A route meteorological guarantee method and system based on meteorological numerical prediction

Abstract

The application provides a route weather guarantee method and system based on meteorological numerical prediction, relates to the low-altitude flight safety guarantee field, and solves the technical problems that the meteorological information is disconnected with the flight task and the risk assessment lacks continuity in the prior art. The method comprises the following steps: acquiring observation and prediction data of a route area, and generating meteorological field data by using a four-dimensional variational assimilation method; correcting errors of the meteorological field data by using an AI corrector constructed by a space-time graph convolution network to obtain an optimized prediction result; generating route-height meteorological profile data by using a spatial interpolation algorithm according to user-input route information; performing disaster risk assessment on the profile data based on a risk propagation network model to generate a risk heat map and superimpose the risk heat map on the profile to form a route guarantee map; and finally, performing route dynamic planning and height layer optimization on the guarantee map by using a multi-objective optimization algorithm to output an optimized flight path.

Description

Technical Field

[0001] This invention belongs to the field of low-altitude flight safety assurance, specifically a method and system for airway meteorological support based on numerical meteorological forecasting. Background Technology

[0002] Aviation meteorological support is a crucial link in ensuring safe and efficient flight operations. With advancements in numerical weather prediction technology and meteorological observation methods, the availability of meteorological data has significantly increased in both spatiotemporal resolution and the variety of elements. However, existing technological systems still have significant shortcomings in transforming massive amounts of meteorological data into precise route risk information directly usable for flight decision-making and in formulating optimal flight plans, thus hindering further improvements in support effectiveness.

[0003] In the meteorological data assimilation and analysis stage, traditional methods, when dealing with complex terrain areas, typically employ isotropic or simple anisotropic assumptions in their background error covariance models. This fails to adequately consider the impact of terrain slope and undulation on the spatial structure of meteorological element errors, leading to significant biases in the initial field in mountainous regions and introducing systematic error sources into subsequent forecasts. Furthermore, in the meteorological forecast error correction stage, existing correction methods based on statistical or simple machine learning models mostly focus on reducing statistical errors between the forecast and actual fields, lacking explicit constraints on the inherent conservation laws of atmospheric physical processes. This can easily result in physically inaccurate or inconsistent corrected meteorological fields, such as the generation of false mass sources or sinks, affecting the reliability of downstream risk assessments.

[0004] In terms of airway meteorological hazard risk assessment, existing technologies are mostly based on threshold judgments of single-point meteorological elements or use simple spatial smoothing methods. These methods fail to effectively model the dynamic propagation and cumulative effects of risks in three-dimensional space, particularly along the prevailing airflow direction, and also ignore the causal relationships between different hazard factors. This results in poor spatial continuity of risk assessment results, short warning lead times, and insufficient ability to identify transmissible risks. Furthermore, current decision support systems largely rely on pilots or dispatchers manually analyzing meteorological charts and making subjective route selections based on experience. Therefore, it is difficult to quickly generate optimal route plans under complex airspace constraints and real-time changing meteorological conditions, and the scientific rigor, consistency, and efficiency of decision-making need to be improved.

[0005] Therefore, existing airway meteorological support technologies suffer from a series of problems, such as insufficient data processing accuracy, lack of physical consistency, simplistic risk assessment models, and low level of intelligence in airway planning. These issues result in a break in the chain from raw data to final flight decisions, making it difficult to meet the urgent needs of modern aviation for high safety and high efficiency. Summary of the Invention

[0006] This application provides a method and system for airway meteorological support based on numerical weather forecasting, which solves the technical problems in the prior art such as the disconnect between meteorological information and flight missions, the lack of continuity and correlation in risk assessment, and the insufficient safety of airway planning.

[0007] To achieve the above objectives, this application adopts the following technical solution:

[0008] Firstly, a method for providing airway meteorological support based on numerical weather forecasting is provided, including:

[0009] Meteorological observation data and numerical forecast data for the flight route area are acquired, and the meteorological observation data is assimilated into the numerical forecast data using a four-dimensional variational assimilation method to generate meteorological field data.

[0010] An AI corrector built based on a spatiotemporal graph convolutional network is used to correct errors in the meteorological field data to obtain optimized weather forecast results.

[0011] Based on the route information input by the user and the optimized weather forecast results, route-altitude meteorological profile data are generated using a spatial interpolation algorithm;

[0012] Disaster risk assessment is performed on the meteorological profile data based on the risk propagation network model, a risk heat map is generated, and the risk heat map is superimposed on the meteorological profile data to obtain a route support map;

[0013] By combining a multi-objective optimization algorithm, the route support map is subjected to dynamic route planning and altitude layer optimization to obtain an optimized flight path; the optimized flight path is used to ensure low-altitude flight safety.

[0014] Based on the above technical solutions, the airway meteorological support method based on meteorological numerical forecasting provided in this application improves the accuracy and reliability of meteorological field data in the airway area by fusing multi-source data through four-dimensional variational assimilation and combining it with an AI corrector for error correction, thus providing a high-quality data foundation for subsequent support. Then, by generating route-altitude meteorological profile data and performing disaster risk assessment and visualization overlay, complex meteorological field information is transformed into an airway support map closely related to flight missions and with clearly defined risks, lowering the threshold for interpreting meteorological information and improving the intuitiveness of situational awareness and decision-making efficiency. Finally, by introducing a multi-objective optimization algorithm to automate route planning on the airway support map containing risk information, this method can seek optimal or satisfactory flight paths in multiple dimensions such as safety, economy, and efficiency, transforming the traditional decision-making process relying on human experience into intelligent assisted decision-making based on quantitative models, thereby systematically improving the scientific, forward-looking, and intelligent level of low-altitude flight meteorological safety support.

[0015] Furthermore, the assimilation of the meteorological observation data into the numerical forecast data using the four-dimensional variational assimilation method includes:

[0016] The meteorological observation data is processed using an anisotropic background error covariance matrix, and the elements of the covariance matrix are calculated by taking into account the similarity of terrain slope; wherein, the meteorological observation data includes satellite meteorological data, ground sensor data and radar observation data;

[0017] The expected value of terrain correction for each ground sensor location is calculated using a digital elevation model. Then, the deviation between the observed value of the ground sensor and the expected value of terrain correction is calculated. When the deviation exceeds a preset threshold, the observed value is downweighted to generate meteorological observation data after quality control.

[0018] A four-dimensional variational assimilation objective function incorporating terrain constraints is constructed. By minimizing this objective function, the quality-controlled meteorological observation data is assimilated into numerical weather prediction data, generating a high-resolution analytical field as meteorological field data. The background error covariance matrix is ​​used to quantify the correlation and spatial structure characteristics of meteorological element errors between different spatial points in the background field within the four-dimensional variational assimilation objective function. Its inverse matrix is ​​used to weight the deviation between the background field and the observation field, thereby controlling the reasonable adjustment magnitude and direction of the analytical field to the background field when absorbing observation information. This ensures that the assimilation process can produce a physically coordinated and spatially structurally reasonable analytical field in complex terrain areas.

[0019] Furthermore, the expression for the four-dimensional variational assimilation objective function including terrain constraints is: ;in, express The analysis field at any given time, wherein the analysis field includes at least the wind field, temperature field, and humidity field; The background field represents the initial guess field provided by the numerical weather prediction model; Represents the background error covariance matrix; The vector of observation data at time i includes, but is not limited to, satellite radiation brightness temperature, temperature / pressure / humidity / wind speed observations from ground weather stations, radar reflectivity factor, and radial velocity. For the observation operator at time i, it is used to map the model variables to the observation space; Let be the observation error covariance matrix at time i; λ is the terrain-constrained reference field, calculated based on the digital elevation model; λ is the terrain constraint weight coefficient, used to control the intensity of terrain influence. This is the terrain constraint weight matrix, which is adjusted according to the terrain complexity. This represents the total number of moments within the assimilation time window.

[0020] Furthermore, the calculation formula for processing the meteorological observation data using the anisotropic background error covariance matrix is ​​as follows: ;in, The element in the i-th row and j-th column of the background error covariance matrix; , This represents the standard deviation of the background error for the i-th and j-th grid points; The spatial correlation coefficient is used to characterize the error correlation between two points. This represents the horizontal and vertical distances between the i-th and j-th grid points; The scales of the background error in the x, y, and z directions represent anisotropy; This is a terrain slope similarity function.

[0021] Furthermore, the AI ​​corrector constructed based on the spatiotemporal graph convolutional network performs error correction on the meteorological field data, including:

[0022] Multiple two-dimensional horizontal fields of the input meteorological field data are processed by a two-dimensional convolutional encoder with shared weights to extract spatial feature maps of each meteorological element; wherein, the two-dimensional horizontal field represents the gridded data of a single meteorological element on any isobaric surface or model layer.

[0023] The spatial feature maps of each extracted meteorological element are stitched together along the channel dimension and input into the physical information spatiotemporal Transformer module, which outputs spatiotemporal features. The physical information spatiotemporal Transformer module embeds the spatial coordinates and forecast lead time information of each grid point through learnable position encoding, and then captures the global long-range spatiotemporal dependence through a multi-head self-attention mechanism to simulate the long-range interaction between weather systems.

[0024] The spatiotemporal features are input into the decoder, and a preliminary error correction field is generated through multiple deconvolutional layers or upsampling convolutional layers; the preliminary error correction field is then added to the meteorological field data to obtain a preliminary correction field.

[0025] The preliminary correction field is input into the physical constraint layer, and the degree of violation of the physical constraint is used to calculate whether the preliminary correction field satisfies the preset physical conservation law. The physical loss is determined by the degree of violation of the physical constraint.

[0026] During the training process of the spatiotemporal graph convolutional network, the mean square error between the corrected field output by the forward pattern and the reference analysis field is used as the main loss, which is weighted and summed with the physical loss to optimize the network parameters through backpropagation. The reference analysis field represents the true value of the meteorological field used to supervise the training of the network, which is obtained by acquiring historical analysis data from the same period.

[0027] When the spatiotemporal graph convolutional network performs inference, the network outputs optimized weather forecast results and forecast uncertainty estimates for each grid point.

[0028] Furthermore, the expression for the loss function of the spatiotemporal graph convolutional network during training is: ;in, Main loss, Let represent the initial corrected field for the i-th training sample. This represents the reference analysis field corresponding to the i-th training sample. The total number of samples in a training batch; For physical loss, This is the physical loss weighting coefficient.

[0029] Furthermore, the degree of violation of physical constraints is obtained by the deviation of the calculation result of applying physical law operators to the preliminary correction field from zero, wherein the physical law operators include gradient operators and divergence operators.

[0030] Furthermore, the process of generating the route-altitude meteorological profile data includes:

[0031] The user-input route is discretized into a sequence of waypoints at a preset step size. The horizontal coordinates, altitude, and circulation feature vector of the grid where the waypoint is located, extracted from the weather forecast results, are concatenated into a conditional vector. The circulation feature vector is obtained by extracting the standardized values ​​of each meteorological element within a preset radius from the grid data of the weather forecast results, centered on the grid point closest to the waypoint, and flattening them into a one-dimensional vector.

[0032] The conditional vector is input into a pre-trained conditional neural radiation field model, which outputs meteorological element values ​​for each waypoint. The meteorological element values ​​represent the meteorological state at the location of the waypoint, including at least temperature, zonal wind, meridional wind, vertical velocity, relative humidity, and turbulent kinetic energy.

[0033] By traversing all waypoints and user-specified altitude layers, and querying the conditional neural radiation field model, a two-dimensional data matrix is ​​output to obtain the route-altitude two-dimensional meteorological profile data. The horizontal axis of the two-dimensional data matrix is ​​the distance along the route, the vertical axis is the altitude, and each element in the matrix is ​​a set of meteorological element values ​​at the location of the waypoint.

[0034] Furthermore, the disaster risk assessment of the meteorological profile data based on the risk propagation network model includes:

[0035] Each grid point in the meteorological profile data is regarded as a graph node, and the feature vector of each graph node is composed of the meteorological element values ​​of the grid point; edges are established between spatially adjacent grid points, and additional directed edges are established downwind of the nodes according to the wind field direction to obtain a dynamic risk propagation map.

[0036] The dynamic risk propagation graph and the feature vectors of the graph nodes are input into the risk propagation network model to perform risk propagation calculation and risk assessment, thereby obtaining the comprehensive risk value of each node;

[0037] Based on the preset risk level threshold, the two-dimensional risk distribution matrix composed of the comprehensive risk value of each node is mapped into a color map to obtain a risk heat map.

[0038] Furthermore, the risk propagation network model is a K-layer graph attention network, including a graph construction module, a feature encoding module, K stacked causal modulation graph attention layers, and a risk assessment module; wherein,

[0039] The graph construction module is used to define each meteorological element grid point in the meteorological profile data as a graph node, extract meteorological disaster-related indices for each graph node to form an initial node feature vector, and construct spatial adjacency edges and causal propagation directed edges between graph nodes to obtain a dynamic risk propagation graph; wherein, the meteorological disaster-related indices include, but are not limited to, turbulence index, icing index, convective available potential energy, vertical wind shear, and visibility;

[0040] The feature encoding module maps the set of initial node feature vectors to latent space features using a multilayer perceptron, and outputs the encoded initial node feature matrix.

[0041] The K stacked causal modulation graph attention layers are used to perform multi-round iterative information aggregation and updating of node features based on the initial node feature matrix and the edge adjacency relationships in the dynamic risk propagation graph, to obtain the final node feature representation; wherein, the internal workflow of the l-th layer of the causal modulation graph attention layer is as follows:

[0042] Perform linear projection on the features of each node. , thus obtaining the projected node features; Let be the learnable weight matrix of the l-th layer. Let i be the feature vector of node i in the input of layer l. The projected feature vector of node i;

[0043] Based on the projected node features, the original attention score is calculated for the node pairs (i,j) with edge connections in the dynamic risk propagation graph. The calculation formula is: ; Let be the feature vector of node j after projection. () is the activation function for a linear rectifier with leakage. This represents vector concatenation. This is a learnable attention vector;

[0044] Introducing the causal prior matrix M and adaptive scaling The original attention score is adjusted to obtain the modulated attention score. The calculation formula is: The elements of the causal prior matrix M , representing the statistical causal strength of the disaster occurring at node i due to the state of node j, learned from historical disaster data using a causal learning algorithm; the adaptive scale between node i and nodes. Calculated based on the terrain differences and wind shear between nodes i and j;

[0045] Based on the modulation attention score, node i is aggregated with its neighboring nodes to obtain the updated features, calculated as follows: ; For activation functions;

[0046] The output of the l-th layer is the updated node feature matrix. ;

[0047] The risk assessment module is used to utilize a fully connected neural network to process the final node feature matrix output by the last causal modulation attention layer. By performing nonlinear transformation and mapping, the comprehensive risk value of each node is obtained, and the comprehensive risk values ​​of all nodes constitute a two-dimensional risk distribution matrix.

[0048] Furthermore, the step of combining a multi-objective optimization algorithm to perform dynamic route planning and altitude-level optimization on the route support chart includes:

[0049] The problem of route dynamic planning and altitude layer optimization is modeled as a multi-objective constrained optimization problem. The decision variables of the multi-objective constrained optimization problem are the flight path R and the flight altitude profile H. The objectives of the multi-objective constrained optimization problem include minimizing the total risk, minimizing the total flight time, and minimizing the total energy consumption cost. The constraints of the multi-objective constrained optimization problem include flight altitude restrictions, climb / descent rate restrictions, path curvature restrictions, airspace restrictions, and fuel quantity constraints.

[0050] The multi-objective constrained optimization problem is then solved using the non-dominated sorting genetic algorithm NSGA-III to obtain a Pareto optimal solution set; the optimal solution set is a set of schemes that are non-dominated in the objective space.

[0051] From the optimal solution set, based on the user-provided preference weights for each objective, the weighted Chebyshev distance from each candidate solution to the ideal point is calculated using the weighted Chebyshev method. The candidate solution with the smallest distance is selected to obtain the final solution, and the optimized flight path is output.

[0052] Secondly, this application provides a route meteorological support system based on meteorological numerical forecasting, comprising: a data fusion module, an intelligent correction module, a profile generation module, a risk assessment module, a route optimization module, and a visualization and interaction module; wherein,

[0053] The data fusion module is communicatively connected to a multi-source meteorological data source to acquire meteorological observation data and numerical forecast data for the airway area, and uses a four-dimensional variational assimilation method to assimilate the meteorological observation data into the numerical forecast data to generate meteorological field data.

[0054] The intelligent correction module is connected to the data fusion module and is used to receive the meteorological field data and perform error correction on the meteorological field data based on the AI ​​corrector constructed based on the spatiotemporal graph convolutional network to obtain the optimized meteorological forecast result.

[0055] The profile generation module, the intelligent correction module, and the visualization interaction module are connected and used to generate route-altitude meteorological profile data through a spatial interpolation algorithm based on user route information from the interaction module and optimized weather forecast results from the intelligent correction module.

[0056] The risk assessment module is connected to the profile generation module and is used to perform disaster risk assessment on the meteorological profile data based on the risk propagation network model, generate a risk heat map, and overlay the risk heat map onto the meteorological profile data to obtain a route support map.

[0057] The route optimization module is connected to the risk assessment module and the visualization interaction module. It is used to combine a multi-objective optimization algorithm to perform dynamic route planning and altitude layer optimization on the route support map to obtain an optimized flight path.

[0058] The visualization interaction module is connected to the profile generation module and the route optimization module. It is used to receive route information and decision preferences input by the user, and to display the route support map and the optimized flight path in a visual form.

[0059] Furthermore, the route support chart and the optimized flight path are displayed through a visual interactive module, which supports interactive operation, including:

[0060] Multi-dimensional rendering: Input the route support map and a 3D earth model. In the 3D earth model, the route support map is rendered vertically along the actual flight path, and a 3D route risk visualization scene is output.

[0061] Scenario simulation: In response to hypothetical altitude or heading instructions input by the user, the background recalculates the profile, risks, and new optimized flight path under the new conditions based on the current weather field, and displays a comparison view.

[0062] Compared with the prior art, the beneficial effects of this application are:

[0063] First, in the data assimilation stage, by introducing a background error covariance matrix that considers the influence of terrain and anisotropy, and combining it with a variational objective function constrained by terrain, the analysis accuracy of the initial meteorological field in complex terrain areas is effectively improved, laying a solid foundation for subsequent forecast correction. In the forecast correction stage, the constructed physical information spatiotemporal Transformer network not only captures the complex long-range spatiotemporal dependencies between meteorological elements, but also, by introducing physical conservation laws as constraints for joint training, enables the AI ​​correction results to significantly reduce statistical errors while strictly adhering to basic physical laws. This results in a more physically reasonable, quantifiable uncertainty high-precision forecast field, overcoming the physical inconsistency problems that may occur in traditional pure data-driven models.

[0064] Secondly, this application utilizes a conditional neural radiation field model to generate flight path-height profiles, enabling continuous and high-resolution reconstruction of detailed meteorological structures in the vertical direction of the flight path based on a large-scale circulation background. Compared to traditional interpolation methods, this method better reflects the true state under complex weather conditions. Furthermore, risk assessment is conducted through a causal modulation graph attention network. This network not only models the natural diffusion of risk in spatially adjacent areas but also simulates the dynamic propagation process of risk along the dominant airflow direction by introducing causal priors and wind-guided adaptive propagation mechanisms. This makes the identification and assessment of hazards such as turbulence and icing more consistent with their physical mechanisms of occurrence and development, thereby improving the forecast accuracy and spatial rationality of the risk heat map.

[0065] Finally, by deeply integrating intuitive and visualized route risk assurance maps with multi-objective optimization algorithms, the system can automatically perform quantitative trade-offs and global optimization among multiple interdependent objectives such as safety, economy, and efficiency, providing users with a series of Pareto-optimal route planning solutions. This approach transforms the traditional qualitative route selection process, which relies on pilots' personal experience, into an assisted decision-making process based on comprehensive quantitative assessment and intelligent algorithms. This fundamentally improves the proactive avoidance capabilities of low-altitude flight in response to severe weather and enhances overall flight efficiency, achieving a leap from risk notification to decision support in aviation meteorological support. Attached Figure Description

[0066] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0067] Figure 1 A system architecture diagram of a route meteorological support system based on numerical weather forecasting is provided for embodiments of this application;

[0068] Figure 2 A flowchart illustrating a route meteorological support method based on numerical meteorological forecasting, provided for an embodiment of this application;

[0069] Figure 3 A flowchart illustrating another method for airway meteorological support based on numerical meteorological forecasting provided in this application embodiment;

[0070] Figure 4 A flowchart illustrating another method for airway meteorological support based on numerical meteorological forecasting provided in this application embodiment;

[0071] Figure 5 A flowchart illustrating another route meteorological support method based on numerical meteorological forecasting provided in this application embodiment. Detailed Implementation

[0072] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0073] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0074] The route meteorological support method based on numerical weather prediction provided in this application embodiment can be applied to, for example... Figure 1The system shown is a route weather support system based on numerical weather prediction, such as... Figure 1 As shown, the system includes: a data fusion module, an intelligent correction module, a profile generation module, a risk assessment module, a route optimization module, and a visualization and interaction module; among which,

[0075] The data fusion module is connected to a multi-source meteorological data source to acquire meteorological observation data and numerical forecast data for the airway area. It uses a four-dimensional variational assimilation method to assimilate the meteorological observation data into the numerical forecast data to generate meteorological field data.

[0076] The intelligent correction module, connected to the data fusion module, is used to receive meteorological field data and perform error correction on the meteorological field data based on the AI ​​corrector built on the spatiotemporal graph convolutional network to obtain optimized weather forecast results.

[0077] The profile generation module, intelligent correction module, and visualization interaction module are connected to generate route-altitude meteorological profile data through spatial interpolation algorithms based on user route information from the interaction module and optimized weather forecast results from the intelligent correction module.

[0078] The risk assessment module, connected to the profile generation module, is used to conduct disaster risk assessment on meteorological profile data based on the risk propagation network model, generate a risk heat map, and overlay the risk heat map onto the meteorological profile data to obtain a route support map.

[0079] The route optimization module, connected to the risk assessment module and the visualization interaction module, is used to combine multi-objective optimization algorithms to perform dynamic route planning and altitude-level optimization on the route support map to obtain the optimized flight path.

[0080] The visualization and interaction module, connected to the profile generation module and the route optimization module, is used to receive route information and decision preferences input by the user, and to display the route support map and optimized flight path in a visual form.

[0081] To address the technical problems in existing technologies, such as the disconnect between meteorological information and flight missions, the lack of continuity and correlation in risk assessment, insufficient safety in route planning, limited data processing accuracy, and inefficiency due to reliance on human experience in decision-making, this application provides a route meteorological support method and system based on numerical meteorological forecasting. The method includes:

[0082] Meteorological observation data and numerical forecast data for the flight route area are acquired, and the meteorological observation data are assimilated into the numerical forecast data using a four-dimensional variational assimilation method to generate meteorological field data.

[0083] An AI corrector built on a spatiotemporal graph convolutional network corrects errors in meteorological field data to obtain optimized weather forecast results.

[0084] Based on the route information input by the user and the optimized weather forecast results, route-altitude meteorological profile data are generated through spatial interpolation algorithms;

[0085] Disaster risk assessment is performed on meteorological profile data based on a risk propagation network model, generating a risk heat map, which is then overlaid onto the meteorological profile data to obtain a route support map.

[0086] By combining multi-objective optimization algorithms, dynamic route planning and altitude-level optimization are performed on the route support map to obtain optimized flight paths, which are used to ensure low-altitude flight safety.

[0087] Based on this, this application transforms massive meteorological data into intuitive and usable flight decision support information through multi-source data fusion, error correction, refined profile generation, scientific risk assessment, and intelligent path optimization. This improves the safety, efficiency, and scientific nature of low-altitude flight decisions and makes up for the shortcomings of traditional technologies in data transformation, risk modeling, and intelligent planning.

[0088] like Figure 2 As shown in the embodiment of this application, a route meteorological support method based on numerical meteorological forecasting is provided, comprising:

[0089] S1. Acquire meteorological observation data and numerical forecast data for the flight route area, and use the four-dimensional variational assimilation method to assimilate the meteorological observation data into the numerical forecast data to generate meteorological field data.

[0090] Among them, the four-dimensional variational data assimilation method is a data analysis method that combines multi-source observational data within a specific time window with numerical weather prediction models. It obtains the optimal initial field by minimizing the objective function. Its core is to comprehensively utilize observational information from both temporal and spatial dimensions to correct the initial state of the prediction model. Meteorological field data refers to a dataset that comprehensively describes the distribution and changes of meteorological elements within a specific region, covering the spatial distribution information of key meteorological elements such as wind field, temperature field, humidity field, and pressure field.

[0091] In some implementations, data assimilation using the four-dimensional variational assimilation method may include: first, collecting meteorological observation data from multiple channels such as satellite remote sensing, ground observation stations, and radar detection, while simultaneously acquiring forecast data output from numerical weather prediction models; then, constructing an objective function that includes observation errors and background errors to quantify the deviation between the observation data and the forecast data; finally, solving for the minimum value of the objective function through a numerical optimization algorithm to integrate the information from the observation data into the forecast data, thereby generating meteorological field data with a better initial state.

[0092] For example, for a certain low-altitude logistics route area, satellite cloud image data, temperature and humidity observation data from ground meteorological stations, wind field detection data from Doppler radar, and numerical forecast data output by the global forecast system can be obtained for the area within 24 hours. These data are then fused using a four-dimensional variational assimilation method to generate three-dimensional meteorological field data of the route area with a resolution of 1km×1km.

[0093] S2. An AI corrector based on a spatiotemporal graph convolutional network is used to correct errors in the meteorological field data, resulting in optimized weather forecasts.

[0094] The meteorological field data is subject to various factors during the generation process, such as the simplification assumptions of the numerical forecast model itself, measurement errors of the observation data, and computational approximations during the assimilation process. These errors can lead to deviations between the meteorological forecast results and the actual meteorological conditions, which may pose safety risks if used directly for flight decision-making.

[0095] In some implementations, the Spatio-Temporal Graph Convolutional Network combines the spatial relationship modeling capabilities of graph convolutional networks with the temporal evolution capture capabilities of time series models. It can uncover the spatial correlation characteristics and temporal trends of meteorological elements. By learning the error patterns between historical observation data and forecast data, it can intelligently correct systematic errors in numerical forecasting models, random errors in observation data, and biases introduced during assimilation.

[0096] It should be noted that the training of the AI ​​corrector needs to be based on a large amount of high-quality historical meteorological data, including historical observation data, historical forecast data and corresponding actual meteorological data, to ensure that the model can learn stable and reliable error correction rules.

[0097] S3. Based on the route information input by the user and the optimized weather forecast results, generate route-altitude meteorological profile data through spatial interpolation algorithm.

[0098] The optimized weather forecast results are usually presented in the form of regular grid points. However, the flight path of low-altitude flight is a continuous spatial path. Grid point data cannot directly reflect the specific meteorological elements at each point along the flight path and at different altitude layers. Spatial interpolation can transform discrete grid point data into meteorological data that is continuously distributed along the flight path, which can meet the needs of flight decision-making for refined meteorological information.

[0099] In some implementations, methods for generating flight path-altitude meteorological profile data include linear interpolation, cubic spline interpolation, and Kriging interpolation. Linear interpolation calculates the linear relationship between adjacent grid points to obtain interpolation point data; cubic spline interpolation constructs a smooth interpolation function, ensuring the spatial continuity and smoothness of meteorological elements; and Kriging interpolation, based on the principle of spatial autocorrelation, can quantify interpolation errors and is suitable for situations where the spatial distribution of meteorological elements has a certain structure.

[0100] For example, after a user inputs the starting point, ending point, and key points along a low-altitude inspection route, the system obtains optimized meteorological forecast grid data for that area. Using cubic spline interpolation, the system divides the route points along the route direction into 30-meter increments and divides the altitude layers vertically into 10-meter intervals. Through interpolation, the system calculates the meteorological element values ​​such as temperature, wind speed, and humidity for each route point at different altitude layers, ultimately generating route-altitude meteorological profile data.

[0101] S4. Based on the risk propagation network model, conduct disaster risk assessment on meteorological profile data, generate risk heat map, and overlay the risk heat map onto the meteorological profile data to obtain the route support map.

[0102] Among them, the risk propagation network model is a model based on neural network algorithms to simulate the spatial propagation, diffusion and accumulation of meteorological disaster risks. Its core is to regard each spatial point in the meteorological profile data as a network node, and to characterize the propagation path and intensity of risks through the correlation between nodes.

[0103] In some implementation methods, common disaster risk assessment approaches include threshold-based assessment methods, statistical model-based assessment methods, and machine learning-based assessment methods. Threshold-based assessment methods determine the existence and level of disaster risk at various points along the flight route by setting critical values ​​corresponding to each meteorological disaster. Statistical model-based assessment methods establish risk assessment models by analyzing the statistical correlation between historical meteorological disaster data and flight safety events. Machine learning-based assessment methods directly predict disaster risk levels using meteorological profile data by training classification or regression models.

[0104] For example, based on the risk propagation network model, each grid point in the route-altitude meteorological profile data is taken as a node. Risk propagation edges between nodes are constructed according to wind field direction and terrain features. A machine learning-based assessment method is adopted, and a risk assessment model trained by inputting historical turbulence and icing disaster data is used to calculate the turbulence risk level and icing risk level of each node. The risk levels are mapped to different colors to generate a risk heat map, which is then overlaid with the meteorological profile data to obtain a route protection map that clearly shows the meteorological elements and corresponding risks at each location of the route.

[0105] S5. Combining multi-objective optimization algorithms, dynamic route planning and altitude layer optimization are performed on the route support map to obtain optimized flight paths for ensuring low-altitude flight safety.

[0106] Low-altitude flight faces complex airspace constraints and weather risks. A single flight path or altitude may not be able to simultaneously meet multiple objectives such as safety, efficiency, and economy. Through dynamic planning and altitude optimization, it is possible to avoid high-risk areas while taking into account requirements such as flight time and energy consumption, thereby maximizing the comprehensive benefits of flight missions.

[0107] In some implementations, route optimization and altitude optimization can be achieved by establishing multiple objective functions and then using methods such as genetic algorithms, particle swarm optimization algorithms, and simulated annealing algorithms to perform optimization operations, thereby finding a Pareto optimal solution that can balance the various objectives while satisfying the constraints.

[0108] Based on the above technical solutions, this application provides a route meteorological support method based on meteorological numerical forecasting. It achieves effective fusion of multi-source meteorological data through a four-dimensional variational assimilation method; reduces meteorological data errors and improves the accuracy and reliability of forecast results by using an AI corrector constructed with a spatiotemporal graph convolutional network; achieves precise matching of meteorological information and flight routes through route-altitude meteorological profile data generated by spatial interpolation algorithms, meeting the needs of refined decision-making; makes complex risk information more intuitive and understandable through disaster risk assessment and route support map generation based on a risk propagation network model, lowering the decision-making threshold; and transforms the traditional decision-making model relying on human experience into a scientific intelligent assisted decision-making model by combining multi-objective optimization algorithms for route dynamic planning and altitude layer optimization, improving the level of low-altitude flight meteorological safety assurance and providing strong support for the development of the low-altitude economy.

[0109] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 3 As shown, the above S1 can be implemented through the following S101, S102 and S103, which are explained in detail below:

[0110] S101. Meteorological observation data are processed using an anisotropic background error covariance matrix. The background error covariance is obtained by calculating the elements of the covariance matrix by considering the similarity of terrain slope.

[0111] Among them, the anisotropic background error covariance matrix is ​​a matrix that can reflect the different characteristics of meteorological element errors in different spatial directions. It is used to quantify the error distribution of the background field (the initial guess field provided by numerical weather prediction models) and solve the problem that the traditional isotropic assumption cannot adapt to the spatial structure of meteorological errors in complex terrain areas. Meteorological observation data is the foundation for ensuring assimilation accuracy and mainly covers satellite meteorological data, ground sensor data, and radar observation data. These data capture the meteorological conditions of the flight route area from different dimensions.

[0112] In some implementations, the specific processing involves calculating matrix elements by combining terrain slope similarity and spatial correlation. The core formula is as follows:

[0113] This formula is used to calculate the element in the i-th row and j-th column of the background error covariance matrix. By integrating error standard deviation, spatial correlation, terrain slope similarity, and spatial distance attenuation effect, a refined quantification of background field error is achieved.

[0114] , These represent the standard deviations of the background error for the i-th and j-th grid points, respectively, used to describe the dispersion of the background field error for a single grid point, and are obtained through historical forecast error statistics.

[0115] Let represent the spatial correlation coefficient between the i-th and j-th grid points, used to characterize the correlation strength of the errors between the two points. Its calculation depends on an adaptive correlation scale that varies with spatial location, and the formula is: ;in To adapt to the relevant scale, it will be dynamically adjusted according to the geographical location and terrain type of grid points i and j. For example, the relevant scale is larger in plains areas and smaller in mountainous areas.

[0116] The terrain slope similarity function, used to quantify the impact of the difference in terrain slope between grid points i and j on error correlation, is expressed by the following formula: ;in , The topographic slope angles of the i-th and j-th grid points are respectively calculated by the digital elevation model; α is the topographic influence adjustment parameter, which usually ranges from 0.1 to 1.0. The more complex the topography, the larger the value of α, so as to strengthen the influence of slope similarity on error covariance.

[0117] These represent the horizontal distance (x-direction, y-direction) and vertical distance (z-direction) between the i-th grid point and the j-th grid point, respectively, used to describe the spatial position difference between the two points;

[0118] These are the scale parameters of the background error in the x, y, and z directions, respectively, used to reflect the decay rate of the error in different spatial directions and to reflect anisotropic characteristics, such as when the x-direction is the dominant airflow direction. Value greater than .

[0119] During the calculation, the terrain slope angle of all grid points in the route area is first obtained through a digital elevation model, and the background error standard deviation of each grid point is determined in combination with historical data. , and adaptive correlation scale Then calculate the spatial distance between any two grid points. Similarity function of terrain slope Finally, substituting the values ​​into the above formula yields the complete anisotropic background error covariance matrix, completing the preliminary processing of the meteorological observation data.

[0120] It should be noted that the introduction of the terrain slope similarity function and spatial correlation coefficient is the core innovation of this step. It is crucial to ensure that the resolution of the digital elevation model matches the grid resolution of the meteorological data, ideally not less than 1 km; otherwise, it will lead to errors in terrain slope calculation or distortion of spatial correlation characterization. Simultaneously, satellite meteorological data, ground sensor data, and radar observation data need to be preprocessed to avoid raw data defects affecting the accuracy of covariance matrix calculation.

[0121] S102. Calculate the expected value of terrain correction for each ground sensor location using a digital elevation model. Then, calculate the deviation between the ground sensor observations and the expected value of terrain correction. When the deviation exceeds a preset threshold, reduce the weight of the observations to generate quality-controlled meteorological observation data.

[0122] The Digital Elevation Model (DEM) is a digital simulation of the ground topography using limited terrain elevation data, providing accurate terrain information for the location of ground sensors. The terrain correction expectation value refers to the theoretically reasonable value of meteorological elements under specific terrain conditions, used to measure the reliability of ground sensor observations. Weight reduction processing reduces the weight of excessively biased observations in subsequent assimilation processes, minimizing the interference of anomalous observation data on the assimilation results.

[0123] In some implementations, the specific operation process is as follows:

[0124] First, based on the latitude and longitude coordinates of the ground sensors, the terrain parameters of the corresponding location are extracted from the digital elevation model, including altitude, terrain slope, and terrain aspect.

[0125] Secondly, based on the statistical relationship model between topographic parameters and meteorological elements, the expected value of topographic correction is calculated. Taking temperature as an example, the statistical relationship model can be obtained through linear regression analysis of the region's ground sensor observation data and topographic parameters from the same historical period (same season, same time period), and the expression is: ,in This is the expected value for temperature and topography correction, where h is the altitude. This represents the tangent of the terrain slope. denoted as the cosine of the terrain aspect, and a, b, c, and d are regression coefficients obtained by fitting historical data.

[0126] Then, the absolute deviation between each ground sensor observation and the expected terrain correction value is calculated. (Taking temperature as an example);

[0127] Finally, preset thresholds are set, such as a temperature deviation threshold of 3℃ and a wind speed deviation threshold of 2m / s. If the deviation exceeds the threshold, the weighting coefficient of the observation value is reduced to 0.3 to 0.5 (the weighting coefficient of the normal observation value is 1.0), and the meteorological observation data after quality control is obtained.

[0128] For other meteorological elements such as air pressure and wind speed, the calculation logic for topographic correction expected values ​​is consistent with that for temperature, requiring only adjustments to the variables and regression coefficients of the statistical relationship model. However, the topographic correction expected value model for air pressure needs to focus on the influence of altitude, as shown in the expression... (e, f, and g are regression coefficients); the wind speed model needs to consider the obstruction and acceleration effects of terrain slope and aspect on airflow, and the expression is: (h, i, j are regression coefficients).

[0129] It should be noted that the statistical relationship model for the expected value of terrain correction needs to be built based on sufficient historical data, and the regression coefficients need to be fitted separately for each season and time period (such as day / night), because the influence of terrain on meteorological elements varies in different seasons and time periods.

[0130] S103. Construct a four-dimensional variational assimilation objective function that includes terrain constraints. By minimizing this objective function, the quality-controlled meteorological observation data is assimilated into the numerical forecast data to generate a high-resolution analysis field as meteorological field data.

[0131] The four-dimensional variational assimilation objective function is used to quantify the deviations between the analysis field and the background field, the analysis field and the observation data, and the analysis field and the terrain constraints. The analysis field corresponding to its minimum value is the optimal initial field (meteorological field data). The introduction of the terrain constraint term is to force the analysis field to adapt to the meteorological patterns of complex terrain regions and avoid analysis field deviations caused by ignoring the influence of terrain.

[0132] In some implementations, the specific operation process is as follows:

[0133] 1. Clearly define the complete form of the objective function and the meaning of each component, including:

[0134] The formula contains three components, each corresponding to the quantification of different deviations:

[0135] The first term: Background field deviation term, used to measure the consistency between the analytical field and the background field, is expressed as follows: .in, for The analysis field at time t, i.e. the optimal initial field to be solved, includes at least the wind field, temperature field, and humidity field; B represents the background field, i.e., the initial guess field provided by the numerical weather prediction model; B is the anisotropic background error covariance matrix obtained in S101. This is its inverse matrix. The smaller this value, the smaller the deviation between the analysis field and the background field, and the more it matches the initial trend of the numerical weather prediction model.

[0136] The second term: Observational bias, used to measure the consistency between the analytical field and the observed data after quality control, is expressed as follows: .in, The vector of post-quality control observation data at time i includes satellite radiative brightness temperature, temperature / pressure / humidity / wind speed observations from ground weather stations, radar reflectivity factor, radial velocity, etc. For the observation operator at time i, it is used to map the mode variables of the analysis field to the observation space, so as to achieve dimensionality matching between the analysis field and the observation data; Let be the observation error covariance matrix at time i, used to quantify the error of the observation data. The diagonal elements represent the error variance of each observed variable, and the off-diagonal elements represent the error covariance between variables, which can be obtained from sensor accuracy calibration data. N is the total number of times within the assimilation time window, and a value of 6 to 12 is recommended, corresponding to a time window length of 1 to 3 hours, balancing assimilation accuracy and computational efficiency. The smaller this value, the smaller the deviation between the analyzed field and the observed data, and the better it reflects the actual meteorological conditions.

[0137] The third item: Terrain constraint term, used to force the analysis field to adapt to terrain conditions, the expression is as follows: .in, The topographic constraint reference field is calculated based on a digital elevation model and generated through the statistical relationship between topographic parameters and meteorological elements, such as the air pressure reference value corresponding to topographic elevation and the wind speed reference value corresponding to slope. λ is the topographic constraint weight coefficient, used to control the intensity of the topographic influence, with a value ranging from 0.1 to 0.5; the more complex the topography, the larger the value. Q is the topographic constraint weight matrix, a diagonal matrix. The diagonal elements are adjusted according to the topographic complexity, with values ​​ranging from 1.0 to 1.5 for grid points in mountainous areas and from 0.5 to 1.0 for grid points in plains. The smaller this value, the smaller the deviation between the analysis field and the topographic constraint reference field, and the stronger the physical rationality.

[0138] 2. Minimize the objective function:

[0139] The minimum value of the objective function is solved using the conjugate gradient method or the L-BFGS algorithm (a commonly used algorithm in unconstrained optimization). Before solving, the analysis field needs to be... The parameters are parameterized and transformed into an optimizable vector form. During the solution process, the gradient of the objective function is calculated iteratively, and the analysis field vector is continuously adjusted until the objective function value converges.

[0140] 3. Generate meteorological field data:

[0141] After the objective function converges, the corresponding analytical field This refers to high-resolution meteorological field data, containing grid points of each route area. The three-dimensional distribution information of key meteorological elements such as wind field (zonal wind, meridional wind, vertical wind), temperature field, humidity field, and pressure field at any given time.

[0142] It should be noted that the observation operator The construction of the model needs to be designed separately according to different types of observation data. For example, the observation operator for satellite radiation brightness temperature needs to consider the atmospheric radiation transfer process, and the observation operator for radar reflectivity needs to consider the scattering characteristics of precipitation particles. Otherwise, it will lead to mapping errors between model variables and observation data.

[0143] For example, the assimilation time window is set to 2 hours (N=8, every 15 minutes), the background field is 1km resolution numerical weather prediction data output by the Global Forecast System (GFS), and the quality-controlled observation data includes radiative brightness temperature data from 10 meteorological satellites, meteorological element data from 50 ground sensors, and reflectivity and radial velocity data from 3 Doppler radars. Observation operators are constructed: the satellite radiative brightness temperature observation operator uses the RTTOV radiative transfer model, the ground observation operator uses linear interpolation mapping, and the radar observation operator uses the ray tracing method. The terrain constraint weight coefficient λ = 0.3 is set, and the diagonal elements of the terrain constraint weight matrix Q are 1.2 for mountain grid points and 0.8 for plain grid points. After 200 iterations using the conjugate gradient method, the objective function value converges to 8.2 × 10⁻⁶.-7 Meteorological field data with a resolution of 1km×1km×100m were obtained for the route area. The deviation between the temperature field and the ground observation data at the mountain grid points was reduced by 45% and the wind field deviation was reduced by 38% compared with before assimilation, which significantly improved the data accuracy.

[0144] Based on the above technical solutions, S1 addresses the insufficient accuracy of traditional data assimilation in complex terrain areas through processes such as anisotropic background error covariance matrix processing, ground observation data quality control, and four-dimensional variational assimilation with terrain constraints. The anisotropic background error covariance matrix, combined with terrain slope similarity, accurately characterizes the impact of terrain on error distribution; ground observation data quality control filters outomas using terrain correction expectation values, avoiding interference from distorted data; and the objective function including terrain constraints forces the analysis field to adapt to terrain patterns, ensuring the physical rationality and reliability of the meteorological field data. The resulting high-resolution meteorological field data provides a high-quality data foundation for subsequent steps such as error correction in S2 and profile generation in S3, serving as a core prerequisite for improving the accuracy of the entire airway meteorological support method.

[0145] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 4 As shown, the above S2 can be implemented through the following S201, S202, S203, S204 and S205, which are explained in detail below:

[0146] S201. The multiple two-dimensional horizontal fields of the input meteorological field data are processed by a two-dimensional convolutional encoder with shared weights to extract the spatial feature maps of each meteorological element.

[0147] In this context, a two-dimensional horizontal field refers to the gridded data of a single meteorological element on any isobaric surface or model layer, such as gridded temperature data on the 500 hPa isobaric surface and gridded wind speed data on the 850 hPa isobaric surface. Each two-dimensional horizontal field corresponds to the spatial distribution of a meteorological element at a specific altitude. A weighted two-dimensional convolutional encoder refers to an encoding structure where all two-dimensional horizontal fields share the same set of convolutional kernel parameters, used to uniformly extract the spatial features of different meteorological elements. A spatial feature map is an abstract representation of the original meteorological field data, highlighting the spatial distribution patterns of meteorological elements, such as high-pressure centers, low-pressure troughs, and wind speed jet streams.

[0148] In some implementations, the specific operation process is as follows:

[0149] First, the meteorological field data generated by S1 is split into layers, and the two-dimensional horizontal field corresponding to each meteorological element (wind field, temperature field, humidity field, etc.) is extracted according to the isobaric surface or model layer to ensure that the grid resolution of each two-dimensional horizontal field is consistent.

[0150] Secondly, a two-dimensional convolutional encoder with shared weights is constructed. This encoder can contain 3 to 5 convolutional layers. The kernel size of each convolutional layer is set to 3×3 or 5×5, the stride is 1, the padding method is "SAME" to keep the feature map size consistent with the input, and the activation function is ReLU or LeakyReLU.

[0151] Then, each two-dimensional horizontal field is sequentially input into the encoder, and spatial features are extracted through convolution operations. The first layer of convolution captures small-scale spatial details, such as sudden changes in local wind speed, while the deep convolution captures large-scale spatial structures, such as regional air pressure distribution.

[0152] Finally, the spatial feature map corresponding to each meteorological element is output. The number of channels in the feature map is determined by the number of convolutional kernels in the last layer of the encoder.

[0153] For example, the meteorological field data generated by S1 includes four meteorological elements: wind field (zonal wind and meridional wind), temperature field, and humidity field. Each element corresponds to 10 model layers (0 to 3000m, one layer every 300m), for a total of 40 two-dimensional horizontal fields (4 elements × 10 layers). A shared weight two-dimensional convolutional encoder with 4 convolutional layers is constructed, and then the 40 two-dimensional horizontal fields are input into the encoder one by one. Finally, 40 spatial feature maps with a size of 1km × 1km × 256 are output. Among them, the feature map of zonal wind at the 2000m layer clearly highlights the spatial distribution of the jet stream along the route.

[0154] S202. The spatial feature maps of each extracted meteorological element are stitched together in the channel dimension and input into the physical information spatiotemporal Transformer module to output spatiotemporal features.

[0155] Among them, channel dimension stitching refers to merging all meteorological elements and spatial feature maps of all model layers along the channel direction to form a unified high-dimensional feature tensor. For example, stitching together 40 1km×1km×256 feature maps yields a 1km×1km×(40×256) feature tensor. The purpose is to integrate the spatial information of all meteorological elements to prepare for capturing spatiotemporal correlations. The Physical Information Spatio-Temporal Transformer module is the core module that integrates physical law constraints and spatiotemporal dependency modeling. It embeds spatial coordinates and forecast lead time information through learnable position encoding, and then uses a multi-head self-attention mechanism to capture the global long-range spatiotemporal dependencies between meteorological elements, thereby simulating the long-range interactions between weather systems, such as the impact of an upstream low-pressure system on downstream wind speed. Spatiotemporal features are high-dimensional feature representations that integrate spatial distribution patterns and temporal evolution trends, and can comprehensively characterize the dynamic changes of the meteorological field.

[0156] In some implementations, the specific operation process is as follows:

[0157] 1. Feature map stitching: Stack all spatial feature maps output by S201 along the channel dimension. If there are feature maps of different mode layers, layer index identifiers need to be added before stitching.

[0158] 2. Location Encoding Embedding: Construct a learnable location encoding vector, which includes spatial coordinate encoding and forecast timeliness encoding. The spatial coordinate encoding is calculated based on the latitude, longitude, and altitude of each grid point, while the forecast timeliness encoding is calculated based on the time difference between the current forecast time and the initial time. The dimension of the location encoding vector is consistent with the number of channels in the feature map. It is then added to the concatenated feature tensor to achieve location information embedding.

[0159] 3. Multi-head self-attention calculation: Set 8 to 16 attention heads, project the feature tensor after embedding position encoding into three spaces: query, key, and value. Calculate the attention weight of each grid point with all other grid points through matrix operations. The larger the weight value, the stronger the spatiotemporal correlation between the two points.

[0160] 4. Feature aggregation and output: The outputs of multiple attention heads are concatenated and feature fusion is performed through a fully connected layer. The output is a spatiotemporal feature tensor with the same dimension as the input feature tensor.

[0161] It's important to note that the core advantage of the Physical Information Spatiotemporal Transformer module lies in its physical information embedding and long-range dependency capture. Compared to traditional Transformers, its location encoding not only includes spatiotemporal coordinates but also implicitly contains the physical relationships between meteorological elements, such as the correspondence between altitude and air pressure. Simultaneously, the choice of the number of attention heads needs to balance model performance and computational cost; too many heads will lead to a surge in computation, while too few will fail to fully capture complex spatiotemporal dependencies. Furthermore, forecast lead time encoding allows the module to distinguish changes in the meteorological field at different time points, providing a temporal dimension basis for subsequent error correction.

[0162] S203. Input the spatiotemporal features into the decoder, generate a preliminary error correction field through multiple deconvolutional layers or upsampling convolutional layers, and add the preliminary error correction field to the meteorological field data to obtain the preliminary correction.

[0163] The decoder is the inverse structure corresponding to the encoder, used to map high-dimensional spatiotemporal features back to an error correction field with the same dimension as the original meteorological field data. Deconvolutional or upsampling convolutional layers are used to improve the spatial resolution of the feature map, ensuring that the preliminary error correction field corresponds to the original meteorological field data grid by grid. The preliminary error correction field is the meteorological field error distribution predicted by the model, with the value of each grid point representing the error correction amount for the meteorological element at that location. The preliminary correction field is the superposition result of the original meteorological field data and the preliminary error correction field; that is, it offsets part of the error in the original meteorological field through the error correction amount, obtaining an intermediate result that is closer to the actual meteorological conditions.

[0164] In some implementations, the specific operation process is as follows:

[0165] First, construct the decoder. The decoder can contain 3 to 5 deconvolutional layers or upsampling convolutional layers, which correspond one-to-one with the convolutional layers of the encoder. The kernel size of the deconvolutional layer is the same as that of the encoder, the stride is 1, and the number of channels gradually decreases from high to low.

[0166] Then, the spatiotemporal feature tensor output by S202 is input into the decoder, and the feature map resolution is gradually improved through deconvolution operation. The final deconvolution layer outputs a preliminary error correction field with dimensions completely consistent with the original meteorological field data.

[0167] Next, the preliminary error correction field is added to the original meteorological field data generated by S1 grid by grid and element by element. That is, for each meteorological element at each grid point, the corrected value = original value + error correction amount, and finally the preliminary correction field is obtained.

[0168] S204. Input the preliminary correction field into the physical constraint layer, calculate whether the preliminary correction field satisfies the preset physical conservation law through differentiable programming, and determine the physical loss according to the degree of violation of the physical constraint.

[0169] The physical constraint layer is the core module ensuring the physical consistency of the corrected meteorological field. The preset physical conservation laws mainly include fundamental atmospheric physical laws such as the law of conservation of mass and the law of conservation of thermodynamics. Differentiable programming refers to the quantitative calculation of these physical conservation laws using differentiable mathematical operators, ensuring that physical losses can participate in the backpropagation optimization of the model. Physical loss is an indicator that quantifies the degree to which the initial corrected field violates physical conservation laws; a larger loss value indicates poorer physical consistency, and vice versa. It guides the model to reduce statistical errors while avoiding the generation of physically unrealistic meteorological fields.

[0170] In some implementations, the specific operation process is as follows:

[0171] First, implement physical law operators such as gradient operators and divergence operators based on differentiable programming frameworks (such as TensorFlow and PyTorch) to calculate the degree of violation of mass conservation, thermodynamic conservation, etc. in the preliminary correction field;

[0172] Then, the losses from violations of mass conservation and thermodynamic conservation are calculated together, and the physical losses are quantified using the following formula: ;in,

[0173] N represents the total number of samples in a training batch, used to normalize the loss value;

[0174] The weighting coefficient for mass conservation ranges from 0.3 to 0.7 and is used to adjust the strength of the mass conservation constraint. It can be appropriately increased in areas with complex terrain.

[0175] The weighting coefficient representing the thermodynamic constraint, ranging from 0.3 to 0.7, is related to... The sum is usually 1.0;

[0176] This represents the calculated value of the mass divergence term in the mass conservation equation. denoted as air density (derived from temperature and air pressure data of the preliminary correction field), and v is a three-dimensional wind vector (including zonal wind, meridional wind, and vertical wind). If mass conservation holds, the theoretical value of this term is 0. The larger the square of its L2 norm, the more severe the violation of mass conservation.

[0177] Indicates potential temperature Thermodynamic equation residuals This represents the rate of change of potential temperature over time (calculated based on forecast lead time). The term represents the advection variation of potential temperature, and Q represents the non-adiabatic heating rate (derived from historical data statistics). The theoretical value of this term is 0, and the larger the square of its L2 norm, the more severe the violation of thermodynamic conservation.

[0178] Finally, the calculated physical loss value is passed to the model's loss function for subsequent network parameter optimization.

[0179] It should be noted that the choice of physical conservation laws needs to be combined with the type of meteorological field data. In addition to mass conservation and thermodynamic conservation, constraints such as momentum conservation and water vapor conservation can also be added as needed. The implementation of differentiable programming must ensure the differentiability of the physical law operators; otherwise, physical losses cannot participate in backpropagation, leading to the failure of physical constraints. At the same time, the values ​​of α and β need to be determined through experimental debugging.

[0180] S205. During the training of the spatiotemporal graph convolutional network, the mean square error between the corrected field output by the forward pattern and the reference analysis field is used as the main loss, which is weighted and summed with the physical loss to optimize the network parameters through backpropagation. When the network performs inference, it outputs the optimized weather forecast results and the forecast uncertainty estimate for each grid point.

[0181] The training process of the spatiotemporal graph convolutional network involves adjusting network parameters using a large amount of historical data to enable the model to have stable error correction capabilities. The reference analysis field is the true meteorological field used to supervise network training. It is obtained by acquiring high-precision analysis data from the same historical period (such as reanalysis data released by meteorological departments), and its resolution and element types are consistent with the training samples. The main loss, i.e., the data loss, is used to quantify the statistical error between the correction field and the true meteorological field, while the physical loss is used to quantify physical consistency. The weighted sum of the two yields the total loss, which is the core indicator guiding the optimization of network parameters.

[0182] In some implementations, the specific operation process is as follows:

[0183] 1. Construction of the total loss function: The formula for the total loss function is: ,in, The total loss reflects both statistical error and physical consistency. This is the physical loss weighting coefficient, with a value ranging from 0.1 to 0.5, used to adjust the proportion of physical loss in the total loss;

[0184] The main loss is calculated using the mean squared error (MSE), and the formula is as follows: ;in This is the initial correction field for the i-th training sample. Let i be the reference analysis field corresponding to the i-th training sample. The square of the L2 norm is used to quantify the grid-by-grid error between the correction field and the reference analysis field; This represents physical loss, specifically the loss value calculated by S204.

[0185] 2. Network training optimization: Use Adam or SGD optimizer, set the learning rate to 1e-4 to 1e-3, and the number of training iterations to 100 to 300 rounds. In each training round, input the training samples into the network, calculate the total loss, calculate the gradient through backpropagation, and adjust the parameters of each layer such as the convolutional encoder, the physical information spatiotemporal Transformer module, and the decoder until the total loss converges (the convergence threshold is usually set to 1e-5).

[0186] 3. Reasoning process: Input the new meteorological field data (data not used in training) into the trained network, and obtain the preliminary corrected field according to the process of S201 to S204. At the same time, calculate the forecast uncertainty (such as standard deviation) of each grid point through Monte Carlo dropout or Bayesian neural network methods, and finally output the optimized meteorological forecast results and uncertainty estimates.

[0187] It should be noted that the quality of the reference analysis field directly determines the model training effect. It is necessary to ensure that its resolution is not lower than that of the training samples and that the accuracy of the data has been verified, such as by using an internationally recognized reanalysis dataset. The value of needs to be determined through cross-validation. If the value is too small, the physical constraint effect will not be obvious, which may lead to inconsistent physical results in the model output; if the value is too large, the effect of reducing statistical error may be sacrificed.

[0188] Based on the above technical solution, S2 addresses the problems of insufficient physical consistency and the disconnect between statistical error and physical laws in traditional error correction methods through spatial feature extraction, spatiotemporal feature fusion, preliminary error correction, physical constraint verification, and network training and inference. Specifically, the spatial feature extraction and spatiotemporal feature fusion modules accurately capture the spatial distribution patterns and long-range spatiotemporal dependencies of meteorological elements, providing comprehensive feature support for error correction. Preliminary error correction achieves precise error mapping through an encoder-decoder structure. The physical constraint layer forces the correction field to follow fundamental atmospheric physical laws by quantifying the degree of violation of physical conservation laws. Weighted summation and backpropagation optimization of the total loss function achieve a balance between statistical accuracy and physical rationality. Ultimately, the trained AI corrector outputs physically reasonable, statistically inefficient, and quantifiable optimized weather forecast results, providing high-quality meteorological data support for subsequent steps such as route-altitude meteorological profile generation and risk assessment. This is a core component in improving the scientific rigor and reliability of the entire route meteorological support method.

[0189] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 5 As shown, the above S3 can be implemented through the following S301, S302 and S303, which are explained in detail below:

[0190] S301. Discretize the user-input route into a sequence of N waypoints according to a preset step size, and concatenate the horizontal coordinates, altitude, and circulation feature vector extracted from the optimized weather forecast results of each waypoint into a condition vector.

[0191] The user-input route information includes the geographical coordinates (latitude and longitude) of the origin, destination, and key points along the route. This information is transmitted from the visualization interaction module to the profile generation module, forming the spatial basis for generating profile data. The preset step size is a distance interval set based on route length, meteorological data resolution, and actual flight accuracy requirements. It is used to discretize the continuous route into a finite number of computable waypoints, and the resulting sequence of waypoints must completely cover the entire route. The horizontal coordinates are the latitude and longitude information of the waypoints, while the altitude is a user-preset flight altitude range or a specific altitude value. Both together determine the three-dimensional spatial location of the waypoints. The circulation feature vector is a one-dimensional vector reflecting the overall state of the meteorological environment surrounding the waypoint. It is obtained by extracting meteorological element information from a specific range in the optimized weather forecast results and is used to provide large-scale circulation background support for subsequent model predictions. The conditional vector is a high-dimensional vector integrating the spatial location of the waypoint and the surrounding meteorological background. It serves as the input to the conditional neural radiation field model, ensuring that the model can accurately predict the meteorological element values ​​at that point.

[0192] In some implementations, the specific operation process is as follows:

[0193] First, obtain the total route length L input by the user. Then, calculate the number of waypoints N based on the preset step size d (usually set to 100m~500m). The calculation formula is as follows: , This indicates rounding up, with +1 to include the start and end points; then, using linear interpolation, the horizontal coordinates (latitude and longitude) of N waypoints are calculated sequentially based on the coordinates of the start, end, and key points along the route, ensuring that the waypoints are evenly distributed along the route;

[0194] Next, taking the nearest weather forecast grid point as the center, a preset radius r (usually 1km to 3km, which needs to cover the key meteorological influence area around the waypoint) is set, and meteorological elements of all grid points within this radius are extracted, including at least temperature, zonal wind, meridional wind, vertical velocity, relative humidity, and turbulent kinetic energy. Each meteorological element is standardized to eliminate the influence of differences in the dimensions of different elements. All standardized meteorological element values ​​are flattened into a one-dimensional vector in a preset order (such as temperature → zonal wind → meridional wind → … → turbulent kinetic energy), which is the circulation characteristic vector of the waypoint.

[0195] Then, the horizontal coordinates, altitude, and circulation feature vector of each waypoint are concatenated in sequence to form a conditional vector.

[0196] S302. Input the conditional vector into the pre-trained conditional neural radiation field model and output the meteorological element values ​​for each waypoint.

[0197] The Conditional Neural Radiance Fields Model (CNRF Model) is a deep learning model that integrates conditional information and radiation field modeling. Its core principle is to learn the mapping relationship between spatial location, circulation background, and meteorological element values ​​through a neural network. Based on discrete weather forecast grid data, it can accurately predict meteorological element values ​​at any spatial location. A pre-trained model refers to a model that has been trained, validated, and tested based on a large amount of historical meteorological data (including historical waypoint data, corresponding circulation feature vectors, and actual meteorological element observations), and possesses stable predictive capabilities. The output meteorological element values ​​represent the true meteorological state at the waypoint location, including at least temperature, zonal wind, meridional wind, vertical velocity, relative humidity, and turbulent kinetic energy, providing data for generating flight path-altitude meteorological profiles.

[0198] In some implementations, the specific operation process is as follows:

[0199] First, the conditional neural radiation field model consists of an input layer, a feature encoding layer, a hidden layer, and an output layer. The input layer receives the conditional vector generated by S301; the feature encoding layer uses a multilayer perceptron (MLP) to map the conditional vector to high-dimensional latent space features, with 3 to 5 encoding layers and 256 to 512 neurons per layer, using ReLU as the activation function; the hidden layer uses a radiation field modeling module to capture the spatial continuity of meteorological elements by learning the radiation field distribution of spatial location and meteorological elements; the output layer is a fully connected layer, with the output dimension consistent with the types of meteorological elements, such as 6 elements, which output a 6-dimensional vector, and the output value is the predicted value of each meteorological element.

[0200] Next, the training dataset uses historical data from the past 3-5 years. Each training sample contains a "conditional vector-real meteorological element value" pair, with the real meteorological element values ​​coming from high-precision observation data from ground observation stations, radar detection, etc. Mean squared error (MSE) is used as the loss function, with the formula: Where K is the number of training samples and M is the number of meteorological element types. For the k-th sample, the m-th element is the true observed value. The predicted value is used; the Adam optimizer is used, the learning rate is set to 1e-4 to 1e-3, and the number of training iterations is 200 to 500 rounds until the validation set loss converges (the convergence threshold is set to 1e-4).

[0201] Then, the condition vector of each waypoint generated by S301 is sequentially input into the trained model. The model calculates through forward propagation and outputs the predicted values ​​of the six meteorological elements corresponding to the waypoint, thus completing the mapping from the condition vector to the meteorological element values.

[0202] S303. Traverse all waypoints and user-specified altitude layers, and output a two-dimensional data matrix by querying the conditional neural radiation field model to obtain the route-altitude two-dimensional meteorological profile data.

[0203] The user-specified altitude layer refers to the vertical altitude range and layer interval set by the user according to the flight mission requirements, such as an altitude range of 0-3000m, with each altitude layer being 100m, covering the key altitude range for low-altitude flight. The traversal operation refers to sequentially processing the meteorological element predictions for each waypoint at each altitude layer. The query conditional neural radiation field model refers to regenerating the corresponding conditional vector for each "waypoint-altitude layer" combination (only the altitude dimension is updated, while the horizontal coordinates and circulation feature vectors remain unchanged), inputting it into the model to obtain the meteorological element values ​​corresponding to that combination. The two-dimensional data matrix is ​​a structured representation of the flight path-altitude meteorological profile data. The horizontal axis represents the distance along the flight path (derived from the order and spacing of the waypoint sequence), and the vertical axis represents the altitude (user-specified altitude layer). Each element in the matrix is ​​a set of meteorological element values ​​corresponding to the "distance along the flight path-altitude" position, which can intuitively reflect the distribution pattern of meteorological elements along the flight path and vertical altitude.

[0204] In some implementations, the specific operation process is as follows:

[0205] First, receive the height layer parameters input by the user, including the starting height. Termination height Calculate the number of height layers based on the layering interval Δh. Generate a sequence of height values ​​for K height layers. .

[0206] Then, a double traversal is performed according to the order of waypoints (from start to finish) and the order of altitude layers (from low to high or from high to low). For each waypoint i (i=1,2,...,N) and each altitude layer k (k=1,2,...,K), the horizontal coordinates of the waypoint, the altitude value of the current altitude layer, and the original circulation feature vector are reassembled into a new condition vector. The new condition vector is then input into the conditional neural radiation field model to query the set of meteorological element values ​​corresponding to "waypoint i-altitude layer k".

[0207] Finally, using the distance along the flight path as the horizontal axis, the distance along the flight path (the cumulative distance from the origin to the flight path) corresponding to each waypoint is used as the horizontal axis coordinate; the altitude value of the altitude layer is used as the vertical axis coordinate; the set of meteorological element values ​​corresponding to each "distance along the flight path - altitude" coordinate is used as matrix elements to construct a two-dimensional data matrix of dimension (N×K). The rows of the matrix correspond to waypoints (distance along the flight path), the columns correspond to altitude layers, and each element is a 6-dimensional vector, corresponding to 6 meteorological elements.

[0208] It should be noted that in spatial interpolation algorithms, traditional interpolation methods rely on the assumption of local smoothness, using only adjacent grid data for linear or low-order nonlinear inference. This fails to capture the correlation of global meteorological elements in complex weather, and is also difficult to accurately characterize the strong nonlinear abrupt changes of meteorological elements such as turbulence and wind shear. Furthermore, it lacks consideration of meteorological physical laws, merely performing mathematical spatial filling, which may generate physically unreasonable results and easily lead to distorted interpolation results. In contrast, the conditional neural radiation field model, with its strong fitting ability of deep learning, integrates circulation feature vectors into the large-scale meteorological background, enabling it to model the dependencies of global meteorological elements. Through multi-layer nonlinear transformation, it accurately learns and reproduces the nonlinear variation laws of complex weather. It can implicitly learn the physical logic of weather evolution during training with massive historical meteorological data. In addition, the strong robustness brought by global feature fusion and data redundancy learning allows it to maintain prediction stability even when faced with data noise interference in complex weather, thus more accurately restoring the real meteorological state under complex weather conditions.

[0209] For example, the user specifies the height layer parameter as the starting height. =100m, Termination Height =3000m, layer interval Δh=100m, calculate the number of height layers. There are 111 waypoints and 30 altitude layers, with altitude values ​​of 100m, 200m, ..., 3000m. These are traversed sequentially, forming 111 × 30 = 3330 "waypoint-altitude layer" combinations. For each combination, the altitude dimension of the condition vector is updated, and the meteorological element value set is obtained by querying the model. A 111 × 30 two-dimensional data matrix is ​​constructed with the distance along the flight path as the horizontal axis and altitude as the vertical axis. Each element in the matrix represents 6 meteorological element values ​​(e.g., (24.8℃, 7.6m / s, 3.2m / s, 0.1m / s, 65%, 0.02m² / s²)). Finally, the two-dimensional meteorological profile data of the flight path and altitude is obtained.

[0210] Based on the above technical solution, S3 solves the problem that traditional spatial interpolation methods cannot accurately depict the fine meteorological structure in the vertical direction of flight routes by performing operations such as route discretization and conditional vector concatenation, conditional neural radiation field model prediction, and two-dimensional data matrix construction. Specifically, the route discretization and conditional vector concatenation stage integrates the spatial location of waypoints with the surrounding circulation background, providing comprehensive input information for model prediction. The conditional neural radiation field model, leveraging the strong fitting capability of deep learning, achieves accurate mapping from discrete grid data to continuous waypoint meteorological element values, reflecting the true state under complex weather conditions better than traditional interpolation methods. The two-dimensional data matrix construction stage organizes scattered meteorological element values ​​into structured profile data, intuitively presenting the distribution patterns of meteorological elements along the flight route and altitude. The resulting flight route-altitude meteorological profile data provides refined and continuous meteorological foundation data for disaster risk assessment, effectively reducing the disconnect between meteorological information and flight missions.

[0211] In one possible implementation of this application embodiment, the above-mentioned S4 specifically includes the following S401 to S403:

[0212] S401. Treat each grid point in the meteorological profile data as a graph node, extract the node feature vector, construct spatial adjacency edges and causal propagation directed edges, and obtain a dynamic risk propagation graph.

[0213] In this dynamic risk propagation map, grid points from meteorological profile data are the basic units, each corresponding to a specific spatial location defined by "distance along the flight path - altitude". Graph nodes are abstract representations of grid points in the network model, and the feature vector of each node is the core basis for meteorological disaster risk assessment at that location. The dynamic risk propagation map is a network structure that integrates spatial adjacency relationships and wind-guided causal relationships. It is used to depict the natural diffusion of risk in three-dimensional space and its dynamic propagation path along the airflow direction, providing a structured foundation for subsequent risk calculations. Spatial adjacency edges characterize the risk diffusion association between adjacent grid points, while causal propagation directed edges reflect the guiding role of the wind field on risk propagation. The combination of these two elements makes the risk propagation simulation more closely resemble the actual evolution of meteorological disasters.

[0214] In some implementations, the specific operation process is as follows:

[0215] First, define the initial node feature vector. Each element in the two-dimensional data matrix output by S3 (i.e., each "distance-altitude along the flight path" grid point) is defined as a graph node. Meteorological hazard-related indices are extracted from each node to form the initial node feature vector. These indices include, but are not limited to, turbulence index, icing index, convective available potential energy, vertical wind shear, and visibility. Extraction requires direct reading from meteorological profile data or calculation using element values. For example, vertical wind shear can be calculated by dividing the wind speed difference between adjacent height layers by the height interval. These indices are then combined in a preset order (e.g., turbulence index → ​​icing index → ​​convective available potential energy → vertical wind shear → visibility) to form the initial node feature vector.

[0216] Next, spatial adjacency edges are constructed. The "8-neighborhood rule" or "4-neighborhood rule" is used to determine the spatial adjacency relationship between grid points. That is, for the grid matrix of two-dimensional meteorological profile data, if two nodes are adjacent on the horizontal or vertical axis (including diagonal adjacency, i.e., 8-neighborhood), an undirected spatial adjacency edge is established between them to simulate the natural diffusion of risk in adjacent areas.

[0217] Finally, causal propagation directed edges are constructed. Based on the wind field information (zonal wind, meridional wind, and vertical wind) in the meteorological profile data, the dominant airflow direction of each node is determined. Taking node A as an example, if the airflow points from node B to node A, it indicates that node B is upwind of node A. Therefore, a directed edge from B to A is established between node B and node A, resulting in a causal propagation directed edge, which is used to simulate the dynamic propagation of risk along the dominant airflow direction. The weight of the directed edge is initially set to 1.0, and can be dynamically adjusted according to the wind field intensity.

[0218] S402. Input the dynamic risk propagation graph and node feature vectors into the risk propagation network model. Through feature encoding, multi-round causal modulation attention aggregation and nonlinear mapping, the comprehensive risk value of each node is obtained.

[0219] The risk propagation network model is a K-layer deep learning model improved from the Graph Attention Network (GAT). It adaptively allocates the association weights between nodes through an attention mechanism, while simultaneously integrating causal priors and wind field guidance information to accurately calculate the comprehensive risk value of each node. The feature encoding module maps the initial node feature vectors to a high-dimensional latent space, enhancing feature representation. The causal modulation graph attention layer is the core of the model, used to achieve multi-round iterative aggregation of node features, capturing the complex correlations of risk propagation. The risk assessment module is responsible for mapping the aggregated features into a quantified comprehensive risk value. The final comprehensive risk value is the quantified result of the meteorological disaster risk at each grid point; the higher the value, the higher the risk, used to intuitively reflect the degree of safety threat to low-altitude flight at that location.

[0220] In some implementations, the specific operation process is as follows:

[0221] First, the risk propagation network model includes a graph construction module, a feature encoding module, K stacked causal modulation graph attention layers, and a risk assessment module. The feature encoding module employs a 2-3 layer Multilayer Perceptron (MLP), with 64-128 neurons per layer and ReLU activation function. This is used to map the initial node feature vectors to high-dimensional latent space feature vectors through linear transformation and nonlinear activation of the MLP, enhancing the nonlinear expressive power of the features and outputting the encoded initial node feature matrix. The K causal modulation graph attention layers (K typically ranges from 3 to 5) are used to progressively aggregate node features, with each layer capturing more complex risk propagation relationships. The risk assessment module employs a 1-2 layer fully connected neural network, with Sigmoid or Linear activation functions used in the output layer, mapping the final node features to a comprehensive risk value.

[0222] In some implementations, taking layer l as an example, the internal workflow of multi-round causal modulation attention aggregation is as follows:

[0223] First, linear projection is performed: the encoded node feature vectors are linearly projected onto the attention computation space, as shown in the formula. ,in Let be the learnable weight matrix of the l-th layer. Let i be the feature vector of node i in the input of layer l. The projected feature vector of node i;

[0224] Then, the original attention score is calculated: for node pairs (i,j) connected by an edge in the dynamic risk propagation graph, their original attention scores are calculated to quantify the association strength between node i and node j. The formula is as follows: ,in Let be the projected feature vector of node j, LeakyReLU (Leaky Rectified Activation Function) is the activation function, "‖" indicates vector concatenation, and a is the learnable attention vector;

[0225] Next, the modulation attention score is calculated: a causal prior matrix M and an adaptive scaling are introduced. The original attention score is adjusted using the following formula: ,in These are elements of the causal prior matrix M, representing the statistical causal strength of the state of node j leading to the disaster at node i. They are learned from historical disaster data through causal learning algorithms (such as Bayesian networks and structural equation models). The adaptive scale between nodes i and j is calculated based on their terrain differences (such as altitude difference) and wind shear (such as the magnitude of the wind speed vector difference). The greater the terrain difference, the stronger the wind shear. A larger value is used to weaken the risk propagation weight in non-stationary environments; the denominator is the sum of the adjusted scores of all neighboring nodes of node i (N(i) represents the set of neighboring nodes of node i), which normalizes the attention score, and finally obtains... To modulate the attention score, which reflects the contribution of node j to the risk propagation of node i;

[0226] Finally, feature aggregation is performed: based on the modulation attention score, the features of node i are aggregated with those of its neighboring nodes, using the following formula: ,in The activation function (either ReLU or ELU) works by weighted summation to integrate the feature information of neighboring nodes, updating the feature vector of node i so that the node features can reflect the impact of the propagation of surrounding risks; the output of the l-th layer is the updated node feature matrix. .

[0227] After multiple rounds of causal modulation attention aggregation and node feature iterative updates, the final node feature matrix output by the last causal modulation attention layer is... The input risk assessment module performs nonlinear transformation and mapping through a fully connected neural network, outputting the comprehensive risk value of each node. The comprehensive risk values ​​of all nodes are arranged in the order of the original grid matrix to form a two-dimensional risk distribution matrix.

[0228] S403. Based on the preset risk level threshold, the two-dimensional risk distribution matrix is ​​mapped into a color map to generate a risk heat map.

[0229] Among them, the risk level threshold is the critical value for dividing different risk intensities. It is used to discretize continuous comprehensive risk values ​​into a finite number of risk levels to simplify risk assessment and enable operators to quickly identify high-risk areas.

[0230] In some implementations, the specific operation process is as follows:

[0231] First, based on low-altitude flight safety standards, aircraft performance parameters, and historical accident statistics, the comprehensive risk value is divided into several risk levels, typically 4 to 6, such as low risk, lower risk, medium risk, higher risk, and high risk, with corresponding thresholds set. For example, when the comprehensive risk value ranges from 0 to 10, the thresholds can be set as follows: low risk (0-2), lower risk (2-4), medium risk (4-6), higher risk (6-8), and high risk (8-10). The thresholds need to be reviewed by industry experts to ensure they match the actual flight risk. For example, for drones with weak wind resistance, the high-risk threshold can be appropriately lowered (e.g., 7-10 is high risk).

[0232] Next, each element (comprehensive risk value) in the two-dimensional risk distribution matrix is ​​traversed, and its risk level is determined according to a preset threshold. For example, when the comprehensive risk value is 3.5, it corresponds to the "lower risk" level; thus forming a two-dimensional risk level matrix with the same dimensions as the two-dimensional risk distribution matrix.

[0233] Then, a specific color is assigned to each risk level, following the principle of gradually deepening the color from "low risk" to "high risk," for example, low risk (blue), lower risk (cyan), medium risk (yellow), higher risk (orange), and high risk (red). An interpolation algorithm (such as bilinear interpolation) is then used to smooth the two-dimensional risk level matrix to avoid color abrupt changes caused by grid discretization. The color-coded matrix is ​​then rendered into a color image, i.e., a risk heatmap, through a visualization rendering engine. Its horizontal axis represents the distance along the flight path, and its vertical axis represents the altitude. The color distribution corresponds one-to-one with the risk level, intuitively presenting the risk intensity at each spatial location along the flight path.

[0234] S403. Overlay the risk heat map onto the meteorological profile data to obtain the route support map.

[0235] The overlay operation involves aligning the risk heat map and the visualized image of the meteorological profile data according to their spatial location, and then merging them into a single image in a semi-transparent overlay manner. This is used to simultaneously present the distribution of meteorological elements and their corresponding risk levels, enabling rapid acquisition of the information needed for decision-making.

[0236] In some implementations, the specific operation process is as follows:

[0237] First, ensure that the risk heatmap and the meteorological profile data visualization images are completely consistent in scale and range on the horizontal and vertical axes. Then, render the S3 2D meteorological profile data as a contour map or color-filled map to generate a meteorological profile visualization. Next, set the transparency of the risk heatmap to 30%–50% and overlay it onto the meteorological profile visualization using an image overlay algorithm. During overlay, ensure that the risk color and meteorological color of each "distance-altitude along the flight path" grid point correspond precisely. For example, if a grid point has a meteorological color of red (high temperature) and a risk color of orange (higher risk), the overlay will present a composite color representing both high temperature and higher risk. After overlay, add a legend, axis labels, and color bars to generate a route assurance map.

[0238] Based on the aforementioned technical solutions, S4 addresses the shortcomings of traditional risk assessment methods, such as poor spatial continuity, short early warning periods, and insufficient identification of transmissible risks, by constructing a dynamic risk propagation map, calculating comprehensive risk values, and generating risk heat maps. The dynamic risk propagation map integrates spatial adjacency relationships with wind field-oriented causal relationships, making risk propagation simulation more closely aligned with the physical mechanisms of meteorological disaster evolution. The risk propagation network model, through multi-round causal modulation and attention aggregation, accurately quantifies the risk propagation contribution between nodes, improving the accuracy of comprehensive risk value calculation. The overlay of risk heat maps and meteorological profile data integrates complex meteorological and risk information into an intuitive route support map, lowering the barrier to information interpretation. The final generated route support map not only fully preserves the detailed distribution of meteorological elements along the route and altitude but also clearly presents the spatial changes in risk levels, achieving efficient transformation of meteorological information into risk decision-making information.

[0239] In one possible implementation of this application embodiment, the above-mentioned S5 specifically includes the following S501 to S503:

[0240] S501. Model the route dynamic planning and altitude layer optimization problem as a multi-objective constrained optimization problem, and clarify the decision variables, objective function and constraints.

[0241] Among them, route dynamic planning refers to finding the optimal flight trajectory along the route given the starting point, ending point and airspace constraints; altitude layer optimization refers to selecting the optimal flight altitude range in the vertical direction, and the two are combined to form a complete flight plan.

[0242] In some implementations, the specific operation process is as follows:

[0243] First, define the decision variables. The decision variables are the flight path R and the flight altitude profile H. The flight path R can be expressed as a function of the arc length coordinates s along the flight path. , where lon(s) and lat(s) are the longitude and latitude at arc length s, respectively, used to describe the horizontal flight trajectory; the flight altitude profile H is expressed as a function of arc length coordinates s, H(s) = h(s), where h(s) is the flight altitude at arc length s, used to describe the vertical altitude change.

[0244] Then, with the objectives of "minimizing total risk", "minimizing total flight time", and "minimizing total energy cost", a vector objective function is constructed. The complete form and explanation of each objective function are as follows:

[0245] Overall risk objective function: ;in The comprehensive risk value for position (s, h) in the route support map generated for S4, where s is the arc length coordinate along the route and h is the altitude. The integration process covers the entire flight path R and altitude profile H. The higher the risk value and the wider the coverage, the larger the integration result. The goal is to minimize this value through optimization.

[0246] Total flight time objective function: ;in To account for the influence of wind, the ground velocity is synthesized from wind speed, wind direction, and aircraft airspeed in the S3 meteorological profile data. The higher the ground velocity, the shorter the flight time per unit arc length, and the smaller the integral result. The goal is to maximize the ground velocity by optimizing the flight path and altitude to shorten the total time.

[0247] Total energy cost objective function: Where P is the power consumption model, and its inputs include flight path R, flight altitude H, aircraft weight W, and fuel flow rate. (Or battery energy consumption rate). The power consumption model needs to be pre-calibrated based on the aircraft's performance parameters. The smaller the integral result, the lower the energy consumption. The goal is to reduce the total energy consumption through optimization.

[0248] At the same time, constraints are set, including:

[0249] Flight altitude restrictions: ,in , These are the minimum and maximum permitted flight altitudes for the aircraft, respectively.

[0250] Climb / descent rate limit: ,in The maximum permissible climb / descent gradient is determined by the aircraft's performance.

[0251] Flight path curvature limit: ,in For the curvature of the flight path, The maximum permissible curvature of the flight path;

[0252] Airspace restrictions: ,in To restrict airspace, such as airport clear zones and military restricted areas;

[0253] Fuel quantity constraints: ,in This refers to the remaining fuel (or remaining electricity) along the route.

[0254] S502. The non-dominated sorting genetic algorithm III (NSGA-III) is used to solve the multi-objective constrained optimization problem and obtain the Pareto optimal solution set.

[0255] Among them, the Non-dominated Sorting Genetic Algorithm III (NSGA-III) is an evolutionary algorithm for multi-objective optimization problems. Through non-dominated sorting and reference point guidance, it finds an equilibrium solution among multiple mutually constrained objectives, avoiding getting trapped in local optima. The Pareto optimal solution set refers to a set of non-dominated solutions in the objective space; that is, for any solution in the set, no other solution is superior to it on all objectives. Each solution is the optimal solution under specific objective weights, providing users with diverse choices.

[0256] In some implementations, the specific operation process is as follows:

[0257] First, the algorithm parameters are initialized, including population size (usually 100–200), number of iterations (usually 200–500), crossover probability (0.7–0.9), and mutation probability (0.01–0.05). Each individual in the population corresponds to a potential "path + altitude profile" scheme, using a real-number encoding method: the path is encoded as the longitude of the arc length s along the route and the latitude offset relative to the initial route; the altitude profile is encoded as the altitude value of each arc length node, ensuring a one-to-one correspondence between the encoding and the decision variables.

[0258] Next, population initialization is performed. An initial population is randomly generated based on the initial flight path (the user-input route) and altitude range. All constraints must be met during generation.

[0259] Next, for each individual in each generation of the population, the values ​​of its three objective functions are calculated. The population is then sorted into non-dominated levels. The sorting rule is as follows: if all objective function values ​​of individual A are less than or equal to those of individual B, and at least one objective function value is strictly less than that of B, then A dominates B. The population is divided into different non-dominated levels, with the first level (Rank=1) being the set of the best non-dominated individuals in the current population, and subsequent levels decreasing in size.

[0260] Next, based on the value range of the three objectives, reference points are evenly set in the target space, such as setting 12 reference points to cover different preference directions such as safety priority, efficiency priority, and economic priority; the association between each non-dominant individual and the reference points is calculated to ensure that the population can spread to different regions of the target space.

[0261] Then, a tournament selection method was used to select individuals with high fitness (highest non-dominant level and close association with the reference point) from the population to enter the mating pool; the flight path code and altitude code of the selected individuals were cross-linked by simulated binary crossover (SBX) to generate offspring individuals, retaining the excellent characteristics of the parents; the codes of some offspring individuals were randomly perturbed by multinomial mutation to introduce new gene mutations and enhance population diversity.

[0262] Finally, the parent and offspring populations are merged, the objective function value of the merged population is calculated and non-dominated sorting is performed, and the optimal individuals with the same size as the parent population are selected based on the correlation of the reference points to form a new generation population. The above steps are repeated until the preset number of iterations is reached. At this point, the first layer of non-dominated individuals in the final population is the Pareto optimal solution set.

[0263] S503. Based on the target preference weights provided by the user, calculate the distance from the candidate solution to the ideal point using the weighted Chebyshev method, and select the solution with the smallest distance as the final optimized flight path.

[0264] The user objective preference weight refers to the importance weight (the sum of the weights is 1) assigned by the user to the three objectives of "minimizing total risk," "minimizing total flight time," and "minimizing total energy consumption cost" based on the flight mission requirements. For example, an emergency rescue mission might be assigned a risk weight of 0.6, a time weight of 0.3, and an energy consumption weight of 0.1, while a logistics transportation mission might be assigned a time weight of 0.5, an energy consumption weight of 0.4, and a risk weight of 0.1. The weighted Chebyshev method is a multi-objective decision-making method. Its core principle is to construct an ideal point (the theoretical optimal value of each objective) and calculate the weighted distance from each candidate solution to the ideal point. The smaller the distance, the closer the solution is to the optimal state preferred by the user. This method is used to select the final solution that meets the user's personalized needs from multiple solutions in the Pareto optimal solution set.

[0265] In some implementations, the specific operation process is as follows:

[0266] First, determine the ideal point. The ideal point is the set of theoretical optimal values ​​for each objective function, i.e. ,in , , M represents the number of solutions in the Pareto optimal solution set. The ideal point is the optimal benchmark in the objective space, which cannot be reached by practical solutions, but can serve as a reference for evaluating the quality of solutions.

[0267] Next, the user inputs the weights of the three objectives through a visual interactive module. , , The sum is 1; if the user does not explicitly input the weights, the system will assign a default balance weight. =0.4、 =0.3、 =0.3, balancing safety, efficiency and economy.

[0268] Then, the weighted Chebyshev distance is calculated. For each candidate solution k in the Pareto optimal solution set, the corresponding objective function value is... , , to the ideal point The weighted Chebyshev distance formula is:

[0269] ;

[0270] in, , , These represent the absolute deviations of scheme k from the ideal point on the three objectives; , , These represent the ranges of values ​​for the three objectives in the Pareto optimal solution set, used to normalize the absolute deviation and eliminate the influence of differences in the dimensions of different objectives; Let be the weighted Chebyshev distance from solution k to the ideal point. The smaller the distance, the closer the solution is to the optimal state preferred by the user.

[0271] Finally, by iterating through all the solutions in the Pareto optimal solution set, the solution with the smallest weighted Chebyshev distance D_k is selected. This solution is the final optimized flight path that meets the user's preferences, and includes specific horizontal flight path coordinates and vertical altitude profiles.

[0272] Based on the aforementioned technical solution, S5 addresses the problems of traditional route planning relying on human experience and failing to balance multiple objectives by employing a multi-objective constraint optimization modeling process, using NSGA-III to solve for the Pareto optimal solution set, and employing the weighted Chebyshev method to select the final solution. The modeling stage transforms flight requirements into a quantitative mathematical model, closely linking it to meteorological data and risk assessment results from previous steps to ensure the accuracy of the model input. The solution stage efficiently finds the Pareto optimal solution set that balances multiple objectives using the NSGA-III algorithm, providing users with diverse choices. The selection stage combines user-specific preferences with scientific decision-making methods to lock in the optimal solution, achieving an organic combination of algorithmic optimization and human preference. The final optimized flight path output satisfies the balance between safety, efficiency, and economy, while also adapting to the airspace and physical constraints of low-altitude flight. This transforms traditional qualitative human decision-making into quantitative intelligent assisted decision-making, significantly improving the safety, foresight, and intelligence of low-altitude flight, and providing core technical support for the leap from risk notification to decision support in aviation meteorological support.

[0273] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0274] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of this application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and modifications of this application fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and modifications.

Claims

1. A method for providing en-route meteorological support based on numerical weather prediction, characterized in that, include: Meteorological observation data and numerical forecast data for the flight route area are acquired, and the meteorological observation data is assimilated into the numerical forecast data using a four-dimensional variational assimilation method to generate meteorological field data. An AI corrector built based on a spatiotemporal graph convolutional network is used to correct errors in the meteorological field data to obtain optimized weather forecast results. Based on the route information input by the user and the optimized weather forecast results, route-altitude meteorological profile data are generated using a spatial interpolation algorithm; Disaster risk assessment is performed on the meteorological profile data based on the risk propagation network model, a risk heat map is generated, and the risk heat map is superimposed on the meteorological profile data to obtain a route support map; By combining a multi-objective optimization algorithm, the route support map is subjected to dynamic route planning and altitude-level optimization to obtain an optimized flight path; The optimized flight path is used to ensure low-altitude flight safety; The method of assimilating the meteorological observation data into the numerical forecast data using a four-dimensional variational assimilation method includes: The meteorological observation data is processed using an anisotropic background error covariance matrix. The background error covariance is obtained by calculating the elements of the covariance matrix by considering the similarity of terrain slope. The meteorological observation data includes satellite meteorological data, ground sensor data, and radar observation data. The background error covariance matrix is ​​used to quantify the correlation and spatial structure characteristics of meteorological element errors between different spatial points in the background field in a four-dimensional variational assimilation objective function. The expected value of terrain correction for each ground sensor location is calculated using a digital elevation model. Then, the deviation between the observed value of the ground sensor and the expected value of terrain correction is calculated. When the deviation exceeds a preset threshold, the observed value is downweighted to generate meteorological observation data after quality control. A four-dimensional variational assimilation objective function incorporating terrain constraints is constructed. By minimizing the four-dimensional variational assimilation objective function, the quality-controlled meteorological observation data is assimilated into the numerical forecast data, generating a high-resolution analysis field as meteorological field data.

2. The method for providing airway meteorological support based on numerical weather prediction according to claim 1, characterized in that, The AI ​​corrector built based on a spatiotemporal graph convolutional network performs error correction on the meteorological field data, including: Multiple two-dimensional horizontal fields of the input meteorological field data are processed by a two-dimensional convolutional encoder with shared weights to extract spatial feature maps of each meteorological element; wherein, the two-dimensional horizontal field represents the gridded data of a single meteorological element on any isobaric surface or model layer. The spatial feature maps of each extracted meteorological element are stitched together along the channel dimension and input into the physical information spatiotemporal Transformer module, which outputs spatiotemporal features. The physical information spatiotemporal Transformer module embeds the spatial coordinates and forecast lead time information of each grid point through learnable position encoding, and then captures the global long-range spatiotemporal dependence through a multi-head self-attention mechanism to simulate the long-range interaction between weather systems. The spatiotemporal features are input into the decoder, and a preliminary error correction field is generated through multiple deconvolutional layers or upsampling convolutional layers; the preliminary error correction field is then added to the meteorological field data to obtain a preliminary correction field. The preliminary correction field is input into the physical constraint layer, and the degree of violation of the physical constraint is used to calculate whether the preliminary correction field satisfies the preset physical conservation law. The physical loss is determined by the degree of violation of the physical constraint. During the training process of the spatiotemporal graph convolutional network, the mean square error between the corrected field output by the forward pattern and the reference analysis field is used as the main loss, which is weighted and summed with the physical loss to optimize the network parameters through backpropagation. The reference analysis field represents the true value of the meteorological field used to supervise the training of the network, which is obtained by acquiring historical analysis data from the same period. When the spatiotemporal graph convolutional network performs inference, the network outputs optimized weather forecast results and forecast uncertainty estimates for each grid point.

3. The method for providing airway meteorological support based on numerical weather prediction according to claim 2, characterized in that, The expression for the loss function of the spatiotemporal graph convolutional network during training is as follows: ;in, Main loss, Let represent the initial corrected field for the i-th training sample. This represents the reference analysis field corresponding to the i-th training sample. The total number of samples in a training batch; For physical loss, This is the physical loss weighting coefficient.

4. The method for providing airway meteorological support based on numerical weather prediction according to claim 3, characterized in that, The degree of violation of physical constraints is obtained by the deviation of the calculation result of applying physical law operators to the preliminary correction field from zero. The physical law operators include gradient operators and divergence operators. The calculation formula is: ;in, This represents the calculated value of the mass divergence term in the mass conservation equation. air density, It is a three-dimensional wind vector. Let θ represent the thermodynamic equation residuals, Q be the non-adiabatic heating rate, α be the weighting coefficient for mass conservation, and β be the weighting coefficient for thermodynamic constraints.

5. A method for providing airway meteorological support based on numerical weather prediction according to claim 1, characterized in that, The process of generating the route-altitude meteorological profile data includes: The user-input route is discretized into a sequence of waypoints at a preset step size. The horizontal coordinates, altitude, and circulation feature vector of the grid where the waypoint is located, extracted from the weather forecast results, are concatenated into a conditional vector. The circulation feature vector is obtained by extracting the standardized values ​​of each meteorological element within a preset radius from the grid data of the weather forecast results, centered on the grid point closest to the waypoint, and flattening them into a one-dimensional vector. The conditional vector is input into a pre-trained conditional neural radiation field model, which outputs meteorological element values ​​for each waypoint. The meteorological element values ​​represent the meteorological conditions at the waypoint location, including temperature, zonal wind, meridional wind, vertical velocity, relative humidity, and turbulent kinetic energy. By traversing all waypoints and user-specified altitude layers, and querying the conditional neural radiation field model, a two-dimensional data matrix is ​​output to obtain the route-altitude two-dimensional meteorological profile data. The horizontal axis of the two-dimensional data matrix is ​​the distance along the route, the vertical axis is the altitude, and each element in the matrix is ​​a set of meteorological element values ​​at the location of the waypoint.

6. A method for providing airway meteorological support based on numerical weather prediction according to claim 5, characterized in that, The disaster risk assessment based on the meteorological profile data using the risk propagation network model includes: Each grid point in the meteorological profile data is regarded as a graph node, and the feature vector of each graph node is composed of the meteorological element values ​​of the grid point; edges are established between spatially adjacent grid points, and additional directed edges are established downwind of the nodes according to the wind field direction to obtain a dynamic risk propagation map. The dynamic risk propagation graph and the feature vectors of the graph nodes are input into the risk propagation network model to perform risk propagation calculation and risk assessment, thereby obtaining the comprehensive risk value of each node; Based on the preset risk level threshold, the two-dimensional risk distribution matrix composed of the comprehensive risk value of each node is mapped into a color map to obtain a risk heat map.

7. A method for providing airway meteorological support based on numerical weather prediction according to claim 6, characterized in that, The risk propagation network model is a K-layer graph attention network, including a graph construction module, a feature encoding module, K stacked causal modulation graph attention layers, and a risk assessment module; wherein, The graph construction module is used to define each meteorological element grid point in the meteorological profile data as a graph node, extract the meteorological disaster-related index of each graph node to form an initial node feature vector, and construct spatial adjacency edges and causal propagation directed edges between graph nodes to obtain a dynamic risk propagation graph; wherein, the meteorological disaster-related index includes turbulence index, icing index, convective available potential energy, vertical wind shear and visibility; The feature encoding module maps the set of initial node feature vectors to latent space features using a multilayer perceptron, and outputs the encoded initial node feature matrix. The K stacked causal modulation graph attention layers are used to perform multi-round iterative information aggregation and updating of node features based on the initial node feature matrix and the edge adjacency relationships in the dynamic risk propagation graph, to obtain the final node feature representation; wherein, the internal workflow of the l-th layer of the causal modulation graph attention layer is as follows: Perform linear projection on the features of each node. , thus obtaining the projected node features; Let be the learnable weight matrix of the l-th layer. Let i be the feature vector of node i in the input of layer l. The projected feature vector of node i; Based on the projected node features, the original attention score is calculated for the node pairs (i,j) with edge connections in the dynamic risk propagation graph. The calculation formula is: ; Let be the feature vector of node j after projection. () is the activation function for a linear rectifier with leakage. This represents vector concatenation. This is a learnable attention vector; Introducing the causal prior matrix M and adaptive scaling The original attention score is adjusted to obtain the modulated attention score. The calculation formula is: ;in, The original attention score between node i and node k; the elements of the causal prior matrix M. , representing the statistical causal strength of the disaster occurring at node i due to the state of node j, learned from historical disaster data using a causal learning algorithm; the adaptive scale between node i and node j. Calculated based on the terrain differences and wind shear between nodes i and j; Based on the modulation attention score, node i is aggregated with its neighboring nodes to obtain the updated features, calculated as follows: ; For activation functions; Let j be the feature vector of node j in the input of layer l; The output of the l-th layer is the updated node feature matrix. ; The risk assessment module is used to utilize a fully connected neural network to process the final node feature matrix output by the last causal modulation attention layer. By performing nonlinear transformation and mapping, the comprehensive risk value of each node is obtained, and the comprehensive risk values ​​of all nodes constitute a two-dimensional risk distribution matrix.

8. A method for providing airway meteorological support based on numerical weather prediction according to claim 1, characterized in that, The method of combining multi-objective optimization algorithms to perform dynamic route planning and altitude-level optimization on the route support map includes: The problem of route dynamic planning and altitude layer optimization is modeled as a multi-objective constrained optimization problem. The decision variables of the multi-objective constrained optimization problem are the flight path R and the flight altitude profile H. The objectives of the multi-objective constrained optimization problem include minimizing the total risk, minimizing the total flight time, and minimizing the total energy consumption cost. The constraints of the multi-objective constrained optimization problem include flight altitude restrictions, climb / descent rate restrictions, path curvature restrictions, airspace restrictions, and fuel quantity constraints. The non-dominated sorting genetic algorithm NSGA-III is then used to solve the multi-objective constrained optimization problem to obtain the Pareto optimal solution set; From the optimal solution set, based on the user-provided preference weights for each objective, the weighted Chebyshev distance from each candidate solution to the ideal point is calculated using the weighted Chebyshev method. The candidate solution with the smallest distance is selected to obtain the final solution, and the optimized flight path is output.

9. A route meteorological support system based on numerical weather prediction, characterized in that, include: The system includes a data fusion module, an intelligent correction module, a profile generation module, a risk assessment module, a route optimization module, and a visualization and interaction module; among them, The data fusion module is communicatively connected to a multi-source meteorological data source to acquire meteorological observation data and numerical forecast data for the airway area, and uses a four-dimensional variational assimilation method to assimilate the meteorological observation data into the numerical forecast data to generate meteorological field data. The intelligent correction module is connected to the data fusion module and is used to receive the meteorological field data and perform error correction on the meteorological field data based on the AI ​​corrector constructed based on the spatiotemporal graph convolutional network to obtain the optimized meteorological forecast result. The profile generation module, the intelligent correction module, and the visualization interaction module are connected and used to generate route-altitude meteorological profile data through a spatial interpolation algorithm based on user route information from the interaction module and optimized weather forecast results from the intelligent correction module. The risk assessment module is connected to the profile generation module and is used to perform disaster risk assessment on the meteorological profile data based on the risk propagation network model, generate a risk heat map, and overlay the risk heat map onto the meteorological profile data to obtain a route support map. The route optimization module is connected to the risk assessment module and the visualization interaction module. It is used to combine a multi-objective optimization algorithm to perform dynamic route planning and altitude layer optimization on the route support map to obtain an optimized flight path. The visualization interaction module is connected to the profile generation module and the route optimization module, and is used to receive route information and decision preferences input by the user, and to display the route support map and the optimized flight path in a visual form; The method of assimilating the meteorological observation data into the numerical forecast data using a four-dimensional variational assimilation method includes: The meteorological observation data is processed using an anisotropic background error covariance matrix. The background error covariance is obtained by calculating the elements of the covariance matrix by considering the similarity of terrain slope. The meteorological observation data includes satellite meteorological data, ground sensor data, and radar observation data. The background error covariance matrix is ​​used to quantify the correlation and spatial structure characteristics of meteorological element errors between different spatial points in the background field in a four-dimensional variational assimilation objective function. The expected value of terrain correction for each ground sensor location is calculated using a digital elevation model. Then, the deviation between the observed value of the ground sensor and the expected value of terrain correction is calculated. When the deviation exceeds a preset threshold, the observed value is downweighted to generate meteorological observation data after quality control. A four-dimensional variational assimilation objective function incorporating terrain constraints is constructed. By minimizing the four-dimensional variational assimilation objective function, the quality-controlled meteorological observation data is assimilated into the numerical forecast data, generating a high-resolution analysis field as meteorological field data.

Images