimized forecast data and error features compensated by a generative adversarial network. The graph neural network aggregates the features between nodes through graph convolution operations to achieve the sharing and transmission of global and local information. The graph convolution network adopts the standard graph convolution formula:

[0085]

[0086] in Indicates the Layer node feature matrix, is the adjacency matrix including self-connections, is the corresponding degree matrix, For the The weight matrix of the layer, is the activation function, which implements a normalized weighted aggregation of node features. Building on this, the reinforcement learning component introduces a reward function to measure the degree to which each node's forecast error is reduced after correction. Specifically, the reward function is set as the reduction in node forecast error, which is obtained by grid-by-grid comparison with observed baseline data. The reinforcement learning algorithm uses a policy gradient method to adjust the correction weights based on node characteristics, with the goal of minimizing the overall forecast error.

[0087] During the training process, each node receives a corresponding reward based on its current state and correction weight, and the network continuously updates its policy parameters to maximize the expected reward. The core of this method is to use graph neural networks to capture the complex spatial correlations between nodes, and dynamically adjust the correction weights of each node through a reinforcement learning mechanism, so that the final node-level adaptive correction result can effectively reduce the overall prediction error. During implementation, the present invention specifies in detail the number of layers of the graph convolutional network, the feature dimensions of each node in each layer, and the adjacency matrix construction method. At the same time, for the reinforcement learning part, key parameters such as the reward function form, learning rate, and discount factor are set.

[0088] For example, in a certain forecast area, if the local error is large in 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 reward function feedback, so that the forecast data after correction is closer to the real observation data. The present invention not only improves the accuracy of node-level correction, but also realizes the dynamic optimization of the overall forecast data through an adaptive mechanism. In general, the graph neural network combined with reinforcement learning provides a new node-level adaptive correction method for the present invention. Its principle is to use graph convolution to extract spatial features, and then optimize the correction strategy through reinforcement learning to achieve fine-grained adjustment of forecast errors. Through the present invention, the system can automatically adjust the correction weights of each grid node in a variety of meteorological scenarios, thereby significantly improving the accuracy and stability of the overall meteorological downscaling forecast. This technical solution has efficient operability. Its key steps include grid division, graph convolution operation, reward function setting and strategy update. Each step can be implemented through existing deep learning and reinforcement learning platforms, providing reliable theoretical and technical support for practical applications.

[0089] like Figure 5 As shown, the corrected forecast data is formatted to form the final forecast data, the grid-level forecast error is calculated using the observation benchmark data, and the forecast error is converted into a feedback gradient. The feedback gradient is used to update the parameters of local compensation, variational assimilation and adaptive correction, and the updated data is generated through closed-loop feedback iteration.

[0090] The closed-loop feedback iteration part of the present invention aims to achieve dynamic updating and optimization of the parameters of each link of local compensation, variational assimilation and adaptive correction by formatting the correction forecast data, calculating the errors and generating the feedback gradient, and finally generating continuously improved updated forecast data.

[0091] Specifically, the corrected forecast data obtained after various levels of processing are first formatted and unified into a standardized data format, which includes a unified spatial grid, timestamp, and standardized meteorological physical quantity values, providing a consistent data basis for subsequent error calculations. Next, using the pre-collected observational benchmark data, which has high spatial and numerical accuracy after rigorous preprocessing, a grid-by-grid calculation method is used to compare the final forecast data with the observational benchmark data to obtain the forecast error for each grid cell. The calculated forecast error is then converted into a feedback gradient through a specific algorithm. The feedback gradient is a key signal guiding the update of model parameters. It reflects the changing law of the error distribution on the spatial grid and its sensitivity to the parameters of each module of local compensation, variational assimilation, and adaptive correction. Finally, a closed-loop feedback mechanism is used to iteratively update the parameters of each module according to the feedback gradient. Through multiple feedback iterations, the grid-level forecast error is gradually reduced, so that the entire system can continuously approach the real observation during the downscaling forecast process. The final output updated forecast data reaches an ideal state in terms of spatial resolution and detail recovery.

[0092] Overall, this closed-loop feedback iteration process not only ensures error correction during the data fusion process, but also achieves system adaptive optimization through continuous parameter updates, thereby gradually improving the accuracy and stability of meteorological downscaling forecasts over multiple iterations. The core advantage of this invention lies in the use of feedback gradients to achieve error propagation and accurately adjust each processing module based on the error information in each iteration. This allows the entire forecast system to 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 forecasts.

[0093] Preferably, the closed-loop feedback iteration step calculates the forecast error grid by grid by using the observation baseline data and the final forecast data, and converts the obtained forecast error into a feedback gradient. The feedback gradient is used to continuously update the parameters of each link of local compensation processing, variational assimilation and adaptive correction, so that the grid-level forecast error is gradually reduced during the iteration process.

[0094] In the closed-loop feedback iteration process, the key step is to calculate the forecast error grid by grid using the observed baseline data and the final forecast data. The principle is to carefully compare the forecast value in each grid cell with the actual observed value to obtain the error distribution. In the specific implementation, the forecast error of each grid cell is first defined as the difference between the forecast value in the grid and the observed baseline value, using the formula:

[0095]

[0096] in, Indicates that the row List The grid cell prediction error is is the observation benchmark data value of the grid cell, is the value in the final forecast data. To ensure the accuracy of the error calculation, it is required to calculate all grid cells one by one to obtain a complete error matrix, which reflects the error distribution of the forecast data in each region in space. Next, the error matrix is ​​converted into a feedback gradient. This process uses the principle of the back propagation algorithm to map each error value into gradient information. The gradient information describes the sensitivity of the current error to the local model parameters. In the specific calculation process, the gradient function The error Converted into feedback gradient , the feedback gradient can be expressed as:

[0097]

[0098] in represents the loss function, represents the model parameters affecting the forecast output in the corresponding grid cell. Using this formula, the error for each grid cell is propagated through the corresponding gradient function to generate a feedback gradient matrix, which is used to guide subsequent model parameter updates. Operationally, error data is batch processed using a high-precision computing platform, and gradients are accurately calculated using automatic differentiation techniques, resulting in high numerical stability of the feedback gradients. For example, when the forecast error for a grid cell in a certain region is large, the feedback gradient will also be large, indicating that significant adjustment of model parameters in that region is required. Conversely, the feedback gradient in a region with smaller errors will also be small, indicating that the model has converged well in that region. This process ensures that error information is accurately transmitted to each parameter update module, providing specific adjustment directions and quantitative basis for subsequent local compensation, variational assimilation, and adaptive correction modules. This error conversion mechanism not only enables precise measurement of the global forecast error but also provides actionable gradient signals for local parameter adjustments, thereby facilitating the effective execution of the entire closed-loop feedback iteration process, gradually reducing forecast error and improving overall forecast quality. In practical applications, this step requires strict control of data resolution, numerical accuracy, and gradient calculation methods to ensure the accuracy and efficiency of the error conversion process, thereby laying a solid foundation for improving the overall performance of the system.

[0099] In the closed-loop feedback iteration process, the parameter update mechanism is the core link in achieving model adaptive optimization. In the specific operation process, based on the feedback gradient obtained by converting the prediction error, each model parameter is updated according to the gradient descent algorithm to ensure that the prediction error can be effectively reduced in the next iteration. The parameter update formula usually adopts the standard gradient descent formula, namely:

[0100]

[0101] in, Indicates in The parameters of the model at the iteration, is the preset learning rate, is the loss function calculated by feedback gradient in the current iteration for the parameter The application of this formula ensures that each parameter update is carried out in the direction of reducing the overall forecast error. To achieve 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 impact of each grid unit error on the local parameters, but also takes into account the manifestation of physical constraints in the forecast data. In operation, by setting a reasonable learning rate parameter It can balance the amplitude of parameter updates to avoid excessive parameter fluctuations due to excessive learning rate or slow convergence due to too small learning rate.

[0102] In practical implementation, parameter updates can be implemented using batch updates or mini-batch stochastic gradient descent strategies, thereby ensuring update stability while improving iteration efficiency. For example, in practical applications, when the forecast error feedback gradient for a particular region is large, the corresponding local compensation parameters are updated significantly to quickly correct forecast deviations. Conversely, when the error feedback gradient is small, the parameter updates are smaller, ensuring an optimal balance between local and global forecast data. This update mechanism not only ensures that each closed-loop feedback loop effectively reduces forecast error, but also, by continuously adjusting model parameters, continuously optimizes the entire downscaling forecast system over multiple iterations, gradually achieving precise capture of complex meteorological data. The present invention can also incorporate adaptive learning rate adjustment technology to dynamically adjust the learning rate parameter based on the changing trend of the feedback gradient, further improving the stability and convergence speed of the update process. Overall, the parameter update mechanism, through a rigorous gradient descent algorithm and automatic differentiation techniques, achieves continuous optimization of the parameters of the local compensation, variational assimilation, and adaptive correction modules, providing strong technical support for improving the overall forecast performance of the system and ensuring a continuous reduction in forecast error over multiple iterations.

[0103] 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 benchmark data is gradually reduced. To this end, the present invention stipulates a strict iteration strategy and convergence standard. Specifically, the closed-loop feedback iteration process starts with the initial feedback gradient calculation. After each parameter update, the updated forecast data is regenerated. Then, the observed benchmark data is used to calculate the new forecast error for the updated forecast data grid by grid, and the new error is converted into a new feedback gradient, and the cycle continues. Mathematically, the process can be described as: let the initial forecast error be , generate the first The error corresponding to the forecast data is , the feedback gradient is , when: or When , the iterative process is considered to converge, where is the preset convergence threshold. This iterative process uses automatic differentiation techniques to continuously calculate error gradients and update parameters. The error trend is recorded at each iteration to ensure that the system parameters converge near the global optimal solution. In practical applications, to improve iteration efficiency, an early stopping strategy can be adopted. This strategy stops iterations early when the error change falls below a preset threshold over several consecutive iterations, saving computing resources.

[0104] 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 conditions. This process not only reflects the trend of gradual error reduction, but also ensures that both local and overall forecasts are effectively corrected through the dynamic propagation of feedback gradients throughout the forecast area. During the iterative process, the parameter updates of each link are interrelated, and the parameter adjustment of the local compensation module directly affects the effect of variational assimilation and adaptive correction, thereby forming an adaptive optimization closed-loop system. The entire process needs to be implemented in conjunction with a high-performance computing platform, and strict time step and numerical accuracy control is adopted to ensure data consistency and calculation accuracy in multiple iterations. Through this closed-loop feedback iteration mechanism, the forecast system can continuously self-adjust and optimize, thereby maintaining high forecast accuracy and stability in the face of complex and changeable meteorological fields, and achieving an organic balance between data-driven and physical constraints.

[0105] In the actual implementation of the present invention, closed-loop feedback iteration is used to continuously update forecast data and continuously improve system performance, and its effects have been verified in multiple meteorological field forecast cases. In specific implementation, the system first formats the initially generated correction forecast data to ensure that the data has a uniform spatial resolution, time identifier, and numerical unit. Then, the preprocessed observation benchmark data is used to calculate the error of each grid cell to obtain a refined forecast error matrix. Through automatic differentiation technology, the error matrix is ​​converted into a feedback gradient matrix. Each value in the matrix directly reflects the sensitivity of the corresponding grid cell error to the model parameters.

[0106] In one example, the initial average forecast error for a certain region was 4 degrees Celsius. Through closed-loop feedback iteration, the average error gradually decreased with each iteration. After the fifth iteration, the average error dropped to 1.2 degrees Celsius, and after the tenth iteration, the average error dropped to 0.6 degrees Celsius, meeting the preset convergence criteria. The parameter updates during this iterative process not only improved forecast accuracy globally but also targeted forecast anomalies in local areas, significantly reducing local errors. To achieve this, the system employed batch parameter updates and an adaptive learning rate adjustment strategy during the iteration process. The parameter updates and error reduction were recorded at each iteration, and the learning rate was adjusted based on the error changes, ensuring that parameter updates were neither excessive nor slow. Furthermore, to ensure the accuracy of the feedback gradient calculation, the system employed high-precision floating-point arithmetic and incorporated an error smoothing algorithm into the feedback process to avoid parameter fluctuations caused by numerical instability. This closed-loop feedback iteration mechanism demonstrated high robustness and adaptability in practical applications. Even in the face of sudden meteorological events or high data noise, the system was able to effectively correct forecast data through continuous iteration, improving the reliability of forecast results.

[0107] For example, in a weather forecast involving enhanced local convective activity, the initial forecast error was large due to drastic changes in local temperature and humidity data. However, through closed-loop feedback iteration, the system completed parameter adjustments in a short period of time, achieving timely correction of the local anomaly forecast. Overall, this embodiment demonstrates the effectiveness of the closed-loop feedback iteration mechanism in improving the accuracy of meteorological downscaling forecasts. Its operational details cover error calculation, feedback gradient generation, parameter update, and iterative convergence strategy, providing a detailed and replicable optimization process for actual forecast systems, significantly enhancing the system's adaptability and forecast stability in complex meteorological environments.

[0108] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware.

[0109] The above are merely embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included within the scope of the claims of the present application.

Claims

1. A meteorological downscaling method integrating numerical weather forecast and AI weather forecast, characterized in that: The following steps are involved: Collect forecast data output by the numerical model and satellite observation data within the corresponding forecast area, preprocess the data to unify the data format, and use interpolation and mapping techniques to obtain refined forecast data and observation benchmark data respectively; The refined forecast data is subjected to local compensation processing based on a deep neural network. 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 including mean square error and physical constraints is constructed. Variational assimilation and joint optimization are performed on the preliminary refined data. Error compensation is performed under physical constraints using a generative adversarial network. Node-level adaptive correction is achieved using a graph neural network combined with reinforcement learning to obtain corrected forecast data. The meta-learning technology uses a model-independent meta-learning method to pre-train deep neural network parameters on no less than three independent historical data sets, and the network parameters are fine-tuned during the regional adaptation stage to generate local compensation data. The corrected forecast data are formatted to form final forecast data, the grid-level forecast error is calculated using the observational benchmark data, and the forecast error is converted into a feedback gradient. The feedback gradient is used to update the parameters of local compensation, variational assimilation, and adaptive correction, and updated data is generated through closed-loop feedback iteration.

2. The method according to claim 1, characterized in that The preprocessing operation uses 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 converting 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 technology uses a bilinear interpolation method to calculate the weighted average of the adjacent pixel values ​​in the forecast data and satellite observation data, and combines it with a nonlinear mapping algorithm based on convolution operation to improve the spatial resolution of the forecast data and satellite observation data, thereby obtaining refined forecast data and observation benchmark data respectively.

4. The method according to claim 1, wherein The local compensation processing 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 jump connections. A self-attention module is set in the network to enhance local feature responses.

5. The method according to claim 1, wherein 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 conservation of momentum and energy, and sets a fixed value as the weight coefficient to constrain the physical characteristics.

6. The method according to claim 1, characterized in that 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 achieve the optimization of the network parameters.

7. The method according to claim 1, characterized in that The generative adversarial network uses a convolutional autoencoder structure to form a generator, extracts the error features in the joint optimization forecast data through encoding and reconstructs the error compensation data through decoding, while the discriminator uses a convolutional neural network structure to distinguish the generated error compensation data from the real data, thereby achieving error compensation under physical constraints.

8. The method according to claim 1, characterized in that Graph neural networks combined with reinforcement learning use graph convolutional networks to aggregate the features of grid nodes divided into forecast areas, and set 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.

9. 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 baseline data and the final forecast data, and converts the resulting forecast error into a feedback gradient. The feedback gradient is used to continuously update the parameters of each link of local compensation processing, variational assimilation and adaptive correction, so that the grid-level forecast error is gradually reduced during the iteration process.

Images