A method and system for fine reconstruction of tropospheric vertical structure considering regional characteristics

Abstract

The application discloses a troposphere vertical structure fine reconstruction method considering regional characteristics, and comprises the following steps: obtaining troposphere wet refractive index errors of each isobaric layer by using a pre-trained meteorological large model, and obtaining gridded vertical error altitudes based on the troposphere wet refractive index errors of each isobaric layer; based on the gridded vertical error altitudes, specific regional terrain factors and physical constraints, a freezing-fine tuning strategy is used to regionally refine the pre-trained meteorological large model to obtain a regionally refined meteorological large model, and the troposphere wet refractive index of each isobaric layer is reconstructed; the troposphere wet refractive index of each isobaric layer is mapped to the troposphere wet refractive index of a fixed geometric height layer, and a numerical integration method and an error propagation law are used to generate a real-time troposphere wet delay vertical structure grid product covering a specific region and having uncertainty identification. Based on this, the application realizes fine reconstruction of a specific regional real-time high-resolution wet delay vertical structure.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting technology, specifically to a method and system for fine reconstruction of the vertical structure of the troposphere that takes into account regional characteristics. Background Technology

[0002] The tropospheric atmosphere exhibits significant non-uniform stratification and time-varying characteristics in the vertical direction, and accurate representation of these characteristics is the core foundation for high-precision three-dimensional tropospheric wet-delay modeling. Traditional research relies heavily on empirical physical models, reanalysis data, or numerical weather prediction models to obtain temperature, pressure, humidity fields, and tropospheric parameters. However, while empirical models are simple in form, they struggle to capture transient characteristics under complex terrain and drastic weather changes; reanalysis data, such as ERA5, offers high accuracy but suffers from data release delays; and while numerical weather prediction models possess complete physical mechanisms, their assimilation and computation processes are complex and time-delayed, making it difficult to meet the demands for detailed real-time vertical structure modeling of the troposphere. In recent years, large-scale meteorological models, such as GraphCast and Pangu-Weather, have rapidly developed, providing detailed three-dimensional meteorological fields globally, offering an opportunity for real-time, high-resolution tropospheric vertical structure modeling. Related research indicates that large-scale meteorological models can not only learn statistical patterns from massive historical samples but also implicitly express the physical characteristics of atmospheric vertical structure to a certain extent, thus providing more refined prior constraints for wet-delay tropospheric vertical structure modeling. GraphCast supports high-resolution pressure layer inputs of up to 37 layers in ERA5, outperforming large meteorological models with fewer layers in terms of vertical structure representation. Studies have found that tropospheric delay calculated using GraphCast outperforms traditional empirical models and large meteorological models with fewer vertical layers in terms of global bias and root mean square error. Therefore, GraphCast offers new possibilities for real-time, high-vertical-resolution representation of tropospheric wet delay structure.

[0003] While large-scale meteorological models offer the possibility of finely representing the vertical structure of tropospheric wet delay, several key issues remain in current research. First, current evaluations of large-scale meteorological models largely focus on the accuracy of tropospheric delay across the entire layer, lacking a description of how deviations vary with altitude across the entire vertical hierarchy, particularly struggling to identify structural representation defects within key altitude ranges such as the boundary layer. Second, existing large-scale meteorological models are mostly trained on global-scale data, and there are significant differences between the global prior background and specific complex terrains and climates. This results in deviations between the output regional temperature, pressure, and humidity fields and the actual data, lacking refinement and transfer mechanisms tailored to specific regional characteristics. Third, large-scale meteorological models typically output temperature, pressure, and humidity fields at dynamic pressure layers, while high-precision GNSS applications are based on geometrical altitude layers. Existing altitude conversion methods often ignore the spatial variations and dynamic fluctuations of isobaric surface altitudes under complex terrain and different latitudes and longitudes, limiting the accuracy of vertical modeling.

[0004] In summary, while current large-scale meteorological models offer new approaches to representing high-vertical-resolution wet delay structures, significant shortcomings remain in characterizing complete vertical hierarchical biases, adapting to regional migrations, and accurately mapping pressure layers to geometrical height layers. Therefore, revealing the variation patterns of vertical structural biases in large-scale meteorological models under different altitudes and complex terrain conditions, and utilizing the global forecast prior information and regional refinement techniques of large-scale meteorological models to achieve real-time, high-resolution wet delay vertical structure modeling for specific regions, has become a crucial area requiring breakthroughs in three-dimensional tropospheric wet delay modeling. Summary of the Invention

[0005] To overcome the problem of insufficient accuracy of tropospheric wet-delay vertical structure under specific complex terrain and large elevation differences in existing technologies, this invention provides a method and system for fine reconstruction of tropospheric vertical structure that takes into account regional characteristics. By revealing the variation law of vertical structure deviation of meteorological large models under different altitude layers and complex terrain conditions, and utilizing the global forecast prior information of meteorological large models and regional fine-tuning technology, the fine reconstruction of real-time high-resolution wet-delay vertical structure of target areas is achieved.

[0006] According to one aspect of the present invention, a method for fine reconstruction of tropospheric vertical structure considering regional characteristics is provided, comprising: step S1, inputting global meteorological data into a pre-trained meteorological large model, outputting tropospheric wet refractive index prediction data for each isobaric layer, obtaining the tropospheric wet refractive index error of each isobaric layer based on the tropospheric wet refractive index prediction data for each isobaric layer, and obtaining the gridded vertical error elevation based on the tropospheric wet refractive index error of each isobaric layer; step S2, based on the gridded vertical error elevation, the topographic factor of the target area, and physical constraints, using a freeze-fine-tuning strategy to refine the pre-trained meteorological large model regionally to obtain a regionally refined meteorological large model, and using the regionally refined meteorological large model to reconstruct the tropospheric wet refractive index of each isobaric layer; step S3, mapping the reconstructed tropospheric wet refractive index of each isobaric layer to the tropospheric wet refractive index of a fixed geometric height layer, and using a numerical integration method and the error propagation law to generate a real-time tropospheric wet delay vertical structure grid product covering the target area and having an uncertainty indicator.

[0007] Furthermore, global meteorological data is input into a pre-trained meteorological model, which outputs tropospheric wet refractive index prediction data for each isobaric layer. Based on the tropospheric wet refractive index prediction data for each isobaric layer, the tropospheric wet refractive index error for each isobaric layer is obtained. This includes: using global meteorological data to drive the pre-trained meteorological model to perform autoregressive rolling predictions, generating a global real-time meteorological prediction grid for a preset future time period, and calculating the tropospheric wet refractive index prediction grid based on atmospheric physics equations; using bilinear and log-linear interpolation techniques, the tropospheric wet refractive index prediction grid is mapped with high precision to the radiosonde stations and each isobaric layer in the target area to obtain the tropospheric wet refractive index prediction data for each radiosonde station on each isobaric layer; by comparing the difference between the tropospheric wet refractive index prediction data for each radiosonde station on each isobaric layer and the measured data of the tropospheric wet refractive index station, the tropospheric wet refractive index error for each isobaric layer is obtained.

[0008] Furthermore, the gridded vertical error elevation is obtained based on the tropospheric wet refractive index error of each isobaric layer, including: fitting the tropospheric wet refractive index error of each isobaric layer with an exponential function to obtain the error elevation parameters at each discrete station; and using a three-dimensional inverse distance weighting method to spatially extrapolate the error elevation parameters extracted at each discrete station to obtain the gridded vertical error elevation.

[0009] Further, step S2 includes: calculating the topographic factors of the target area based on the digital elevation model data of the target area, and inputting the topographic factors of the target area into the pre-trained meteorological big model, so that the meteorological big model constructs a static topographic feature tensor based on the topographic factors of the target area, and fuses the static topographic feature tensor with the original static geographic features of the meteorological big model to obtain fused static geographic features; using the gridded vertical error elevation as prior information, constructing a weighted loss matrix that adaptively decays with height, and then constructing a custom loss function that fuses physical constraints and height weighting based on the weighted loss matrix; combining the fused static geographic features and the custom loss function, using a freeze-fine-tuning strategy to refine the pre-trained meteorological big model regionally to obtain a regionally refined meteorological big model; using the regionally refined meteorological big model to obtain a real-time high-precision meteorological field, and combining the real-time high-precision meteorological field and atmospheric physics equations to reconstruct the tropospheric wet refractive index of each isobaric layer.

[0010] Furthermore, combining the fused static geographic features and the custom loss function, a freeze-fine-tuning strategy is used to refine the pre-trained meteorological big model regionally, resulting in a regionally refined meteorological big model. This includes: freezing the intermediate layer representing global fluid dynamics in the meteorological big model, unlocking only the input encoder and output decoder, and fine-tuning the model parameters in the meteorological big model using the fused static geographic features and the custom loss function to obtain the regionally refined meteorological big model. The meteorological big model includes an input encoder, an output decoder, and an intermediate layer located between them.

[0011] Furthermore, using the gridded vertical error elevation as prior information, a weighted loss matrix that adaptively decays with height is constructed. Then, based on this weighted loss matrix, a custom loss function fusing physical constraints and height weighting is constructed, with the corresponding formula being:

[0012] ,

[0013] ,

[0014] In the formula, h is the geometric height corresponding to the current pressure layer. To represent the vertical error elevation in the grid, To adjust the hyperparameters of physical constraint strength, To integrate physical constraints with a highly weighted custom loss function, For mean square error loss, and These are two major physical constraints. This is the weighted loss matrix.

[0015] Further, step S3 includes: step S31, obtaining the prior error covariance matrix before adjustment by statistically analyzing the tropospheric wet refractive index error of each isobaric layer under the same verification time and different forecast lead times, and inverting the prior error covariance matrix to generate the prior weight matrix required for least squares solution; step S32, performing weighted least squares adjustment based on the reconstructed tropospheric wet refractive index and prior weight matrix of each isobaric layer to obtain the tropospheric wet refractive index and posterior error covariance matrix at a fixed geometric height layer; step S33, calculating the interlayer tropospheric wet delay of each fixed geometric height layer using a numerical integration method based on the tropospheric wet refractive index at the fixed geometric height layer, and simultaneously calculating the uncertainty of each interlayer tropospheric wet delay according to the error propagation law and the posterior error covariance matrix, thereby generating a real-time tropospheric wet delay vertical structure grid product covering the target area and having uncertainty indicators.

[0016] According to one aspect of the present invention, a fine-scale reconstruction system for the vertical structure of the troposphere that takes into account regional characteristics is provided, comprising: a priori error quantitative assessment module, used to input global meteorological data into a pre-trained meteorological large model, output tropospheric wet refractive index prediction data for each isobaric layer, obtain the tropospheric wet refractive index error of each isobaric layer based on the tropospheric wet refractive index prediction data for each isobaric layer, and obtain the gridded vertical error elevation based on the tropospheric wet refractive index error of each isobaric layer; and a model retraining fine-tuning module, used to obtain the gridded vertical error elevation, target vertical error elevation, and other relevant data. Based on the topographic factors and physical constraints of the target area, a freeze-fine-tuning strategy is used to refine the pre-trained meteorological model to obtain a refined meteorological model. The refined meteorological model is then used to reconstruct the tropospheric wet refractive index of each isobaric layer. The wet delay mapping reconstruction module is used to map the reconstructed tropospheric wet refractive index of each isobaric layer to the tropospheric wet refractive index of a fixed geometric height layer. A real-time tropospheric wet delay vertical structure grid product covering the target area and equipped with uncertainty indicators is generated using numerical integration method and error propagation law.

[0017] According to one aspect of the present invention, an electronic device is provided, including a memory and a processor, the memory storing program instructions executed by the processor, the processor invoking the program instructions to perform the aforementioned method for fine reconstruction of tropospheric vertical structure taking into account regional characteristics.

[0018] According to one aspect of the present invention, a non-transitory computer-readable storage medium is provided, the non-transitory computer-readable storage medium storing computer instructions that cause the computer to perform the aforementioned method for fine reconstruction of tropospheric vertical structure taking into account regional characteristics.

[0019] The above technical solution first utilizes a pre-trained meteorological large-scale model to quantitatively evaluate the tropospheric wet refractive index error of each isobaric layer, and then obtains a vertical error elevation model based on the tropospheric wet refractive index error of each isobaric layer. This vertical error elevation model reveals the variation law of the vertical structure deviation of the meteorological large-scale model at different altitudes. Next, based on the vertical error elevation model, a gridded vertical error elevation is obtained. This gridded vertical error elevation will serve as a key weight parameter in the subsequent re-fine-tuning training of the meteorological large-scale model, guiding the model to implement differentiated corrections at different geographical locations and altitudes. Finally, based on the gridded vertical error... The differential elevation, based on the gridded vertical error elevation, target area topographic factors, and physical constraints, employs a freeze-fine-tuning strategy to refine the pre-trained meteorological model regionally, resulting in a refined meteorological model. The tropospheric wet refractive index of each isobaric layer is then reconstructed using this model. While the pre-trained meteorological model is adapted to global weather forecasts, significant differences exist between the global prior background and specific complex terrains and climates. Consequently, the output temperature, pressure, and humidity fields for the target area deviate from the actual data. Therefore, this step achieves regional adaptation of the meteorological model parameters from a globally universal form to the local atmospheric characteristics of the target area. Finally, the reconstructed tropospheric wet refractive index of each isobaric layer is mapped to the tropospheric wet refractive index of a fixed geometric height layer. Numerical integration and the error propagation law are then used to generate a real-time tropospheric wet delay vertical grid product covering the target area and equipped with uncertainty indicators.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0021] (1) Based on radiosonde data, the prediction bias of the GraphCast prior field in the target area was evaluated, its systematic bias and distribution characteristics in 37 isobaric layers were analyzed, a vertical error elevation model was constructed, and the variation law of vertical structural bias of the meteorological large model at different altitude layers was revealed.

[0022] (2) Construct a transfer learning framework that integrates physical constraints with topographic factors of the target area such as elevation and slope. Update the weights of the input encoder and output decoder in the meteorological big model through the freeze-fine-tuning strategy to achieve regional adaptation of the meteorological big model parameters from the global general form to the local atmospheric characteristics of the target area.

[0023] (3) By constructing an accurate mapping algorithm from dynamic isobaric layers to fixed geometric height layers, and by using weighted least squares and the error propagation law, the real-time tropospheric wet delay vertical structure product and its accuracy information of the target area are generated. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram illustrating the principle of a fine reconstruction method for the vertical structure of the troposphere that takes into account regional characteristics, provided in an embodiment of the present invention.

[0026] Figure 2 This is a flowchart of a method for fine reconstruction of the tropospheric vertical structure that takes into account regional characteristics, provided in an embodiment of the present invention. Detailed Implementation

[0027] It should be noted that:

[0028] The terms “comprising” and “having”, and any variations thereof, in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, product, or apparatus that includes a series of steps or units, not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0029] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices. The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be decomposed, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form new technical solutions. Such combinations are not bound by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0031] Please refer to the appendix. Figure 1 and Figure 2 This invention provides a method for fine reconstruction of tropospheric vertical structure that takes into account regional characteristics. Addressing the problem of insufficient accuracy in reconstructing the wet-delayed tropospheric vertical structure under specific complex terrain and large elevation differences, this invention proposes to conduct research on a real-time wet-delayed tropospheric vertical structure fine reconstruction method based on a large meteorological model (in this embodiment, the GraphCast model is used). ① Based on radiosonde data, the prediction bias of the GraphCast model's prior field in the target region is evaluated, and its systematic bias and distribution characteristics across 37 isobaric layers are analyzed. A vertical error elevation model is constructed, revealing the law of model error decay with altitude. ② A transfer learning framework integrating physical constraints and topographic factors such as elevation and slope is constructed. The weights of the input encoder and output decoder are updated through a freeze-fine-tuning strategy, achieving regional adaptation of the GraphCast model parameters from a globally universal form to the local atmospheric characteristics of the target region. ③ An accurate mapping algorithm from dynamic isobaric layers to fixed geometric height layers is studied. Through weighted least squares and the error propagation law, a gridded real-time tropospheric wet-delayed vertical structure product and its accuracy information for the target region are generated.

[0032] The GraphCast large-scale model is a data-driven weather forecasting model based on graph neural networks. It can efficiently predict meteorological parameters of 37 isobars at a resolution of 0.25°, and its high-precision prediction results provide reliable prior information for fine tropospheric reconstruction. However, this GraphCast large-scale model is mainly trained on ERA5 reanalysis data at a global scale, and has systematic biases when dealing with complex terrain in the target area. To address this, this invention constructs a transfer learning framework that integrates physical constraints and topographic factors to correct the meteorological field of the target area, rigorously mapping dynamic isobaric parameters to a fixed geometric height layer, and outputting an interlayer zenith tropospheric wet delay with uncertainty estimation. The specific technical approach is as follows:

[0033] Step S1: Input global meteorological data into the pre-trained meteorological model, output tropospheric wet refractive index prediction data for each isobaric layer, obtain the tropospheric wet refractive index error for each isobaric layer based on the tropospheric wet refractive index prediction data for each isobaric layer, and obtain the gridded vertical error elevation based on the tropospheric wet refractive index error for each isobaric layer.

[0034] It should be noted that the troposphere is the entire atmospheric region, while the isobaric layer is a vertical stratification of the troposphere based on air pressure. The two are related as a whole and its stratified units. Tropospheric wet refractive index error refers to the error between the tropospheric wet refractive index predicted by the large meteorological model and the actual tropospheric wet refractive index. The specific steps of step S1 are as follows:

[0035] Step S11: Input global meteorological data into the pre-trained meteorological model, output the predicted tropospheric wet refractive index of each isobaric layer, and obtain the tropospheric wet refractive index error of each isobaric layer based on the predicted tropospheric wet refractive index of each isobaric layer.

[0036] In step S11, pre-training refers to global-scale pre-training. During the pre-training phase, the large-scale meteorological model uses historical meteorological data from around the world, including the spatiotemporal distribution of meteorological elements such as temperature, humidity, air pressure, and wind fields in various regions, enabling the model to learn the physical laws governing global atmospheric dynamics. Based on this, this invention first uses global meteorological data (including but not limited to ERA5 barosphere data) to drive the pre-trained large-scale meteorological model to perform autoregressive rolling predictions, generating a global real-time meteorological prediction grid for a future preset time period (the preset time period can be limited according to actual needs, such as the next 120 hours). Then, it calculates the tropospheric wet refractive index prediction grid based on atmospheric physics equations. Using bilinear and log-linear interpolation techniques, the tropospheric wet refractive index prediction grid is mapped with high precision to the radiosonde stations and 37 isobars in the target area to obtain the tropospheric wet refractive index prediction data for each station on each isobar. Finally, the measured tropospheric wet refractive index data is extracted from the measured meteorological data of the radiosonde stations in the target area. By comparing the model prediction values ​​(i.e., predicted tropospheric wet refractive index data) and the actual measured values ​​(i.e., measured tropospheric wet refractive index data) of each isobaric layer at the radiosonde station, the tropospheric wet refractive index error of each isobaric layer can be obtained.

[0037] Step S12: The tropospheric wet refractive index error of each isobaric layer is fitted using an exponential function to obtain the error elevation parameters at each discrete station.

[0038] In step S12, the tropospheric wet refractive index error of each isobaric layer is a discrete value, which can only reflect the magnitude of the error at a specific altitude and cannot directly describe how the error evolves with altitude. Furthermore, since the forecast error of meteorological elements usually shows an exponential decay trend with increasing altitude, this invention introduces the concept of error elevation and uses an exponential function to fit the error of each isobaric layer to obtain a vertical error elevation model. This vertical error elevation model is used to quantify the vertical evolution characteristics of the forecast error of meteorological elements (i.e., the error of the tropospheric wet refractive index predicted by the large meteorological model) with altitude.

[0039] In this embodiment, an exponential function is used to fit the root mean square error (RMSE) of each isobaric layer, and the formula is as follows:

[0040] (1)

[0041] In the formula, RMSE(h) represents the root mean square error at height h, and the core parameter is... The error elevation parameter represents the station, where h represents the height, and A and B are the fitting coefficients.

[0042] Step S13: Using the three-dimensional inverse distance weighting method, the error elevation parameters extracted at each discrete station are spatially extrapolated to obtain the gridded vertical error elevation.

[0043] Step S13 aims to perform spatial extrapolation of error features. Considering the sparsity of the spatial distribution of radiosonde stations, this invention uses elevation as the third-dimensional coordinate and employs a three-dimensional inverse distance weighting method to spatially extrapolate the error elevation parameters extracted from each discrete station (e.g., extrapolate to a 0.25° continuous grid). Through step S13, this invention constructs a priori field of error features (i.e., gridded vertical error elevation) covering the target area and possessing regional difference information. This gridded vertical error elevation will be directly input as a key weight parameter into the transfer learning framework of step S2, guiding the meteorological large model to implement differentiated corrections at different geographical locations and altitude levels.

[0044] Step S2: Based on the gridded vertical error elevation, the topographic factor of the target area, and physical constraints, the pre-trained meteorological large model is refined regionally using a freeze-fine-tuning strategy to obtain a refined meteorological large model. The refined meteorological large model is then used to reconstruct the tropospheric wet refractive index of each isobaric layer.

[0045] In step S2, to achieve accurate adaptation of global forecast parameters to specific complex terrain regions, this invention constructs a transfer learning framework that integrates topographic factors, physical constraints, and height weighting of the target region. The regional refinement of the large meteorological model includes:

[0046] Step S21: Calculate the topographic factors of the target area based on the digital elevation model data of the target area, and use the topographic factors of the target area as the input features of the meteorological big model, so that the meteorological big model can construct a static topographic feature tensor based on the topographic factors of the target area, and fuse the static topographic feature tensor with the original static geographic features of the meteorological big model to obtain fused static geographic features.

[0047] In this embodiment, based on the target area digital elevation model data (i.e., the target area DEM), the target area topographic factors such as slope, aspect, and topographic relief are calculated. The target area topographic factors are then input into the pre-trained meteorological big model (i.e., the global big model) so that the meteorological big model can construct a 4-channel static topographic feature tensor based on the target area topographic factors. This static topographic feature tensor is then fused with the original static geographic features of the meteorological big model, so that the meteorological big model can capture prior information about local topography while retaining the global-scale pre-trained features.

[0048] Step S22: Using the gridded vertical error elevation as prior information, construct a weighted loss matrix that adaptively decays with height, and then construct a custom loss function that integrates physical constraints and height weighting based on the weighted loss matrix.

[0049] Step S22 aims to use the gridded vertical error elevation from step S1 as prior information to construct a weighted loss matrix that adaptively decays with height, and incorporate it into the loss function. The corresponding formula is:

[0050] (2)

[0051] (3)

[0052] In the formula, h is the geometric height corresponding to the current pressure layer. To represent the vertical error elevation in the grid, To adjust the hyperparameters of physical constraint strength, To integrate physical constraints with a highly weighted custom loss function, For mean square error loss, and These are two major physical constraints. This is the weighted loss matrix.

[0053] Step S23: Combining the fused static geographic features and the custom loss function, the meteorological big model is refined regionally using a freeze-fine-tuning strategy to obtain the regionally refined meteorological big model.

[0054] It should be noted that the meteorological big data model includes an input encoder, an output decoder, and an intermediate layer between them. The core idea of ​​step S23 is to freeze the intermediate layer that represents the global fluid dynamics, and only unlock the input encoder and output decoder. By combining static geographical features and a custom loss function, the model parameters in the pre-trained meteorological big data model are re-tuned and trained. In other words, the meteorological big data model is regionally refined to obtain a meteorological big data model that takes into account regional characteristics.

[0055] It should also be noted that this invention combines a custom loss function that integrates physical constraints and height weighting during the re-fine-tuning training process, achieving synergistic optimization of data-driven and physical laws. By forcing the model to follow the basic laws of atmospheric motion through physical mechanisms, the model's adaptability and physical consistency in complex terrain regions are enhanced.

[0056] Step S24: Use the refined meteorological model of the region to obtain a real-time high-precision meteorological field, and combine the real-time high-precision meteorological field and atmospheric physics equations to reconstruct the tropospheric wet refractive index of each isobaric layer.

[0057] In step S24, a real-time high-precision meteorological field (i.e., pressure layer data that dynamically fluctuates with weather systems) is obtained using a refined regional meteorological model. The geopotential height of each grid point is extracted from this field and converted into geometric height, thus obtaining the dynamic time-varying geometric height-pressure correspondence for each grid point. Based on this, the tropospheric wet refractive index of each dynamic isobaric layer is calculated using atmospheric physics equations, achieving the same vertical resolution as the original model with improved accuracy. Geopotential height is a hypothetical height in meteorology representing the potential energy per unit mass of air relative to sea level, measured in geopotential meters, and primarily used for isobaric weather map analysis.

[0058] Step S3: Map the tropospheric wet refractive index of each isobaric layer obtained from the reconstruction to the tropospheric wet refractive index of the fixed geometric height layer, and use numerical integration method and error propagation law to generate a real-time tropospheric wet delay vertical structure grid product covering the target area and with uncertainty identification.

[0059] In step S3, since the output of the refined meteorological model is barometric data that dynamically fluctuates with weather systems, the tropospheric wet refractive index of each isobaric layer reconstructed based on the barometric data also dynamically fluctuates with weather systems. GNSS positioning and meteorological applications typically require a unified geometrical height layer product. Therefore, this invention constructs a weighted least squares adjustment model to map the non-uniform, fluctuating tropospheric wet refractive index of each isobaric layer to a spatial geometrical grid with a fixed geometrical height layer. The specific steps of step S3 are as follows:

[0060] Step S31: Obtain the prior error covariance matrix before adjustment by statistically analyzing the tropospheric wet refractive index error of each isobaric layer under the same verification time and different forecast lead times. Invert the prior error covariance matrix to generate the prior weight matrix required for least squares solution.

[0061] In step S31, the NMC (National Meteorological Center) method is introduced to calculate the prior error covariance matrix before adjustment. Specifically, this method obtains the prior error covariance matrix of the forecast wet refractive index of the meteorological large model for the target area by statistically analyzing the tropospheric wet refractive index errors of each isobaric layer corresponding to the same validation time and different forecast lead times (e.g., 24h and 12h) in step S1. By inverting this prior error covariance matrix to generate the prior weight matrix required for least squares solution, the model can achieve reasonable weight allocation based on the prior accuracy of GraphCast at each altitude layer.

[0062] Step S32: Based on the tropospheric wet refractive index and prior weight matrix of each isobaric layer obtained by reconstruction, a weighted least squares adjustment is performed to obtain the tropospheric wet refractive index and posterior error covariance matrix of the layer with a fixed geometric height.

[0063] In step S32, after determining the prior weight matrix, this invention constructs a linear observation equation using the tropospheric wet refractive index at a fixed geometric height layer as the parameter to be determined. This linear observation equation uses an interpolation function to describe the interpolation mapping relationship from the fixed geometric height layer to the height of the GraphCast dynamic isobaric layer, and the linear observation equation is solved analytically according to the least squares criterion to obtain the tropospheric wet refractive index and the posterior error covariance matrix at the fixed geometric height layer.

[0064] Step S33: Based on the wet refractive index of the troposphere at a fixed geometric height, the interlayer wet tropospheric delay of each fixed geometric height layer is calculated using a numerical integration method. Then, based on the error propagation law and combined with the posterior error covariance matrix, the uncertainty of the wet tropospheric delay of each layer is calculated simultaneously, thereby generating a real-time tropospheric wet delay vertical structure grid product covering the target area and equipped with uncertainty indicators.

[0065] In step S33, based on the wet refractive index of the fixed geometric height layer, the interlayer zenith tropospheric wet delay is calculated using a numerical integration method. Simultaneously, according to the error propagation law and combined with the posterior error covariance matrix obtained from the adjustment calculation, the uncertainty of the interlayer zenith tropospheric wet delay is calculated synchronously. Finally, a real-time tropospheric wet delay vertical structured grid product covering the target area and equipped with uncertainty indicators is generated. Understandably, the real-time tropospheric wet delay vertical structured grid product includes the interlayer zenith tropospheric wet delay and its uncertainty (i.e., accuracy information).

[0066] Based on the same technical concept as the aforementioned embodiments, this invention also provides a fine-scale reconstruction system for the tropospheric vertical structure that takes into account regional characteristics, comprising: a priori error quantitative assessment module, used to input global meteorological data into a pre-trained meteorological large model, output tropospheric wet refractive index prediction data for each isobaric layer, obtain the tropospheric wet refractive index error of each isobaric layer based on the tropospheric wet refractive index prediction data for each isobaric layer, and obtain the gridded vertical error elevation based on the tropospheric wet refractive index error of each isobaric layer; and a model retraining fine-tuning module, used to obtain the gridded vertical error elevation based on the gridded vertical error elevation. Based on the terrain factors and physical constraints of the target area, a freeze-fine-tuning strategy is used to refine the pre-trained meteorological model to obtain a refined meteorological model. The refined meteorological model is then used to reconstruct the tropospheric wet refractive index of each isobaric layer. The wet delay mapping reconstruction module is used to map the reconstructed tropospheric wet refractive index of each isobaric layer to the tropospheric wet refractive index of a fixed geometric height layer. A real-time tropospheric wet delay vertical structure grid product covering the target area and equipped with uncertainty indicators is generated using numerical integration method and error propagation law.

[0067] Based on the same technical concept as the foregoing embodiments, the present invention also provides an electronic device, including a memory and a processor, wherein the memory stores program instructions that are executed by the processor, and the processor invokes the program instructions to execute a fine reconstruction method for tropospheric vertical structure that takes into account regional characteristics.

[0068] Based on the same technical concept as the foregoing embodiments, the present invention also provides a non-transitory computer-readable storage medium, characterized in that the non-transitory computer-readable storage medium stores computer instructions that cause the computer to execute a fine reconstruction method of tropospheric vertical structure taking into account regional characteristics.

[0069] In summary, the above embodiments first utilize a pre-trained meteorological model to quantitatively evaluate the tropospheric wet refractive index error of each isobaric layer. Based on the tropospheric wet refractive index error of each isobaric layer, a vertical error elevation model is fitted. This vertical error elevation model reveals the variation law of the vertical structural deviation of the meteorological model at different altitudes. Then, based on the vertical error elevation model, a gridded vertical error elevation is obtained. This gridded vertical error elevation will serve as a key weight parameter in the subsequent fine-tuning training of the meteorological model, guiding the meteorological model to implement differentiated training at different geographical locations and altitudes. The model is then corrected. Next, based on the gridded vertical error elevation, the pre-trained meteorological model is fine-tuned to obtain a model that considers regional characteristics. The tropospheric wet refractive index of each isobaric layer is reconstructed using this model. Understandably, the pre-trained meteorological model is adapted to global weather forecasts, but there are significant differences between the global prior background and specific complex terrain and climate. The output temperature, pressure, and humidity field of the target area deviates from the actual data. Therefore, this step achieves regional adaptation of the meteorological model parameters from a globally general form to the local atmospheric characteristics of the target area. Finally, the reconstructed tropospheric wet refractive index of each isobaric layer is mapped to the tropospheric wet refractive index of a fixed geometric height layer, and a real-time tropospheric wet delay vertical grid product covering the target area and equipped with uncertainty indicators is generated using numerical integration methods and the error propagation law.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for fine reconstruction of the tropospheric vertical structure that takes into account regional characteristics, characterized in that, include: Step S1: Input global meteorological data into the pre-trained meteorological big model, output the tropospheric wet refractive index prediction data of each isobaric layer, obtain the tropospheric wet refractive index error of each isobaric layer based on the tropospheric wet refractive index prediction data of each isobaric layer, and obtain the gridded vertical error elevation based on the tropospheric wet refractive index error of each isobaric layer. Step S2: Based on the gridded vertical error elevation, the topographic factor of the target area, and physical constraints, a freeze-fine-tuning strategy is used to refine the pre-trained meteorological model regionally to obtain a refined meteorological model. The refined meteorological model is then used to reconstruct the tropospheric wet refractive index of each isobaric layer. The meteorological model includes an input encoder, an output decoder, and an intermediate layer between them. The freeze-fine-tuning strategy includes freezing the intermediate layer in the meteorological model that represents the global fluid dynamics, unlocking only the input encoder and the output decoder, and fine-tuning the model parameters in the meteorological model. Step S3: Map the tropospheric wet refractive index of each isobaric layer obtained from the reconstruction to the tropospheric wet refractive index of the fixed geometric height layer, and use numerical integration method and error propagation law to generate a real-time tropospheric wet delay vertical structure grid product covering the target area and with uncertainty identification.

2. The method for fine reconstruction of tropospheric vertical structure considering regional characteristics as described in claim 1, characterized in that, Global meteorological data is input into a pre-trained meteorological model, which outputs predicted tropospheric wet refractive index data for each isobaric layer. Based on these predicted data, the tropospheric wet refractive index error for each isobaric layer is obtained, including: Using global meteorological data to drive a pre-trained meteorological model for autoregressive rolling prediction, a global real-time meteorological prediction grid for a future preset time period is generated, and a tropospheric wet refractive index prediction grid is calculated based on atmospheric physics equations. By using bilinear and log-linear interpolation techniques, the tropospheric wet refractive index prediction grid is mapped with high precision to the radiosonde stations and isobars in the target area, so as to obtain the tropospheric wet refractive index prediction data of the radiosonde stations on each isobar. By comparing the predicted tropospheric wet refractive index data from radiosonde stations with the measured tropospheric wet refractive index data from radiosonde stations at each isobaric layer, the error of the tropospheric wet refractive index for each isobaric layer can be obtained.

3. The method for fine reconstruction of tropospheric vertical structure considering regional characteristics as described in claim 1, characterized in that, The gridded vertical error elevation is obtained based on the tropospheric wet refractive index error of each isobaric layer, including: An exponential function was used to fit the tropospheric wet refractive index error of each isobaric layer to obtain the error elevation parameters at each discrete station. The three-dimensional inverse distance weighting method is used to spatially extrapolate the error elevation parameters extracted from each discrete station to obtain the gridded vertical error elevation.

4. The method for fine reconstruction of tropospheric vertical structure considering regional characteristics as described in claim 1, characterized in that, Step S2 includes: The topographic factors of the target area are calculated based on the digital elevation model data of the target area, and the topographic factors of the target area are input into the pre-trained meteorological big model so that the meteorological big model constructs a static topographic feature tensor based on the topographic factors of the target area. The static topographic feature tensor is then fused with the original static geographic features of the meteorological big model to obtain fused static geographic features. Using the gridded vertical error elevation as prior information, a weighted loss matrix that adaptively decays with height is constructed, and then a custom loss function that integrates physical constraints and height weighting is constructed based on the weighted loss matrix. Combining the aforementioned fusion of static geographic features and a custom loss function, a freeze-fine-tuning strategy is employed to refine the pre-trained meteorological model regionally, resulting in a regionally refined meteorological model. Real-time high-precision meteorological fields are obtained using a regionally refined meteorological model, and the tropospheric wet refractive index of each isobaric layer is reconstructed by combining the real-time high-precision meteorological fields with atmospheric physics equations.

5. The method for fine reconstruction of tropospheric vertical structure considering regional characteristics as described in claim 4, characterized in that, Using the gridded vertical error elevation as prior information, a weighted loss matrix that adaptively decays with height is constructed. Then, based on this weighted loss matrix, a custom loss function fusing physical constraints and height weighting is constructed, with the corresponding formula as follows: ， ， In the formula, h is the geometric height corresponding to the current pressure layer. To represent the vertical error elevation in the grid, To adjust the hyperparameters of physical constraint strength, To integrate physical constraints with a highly weighted custom loss function, For mean square error loss, and These are two major physical constraints. This is the weighted loss matrix.

6. The method for fine reconstruction of tropospheric vertical structure considering regional characteristics as described in claim 1, characterized in that, Step S3 includes: Step S31: Obtain the prior error covariance matrix before adjustment by statistically analyzing the tropospheric wet refractive index error of each isobaric layer under the same verification time and different forecast lead time. Invert the prior error covariance matrix to generate the prior weight matrix required for least squares solution. Step S32: Based on the tropospheric wet refractive index and prior weight matrix of each isobaric layer obtained by reconstruction, perform weighted least squares adjustment to obtain the tropospheric wet refractive index and posterior error covariance matrix of the fixed geometric height layer. Step S33: Based on the tropospheric wet refractive index of the fixed geometric height layer, the interlayer tropospheric wet delay of each fixed geometric height layer is calculated using a numerical integration method. Based on the error propagation law and the posterior error covariance matrix, the uncertainty of the interlayer tropospheric wet delay is calculated simultaneously, thereby generating a real-time tropospheric wet delay vertical structure grid product covering the target area and equipped with uncertainty indicators.

7. A fine-scale reconstruction system for the vertical structure of the troposphere that takes into account regional characteristics, characterized in that, include: The prior error quantitative assessment module is used to input global meteorological data into the pre-trained meteorological big model, output the tropospheric wet refractive index prediction data of each isobaric layer, obtain the tropospheric wet refractive index error of each isobaric layer based on the tropospheric wet refractive index prediction data of each isobaric layer, and obtain the gridded vertical error elevation based on the tropospheric wet refractive index error of each isobaric layer. The model retraining and fine-tuning module is used to refine the pre-trained meteorological large model regionally based on gridded vertical error elevation, target area topographic factors, and physical constraints, using a freeze-fine-tuning strategy to obtain a regionally refined meteorological large model. The refined meteorological large model is then used to reconstruct the tropospheric wet refractive index of each isobaric layer. The meteorological large model includes an input encoder, an output decoder, and an intermediate layer between them. The freeze-fine-tuning strategy involves freezing the intermediate layer representing global hydrodynamic laws in the meteorological large model, unlocking only the input encoder and output decoder, and then fine-tuning the model parameters within the meteorological large model. The wet delay mapping reconstruction module is used to map the tropospheric wet refractive index of each isobaric layer obtained by reconstruction to the tropospheric wet refractive index of a fixed geometric height layer, and uses numerical integration method and error propagation law to generate a real-time tropospheric wet delay vertical structure grid product covering the target area and with uncertainty identification.

8. An electronic device, characterized in that, The system includes a memory and a processor, the memory storing program instructions that are executed by the processor, which invokes the program instructions to execute a fine reconstruction method for tropospheric vertical structure that takes into account regional characteristics, as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions that cause the computer to execute a fine reconstruction method for tropospheric vertical structure that takes into account regional characteristics, as described in any one of claims 1 to 6.

Images