Meteorological forecasting method based on cooperative driving of Beidou three-dimensional water vapor and large meteorological model

Abstract

The invention provides a weather forecasting method based on big Beidou three-dimensional water vapor and weather model cooperative driving, and the method comprises the steps: carrying out the point-by-point registration of the obtained prediction data of a GraphCast model and ERA5 measured data, obtaining residual field data, taking the residual field data as the input of a deep learning residual correction network, and obtaining predicted residual field data; a terrain adaptive algorithm is adopted, predicted ERA5 meteorological parameters corresponding to the predicted residual field data are converted into a three-dimensional grid structure under height coordinates, and zenith troposphere delay is determined through an integration method; based on the dynamic correction term, converting the zenith troposphere delay into a precipitable water amount through a space-time inversion network; inputting the predicted ERA5 meteorological parameters and the rainfall data into a parameter efficient adapter and a big ancient meteorological model to obtain meteorological forecast factors output by the big ancient meteorological model; through the above technical scheme, real-time positioning enhancement and gradual improvement of meteorological element forecasting capability can be realized.

Description

Technical Field

[0001] This application relates to the field of meteorological monitoring and forecasting technology, and in particular to a meteorological forecasting method based on the collaborative driving of BeiDou three-dimensional water vapor and meteorological large model. Background Technology

[0002] Tropospheric delay is a key factor affecting the accuracy of GNSS positioning and radio astronomy observations, especially in Very Long Baseline Interferometry (VLBI) and high-frequency radio measurements. Furthermore, tropospheric delay is closely related to atmospheric water vapor content. Monitoring tropospheric delay can indirectly provide information on atmospheric water vapor content, which is of great significance for the early warning and monitoring of extreme weather events. While the BeiDou CORS (Continuous Operation Reference Station) has the advantage of acquiring high temporal resolution tropospheric delay data, it can only monitor at the station location and cannot acquire high spatial resolution data.

[0003] While existing tropospheric delay models have made some progress, they still have shortcomings in terms of spatial resolution, real-time update capability, and characterization of vertical stratification, making it difficult to meet the requirements of high dynamics, real-time performance, and fine-grained positioning. Summary of the Invention

[0004] This application provides a weather forecasting method based on the collaborative driving of BeiDou three-dimensional water vapor and meteorological large model, which solves the problems of insufficient spatial resolution, real-time update capability and vertical layering structure in the existing technology, and realizes the gradual improvement of real-time positioning enhancement and meteorological element forecasting capability.

[0005] To achieve the above objectives, this application provides a weather forecasting method based on the collaborative driving of BeiDou three-dimensional water vapor and meteorological large model, the method comprising:

[0006] The predicted data from the obtained GraphCast model and the measured data from ERA5 were registered point by point to obtain the residual field data.

[0007] The residual field data is used as input to a deep learning residual correction network to obtain the predicted residual field data output by the deep learning residual correction network. The deep learning residual correction network includes a 3D convolution module, a ResNet cross-layer residual connection structure and a physical constraint module. Each residual block in the ResNet cross-layer residual connection structure embeds a channel attention mechanism and a spatial attention module, and a ConvLSTM structure is introduced at the back end.

[0008] A terrain-adaptive algorithm is used to transform the predicted ERA5 meteorological parameters corresponding to the predicted residual field data into a three-dimensional grid structure in height coordinates, and the integral method is used to determine the zenith tropospheric delay.

[0009] The coupling relationship between temperature, humidity and air pressure is nonlinearly fitted through a multi-layer sensing mechanism to form a dynamic correction term for the traditional π factor; based on the dynamic correction term, the zenith tropospheric delay is converted into precipitable water through a spatiotemporal inversion network; the spatiotemporal inversion network includes a CNN module and a Transformer module.

[0010] The predicted ERA5 meteorological parameters and the precipitable water data input parameter efficient adapter are adjusted, and the adjusted predicted ERA5 meteorological parameters and the precipitable water data are used as input to the Pangu meteorological big model to obtain the meteorological forecast factors output by the Pangu meteorological big model. The parameter efficient adapter adopts a bottleneck residual structure and uses LoRA low-rank decomposition on key convolutional layers or self-attention layers. It is pre-trained through a physical-uncertainty hybrid loss, meta-learning initialization, and progressive-contrastive fine-tuning strategy.

[0011] In one possible implementation, the method further includes:

[0012] In response to receiving geographic coordinate information of the target location input by the user; the geographic coordinate information includes the longitude parameter, latitude parameter, elevation parameter and time parameter of the target location;

[0013] The geographic coordinate information is determined by identifying four adjacent reference grid points within the three-dimensional grid structure.

[0014] The tropospheric delay of the target location is determined at each altitude using the horizontal inverse distance weighted interpolation method.

[0015] The vertical distribution of tropospheric delay at the target location was fitted using an exponential function.

[0016] Based on the elevation parameters of the target location, the corresponding target tropospheric delay is obtained from the zenith tropospheric delay.

[0017] In one possible implementation, the physical constraint module includes tropospheric refractive index constraints, tropospheric delay constraints, integrated loss function constraints, and terrain correction operator constraints.

[0018] The tropospheric refractive index constraint is: Where N is the tropospheric refractive index, P is the air pressure, T is the temperature, e is the water vapor pressure, and constants k1, k2, and k3 are empirical coefficients.

[0019] The tropospheric delay constraint is: Where ZTD is the tropospheric retardation, N is the tropospheric refractive index, h1 is the bottom height, and h2 is the top height;

[0020] The constraints of the comprehensive loss function are as follows: Where λ1, λ2, and λ3 are weighting coefficients. Indicates the mean square error loss. Represents structural similarity loss. Represents the loss due to physical consistency constraints: Among them, ZTD calc ( (ZTD) represents the zenith tropospheric delay obtained from the refractive index field predicted by the model. ERA5 This represents the actual zenith tropospheric delay corresponding to the ERA5 dataset;

[0021] The terrain correction operator is constrained as follows: Where H is the weight matrix in the terrain correction module. H is the terrain correction factor. c For the comprehensive terrain correction factor, H ERA5 The elevation field provided for the ERA5 dataset, H DEM Provides the actual surface elevation for the Digital Elevation Model (DEM).

[0022] In one possible implementation, the deep learning residual correction network is:

[0023] Among them, F out The output of the deep learning residual correction network is defined as follows: Attention represents the multidimensional attention weighting function used to fuse channel and spatial weights; ConvLSTM is responsible for temporal dimension modeling; ReLU is a non-linear activation function; Conv3D represents a three-dimensional convolutional module; and F... in The three-dimensional meteorological feature tensor input to the deep residual correction network originates from the residual field data and is used to predict and correct the residual field by means of convolutional feature extraction and temporal modeling.

[0024] In one possible implementation, the step of employing a terrain-adaptive algorithm to transform the predicted ERA5 meteorological parameters corresponding to the predicted residual field data into a three-dimensional grid structure in height coordinates, and using an integral method to determine the zenith tropospheric delay, includes:

[0025] Based on the predicted residual field data and the obtained ERA5 meteorological parameters, the predicted ERA5 meteorological parameters are obtained.

[0026] The terrain adaptive interpolation algorithm is used to introduce a terrain gradient correction term in the interpolation process of air pressure and temperature, and to perform vertical layer reconstruction to obtain a three-dimensional mesh structure in height coordinates.

[0027] Based on the predicted ERA5 meteorological parameters, the zenith tropospheric delay is determined using an integral method.

[0028] In one possible implementation, the method further includes:

[0029] The posterior probability distribution of the precipitable water volume is obtained through Bayesian uncertainty estimation.

[0030] The posterior probability distribution is sampled using the MCMC method to obtain the confidence interval of the precipitable water amount, thereby achieving an assessment of the accuracy of the precipitable water amount.

[0031] In one possible implementation, the bottleneck residual structure in the parameter-efficient adapter is as follows: ;in, , This is the weight matrix. , For bias terms, The Swish activation function is used; the LoRA low-rank decomposition is... Where A and B are low-rank matrices of LoRA, and the dimension of A is . r d, used for dimensionality reduction, where the dimension of B is... r d' is used for dimensionality recovery, and r is the rank of the low-rank decomposition, used to control the parameter size and adaptation speed.

[0032] In one possible approach, the physical-uncertainty hybrid loss is:

[0033] ;in, These are the weighting coefficients for the mean squared error loss term. λ is the mean square error loss; SSIM These are the weighting coefficients for the structural similarity loss term. For structural similarity loss; The weighting coefficients for the physical consistency loss term, For physical consistency loss: Among them, ZTD ( (ZTD) represents the zenith tropospheric delay calculated based on the model-predicted refractive index field. obs For the actual zenith tropospheric delay from ERA5; The weighting coefficients for the uncertainty loss term. Loss due to uncertainty: , where y i For the actual observed value, μ i Output values ​​for the model. The variance of the model prediction is represented by i, which represents the sample index, i.e., the summation over each grid point.

[0034] In one possible implementation, the meta-learning initialization involves pre-training an initialization parameter on historical multi-region / multi-time period data. This allows it to quickly adapt to new sites using a small number of gradient steps; the formal meta-objective is: Among them, the task Indicates different local subdomains or time periods. For the inner loop learning rate, For the formalized meta-objective, The training loss is used; during the meta-learning phase, only the parameters of the efficient adapter / LoRA are updated.

[0035] In one possible implementation, the progressive-contrast fine-tuning strategy includes a coarse-tuning strategy, a fine-tuning strategy, and a stabilization strategy; wherein the coarse-tuning strategy is characterized by short steps and a large learning rate, enabling the adapter to learn the overall bias field; the fine-tuning strategy is characterized by a small learning rate combined with... To refine alignment with physical consistency constraints; the stabilization strategy incorporates contrastive learning to construct observation-consistent / inconsistent sample pairs, and applies contrastive loss to the output of the parameter-efficient adapter to maintain local structure without distortion; the contrastive loss is expressed as: ;in, For temperature parameters, The feature representation of the current sample, As a positive sample, This is a negative sample.

[0036] The technical solutions provided in this application embodiment have at least the following technical effects or advantages:

[0037] Based on the above technical solution, multi-scale meteorological observation data and physical mechanisms are integrated to construct a dynamic response meteorological element three-dimensional fine distribution and minute-level vertical decline function model, breaking through the accuracy bottleneck of empirical modeling of tropospheric delay under complex meteorological conditions, and realizing real-time high-precision three-dimensional tropospheric forecasting under complex environments; in order to address the problem that the Pangoal Infinite Model cannot identify three-dimensional water vapor characteristics, a three-dimensional water vapor characteristic input adapter and a lightweight forecast post-processing module are constructed to realize multi-element meteorological forecasting driven by three-dimensional water vapor and the Pangoal Infinite Model. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments of this application or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 A flowchart illustrating a weather forecasting method based on the collaborative driving of BeiDou three-dimensional water vapor and meteorological large model, provided for embodiments of this application;

[0040] Figure 2 A flowchart of another weather forecasting method based on the collaborative driving of BeiDou three-dimensional water vapor and meteorological large model, provided for embodiments of this application. Detailed Implementation

[0041] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0042] In the description of the embodiments of this application, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the embodiments of this application and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application. The terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; a mechanical connection or an electrical connection; a direct connection or an indirect connection through an intermediate medium; or a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in the embodiments of this application according to the specific circumstances.

[0043] To address the shortcomings of existing tropospheric delay models in terms of spatial resolution, real-time update capability, and characterization of vertical stratification, which make it difficult to meet the requirements of high dynamics, real-time performance, and refined positioning, this application leverages the GNSS satellite positioning infrastructure already built by the CORS system. It fully utilizes the resource advantages and network coverage capabilities of CORS in regional high-precision positioning services and integrates multi-scale meteorological observation data and physical mechanisms to construct a dynamic response meteorological element three-dimensional refined distribution and minute-level vertical decline function model. This breaks through the accuracy bottleneck of empirical modeling of tropospheric delay under complex meteorological conditions and enables real-time high-precision three-dimensional tropospheric forecasting under complex environments.

[0044] Figure 1 A flowchart illustrating a weather forecasting method based on the collaborative driving of BeiDou three-dimensional water vapor and meteorological large model, provided for embodiments of this application. Figure 1 As shown, the method may include the following steps.

[0045] S101. Perform point-by-point registration between the obtained GraphCast model prediction data and ERA5 measured data to obtain residual field data.

[0046] For example, ERA5 could be the fifth-generation global climate reanalysis dataset released by the European Centre for Medium-Range Weather Forecasts.

[0047] S102. Use the residual field data as input to the deep learning residual correction network to obtain the predicted residual field data output by the deep learning residual correction network.

[0048] The deep learning residual correction network includes a 3D convolution module, a ResNet cross-layer residual connection structure, and a physical constraint module. Each residual block in the ResNet cross-layer residual connection structure embeds a channel attention mechanism and a spatial attention module, and a ConvLSTM structure is introduced at the back end.

[0049] S103. Using a terrain-adaptive algorithm, the predicted ERA5 meteorological parameters corresponding to the predicted residual field data are transformed into a three-dimensional grid structure in height coordinates, and the zenith tropospheric delay is determined by the integral method.

[0050] For example, the zenith total delay is ZTD.

[0051] S104. The coupling relationship between temperature, humidity and air pressure is nonlinearly fitted through a multi-layer sensing mechanism to form a dynamic correction term for the traditional Π factor. Based on this dynamic correction term, the zenith tropospheric delay is converted into precipitable water through a spatiotemporal inversion network.

[0052] The spatiotemporal inversion network includes a CNN module and a Transformer module. The precipitable water volume is PWV (Precipitable Water Vapor).

[0053] S105. Adjust the high-efficiency adapter of the predicted ERA5 meteorological parameters and the precipitable water data input parameters, and use the adjusted predicted ERA5 meteorological parameters and the precipitable water data as input to the Pangu meteorological model to obtain the meteorological forecast factors output by the Pangu meteorological model.

[0054] The parameter-efficient adapter employs a bottleneck residual structure and uses LoRA low-rank decomposition on key convolutional layers or self-attention layers. It is pre-trained using a physical-uncertainty hybrid loss, meta-learning initialization, and progressive-contrastive fine-tuning strategy.

[0055] In some embodiments, the weather forecast factors may include temperature, humidity, PM2.5 concentration, air pressure, wind speed, cloud cover, cloud type, visibility, precipitation, etc. Among them, wind speed can affect the attitude stability and path deviation of the aircraft, while precipitation and PM2.5 concentration can significantly affect visibility and sensor perception accuracy, posing a threat to flight safety. Therefore, the weather forecast factors obtained in this application can also be used to assess the safety risks of aircraft.

[0056] Based on the above technical solution, multi-scale meteorological observation data and physical mechanisms are integrated to construct a dynamic response meteorological element three-dimensional fine distribution and minute-level vertical decline function model, breaking through the accuracy bottleneck of empirical modeling of tropospheric delay under complex meteorological conditions, and realizing real-time high-precision three-dimensional tropospheric forecasting under complex environments; in order to address the problem that the Pangoal Infinite Model cannot identify three-dimensional water vapor characteristics, a three-dimensional water vapor characteristic input adapter and a lightweight forecast post-processing module are constructed to realize multi-element meteorological forecasting driven by three-dimensional water vapor and the Pangoal Infinite Model.

[0057] In one possible implementation, the method further includes: responding to receiving geographic coordinate information of a target location input by a user; the geographic coordinate information includes longitude parameters, latitude parameters, elevation parameters, and time parameters of the target location; determining four adjacent reference grid points of the geographic coordinate information in the three-dimensional grid structure, and using a horizontal inverse distance weighted interpolation method to determine the tropospheric delay of the target location at each altitude layer; fitting the vertical distribution of the tropospheric delay of the target location using an exponential function; and obtaining the corresponding target tropospheric delay from the zenith tropospheric delay based on the elevation parameters of the target location.

[0058] In one possible implementation, the physical constraint module includes tropospheric refractive index constraints, tropospheric delay constraints, integrated loss function constraints, and terrain correction operator constraints; the tropospheric refractive index constraints are as follows: Where N is the tropospheric refractive index, P is the air pressure, T is the temperature, e is the water vapor pressure, and constants k1, k2, and k3 are empirical coefficients; the tropospheric delay constraint is: Where ZTD is the tropospheric delay, N is the tropospheric refractive index, h1 is the bottom height, and h2 is the top height; the constraints of this comprehensive loss function are: Where λ1, λ2, and λ3 are weighting coefficients. Indicates the mean square error loss. Represents structural similarity loss. Represents the loss due to physical consistency constraints: Among them, ZTD calc ( (ZTD) represents the zenith tropospheric delay obtained from the refractive index field predicted by the model. ERA5 The terrain correction operator is constrained to represent the true zenith tropospheric delay corresponding to the ERA5 dataset, and is defined as follows: Where H is the weight matrix in the terrain correction module. H is the terrain correction factor. c For the comprehensive terrain correction factor, H ERA5 The elevation field provided for the ERA5 dataset, H DEM Provides the actual surface elevation for the Digital Elevation Model (DEM).

[0059] In one possible implementation, the deep learning residual correction network is: Among them, F out The output of this deep learning residual correction network is represented by Attention, which represents the multidimensional attention weighting function used to fuse channel and spatial weights. ConvLSTM is responsible for temporal dimension modeling, ReLU is a non-linear activation function, and Conv3D represents a three-dimensional convolutional module. in This represents the three-dimensional meteorological feature tensor input to the deep residual correction network, which originates from the residual field data and is used to predict and correct the residual field by means of convolutional feature extraction and temporal modeling.

[0060] In one possible implementation, a terrain-adaptive algorithm is used to transform the predicted ERA5 meteorological parameters corresponding to the predicted residual field data into a three-dimensional grid structure in height coordinates, and an integral method is used to determine the zenith tropospheric delay. This includes: obtaining the predicted ERA5 meteorological parameters based on the predicted residual field data and the acquired ERA5 meteorological parameters; using the terrain-adaptive interpolation algorithm, introducing a terrain gradient correction term during the pressure and temperature interpolation process, performing vertical layer reconstruction to obtain a three-dimensional grid structure in height coordinates; and determining the zenith tropospheric delay based on the predicted ERA5 meteorological parameters using an integral method.

[0061] In one possible implementation, the method further includes: obtaining the posterior probability distribution of the precipitable water amount through Bayesian uncertainty estimation; and sampling the posterior probability distribution using the MCMC method to obtain the confidence interval of the precipitable water amount, so as to achieve the assessment of the accuracy of the precipitable water amount.

[0062] In one possible implementation, the bottleneck residual structure in this parameter-efficient adapter is as follows: ;in, , This is the weight matrix. , For bias terms, The Swish activation function is used; the LoRA low-rank decomposition is... Where A and B are low-rank matrices of LoRA, and the dimension of A is . r d, used for dimensionality reduction, where the dimension of B is... r d' is used for dimensionality recovery, and r is the rank of the low-rank decomposition, used to control the parameter size and adaptation speed.

[0063] In one possible implementation, the physical-uncertainty hybrid loss is: ;in, These are the weighting coefficients for the mean squared error loss term. λ is the mean square error loss; SSIM These are the weighting coefficients for the structural similarity loss term. For structural similarity loss; The weighting coefficients for the physical consistency loss term, For physical consistency loss: Among them, ZTD ( (ZTD) represents the zenith tropospheric delay calculated based on the model-predicted refractive index field. obs For the actual zenith tropospheric delay from ERA5; The weighting coefficients for the uncertainty loss term. Loss due to uncertainty: , where y i For the actual observed value, μ i Output values ​​for the model. The variance of the model prediction is represented by i, which represents the sample index, i.e., the summation over each grid point.

[0064] In one possible implementation, the meta-learning is initialized by pre-training an initialization parameter on historical multi-region / multi-time period data. This allows it to quickly adapt to new sites using a small number of gradient steps; the formal meta-objective is: Among them, the task Indicates different local subdomains or time periods. For the inner loop learning rate, For the formalized meta-objective, The training loss is used; during the meta-learning phase, only the parameters of the efficient adapter / LoRA are updated.

[0065] In one possible implementation, the progressive-contrast fine-tuning strategy includes a coarse-tuning strategy, a fine-tuning strategy, and a stabilization strategy; wherein the coarse-tuning strategy is characterized by short steps and a large learning rate, enabling the adapter to learn the overall bias field; the fine-tuning strategy is characterized by a small learning rate combined with... To refine alignment with physical consistency constraints, this stabilization strategy incorporates contrastive learning to construct observation-consistent / inconsistent sample pairs, applying contrastive loss to the output of the parameter-efficient adapter to maintain local structure integrity. This contrastive loss is expressed as: ;in, For temperature parameters, The feature representation of the current sample, As a positive sample, This is a negative sample.

[0066] In one possible implementation, this application also provides a weather forecasting platform based on the collaborative driving of a BeiDou 3D water vapor and meteorological big data model. This platform is used to execute the aforementioned weather forecasting method based on the collaborative driving of a BeiDou 3D water vapor and meteorological big data model, and includes a low-altitude sensing module, a data augmentation module, a 3D modeling module, a real-time positioning module, and a local short-term nowcasting module. The platform may also include a flight meteorological risk visualization interface, path planning assistance functions, and a flight mission dynamic adjustment module, supporting real-time decision-making by end users and ensuring the safety and stability of low-altitude UAV missions under complex weather conditions. The construction of this platform can provide real-time support for the refined operation of low-altitude UAVs under complex weather conditions, improving their autonomy and safety in typical scenarios such as power line inspection, urban logistics, and emergency response, and providing highly reliable environmental perception and flight risk early warning support for the development of the low-altitude economy.

[0067] Figure 2 A flowchart illustrating another weather forecasting method based on the collaborative driving of BeiDou three-dimensional water vapor and meteorological large model, provided for embodiments of this application. This method may include the following steps.

[0068] S201. Construction of a deep learning residual correction network that integrates spatiotemporal features.

[0069] The GraphCast model is an autoregressive model trained on ERA5 historical weather data with a spatial resolution of 0.25°×0.25°. It can predict five surface variables and six atmospheric variables across 37 vertical pressure levels. However, this model still exhibits certain errors in practical applications, particularly in medium-range forecasts, where the errors increase significantly, directly reducing the accuracy and reliability of ZTD calculations.

[0070] To this end, this application introduces an attention mechanism and a temporal modeling module into the existing 3D convolutional and residual network structure to construct a deep learning residual correction network (3D-CNN-ResNet-Attention-ConvLSTM) that integrates spatiotemporal features. This improved model aims to achieve more efficient joint feature extraction and correction in the spatial, vertical, and temporal directions.

[0071] First, the input to this deep learning residual correction network uses the prediction data from the GraphCast model and performs point-by-point registration with the ERA5 measured data to obtain residual field data as the supervision target. In the spatial dimension, multi-scale three-dimensional convolutional kernels (a combination of 3×3×3 and 5×5×5) are used to extract features of variables such as air pressure, temperature, and humidity at different spatial scales. To overcome the gradient decay and feature dilution problems of traditional CNN networks during deep training, this application introduces a ResNet cross-layer residual connection structure and embeds a channel attention mechanism (SE-Block) and a spatial attention module (CBAM) in each residual block to achieve adaptive weighting of key meteorological variable regions and enhance the model's sensitivity to high-impact areas.

[0072] Secondly, in order to fully capture the evolution of meteorological elements over time, the model introduces a ConvLSTM structure in the back end of the residual module to explicitly model the time dependence and dynamic continuity of ERA5 sequence data, thereby improving the stability and generalization ability of medium- and long-term forecast corrections.

[0073] The overall network structure of this deep residual correction network can be formally represented as:

[0074] Among them, F out For network output, Attention represents the multidimensional attention weighting function used to fuse channel and spatial weights, ConvLSTM is responsible for temporal dimension modeling, ReLU is a non-linear activation function, Conv3D represents the three-dimensional convolution module, and F... in This represents the three-dimensional meteorological feature tensor input to the deep residual correction network, which originates from the residual field data and is used to predict and correct the residual field by means of convolutional feature extraction and temporal modeling.

[0075] Furthermore, to further enhance the physical consistency and interpretability of the model, this application introduces a Physics-Constrained Module (PCM) into the deep learning prediction framework, forming a hybrid modeling system driven by both a physical model and a neural network. This module, based on the prior equations relating atmospheric refractive index and tropospheric delay (ZTD), incorporates the physical constraints between atmospheric state variables (the coupling relationship between temperature, pressure, and humidity) into the network training process in the form of regularization terms.

[0076] Specifically, let the meteorological field output by the neural network be... It includes air pressure ,temperature , specific moisture And other factors. Based on the physical expression of the tropospheric refractive index: In the formula, Where is the water vapor pressure, a constant. These are empirical coefficients. The tropospheric delay ZTD can be obtained by vertical integration: .

[0077] During training, the model not only minimizes the error term between the predicted and measured ERA5, but also additionally minimizes the ZTD error calculated based on the above physical equations, thereby constraining the network output to maintain physical consistency.

[0078] Its comprehensive loss function is: In the formula, The loss due to physical consistency constraints is defined as follows: .

[0079] By backpropagating through the network Calculate the gradient to implement physical-informed backpropagation constraints on the model output, ensuring that the model still conforms to meteorological dynamics while capturing complex spatiotemporal variation characteristics.

[0080] In addition, to reduce the error of the ERA5 grid field in areas with undulating terrain, this application further introduces a terrain correction operator in the PCM module to perform terrain correction on the ERA5 geopotential height field, making the model's inference of meteorological variables in complex terrain smoother and more physically consistent.

[0081] The terrain correction operator is expressed as follows: .

[0082] S202, Real-time 3D Tropospheric Delay Forecasting and Optimization.

[0083] Based on the prediction residuals corresponding to the ERA5 meteorological variables obtained in S201, combined with meteorological parameters, the predicted ERA5 meteorological parameters are obtained. Tropospheric delay calculation is performed to construct a real-time three-dimensional tropospheric model. The ERA5 data uses geopotential and isobaric surface layers to represent meteorological parameters. It is necessary to consider the differences between the ground elevation and the ERA5 dataset and isobaric surface height. A vertical layer reconstruction algorithm is constructed to convert the predicted real-time meteorological data into a three-dimensional grid structure in height coordinates.

[0084] This application employs a terrain-adaptive interpolation (GAI) algorithm, introducing a terrain gradient correction term during the interpolation of air pressure and temperature to achieve accurate profile reconstruction in terrain-complex areas. For temperature and relative humidity, linear interpolation is used to estimate the surface layer values. For air pressure parameters, assuming the surface layer lies between the k-th and (k-1)-th pressure layers, interpolation is performed using the following formula: In the formula, , , These represent the pressure at a specified vertical height, the first... Layer pressure and the first Layer pressure, , , They represent the specified potential heights, the first... The height of the stratum and the first High stratum potential.

[0085] If the geopotential height is lower than the lowest pressure layer in the model, extrapolation is performed using the following formula:

[0086] In the formula, and These represent the air pressure at a specified vertical altitude and the air pressure at the lowest pressure level, respectively. and It refers to the geopotential height corresponding to the highest geopotential and the lowest pressure layer. For a false temperature reading, the following formula can be used for calculation: In the formula, Indicates temperature. This represents the specific humidity. After completing the vertical stratification reconstruction, a 3D-ZTD model is constructed, and the zenith tropospheric delay is calculated using the integral method based on ERA5 data.

[0087] First, based on the calibrated ERA5 meteorological parameters, the water vapor pressure is calculated using the following formula: In the formula, S is the specific humidity, e is the water vapor pressure (hPa), and P is the atmospheric pressure (hPa).

[0088] In traditional physical models, the calculation of ZTD, ZHD, and ZWD relies on fixed constants and linear assumptions. To overcome the errors caused by these simplifying assumptions, this application proposes a machine learning inversion framework based on physical constraints.

[0089] First, the temperature profile Introducing neural network correction terms: ;in This represents a ResNet structure, taking four-dimensional features (air pressure, temperature, humidity, and altitude) as input and outputting correction terms to learn the nonlinear temperature and humidity relationship in historical ERA5 samples.

[0090] The corrected temperature is used to calculate the refractive index. And replace the traditional empirical formula:

[0091] ;in, This is a physical consistency compensation term output by a small-scale convolutional network, used to capture the effects of complex humidity profiles and nonlinear atmospheric structures.

[0092] This method significantly reduces errors while maintaining physical plausibility. Finally, the zenith tropospheric delay is calculated using an integral method, as shown in the following formula: In the formula, For zenith tropospheric delay, This is the height of the bottom layer, in meters. The height is the top layer, in meters. The wet delay in the upper atmosphere is negligible, and the dry delay in the troposphere is calculated using the Saastamoinen model. Thus, a preliminary real-time three-dimensional tropospheric model of the region can be obtained.

[0093] In practical applications, real-time calculation of tropospheric delay is required based on the target location's longitude, latitude, elevation, and time quadruple parameters. The specific workflow is as follows:

[0094] (1) Based on the user's geographic coordinates, locate four adjacent reference grid points in a 0.25°×0.25° grid.

[0095] (2) The horizontal inverse distance weighted interpolation method is used to calculate the tropospheric delay of the target position at each altitude layer.

[0096] (3) Fit the vertical distribution of tropospheric delay at the target location using an exponential function:

[0097] In the formula, This represents the ZTD value for the corresponding layer. Represents the reference elevation value for the corresponding height. The ZTD high normalization factor value is calculated using the Fourier function:

[0098] ;

[0099] In the formula, A represents ( =1,2,3,4) and ( =1,2,3,4) and ( =1,2…7) are model coefficients, doy represents the year-to-date, and hod represents a specific time of day.

[0100] (4) Obtain the tropospheric delay based on the elevation parameters of the target location.

[0101] S203, High-precision three-dimensional water vapor inversion based on three-dimensional tropospheric delay.

[0102] Tropospheric water vapor, as a primary indicator of total water vapor content, plays a crucial role in meteorological early warning and forecasting. However, traditional single-point BeiDou stations can only provide path integral water vapor products, which is insufficient to meet the needs of three-dimensional dynamic environmental perception. Therefore, conducting high spatiotemporal resolution three-dimensional water vapor inversion has become a key technological approach to achieving refined meteorological perception.

[0103] This application, using the three-dimensional ZTD obtained in S202 and combined with the three-dimensional meteorological parameters of multiple pressure layers, further introduces a nonlinear machine learning inversion and spatiotemporal fusion algorithm to overcome the bias caused by the fixed model parameters and linear assumptions, based on the traditional ZTD→ZWD→PWV conversion based on the Saastamoinen model.

[0104] First, a nonlinear fitting of the coupling relationship between temperature, humidity, and air pressure is performed using a multilayer sensing mechanism (MLP) to form a model of traditional... Dynamic correction term of the factor The corrected conversion formula is: .

[0105] Subsequently, to fully utilize the temporal continuity and spatial correlation of ERA5 data, a spatiotemporal inversion network (STFN) integrating CNN and Transformer was constructed. While retaining the Saastamoinen physical framework, it achieved joint modeling of ZTD temporal features and the three-dimensional atmospheric field. In addition, a Bayesian uncertainty estimation module was introduced, and the MCMC method was used to perform posterior sampling on the inversion results to obtain the confidence interval of PWV, thereby achieving a quantitative characterization of the inversion error.

[0106] S204, weather forecasting driven by the synergistic effect of three-dimensional water vapor and meteorological large model.

[0107] In low-altitude operational scenarios supported by GNSS tropospheric three-dimensional water vapor tomography, such as ensuring the flight safety of low-altitude UAVs, changes in meteorological elements are the main external factors leading to flight accidents. Wind speed directly affects aircraft attitude stability and path deviation; while rainfall and PM2.5 concentration significantly affect visibility and sensor accuracy, posing potential safety hazards. Therefore, achieving rapid weather forecasting at the regional scale is crucial. Traditional numerical weather prediction models (such as WRF) can accept external data assimilation, but their computational demands and adaptation difficulties make them unsuitable for rapid regional-scale forecasting. Large-scale meteorological models, such as the Pangu meteorological model, rely on a 3D neural network architecture and possess excellent operational efficiency and prediction accuracy.

[0108] Therefore, this application proposes using inverted 3D water vapor as an external physical prior to collaborate with the Pangoal Meteorological Model for refined regional-scale weather forecasting. While the Pangoal Model possesses global-scale weather forecasting capabilities, its original training data primarily comes from standardized meteorological reanalysis data, lacking sensitivity to complex underlying surfaces and non-standard meteorological inputs (such as 3D water vapor). Therefore, this application uses 3D water vapor as a high-precision label to physically guide the fine-tuning of the Pangoal Meteorological Model, aiming to construct a 3D water vapor-driven weather forecasting model. This not only compensates for the original model's insufficient ability to depict low-level details but also enhances its responsiveness to regional meteorological changes. The Pangoal Meteorological Model unifies and merges meteorological information from different longitudes, latitudes, and altitudes to form a standardized atmospheric state field (such as 3D grid data of isobaric surfaces or isopotential height surfaces) as model input. Since the 3D water vapor field inverted in this application is structured down to the geopotential height layer, it can be directly used for model input enhancement and guidance.

[0109] This application employs a strategy of freezing the model's backbone network and adding a trainable external feature adapter at the input. Specifically, while freezing all layer parameters of the Pangu meteorological model, a lightweight adapter module is added before the model input. The input feature X is transformed by the adapter and output as follows: In the formula, , This is the weight matrix. , For bias terms, This is the Swish activation function.

[0110] To ensure that the model output fully reflects the high-precision information of three-dimensional water vapor, this application defines a loss function based on weighted mean square error to directly utilize the three-dimensional water vapor interpolation results as a supervision signal. The loss function can be expressed as: In the formula, These are the prediction results from the Pangu meteorological model after adapter correction. The data represents the three-dimensional water vapor label data obtained through interpolation, where λ is a weighting coefficient set based on the uncertainty of the impact of each meteorological variable on the regional forecast.

[0111] To further improve the adaptability and generalization ability of the Pangoal Meteorological Model in localized 3D water vapor field and ZTD inversion tasks, this application proposes a parameter-efficient and physically-aware fine-tuning framework—PE-PIML-Meta (Parameter-Efficient, Physics-Informed, Meta-tuning). This framework comprises four innovative modules: parameter-efficient adapters (Adapters + LoRA), physically consistent regularization and uncertainty-weighted loss, meta-learning initialization to enhance few-shot adaptation (MAML-style), and a progressive-contrastive fine-tuning strategy to refine local details in steady state. The specific design is as follows.

[0112] (1) High-efficiency parameter adapter and low-rank fine-tuning

[0113] While keeping most parameters of the Pangu meteorological model frozen, a lightweight adapter layer and low-rank decomposition fine-tuning (LoRA) are introduced to quickly adapt to local three-dimensional water vapor information. The adapter adopts a bottleneck residual structure. LoRA low-rank decomposition is applied to key convolutional layers or self-attention layers. (in , ), only optimize With adapter parameters. This design significantly reduces the number of fine-tuning parameters and accelerates convergence, while retaining the general representational capabilities of the main model.

[0114] (2) Physical-uncertainty mixed loss

[0115] To ensure the physical consistency of the fine-tuned output and its adaptability to different observation qualities, a hybrid loss is introduced: In the formula, the physical consistency term is defined as the difference between the model output after passing through a physical operator (such as the refractive index integral) and the observed ZTD: The uncertainty term is expressed using probability regression / heteroscedastic loss or Bayesian approximation, as follows: In the formula, the network simultaneously outputs the mean. With prediction variance It can automatically weaken the training weights of low-quality observations, thereby improving robustness.

[0116] (3) Meta-learning initialization (fast adaptation with few samples)

[0117] To enhance the rapid adaptation capability under single-site or limited sounding sample data, a two-stage training method using MAML style is adopted: an adapter initialization parameter is trained on historical multi-region / multi-time period data. This allows it to quickly adapt to new sites using a small number of gradient steps. The formal meta-objective is: In the formula, the task Indicates different local subdomains or time periods. This is the learning rate for the inner loop. During the meta-learning phase, only the adapter / LoRA parameters are updated, while the main model parameters remain frozen, thus balancing efficiency and generalization.

[0118] (4) Curriculum + Contrastive Alignment Strategy

[0119] The fine-tuning strategy employs a three-stage training process: coarse-grained → fine-grained → constrained.

[0120] ① Coarse adjustment (short step, large learning rate): enables the adapter to learn the overall bias field;

[0121] ② Fine-tuning (small learning-rate, adding) ): Refine and align with physical consistency constraints;

[0122] ③ Stabilization (Introducing Contrastive Learning): Constructing positive / negative pairs (sample pairs with consistent / inconsistent observations), and applying contrastive loss to the adapter output to maintain local structure without distortion; Contrastive Loss:

[0123] .

[0124] By leveraging the synergistic effect of 3D water vapor and the Pangoal Meteorological Model, not only are the global large-scale features obtained from training the Pangoal Meteorological Model on reanalysis data preserved, but also the shortcomings of regional models in detail description are compensated for through local fine-tuning. This provides solid technical support for building a hybrid forecast that combines generalization ability with regional forecast accuracy.

[0125] Building upon the aforementioned technical solutions, this application also proposes an integrated platform with "sensing-processing-feedback" capabilities. This platform will integrate a 3D tropospheric modeling module with an optimized Pangu weather forecasting module, achieving low-altitude sensing, data augmentation, 3D modeling, real-time positioning, and local short-term nowcasting capabilities. The platform also develops a flight weather risk visualization interface, path planning assistance functions, and a flight mission dynamic adjustment module, supporting real-time decision-making by end users and ensuring the safety and stability of low-altitude UAV missions under complex weather conditions. The construction of this platform can provide real-time support for the refined operation of low-altitude UAVs under complex weather conditions, enhancing their autonomy and safety in typical scenarios such as power line inspection, urban logistics, and emergency response, and providing highly reliable environmental perception and flight risk early warning support for the development of the low-altitude economy.

[0126] The technical solution provided in this application fully integrates multidisciplinary methodologies such as BeiDou navigation and positioning, artificial intelligence modeling, and high-resolution remote sensing analysis. It systematically studies key issues including tropospheric delay 3D modeling and forecasting, forecast correction, and real-time services. It introduces residual compensation networks and AI technology into the GNSS meteorological positioning field, overcoming the technical bottlenecks of traditional models such as low spatial resolution, poor timeliness, and insufficient adaptability. Furthermore, following a technical architecture of "3D modeling and forecasting—3D water vapor-driven forecasting," it progressively constructs a multi-stage processing chain for 3D modeling and dynamic forecasting, achieving a gradual improvement in real-time positioning enhancement and meteorological element forecasting capabilities.

[0127] The various embodiments in this specification are described in a progressive manner. For the same or similar parts between the various embodiments, please refer to each other. Each embodiment focuses on describing the differences from other embodiments.

[0128] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit this application. Although this application 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. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of this application.

Claims

1. A weather forecasting method based on the collaborative driving of BeiDou three-dimensional water vapor and meteorological large model, characterized in that, The method includes: The predicted data from the obtained GraphCast model and the measured data from ERA5 were registered point by point to obtain the residual field data. The residual field data is used as input to a deep learning residual correction network to obtain the predicted residual field data output by the deep learning residual correction network. The deep learning residual correction network includes a 3D convolution module, a ResNet cross-layer residual connection structure and a physical constraint module. Each residual block in the ResNet cross-layer residual connection structure embeds a channel attention mechanism and a spatial attention module, and a ConvLSTM structure is introduced at the back end. A terrain-adaptive algorithm is used to transform the predicted ERA5 meteorological parameters corresponding to the predicted residual field data into a three-dimensional grid structure in height coordinates, and the integral method is used to determine the zenith tropospheric delay. The coupling relationship between temperature, humidity and air pressure is nonlinearly fitted through a multi-layer sensing mechanism to form a dynamic correction term for the traditional π factor; based on the dynamic correction term, the zenith tropospheric delay is converted into precipitable water through a spatiotemporal inversion network; the spatiotemporal inversion network includes a CNN module and a Transformer module. The predicted ERA5 meteorological parameters and the precipitable water data input parameter efficient adapter are adjusted, and the adjusted predicted ERA5 meteorological parameters and the precipitable water data are used as input to the Pangu meteorological big model to obtain the meteorological forecast factors output by the Pangu meteorological big model. The parameter efficient adapter adopts a bottleneck residual structure and uses LoRA low-rank decomposition on key convolutional layers or self-attention layers. It is pre-trained through a physical-uncertainty hybrid loss, meta-learning initialization, and progressive-contrastive fine-tuning strategy.

2. The method according to claim 1, characterized in that, The method further includes: In response to receiving geographic coordinate information of the target location input by the user; the geographic coordinate information includes the longitude parameter, latitude parameter, elevation parameter and time parameter of the target location; The geographic coordinate information is determined by identifying four adjacent reference grid points within the three-dimensional grid structure. The tropospheric delay of the target location is determined at each altitude using the horizontal inverse distance weighted interpolation method. The vertical distribution of tropospheric delay at the target location was fitted using an exponential function. Based on the elevation parameters of the target location, the corresponding target tropospheric delay is obtained from the zenith tropospheric delay.

3. The method according to claim 2, characterized in that, The physical constraint module includes tropospheric refractive index constraints, tropospheric delay constraints, integrated loss function constraints, and terrain correction operator constraints. The tropospheric refractive index constraint is: Where N is the tropospheric refractive index, P is the air pressure, T is the temperature, e is the water vapor pressure, and constants k1, k2, and k3 are empirical coefficients. The tropospheric delay constraint is: Where ZTD is the tropospheric retardation, N is the tropospheric refractive index, h1 is the bottom height, and h2 is the top height; The constraints of the comprehensive loss function are as follows: Where λ1, λ2, and λ3 are weighting coefficients. Indicates the mean square error loss. Represents structural similarity loss. Represents the loss due to physical consistency constraints: Among them, ZTD calc ( (ZTD) represents the zenith tropospheric delay obtained from the refractive index field predicted by the model. ERA5 This represents the actual zenith tropospheric delay corresponding to the ERA5 dataset; The terrain correction operator is constrained as follows: Where H is the weight matrix in the terrain correction module. H is the terrain correction factor. c For the comprehensive terrain correction factor, H ERA5 The elevation field provided for the ERA5 dataset, H DEM Provides the actual surface elevation for the Digital Elevation Model (DEM).

4. The method according to claim 3, characterized in that, The deep learning residual correction network is: Among them, F out The output of the deep learning residual correction network is defined as follows: Attention represents the multidimensional attention weighting function used to fuse channel and spatial weights; ConvLSTM is responsible for temporal dimension modeling; ReLU is a non-linear activation function; Conv3D represents a three-dimensional convolutional module; and F... in The three-dimensional meteorological feature tensor input to the deep residual correction network originates from the residual field data and is used to predict and correct the residual field by means of convolutional feature extraction and temporal modeling.

5. The method according to claim 1, characterized in that, The method employs a terrain-adaptive algorithm to transform the predicted ERA5 meteorological parameters corresponding to the predicted residual field data into a three-dimensional grid structure in height coordinates, and uses an integral method to determine the zenith tropospheric delay, including: Based on the predicted residual field data and the obtained ERA5 meteorological parameters, the predicted ERA5 meteorological parameters are obtained. The terrain adaptive interpolation algorithm is used to introduce a terrain gradient correction term in the interpolation process of air pressure and temperature, and to perform vertical layer reconstruction to obtain a three-dimensional mesh structure in height coordinates. Based on the predicted ERA5 meteorological parameters, the zenith tropospheric delay is determined using an integral method.

6. The method according to claim 1, characterized in that, The method further includes: The posterior probability distribution of the precipitable water volume is obtained through Bayesian uncertainty estimation. The posterior probability distribution is sampled using the MCMC method to obtain the confidence interval of the precipitable water amount, thereby achieving an assessment of the accuracy of the precipitable water amount.

7. The method according to claim 2, characterized in that, The bottleneck residual structure in the parameter-efficient adapter is as follows: ;in, , Let X be the weight matrix, and X be the model input feature tensor. , For bias terms, The Swish activation function is used; the LoRA low-rank decomposition is... Where A and B are low-rank matrices of LoRA, and the dimension of A is . r d, used for dimensionality reduction, where the dimension of B is... r d' is used for dimensionality recovery, and r is the rank of the low-rank decomposition, used to control the parameter size and adaptation speed.

8. The method according to claim 7, characterized in that, The physical-uncertainty hybrid loss is: ;in, These are the weighting coefficients for the mean squared error loss term. λ is the mean square error loss; SSIM These are the weighting coefficients for the structural similarity loss term. For structural similarity loss; The weighting coefficients for the physical consistency loss term, For physical consistency loss: Among them, ZTD ( (ZTD) represents the zenith tropospheric delay calculated based on the model-predicted refractive index field. obs For the actual zenith tropospheric delay from ERA5; The weighting coefficients for the uncertainty loss term. Loss due to uncertainty: , where y i For the actual observed value, μ i Output values ​​for the model. The variance of the model prediction is represented by i, which represents the sample index, i.e., the summation over each grid point.

9. The method according to claim 8, characterized in that, The meta-learning initialization involves pre-training an initialization parameter on historical multi-region / multi-time period data. This allows it to quickly adapt to new sites using a small number of gradient steps; the formal meta-objective is: Among them, the task Indicates different local subdomains or time periods. For the inner loop learning rate, For the formalized meta-objective, The training loss is used; during the meta-learning phase, only the parameters of the efficient adapter / LoRA are updated.

10. The method according to claim 9, characterized in that, The progressive-contrastive fine-tuning strategy includes a coarse-tuning strategy, a fine-tuning strategy, and a stabilization strategy; wherein, the coarse-tuning strategy is characterized by short steps and a large learning rate, enabling the adapter to learn the overall bias field; the fine-tuning strategy is characterized by a small learning rate combined with... To refine alignment with physical consistency constraints; the stabilization strategy combines contrastive learning to construct observation-consistent / inconsistent sample pairs, and applies contrastive loss to the output of the parameter-efficient adapter to maintain local structure without distortion; The contrast loss is expressed as: ;in, For temperature parameters, The feature representation of the current sample, As a positive sample, This is a negative sample.

Images