A meteorological field error correction method and system for a meteorological large model

Abstract

This invention discloses a method and system for correcting meteorological field errors in large-scale meteorological models, relating to the field of meteorological data processing technology. It includes: inputting a multidimensional meteorological input tensor into an error correction model to predict the systematic deviation between the original NWP analysis field and the ERA5 ground truth field, obtaining a systematic residual field; concatenating the systematic residual field with the original NWP analysis field to obtain a meteorological correction field; determining whether the meteorological correction field conforms to the true atmospheric multidimensional spatial distribution characteristics of the ERA5 ground truth field, obtaining asymmetric adversarial loss and pixel reconstruction loss; extracting the core guiding layer variables of the meteorological correction field to obtain dynamic physical loss; and updating the model parameters of the residual generator through backpropagation based on the total loss function to obtain the trained residual generator. This invention enables the model to robustly learn the systematic deviation of extreme small values, significantly improving the correction effect of upper-air variables and the forecasting effect of large-scale meteorological models.

Description

Technical Field

[0001] This invention relates to the field of meteorological data processing technology, and in particular to a method and system for correcting meteorological field errors for large meteorological models. Background Technology

[0002] Numerical Weather Prediction (NWP) analysis fields and forecast products serve as crucial reference factors in meteorological forecasting, meeting the real-time forecasting needs of practical operations. NWP analysis field data can provide the driving field for large-scale meteorological models in real-time forecasting operations. However, due to differences in model data assimilation techniques and the similarity between their own dynamic framework and physical parameterization schemes, NWP suffers from systematic biases. Using NWP analysis fields with systematic biases to drive the training of large-scale meteorological models results in severe data distribution shifts, directly impacting the forecasting performance of these models and causing significant degradation. Therefore, how to intelligently and rapidly correct biases in NWP analysis fields using reanalysis data to improve the forecasting performance of large-scale meteorological models in real-time operational forecasting has become a core problem urgently needing to be solved in the field of meteorological service engineering.

[0003] In existing technologies, when dealing with extremely small variables that approximate physical vacuum states, such as the specific humidity of the upper atmosphere, traditional error evaluation algorithms, in order to measure the performance of the corrected model, directly use the original error as the denominator and finally apply a fixed percentage formula to calculate the improvement rate.

[0004] However, in this existing technology, the denominator tending to zero can lead to the illusion of an explosion in percentage values. This creates a blind spot in the model's generalization and objective evaluation, resulting in the entire evaluation system lacking a robust truncation protection mechanism for upper-level minimum values. This seriously affects the forecasting performance of large meteorological models in real-time operations. Summary of the Invention

[0005] Therefore, it is necessary to provide a method and system for correcting meteorological field errors in large meteorological models to address the aforementioned technical problems.

[0006] This invention provides a meteorological field error correction method for large meteorological models, comprising: Obtain the original NWP analysis field and ERA5 ground truth field at multiple historical moments, and construct a multidimensional meteorological input tensor; The multidimensional meteorological input tensor is used to input the error correction model. The error correction model includes: a residual generative adversarial network and a physical constraint module. The residual generative adversarial network includes: a residual generator and a discriminator. The systematic residual field is obtained by predicting the systematic deviation between the original NWP analysis field and the ERA5 true field using a residual generator; the systematic residual field is then spliced ​​with the original NWP analysis field to obtain the meteorological correction field. The discriminator determines whether the meteorological correction field conforms to the true atmospheric multidimensional spatial distribution characteristics of the ERA5 true field, and obtains the asymmetric adversarial loss and pixel reconstruction loss. The core guiding layer variables of the meteorological correction field are extracted through the physical constraint module to introduce geostrophic balance physical constraints into the meteorological correction field, and obtains the dynamic physical loss. Based on the total loss function obtained by weighted fusion of asymmetric adversarial loss, pixel reconstruction loss and dynamic physical loss, the model parameters of the residual generator are updated through backpropagation to obtain the trained residual generator; The real-time NWP analysis field is input into the trained residual generator to obtain the corrected meteorological field.

[0007] Optionally, the original NWP analysis field and ERA5 ground truth field at multiple historical moments are obtained to construct a multidimensional meteorological input tensor, specifically including: Both the original NWP analysis field and the ERA5 true field include: meteorological variables at different altitudes and surface variables; meteorological variables at different altitudes are used to characterize the thermal, dynamic and water vapor states of the atmosphere in three-dimensional space, while surface variables are used to characterize near-surface meteorological elements and circulation characteristics; By splicing meteorological and surface variables at different altitude levels along the channel dimension and introducing digital elevation model data as an independent channel, a multidimensional meteorological tensor containing both spatial and channel dimensions is obtained.

[0008] Optionally, it also includes: before applying the multidimensional meteorological input tensor to the input error correction model, performing time-stamp-aware adaptive data normalization processing on the multidimensional meteorological input tensor, specifically including: The global mean and standard deviation of each channel of the multidimensional meteorological input tensor are determined based on the following formula: ; ; The multidimensional meteorological input tensor is normalized and mapped based on the following formula: ; in, Let be the global mean of the c-th physical channel. Let c be the standard deviation of the c-th physical channel. H The latitude dimension of the latitude and longitude spatial grid. W For the longitude dimension of the latitude and longitude spatial grid, N The total number of samples in the time dimension. For the normalized result, For multidimensional meteorological input tensors, It is a tensor matrix.

[0009] Optionally, a systematic residual field is obtained by predicting the systematic deviation between the original NWP analysis field and the ERA5 true field using a residual generator, specifically including: The residual generator adopts a fully convolutional encoding and decoding architecture, including: an encoder and a decoder; the encoder includes: a plurality of first two-dimensional convolutional layers, a first instance normalization layer and a first linear unit with leakage correction connected in sequence; the decoder includes: a plurality of deconvolutional layers, a skip connection module and a first linear two-dimensional convolutional layer connected in sequence. The multidimensional meteorological input tensor is encoded by multiple first two-dimensional convolutional layers to extract the spatial texture information of the atmospheric physical field and obtain the encoded features. The encoded features are normalized and nonlinearly activated by the first instance normalization layer and the first leakage correction linear unit to gradually compress the spatial resolution of the encoded features and obtain a deep feature map that represents the large-scale weather modal features. Multiple deconvolutional layers are used to upsample the deep feature map multiple times to gradually restore the spatial resolution of the deep feature map. After each upsampling, the high-dimensional feature map of the same resolution at the corresponding level in the encoder is concatenated with the output feature map of the current decoder in the channel dimension through the skip connection module to supplement micro-scale meteorological spatial information and obtain the highest resolution feature map. The highest resolution feature map is mapped to the same number of channels as the original NWP analysis field through the first linear two-dimensional convolutional layer, thus obtaining a systematic residual field characterizing the error distribution of various meteorological variables at high and low altitudes.

[0010] Optionally, a discriminator is used to determine whether the meteorological correction field conforms to the true atmospheric multidimensional spatial distribution characteristics of the ERA5 ground truth field, resulting in asymmetric adversarial loss and pixel reconstruction loss, specifically including: The discriminator employs a convolutional neural network based on the PatchGAN structure, which includes: multiple second two-dimensional convolutional layers, a second instance normalization layer, a second linear unit with leakage correction, and a second linear two-dimensional convolutional layer connected in sequence. The multidimensional meteorological input tensor is concatenated with the meteorological correction field and the ERA5 ground truth field in the depth channel dimension to constrain the discrimination process to the initial conditions of the original NWP analysis field, thus obtaining a hybrid high-dimensional tensor. Multiple second-dimensional convolutional layers are used to downsample the mixed high-dimensional tensor multiple times to extract features, thereby gradually expanding the receptive field. After each downsampling, a second instance normalization layer is used to eliminate global contrast differences between samples, and a second leak-corrected linear unit is used for nonlinear activation to extract high-frequency meteorological physical textures of multi-scale local spatial features, thus obtaining multi-scale local spatial features. The second linear two-dimensional convolutional layer maps multi-scale local spatial features and outputs a continuous two-dimensional confidence feature matrix. Each pixel value in the two-dimensional confidence feature matrix independently represents the confidence probability that the corresponding local spatial patch in the mixed high-dimensional tensor conforms to the real atmospheric multi-dimensional spatial distribution characteristics in the ERA5 ground truth field. Based on the two-dimensional confidence feature matrix, the asymmetric adversarial loss and pixel reconstruction loss are determined.

[0011] Alternatively, the asymmetric adversarial loss can be determined based on the following formula: ; The pixel reconstruction loss is determined based on the following formula: ; The dynamic physical loss is determined based on the following formula: ; The total loss function, obtained by weighted fusion of asymmetric adversarial loss, pixel reconstruction loss, and dynamic physical loss, is determined based on the following formula: ; in, For the total loss function, For pixel reconstruction loss, For asymmetric adversarial losses, For dynamic physical loss, These are the weighting coefficients for pixel reconstruction loss. These are the weighting coefficients for asymmetric adversarial loss. These are the weighting coefficients for dynamic physical loss. For meteorological true values, For weather correction field, The mapping function representing the discriminator, It is a tensor matrix. This represents the total number of elements in the confidence matrix. This is a comprehensive scaling constant. The core steering layer has a zonal wind pattern at 500 hPa. The core guiding layer is subjected to meridional winds at 500 hPa.

[0012] Optionally, the dynamic physical loss is constructed for the 500 hPa core guiding layer, and the spatial gradient of the height field is approximated using the central finite difference method based on the following equation: ; ; in, The height field corresponding to the core guiding layer at 500 hPa, Δ x For the east-west direction, Δ y It is oriented north-south.

[0013] Optionally, it also includes: after obtaining the corrected meteorological field, performing physical minimum truncation processing through a built-in piecewise minimum truncation determination mechanism based on physical meaning, specifically including: Determine the original root mean square error and the post-prediction error of the target physical channel respectively; If the root mean square error of the original meteorological data is less than the preset physical minimum truncation threshold, then the variable is determined to be close to the physical vacuum state, and the truncation mechanism is directly triggered, forcing the output relative improvement rate to be 0. If the root mean square error of the original meteorological data is greater than the preset physical minimum cutoff threshold, then the variable is determined to have objective physical assessment significance, and the true relative improvement rate is determined based on the following formula: ; in, This represents the root mean square error of the original meteorological data. For the prediction error, This refers to the relative increase rate.

[0014] This invention provides a meteorological field error correction system for large meteorological models, comprising: The tensor construction module is used to obtain the original NWP analysis field and ERA5 truth field at multiple historical moments and construct a multidimensional meteorological input tensor. The input module is used to input the multidimensional meteorological input tensor into the error correction model. The error correction model includes: a residual generative adversarial network and a physical constraint module. The residual generative adversarial network includes: a residual generator and a discriminator. The correction module is used to predict the systematic deviation between the original NWP analysis field and the ERA5 true field through the residual generator to obtain the systematic residual field; the systematic residual field is then spliced ​​with the original NWP analysis field to obtain the meteorological correction field. The loss construction module is used to discriminate whether the meteorological correction field conforms to the real atmospheric multidimensional spatial distribution characteristics of the ERA5 true field through a discriminator, and obtain asymmetric adversarial loss and pixel reconstruction loss; the physical constraint module extracts the core guiding layer variables of the meteorological correction field to introduce geostrophic balance physical constraints into the meteorological correction field, and obtains dynamic physical loss. The update module is used to update the model parameters of the residual generator through backpropagation based on the total loss obtained by weighted fusion of asymmetric adversarial loss, pixel reconstruction loss and dynamic physical loss, so as to obtain the trained residual generator. The inference output module is used to input the real-time NWP analysis field into the trained residual generator to obtain the corrected meteorological correction field.

[0015] The meteorological field error correction method and system for large meteorological models provided in this invention have the following advantages compared with the prior art: This invention constructs a loss function based on asymmetric adversarial loss, pixel reconstruction loss, and dynamic physical loss under geostrophic balance constraints through joint training of a residual generative adversarial network and a physical constraint module. The entire training process completely avoids the dependence on the percentage improvement rate found in existing technologies, directly optimizing the multidimensional spatial distribution differences between the meteorological correction field and the true field. This technique enables the trained residual generator to directly output the corrected meteorological field in real-time operations, eliminating the need for fragile percentage formulas for evaluation. This fundamentally eliminates the artifacts caused by the denominator approaching zero, allowing the model to robustly learn the systematic biases of extreme small values, significantly improving the correction effect of upper-air variables and the forecasting performance of large-scale meteorological models. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a meteorological field error correction method for a large meteorological model, as provided in one embodiment. Figure 2 This is a model training flowchart for a meteorological field error correction method for a large meteorological model, provided in one embodiment. Figure 3 This is a visual comparison of sea level pressure field error corrections provided in one embodiment of a meteorological field error correction method for a large meteorological model. Figure 3 (a) in the original NECP is... Figure 3 (b) in the figure represents the NECP after error correction. Figure 3 (c) in the equation represents the true value of ERA5. Figure 3 In the figure, (d) represents the original NECP root mean square error. Figure 3 In the figure (e), the NECP root mean square error after error correction is represented. Figure 3 (f) in the figure represents the improvement amount, which is the absolute value of the original error minus the absolute value of the corrected error. Red indicates that there is an improvement. Figure 4 This is a visualization comparison of error correction at 2m temperature for a meteorological field error correction method for a large meteorological model provided in one embodiment. Figure 4 (a) in the original NECP is... Figure 4 (b) in the figure represents the NECP after error correction. Figure 4 (c) in the equation represents the true value of ERA5. Figure 4 In the figure, (d) represents the original NECP root mean square error. Figure 4 In the figure (e), the NECP root mean square error after error correction is represented. Figure 4(f) in the figure represents the improvement amount, which is the absolute value of the original error minus the absolute value of the corrected error. Red indicates that there is an improvement. Figure 5 This is a visual comparison of error correction under 10m zonal wind conditions, illustrating a meteorological field error correction method for a large meteorological model provided in one embodiment. Figure 5 (a) in the original NECP is... Figure 5 (b) in the figure represents the NECP after error correction. Figure 5 (c) in the equation represents the true value of ERA5. Figure 5 In the figure, (d) represents the original NECP root mean square error. Figure 5 In the figure (e), the NECP root mean square error after error correction is represented. Figure 5 (f) in the figure represents the improvement amount, which is the absolute value of the original error minus the absolute value of the corrected error. Red indicates that there is an improvement. Figure 6 This is a visual comparison of error correction under 10m meridional wind conditions, illustrating a meteorological field error correction method for a large meteorological model provided in one embodiment. Figure 6 (a) in the original NECP is... Figure 6 (b) in the figure represents the NECP after error correction. Figure 6 (c) in the equation represents the true value of ERA5. Figure 6 In the figure, (d) represents the original NECP root mean square error. Figure 6 In the figure (e), the NECP root mean square error after error correction is represented. Figure 6 (f) in the figure represents the improvement amount, which is the absolute value of the original error minus the absolute value of the corrected error. Red indicates that there is an improvement. Figure 7 This is a visualization comparison of error correction at 1000 hPa geopotential for a meteorological field error correction method for a large meteorological model provided in one embodiment. Figure 7 (a) in the original NECP is... Figure 7 (b) in the figure represents the NECP after error correction. Figure 7 (c) in the equation represents the true value of ERA5. Figure 7 In the figure, (d) represents the original NECP root mean square error. Figure 7 In the figure (e), the NECP root mean square error after error correction is represented. Figure 7 (f) in the figure represents the improvement amount, which is the absolute value of the original error minus the absolute value of the corrected error. Red indicates that there is an improvement. Figure 8 This is a visualization comparison of error correction at 850 hPa temperature for a meteorological field error correction method for a large meteorological model provided in one embodiment. Figure 8 (a) in the original NECP is... Figure 8(b) in the figure represents the NECP after error correction. Figure 8 (c) in the equation represents the true value of ERA5. Figure 8 In the figure, (d) represents the original NECP root mean square error. Figure 8 In the figure (e), the NECP root mean square error after error correction is represented. Figure 8 (f) in the figure represents the improvement amount, which is the absolute value of the original error minus the absolute value of the corrected error. Red indicates that there is an improvement. Figure 9 This is a visual comparison of error correction under 500 hPa zonal wind conditions, illustrating a meteorological field error correction method for a large meteorological model provided in one embodiment. Figure 9 (a) in the original NECP is... Figure 9 (b) in the figure represents the NECP after error correction. Figure 9 (c) in the equation represents the true value of ERA5. Figure 9 In the figure, (d) represents the original NECP root mean square error. Figure 9 In the figure (e), the NECP root mean square error after error correction is represented. Figure 9 (f) in the figure represents the improvement amount, which is the absolute value of the original error minus the absolute value of the corrected error. Red indicates that there is an improvement. Figure 10 This is a visual comparison of error correction under 500 hPa meridional wind conditions for a meteorological field error correction method for a large meteorological model provided in one embodiment. Figure 10 (a) in the original NECP is... Figure 10 (b) in the figure represents the NECP after error correction. Figure 10 (c) in the equation represents the true value of ERA5. Figure 10 In the figure, (d) represents the original NECP root mean square error. Figure 10 In the figure (e), the NECP root mean square error after error correction is represented. Figure 10 (f) in the figure represents the improvement amount, which is the absolute value of the original error minus the absolute value of the corrected error. Red indicates that there is an improvement. Figure 11 This is a visualization comparison of error correction at 500 hPa geopotential for a meteorological field error correction method for a large meteorological model provided in one embodiment. Figure 11 (a) in the original NECP is... Figure 11 (b) in the figure represents the NECP after error correction. Figure 11 (c) in the equation represents the true value of ERA5. Figure 11 In the figure, (d) represents the original NECP root mean square error. Figure 11 In the figure (e), the NECP root mean square error after error correction is represented. Figure 11 (f) in the figure represents the improvement amount, which is the absolute value of the original error minus the absolute value of the corrected error. Red indicates that there is an improvement. Figure 12 This is an error comparison chart showing the sea level pressure as the initial field input to a large meteorological model (such as the Pangoal Meteorological Model) as provided in one embodiment of a meteorological field error correction method for large meteorological models. Figure 12 In the figure, (a) represents the root mean square error of sea level pressure. Figure 12 (b) in the figure represents the spatial anomaly correlation coefficient of sea level pressure; Figure 13 This is an error comparison chart of a meteorological field error correction method for a large meteorological model (such as the Pangoal Inferior Model) provided in one embodiment, using 2m temperature as the initial field input. Figure 13 In the figure, (a) represents the root mean square error at a temperature of 2m. Figure 13 (b) in the figure represents the spatial anomaly correlation coefficient for 2m temperature. Figure 14 This is an error comparison chart of a meteorological field error correction method for a large meteorological model (such as the Pangoal Meteorological Model) provided in one embodiment, using 10m zonal wind as the initial field input. Figure 14 In the figure (a), the root mean square error of the zonal wind at 10m is... Figure 14 (b) in the figure represents the spatial anomaly correlation coefficient of the 10m zonal wind; Figure 15 This is an error comparison chart showing the 1000 hPa geopotential as the initial field input to a large meteorological model (such as the Pangoal Inferior Model) as provided in one embodiment of a meteorological field error correction method for large meteorological models. Figure 15 In the figure, (a) represents the root mean square error of the 1000 hPa geopotential. Figure 15 (b) in the figure represents the spatial anomaly correlation coefficient of the 1000 hPa geopotential. Figure 16 This is an error comparison chart showing the use of 925 hPa specific humidity as the initial field input to a large meteorological model (such as the Pangoal Meteorological Model) as part of a meteorological field error correction method for a large meteorological model, as provided in one embodiment. Figure 16 In the figure (a), the root mean square error of the specific humidity at 925 hPa is... Figure 16 (b) in the figure represents the spatial anomaly correlation coefficient of the specific humidity at 925 hPa; Figure 17 This is an error comparison chart of a meteorological field error correction method for a large meteorological model (such as the Pangoal Meteorological Model) provided in one embodiment, using 850 hPa zonal wind as the initial field input. Figure 17 In the figure (a), the root mean square error of the zonal wind at 850 hPa is... Figure 17 (b) in the figure represents the spatial anomaly correlation coefficient of the zonal wind at 850 hPa; Figure 18This is a comparison chart showing the errors of a meteorological field error correction method for a large meteorological model (such as the Pangoal Inferior Model) provided in one embodiment, using an 850 hPa temperature as the initial field input. Figure 18 In the figure, (a) represents the root mean square error of the temperature at 850 hPa. Figure 18 (b) in the figure represents the spatial anomaly correlation coefficient for the 850 hPa temperature. Figure 19 This is an error comparison diagram of a meteorological field error correction method for a large meteorological model (such as the Pangoal Inferior Model) provided in one embodiment, using the 500 hPa geopotential as the initial field input. Figure 19 In the figure, (a) represents the root mean square error of the 500 hPa geopotential. Figure 19 (b) in the figure represents the spatial anomaly correlation coefficient of the 500 hPa geopotential. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] In recent years, AI-based large-scale meteorological models have made groundbreaking progress in weather forecasting. Currently, mainstream large-scale meteorological models heavily rely on meteorological reanalysis data for training, such as the fifth-generation global atmospheric reanalysis dataset (ECMWF Reanalysis v5, ERA5). This type of reanalysis data integrates data from various observation sources and state-of-the-art assimilation systems, possessing the ability to reflect the true atmospheric conditions with high precision. Therefore, in practical inference, constructing ideal driving fields based on reanalysis data for large-scale meteorological models can significantly improve the model's prediction accuracy, reflecting the ideal state closest to the real atmosphere. However, in actual meteorological public service operations, meteorological reanalysis data suffers from a severe data release lag, making it impossible to provide real-time input to large-scale models.

[0019] With the widespread application of artificial intelligence technology, deep learning has made some progress in data-driven correction of meteorological analysis fields, showing great potential in fitting nonlinear meteorological characteristics. However, traditional pure data-driven deep learning is still a method with opaque physical mechanisms. In the process of minimizing errors, this type of purely data-driven mode often disrupts the physical equilibrium that is crucial to atmospheric dynamics.

[0020] Due to the lack of prior physical conditions such as geostrophic equilibrium, the upper-level wind field and altitude field generated by the model are prone to serious physical and logical conflicts, and such forecast results that violate basic physical laws are unreliable. In deep learning models lacking physical mechanisms, the models often rely heavily on massive amounts of training data spanning multiple decades to search for local optima. In actual meteorological forecasting operations or research on specific meteorological events, when high-quality data samples are lacking, existing models often struggle to converge, leading to model collapse or severe overfitting, and consequently failing to efficiently extract the underlying systematic bias modes of the NWP model.

[0021] This invention provides a method for correcting meteorological field errors in large meteorological models, such as... Figure 1 As shown, the method includes: The original NWP analysis field and ERA5 ground truth field at multiple historical moments are obtained to construct a multidimensional meteorological input tensor. This multidimensional meteorological input tensor is then input into an error correction model, which includes a residual generative adversarial network (GAN) and a physical constraint module. The GAN includes a residual generator and a discriminator.

[0022] The systematic deviation between the original NWP analysis field and the ERA5 true field is predicted using a residual generator to obtain a systematic residual field. This systematic residual field is then concatenated with the original NWP analysis field to obtain the meteorological correction field.

[0023] A discriminator is used to determine whether the meteorological correction field conforms to the true atmospheric multidimensional spatial distribution characteristics of the ERA5 ground truth field, resulting in asymmetric adversarial loss and pixel reconstruction loss. The core guiding layer variables of the meteorological correction field are extracted through a physical constraint module to introduce geostrophic equilibrium physical constraints into the meteorological correction field, resulting in dynamic physical loss.

[0024] Based on the total loss function obtained by weighted fusion of asymmetric adversarial loss, pixel reconstruction loss and dynamic physical loss, the model parameters of the residual generator are updated through backpropagation to obtain the trained residual generator.

[0025] The real-time NWP analysis field is input into the trained residual generator to obtain the corrected meteorological field.

[0026] A specific embodiment of the present invention is provided: (1) Obtain the original meteorological data and construct multidimensional input tensors such as static terrain features.

[0027] Acquire raw NWP analysis field data (such as NCEP) and meteorological reanalysis data (ERA5) from multiple time periods. Strictly isolate the test set and training set in the time dimension. Use consecutive time periods within a certain historical time period as the training set and another completely independent time period as the test set to ensure the continuity in physical time sequence and avoid data leakage problems caused by random partitioning.

[0028] Meteorological variables at different altitudes (characterizing the thermal, dynamic, and water vapor states of the atmosphere in three-dimensional space, including geopotential, temperature, specific humidity, zonal wind, and meridional wind) and surface variables (characterizing near-surface meteorological elements and circulation characteristics, including sea level pressure, 2m temperature, 10m zonal wind, and 10m meridional wind) are spliced ​​together, while static digital elevation model (DEM) data is introduced as an independent channel to construct a shape of... Multidimensional meteorological tensor ,in, C This represents the number of dynamic weather channels. H and W These are the latitude and longitude dimensions of the latitude and longitude spatial grid, respectively.

[0029] The processing method for the Digital Elevation Model (DEM) is as follows: First, the surface geopotential (Geopotential) is extracted from the meteorological data source (ERA5). ), with dimensions of And by dividing by the standard gravitational acceleration constant This is converted into geopotential height (in meters) with practical geographical significance: ; Considering the long-term invariance of terrain, the transformed multidimensional tensor undergoes dimensionality reduction, stripping away the time dimension to extract a 721×1440 two-dimensional static spatial matrix with a resolution of 0.25°×0.25°, which is then converted into a float32 single-precision tensor for storage. Furthermore, to avoid gradient explosion caused by the vast elevation span of global terrain, strict extreme value normalization is performed on the DEM matrix during dataset construction.

[0030] ; Introduced The truncation protection factor prevents the denominator from being zero. Global terrain is smoothly mapped to the standard dimensionless space [0, 1].

[0031] During the data loading stage, the channel dimension of the two-dimensional DEM matrix is ​​expanded and concatenated along the tensor depth direction after 69 dynamic meteorological tensors, serving as the low-level geospatial prior channel input generator. This multimodal fusion mechanism enables the fully convolutional residual network to adaptively incorporate underlying surface forced boundary conditions such as topographic blocking and mountain runoff when correcting low-level atmospheric characteristics.

[0032] (2) Time-aware automatic adaptive data normalization processing.

[0033] During the construction of the multidimensional meteorological input tensor, the underlying data files of the original NWP analysis field and the ERA5 ground truth field, as well as the last modification timestamp of the global statistics cache file, are automatically compared. When the modification time of the underlying data file is detected to be later than that of the global statistics cache file, the cache overwrite mechanism is automatically triggered. Based on the updated multidimensional meteorological input tensor, the multi-channel global mean used for tensor normalization operations is re-extracted and recalculated. with standard deviation .

[0034] For any physical channel in a multidimensional meteorological input tensor Its global mean with standard deviation The statistical expectation formulas are as follows: ; ; in, This represents a tensor matrix containing the original NWP analysis field or the ERA5 truth field. The total number of samples in the time dimension. and The latitude and longitude dimensions of the latitude and longitude spatial grids, respectively. To prevent the minimum value constant of division by zero (e.g.) ).

[0035] The normalized mapping formula is: ; in, Let be the global mean of the c-th physical channel. Let c be the standard deviation of the c-th physical channel. H The latitude dimension of the latitude and longitude spatial grid. W For the longitude dimension of the latitude and longitude spatial grid, N The total number of samples in the time dimension. For the normalized result, For multidimensional meteorological input tensors.

[0036] To reduce computational overhead, a predetermined number of samples are randomly selected when calculating the mean and standard deviation to perform unbiased estimation.

[0037] (3) Construct a residual generation adversarial network, including a residual generator. and discriminator .

[0038] A. Residual Generator The processing flow.

[0039] The residual generator employs a fully convolutional encoder-decoder architecture, comprising an encoder and a decoder. The encoder includes: multiple first 2D convolutional layers, a first instance normalization layer, and a first linear unit with leakage correction, connected in sequence. The decoder includes: multiple deconvolutional layers, a skip connection module, and a first linear 2D convolutional layer, connected in sequence.

[0040] A multidimensional meteorological input tensor is obtained, which contains dynamic meteorological variables and static geographical variables in the depth dimension. The residual generator first performs a segmentation operation on the input tensor in the channel dimension, extracting channels with a total number of channels. The original NWP analysis field matrix This is for use in the subsequent residual addition stage.

[0041] The complete multidimensional meteorological input tensor is fed into the encoder module. The encoder consists of consecutive two-dimensional convolutional layers with a stride of 2. In each forward propagation, the convolutional block extracts the spatial texture of the atmospheric physical field through the first two-dimensional convolutional layer, and then suppresses gradient vanishing through the first instance normalization layer and the first leakage correction linear unit. After multiple levels of downsampling, the model gradually compresses the spatial resolution of the tensor, extracting large-scale weather modal features with a macroscopic receptive field.

[0042] The deep feature map output from the encoder is input into the decoder module. The decoder consists of consecutive deconvolution blocks, which progressively restore the spatial resolution of the tensor. To prevent the loss of microscale meteorological spatial information during downsampling, after each deconvolution operation, a skip connection mechanism is used to hard-join the high-dimensional feature map of the same resolution at the corresponding level in the encoder with the output feature map of the decoder in the channel dimension.

[0043] The highest-resolution feature map output by the decoder is compressed and mapped to the same number of channels as the original NWP analysis field by passing it through the last linear two-dimensional convolutional layer (whose weights and biases are initialized to zero to ensure the stability of the identity mapping in the early stages of training). A systematic residual field characterizing the error distribution of various meteorological variables at high and low altitudes was obtained. ; Due to the topological characteristics of deep residual networks, the residual generator does not directly output the final predicted weather field. Instead, it generates the systematic residual field obtained in step 4 through element-wise addition. Superimposed on the multidimensional meteorological tensor retained in step 1 Above: ; This generates and outputs a final weather correction field with high physical fidelity. .

[0044] B. Discriminator The processing flow.

[0045] In each forward propagation of the adversarial game, the discriminator does not evaluate the output in isolation. Instead, it first receives a multidimensional meteorological input tensor as a large-scale environmental background. The discriminator then physically concatenates this multidimensional meteorological input tensor with either the meteorological correction field or the ERA5 ground truth field in the depth channel dimension. This concatenation mechanism ensures that the discrimination process is strictly constrained by the initial conditions of the original NWP forecast, requiring that the reconstructed meteorological field not only looks realistic but also closely matches the input background in terms of physical dynamics.

[0046] The concatenated high-dimensional tensors are sequentially fed into the network body, which consists of multiple second-dimensional convolutional layers with a stride of 2. A second instance normalization layer eliminates global contrast differences between different meteorological samples. Combined with the activation function in the second-band leakage correction linear unit, high-frequency local meteorological physical textures (such as strong wind shear zones and dramatic local temperature gradients) are extracted. Multi-level downsampling gradually expands the network's receptive field, implicitly dividing the global meteorological field into multiple overlapping local spatial patches.

[0047] Unlike traditional global discriminant networks that output a single scalar probability, this discriminator employs a convolutional neural network based on the PatchGAN structure, comprising: multiple sequentially connected second two-dimensional convolutional layers, a second instance normalization layer, a second linear unit with leakage correction, and a second linear two-dimensional convolutional layer. The final layer uses a 3×3 second linear two-dimensional convolutional layer with a padding parameter of 1, directly mapping and outputting a continuous two-dimensional confidence feature matrix. Each pixel value in this matrix independently represents the confidence probability that the corresponding local spatial patch in the original meteorological field conforms to the true atmospheric multidimensional spatial distribution characteristics of the ERA5 ground truth field. This local decision mechanism significantly enhances the discriminator's sensitivity to microscale meteorological distortions.

[0048] The multidimensional meteorological input tensor is concatenated with the meteorological correction field and the ERA5 ground truth field in the depth channel dimension to constrain the discrimination process to the initial conditions of the original NWP analysis field, resulting in a hybrid high-dimensional tensor. Multiple second-dimensional convolutional layers are used to downsample the hybrid high-dimensional tensor multiple times to extract features, gradually expanding the receptive field. After each downsampling, a second instance normalization layer eliminates global contrast differences between samples, and a second leakage-corrected linear unit is used for nonlinear activation to extract high-frequency meteorological physical textures of multi-scale local spatial features, obtaining multi-scale local spatial features. A second linear two-dimensional convolutional layer maps these multi-scale local spatial features, outputting a continuous two-dimensional confidence feature matrix. Each pixel value in the two-dimensional confidence feature matrix independently represents the confidence probability that the corresponding local spatial patch in the hybrid high-dimensional tensor conforms to the true atmospheric multidimensional spatial distribution characteristics in the ERA5 ground truth field. Based on the two-dimensional confidence feature matrix, asymmetric adversarial loss and pixel reconstruction loss are determined.

[0049] In the backpropagation and parameter optimization stages, an asymmetric label smoothing mechanism is first introduced to calculate the discriminator's own update loss: for the ERA5 ground truth field, the soft label target value is set to 0.9; for the meteorological correction field, the hard label target value is set to 0.0. This mechanism forces the discriminator not to be overconfident in the early stages of the game, thereby avoiding gradient vanishing in the residual generator.

[0050] Secondly, the discriminator performs a true / false mapping on the weather correction field again and calculates the mean square error between it and the target absolute true value of 1.0, thereby calculating the asymmetric adversarial loss across the network to guide the generator. : ; in, The mapping function representing the discriminator, It is a tensor matrix. For meteorological true values, For weather correction field, represents the total number of elements in the confidence matrix. The asymmetric design, where the discriminator is anchored at 0.9 while the generator aims for 1.0, forms the core mathematical basis of the asymmetric adversarial loss. After this loss is calculated, it is stripped and passed to the total loss function for subsequent parameter updates of the residual generator for nonlinear error approximation.

[0051] (4) Introduce geostrophic balance and asymmetric confrontation for joint training.

[0052] like Figure 2 As shown, the generator's total loss function during training. Pixel reconstruction loss Asymmetric adversarial losses and dynamic physical loss Composition, namely: ; in, , , These correspond to the weight coefficients of the loss function.

[0053] Pixel reconstruction loss uses L1 norm Meanwhile, dynamic physical loss Specifically targeting 500 hPa as the core guiding layer, the central finite difference method is used to approximate the height field of this layer. The spatial gradient of is expressed by the difference formula as:

[0054] ; ; Where, Δ x For the east-west direction, Δ y It is oriented north-south.

[0055] The model uses the L1 norm to constrain the zonal winds of the 500hPa core steering layer. and the wind To satisfy the geostrophic balance relationship, the formula is: ; in, It incorporates Coriolis parameters And the comprehensive scaling constant for differential grid unit conversion.

[0056] (5) Apply physical minimum threshold truncation to generate post-processed weather forecasts and perform spatial assessment.

[0057] Real-time acquired business model analysis data is fed into a trained artificial intelligence model to output a high-quality correction field that approximates the distribution of reanalysis data. This correction field serves as the ideal driving field for the large-scale artificial intelligence model to provide meteorological services. During the evaluation phase, a piecewise minimum truncation mechanism based on physical meaning is incorporated, assuming the original meteorological root mean square error is... The prediction error is The cutoff threshold is The relative improvement rate The piecewise calculation model is as follows:

[0058] When the root mean square error of the original meteorological data When the physical variable is determined to be approaching a vacuum state and has lost its mathematical benchmark for relative percentage evaluation, the output is forcibly truncated. ; When the root mean square error of the original meteorological data When the physical variable is deemed to have objective evaluative significance, the actual relative improvement rate is output according to the conventional proportion: .

[0059] A specific embodiment of the present invention is provided: (1) The input data is the NWP (e.g., NCEP) analysis field and ERA5 reanalysis data for a certain historical period (e.g., March 2025). To prevent the discriminator from converging too quickly with a small sample size (approximately 124 time-series samples), which could lead to the generator gradient vanishing, this invention employs a core asymmetric adversarial strategy where the discriminator's learning rate is lower than the generator's learning rate. The generator's learning rate is set to 10. -4 Set the learning rate of the discriminator to 10. -5 At the same time, an asymmetric label smoothing strategy is introduced for the discriminator. The loss function is defined as:

[0060] ; (2) In addition, the design of the physical loss mechanism includes the following implementation details: At 500 hPa, it is located in the middle troposphere and belongs to the free atmosphere. It is not affected by the friction of the ground and is also free from the interference of complex boundary layer thermal processes. Introducing physical constraints at this level can anchor the dynamic framework of large-scale circulation with minimal computational cost, avoiding the serious noise interference caused by introducing near-surface physical constraints.

[0061] Considering that the real atmosphere does not perfectly conform to the original equations, this invention uses geostrophic equilibrium as a soft constraint for directional guidance, and integrates scaling constants. Based on empirical values ​​for mid-latitude regions, the value was set to 0.01. Experiments confirmed that when... At this point, the model reaches the optimal balance between physical constraints and pixel fitting, with all channel variables consistently showing positive optimization. Thanks to the extremely small search space of the residual network and the guidance of the physical formula, the experimental results of this embodiment show that the model achieves optimal generalization performance on the independent test set at the 10th epoch; further training will cause the model to collapse. Therefore, for a single month's data volume, the optimal number of training epochs is fixed at 10 epochs in this embodiment. When expanding to other data scales, those skilled in the art can dynamically determine the optimal number of epochs based on an independent validation set using an early stopping mechanism.

[0062] (3) During the model generalization evaluation stage, the root mean square error (RMSE), mean absolute error (MAE), and spatial anomaly correlation coefficient (ACC) of each channel are calculated. The core task of this invention is to improve the ability to reconstruct the entire spatial structure of each independent time period in meteorological services, and to provide a high-fidelity initial field for a large model, rather than performing time series prediction on a single spatial grid point. By setting the prediction field of a single time period as The truth field is Spatial mean is and The total number of grids is The formula for calculating the spatial anomaly correlation coefficient at a single time interval is:

[0063] ; The ACC is calculated for each independent time interval (e.g., 60 independent test times) in the test set, and then the arithmetic mean is taken. This index accurately reflects the model's ability to correct for the phase position of large-scale weather systems. Meanwhile, for variables such as upper-level specific humidity that approximate a physical vacuum, due to the original error... Extremely small, so set By applying a piecewise threshold function, when the error is extremely small, the relative improvement rate is directly set to 0.0%.

[0064] The error reduction rate and correlation coefficient improvement results of the core meteorological elements are shown in Table 1. Based on the data in Table 1... Figures 3-19 The visualization results show that the absolute errors of various scale variables and near-surface core variables have decreased significantly. When the output meteorological correction field is used as the driving source, its RMSE is significantly lower than that of the original analysis field throughout the entire 24-hour forecast period.

[0065] Table 1. Results of Error Reduction Rate and Correlation Coefficient Improvement for Core Meteorological Elements Figure 3 , Figure 4 , Figure 5 , Figure 6 , Figure 7 , Figure 8 , Figure 9 , Figure 10 and Figure 11 Images (a), (b), and (c) show the original NWP analysis field (original NCEP), the meteorological correction field output by this model, and the true atmospheric observation field (ERA5 ground truth), respectively. The image comparison fully demonstrates that after processing by this model, the spatial texture, high-frequency disturbances, and macro-weather system outlines of meteorological elements closely approximate the true atmospheric state. Figure 3 , Figure 4 , Figure 5 , Figure 6 , Figure 7 , Figure 8 , Figure 9 , Figure 10 and Figure 11Figures (d), (e), and (f) show the spatial distribution of the absolute error before and after the correction, respectively, and indicate the magnitude of the reduction in the global root mean square error (RMSE). The red area in (f) represents the successful reduction of error, clearly revealing the spatial characteristics of the model in eliminating systematic forecast bias.

[0066] Figure 12 , Figure 13 , Figure 14 , Figure 15 , Figure 16 , Figure 17 , Figure 18 and Figure 19 Figures (a) and (b) show the divergence trend of the root mean square error (RMSE) and the decay trend of the spatial anomaly correlation coefficient (ACC) as the forecast lead time increases, respectively.

[0067] Comparing the curves, it can be seen that when the meteorological correction field output by this invention is used as the driving source (red line in the figure), its RMSE is significantly lower than that of the original analysis field (blue line in the figure) throughout the entire 24-hour forecast period, and its ACC is significantly higher. The overall trend of the red line is very close to the theoretical upper limit based on the ideal true value driving (green line in the figure, ERA5 true value).

[0068] (4) Under the premise that the global meteorological field has very strong spatial stability, the global mean is calculated by randomly selecting 100 sample times. with standard deviation Mathematically, this achieves an unbiased estimate of the global expectation, mitigating risks such as memory overflow caused by reading full time series data into memory. Furthermore, the NWP analysis field, as an assimilation product of the numerical model, possesses the inherent terrain parameter bias of that model. The embodiment extracts the error-correcting residual mode of this system bias using only single-month data, which can then be directly applied to other completely independent unknown time series, such as independent test sets across years, for generalization testing. This residual learning mechanism reduces the absolute dependence of traditional deep learning meteorological models on long-sequence, large-sample climatological training data, allowing the model to output a high-quality meteorological field approaching ERA5 accuracy even with only real-time, low-precision forecast data as input, thus possessing significant practical operational deployment value.

[0069] As can be seen from the above embodiments, the beneficial effects that the present invention can achieve include, but are not limited to: (1) By extracting systematic deviation modes from high-precision reanalysis data, intelligent correction is performed on low-precision but real-time operational analysis fields. This effectively solves the problem of degradation of meteorological large model forecasting effect caused by driving field deviation, and realizes the implementation of meteorological service operations with real-time acquisition and high-precision forecasting.

[0070] (2) The physical formula of geostrophic balance relationship is deeply integrated with the loss function of deep learning. Based on the central finite difference method, the disordered search of the pure data-driven model in the solution space is effectively restricted, and the technical pain point of negative optimization in the wind field inversion process is solved.

[0071] (3) By adopting a residual learning architecture, the model avoids fitting to absolute meteorological conditions and instead learns the stable systematic truncation error and terrain parameterization bias of the NWP model. It has strong data temporal stability in this high-frequency error correction mode, which allows the model to achieve high-precision generalization on a completely independent time-series test set with only a small number of samples, greatly reducing the dependence on large-scale historical climate datasets.

[0072] (4) The innovative introduction of a segmented truncation model based on the physical minimum threshold eliminates the false error amplification caused by dividing by near-zero values ​​in the high-altitude water vapor vacuum zone, thus ensuring the rigor of the objective evaluation system.

[0073] Based on the same inventive concept, embodiments of the present invention provide a meteorological field error correction system for large meteorological models, the system comprising: The tensor construction module is used to obtain the original NWP analysis field and ERA5 ground truth field at multiple historical moments and construct a multidimensional meteorological input tensor.

[0074] The input module is used to input the multidimensional meteorological input tensor into the error correction model. The error correction model includes a residual generative adversarial network and a physical constraint module. The residual generative adversarial network includes a residual generator and a discriminator.

[0075] The correction module is used to predict the systematic deviation between the original NWP analysis field and the ERA5 true field through the residual generator to obtain the systematic residual field; the systematic residual field is then spliced ​​with the original NWP analysis field to obtain the meteorological correction field.

[0076] The loss construction module is used to determine whether the meteorological correction field conforms to the true atmospheric multidimensional spatial distribution characteristics of the ERA5 ground truth field through a discriminator, thereby obtaining asymmetric adversarial loss and pixel reconstruction loss. The core guiding layer variables of the meteorological correction field are extracted through the physical constraint module to introduce geostrophic balance physical constraints into the meteorological correction field, thereby obtaining dynamic physical loss.

[0077] The update module is used to update the model parameters of the residual generator through backpropagation based on the total loss obtained by weighted fusion of asymmetric adversarial loss, pixel reconstruction loss and dynamic physical loss, so as to obtain the trained residual generator.

[0078] The inference output module is used to input the real-time NWP analysis field into the trained residual generator to obtain the corrected meteorological correction field.

[0079] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A method for correcting meteorological field errors in large-scale meteorological models, characterized in that, include: Obtain the original NWP analysis field and ERA5 ground truth field at multiple historical moments, and construct a multidimensional meteorological input tensor; The multidimensional meteorological input tensor input error correction model includes: a residual generative adversarial network and a physical constraint module. The residual generative adversarial network includes: a residual generator and a discriminator. The systematic deviation between the original NWP analysis field and the ERA5 true field is predicted by the residual generator to obtain the systematic residual field; the systematic residual field is then spliced ​​with the original NWP analysis field to obtain the meteorological correction field. The discriminator determines whether the meteorological correction field conforms to the true atmospheric multidimensional spatial distribution characteristics of the ERA5 true field, thereby obtaining asymmetric adversarial loss and pixel reconstruction loss. The physical constraint module extracts the core guiding layer variables of the meteorological correction field to introduce geostrophic balance physical constraints into the meteorological correction field, thereby obtaining dynamic physical loss. Based on the total loss function obtained by weighted fusion of the asymmetric adversarial loss, the pixel reconstruction loss, and the dynamic physical loss, the model parameters of the residual generator are updated through backpropagation to obtain the trained residual generator; The real-time NWP analysis field is input into the trained residual generator to obtain the corrected meteorological field.

2. The meteorological field error correction method for large meteorological models as described in claim 1, characterized in that, The process of acquiring the original NWP analysis field and ERA5 ground truth field at multiple historical moments and constructing a multidimensional meteorological input tensor specifically includes: Both the original NWP analysis field and the ERA5 true field include: meteorological variables at different altitudes and surface variables; the meteorological variables at different altitudes are used to characterize the thermal, dynamic and water vapor states of the atmosphere in three-dimensional space, and the surface variables are used to characterize near-surface meteorological elements and circulation characteristics; By splicing meteorological and surface variables at different altitude levels along the channel dimension and introducing digital elevation model data as an independent channel, a multidimensional meteorological tensor containing both spatial and channel dimensions is obtained.

3. The meteorological field error correction method for large meteorological models as described in claim 1, characterized in that, Also includes: Before applying the multidimensional meteorological input tensor to the input error correction model, the multidimensional meteorological input tensor undergoes timestamp-aware adaptive data normalization processing, specifically including: The global mean and standard deviation of each channel of the multidimensional meteorological input tensor are determined based on the following formula: ； ； The multidimensional meteorological input tensor is normalized and mapped based on the following formula: ； in, Let be the global mean of the c-th physical channel. Let c be the standard deviation of the c-th physical channel. H The latitude dimension of the latitude and longitude spatial grid. W For the longitude dimension of the latitude and longitude spatial grid, N The total number of samples in the time dimension. For the normalized result, For multidimensional meteorological input tensors, It is a tensor matrix.

4. The meteorological field error correction method for large meteorological models as described in claim 1, characterized in that, The step of predicting the systematic deviation between the original NWP analysis field and the ERA5 true field using the residual generator to obtain the systematic residual field specifically includes: The residual generator adopts a fully convolutional encoding and decoding architecture, including: an encoder and a decoder; the encoder includes: a plurality of first two-dimensional convolutional layers, a first instance normalization layer and a first linear unit with leakage correction connected in sequence; the decoder includes: a plurality of deconvolutional layers, a skip connection module and a first linear two-dimensional convolutional layer connected in sequence. The multidimensional meteorological input tensor is feature-encoded by multiple first two-dimensional convolutional layers to extract spatial texture information of atmospheric physical field and obtain encoded features; the encoded features are normalized and nonlinearly activated by the first instance normalization layer and the first leakage correction linear unit to gradually compress the spatial resolution of the encoded features and obtain a deep feature map representing large-scale weather modal features. The deep feature map is upsampled multiple times by multiple deconvolution layers to gradually restore the spatial resolution of the deep feature map. After each upsampling, the high-dimensional feature map of the same resolution at the corresponding level in the encoder is concatenated with the output feature map of the current decoder in the channel dimension through the skip connection module to supplement microscale meteorological spatial information and obtain the highest resolution feature map. The highest resolution feature map is mapped to the same number of channels as the original NWP analysis field by the first linear two-dimensional convolutional layer, thus obtaining a systematic residual field characterizing the error distribution of various meteorological variables at high and low altitudes.

5. The meteorological field error correction method for large meteorological models as described in claim 1, characterized in that, The discriminator determines whether the meteorological correction field conforms to the true atmospheric multidimensional spatial distribution characteristics of the ERA5 ground truth field, obtaining asymmetric adversarial loss and pixel reconstruction loss, specifically including: The discriminator employs a convolutional neural network based on the PatchGAN structure, comprising: multiple second two-dimensional convolutional layers, a second instance normalization layer, a second linear unit with leakage correction, and a second linear two-dimensional convolutional layer connected in sequence. The multidimensional meteorological input tensor is concatenated with the meteorological correction field and the ERA5 ground truth field in the depth channel dimension to constrain the discrimination process to the initial conditions of the original NWP analysis field, thus obtaining a hybrid high-dimensional tensor. Multiple downsampling features are extracted from the hybrid high-dimensional tensor through multiple second two-dimensional convolutional layers to gradually expand the receptive field; and after each downsampling, the global contrast difference between samples is eliminated by the second instance normalization layer, and nonlinear activation is performed by the second leakage-corrected linear unit to extract high-frequency meteorological physical texture of multi-scale local spatial features, thereby obtaining multi-scale local spatial features. The multi-scale local spatial features are mapped by the second linear two-dimensional convolutional layer to output a continuous two-dimensional confidence feature matrix; each pixel value in the two-dimensional confidence feature matrix independently represents the confidence probability that the corresponding local spatial patch in the hybrid high-dimensional tensor conforms to the real atmospheric multi-dimensional spatial distribution features in the ERA5 ground truth field. Based on the two-dimensional confidence feature matrix, the asymmetric adversarial loss and the pixel reconstruction loss are determined.

6. The meteorological field error correction method for large meteorological models as described in claim 5, characterized in that, The asymmetric adversarial loss is determined based on the following formula: ； The pixel reconstruction loss is determined based on the following formula: ； The dynamic physical loss is determined based on the following formula: ； The total loss function, obtained by weighted fusion of the asymmetric adversarial loss, the pixel reconstruction loss, and the dynamic physical loss, is determined based on the following formula: ； in, For the total loss function, For pixel reconstruction loss, For asymmetric adversarial losses, For dynamic physical loss, These are the weighting coefficients for pixel reconstruction loss. These are the weighting coefficients for asymmetric adversarial loss. These are the weighting coefficients for dynamic physical loss. For meteorological true values, For meteorological correction field, The mapping function representing the discriminator, It is a tensor matrix. This represents the total number of elements in the confidence matrix. This is a comprehensive scaling constant. The wind is zonal within the 500hPa core steering layer. Meridional winds at 500 hPa core guiding layer.

7. The meteorological field error correction method for large meteorological models as described in claim 6, characterized in that, The dynamic physical loss is constructed for the 500hPa core guiding layer, and the spatial gradient of the height field is approximated using the central finite difference method based on the following formula: ； ； in, The height field corresponding to the core guiding layer at 500 hPa, Δ x For east-west direction, Δ y It is oriented north-south.

8. The meteorological field error correction method for large meteorological models as described in claim 1, characterized in that, Also includes: After obtaining the corrected meteorological field, physical minimum truncation is performed using a built-in piecewise minimum truncation determination mechanism based on physical meaning. This process includes: Determine the original root mean square error and the post-prediction error of the target physical channel respectively; If the original meteorological root mean square error is less than the preset physical minimum truncation threshold, it is determined that the variable is close to the physical vacuum state, and the truncation mechanism is directly triggered, forcing the output relative improvement rate to be 0. If the root mean square error of the original meteorological data is greater than the preset physical minimum cutoff threshold, then the variable is determined to have objective physical evaluation significance, and the true relative improvement rate is determined based on the following formula: ； in, This represents the root mean square error of the original meteorological data. For the prediction error, This refers to the relative increase rate.

9. A meteorological field error correction system for large-scale meteorological models, characterized in that, include: The tensor construction module is used to obtain the original NWP analysis field and ERA5 truth field at multiple historical moments and construct a multidimensional meteorological input tensor. The input module is used to input the multidimensional meteorological input tensor into the error correction model. The error correction model includes a residual generative adversarial network and a physical constraint module. The residual generative adversarial network includes a residual generator and a discriminator. The correction module is used to predict the systematic deviation between the original NWP analysis field and the ERA5 true field through the residual generator to obtain a systematic residual field; and to splice the systematic residual field with the original NWP analysis field to obtain a meteorological correction field. The loss construction module is used to determine whether the meteorological correction field conforms to the real atmospheric multidimensional spatial distribution characteristics of the ERA5 true field through the discriminator, and obtain asymmetric adversarial loss and pixel reconstruction loss; the physical constraint module extracts the core guiding layer variables of the meteorological correction field to introduce geostrophic balance physical constraints into the meteorological correction field, and obtains dynamic physical loss. The update module is used to update the model parameters of the residual generator through backpropagation based on the total loss obtained by weighted fusion of the asymmetric adversarial loss, the pixel reconstruction loss and the dynamic physical loss, so as to obtain the trained residual generator. The inference output module is used to input the real-time NWP analysis field into the trained residual generator to obtain the corrected meteorological correction field.

Images