A low-altitude aviation meteorological service method and system based on multi-source data fusion

Abstract

The present application relates to the field of multi-source data fusion, and more particularly to a low-altitude aviation meteorological service method and system based on multi-source data fusion, which comprises: embedding the stored meteorological data through STGNN and LSTM network architecture, carrying out meteorological prediction with pressure gradient gate, to generate predicted meteorological characteristics; storing site wind field data, elevation geographical image and predicted meteorological characteristics, carrying out wind field meteorological decomposition, meteorological dynamic calculation and turbulence modeling, to generate predicted wind speed field; displaying the predicted wind speed field on the elevation geographical image to generate a full low-altitude airspace wind speed annotation map. Through STGNN and LSTM, the present application captures the long-time sequence dependence and spatial correlation of meteorological data, embeds wavelet decomposition, air temperature difference buoyancy principle and turbulence generation law into the neural network, avoids physical contradictions, and more accurately depicts the turbulence intensity distribution law of the terrain, providing a basis for visual low-altitude aviation meteorological service.

Description

Technical Field

[0001] This invention relates to the field of multi-source data fusion, and in particular to a method and system for low-altitude aviation meteorological services based on multi-source data fusion. Background Technology

[0002] In the field of low-altitude aviation, especially in air logistics, safe operation relies heavily on accurate meteorological services, particularly the dynamic monitoring and forecasting of key parameters such as real-time wind speed and turbulence intensity. However, existing meteorological service methods have the following limitations in low-altitude scenarios:

[0003] Traditional meteorological services rely heavily on ground-based weather stations or radiosonde data, which have limited coverage and low spatiotemporal resolution, making it difficult to capture sudden changes in local wind fields caused by complex low-altitude terrain. Although satellite remote sensing and numerical weather prediction data have been increasingly used in recent years, the heterogeneity of multi-source data, such as station observations, geographic elevation, and historical meteorological time series, makes data fusion difficult. Existing methods often ignore the spatiotemporal correlation between data, resulting in limited prediction accuracy.

[0004] The low-altitude region is significantly affected by surface friction and topographic relief, resulting in high and frequently changing turbulence intensity, which directly threatens aircraft safety. Traditional models often use empirical formulas, such as semi-empirical models based on wind speed gradients, to estimate turbulence. These models struggle to accurately assess the turbulence generation mechanisms under complex terrain. Furthermore, predictions of the spatial distribution of wind speed fields often rely on interpolation algorithms, neglecting the dynamic impact of meteorological factors such as pressure gradients and temperature differences on the wind field.

[0005] Existing meteorological service products are mostly presented in numerical tables or rough contour maps, lacking high-precision spatial annotations for low-altitude aviation scenarios, making it difficult for ground control systems to intuitively identify dangerous areas.

[0006] Therefore, how to realize a visualized low-altitude aviation weather forecast service that integrates multi-source data and physical mechanisms to meet the needs of low-altitude aviation for real-time meteorological parameters and scenario adaptability is a technical problem that needs to be solved. Summary of the Invention

[0007] To this end, this invention provides a low-altitude aviation meteorological service method and system based on multi-source data fusion. By using STGNN and LSTM, it can capture long-term temporal dependencies of meteorological data and model spatial correlations. By embedding wavelet decomposition, the principle of temperature difference buoyancy, and turbulence generation laws into the neural network, it avoids the physical contradictions that are prone to occur in pure data-driven models, and more accurately depicts the distribution law of turbulence intensity in terrain, providing a basis for visualized low-altitude aviation meteorological services.

[0008] To achieve the above objectives, this invention proposes a low-altitude aviation meteorological service method based on multi-source data fusion, comprising:

[0009] The stored meteorological data is used to generate predicted meteorological features by performing meteorological prediction with embedded pressure gradient gating through a meteorological time series prediction model based on STGNN and LSTM network architecture.

[0010] The stored station wind field data, elevation geographic images, and the predicted meteorological features are used to perform wind field meteorological decomposition, meteorological dynamic calculation, and turbulence modeling through a meteorological physics guided model to generate a predicted wind speed field. The wind field meteorological decomposition is used to perform wavelet decomposition on the station wind field data, meteorological dynamic calculation is performed based on the principle of buoyancy due to temperature difference, and turbulence modeling is constructed based on the principle of neural network modeling based on the Navier-Stokes fluid equations.

[0011] The predicted wind speed field is displayed on the elevation geographic image to generate a wind speed annotation map of the entire low-altitude airspace.

[0012] Furthermore, the process of generating the predicted wind speed field includes:

[0013] The wind field data from the aforementioned stations are subjected to wind field meteorological decomposition based on learnable wavelet transform to generate low-frequency wind direction components and high-frequency turbulence components.

[0014] The turbulence components and the predicted meteorological features are fitted to form a turbulence model to generate turbulence prediction values;

[0015] The daily thermal buoyancy prediction value is calculated based on the Archimedes' principle of temperature difference buoyancy using temperature data.

[0016] The predicted thermal buoyancy value is generated by weighting the predicted value of the thermal buoyancy with the valley height gradient, coastline length and urban heat island gradient values ​​obtained from the elevation geographic image.

[0017] The predicted turbulence value, the wind direction component, and the predicted hot wind value are added together to generate the predicted wind speed field;

[0018] The meteorological data includes the temperature data.

[0019] Furthermore, the process of generating turbulence predictions includes:

[0020] The gradient value based on the air pressure data and the summation based on the turbulence components are used to generate the drift term;

[0021] The predicted meteorological features and learnable parameters are fitted and calculated to generate diffusion values, and the diffusion values ​​are multiplied by turbulent kinetic energy to generate diffusion terms.

[0022] The drift term and the diffusion term are substituted into the turbulence modeling model to generate the turbulence prediction value;

[0023] The meteorological data also includes the air pressure data.

[0024] Furthermore, the process of generating drift terms includes:

[0025] Based on the product of the gradient value and the learnable energy transfer, a predicted pressure gradient value is generated.

[0026] The turbulent dissipation prediction is generated based on the product of the turbulent components and the learnable dissipation coefficient.

[0027] The predicted pressure gradient and the predicted turbulent dissipation are summed to generate the drift term.

[0028] Furthermore, the training and optimization process of the meteorological physics-guided model includes:

[0029] A decomposition loss term is constructed based on the differences between the meteorological characteristics of the sample and the wind direction component and the turbulence component;

[0030] A wind speed fitting loss term is constructed based on the difference between the actual wind speed field of the sample and the predicted wind speed field.

[0031] A total loss function is constructed based on the weighted sum of the decomposition loss term and the wind speed fitting loss term, and the total loss function is used for the training and optimization of the meteorological physical guided model.

[0032] In particular, by decomposing station wind field data using learnable wavelet transform, the complex wind field is broken down into a low-frequency wind direction component reflecting macroscopic airflow trends and a high-frequency turbulence component capturing microscale disturbances. Turbulence modeling based on the Navier-Stokes equations, through the coordinated calculation of drift and diffusion terms, overcomes the limitations of empirical formulas in roughly estimating turbulence intensity, resulting in turbulence predictions that closely match fluid dynamics laws. Furthermore, by combining Archimedes' principle of buoyancy based on temperature difference with topographic parameters from elevation geographic images, the influence of thermal factors on the wind speed field is quantified, making the predicted wind speed field more closely reflect actual observation data in its diurnal variations. Finally, through data decomposition, physical modeling, and dynamic fusion, the problems of physical fragmentation and poor scene adaptability in low-altitude aviation meteorological services for wind field prediction are effectively solved.

[0033] Furthermore, the process of generating predicted meteorological features includes:

[0034] Divide the elevation geographic image into STGNN nodes and STGNN edges;

[0035] The meteorological data is processed through a graph convolutional network based on the STGNN edges to perform data fusion between STGNN nodes, thereby generating node-fused meteorological data.

[0036] The node fusion meteorological data is processed by using an LSTM with embedded pressure gradient gating to extract the node's historical meteorological features in order to generate the predicted meteorological features.

[0037] Furthermore, the process of generating predictive meteorological features using an LSTM with embedded pressure gradient gating includes:

[0038] The gating coefficient is calculated based on the learnable weight matrix, the magnitude of the pressure gradient, the angle between the pressure data and the station wind field data, and the difference in temperature data.

[0039] The predicted meteorological features are generated by element-wise multiplication of the gating coefficients and the temporary meteorological features output by the LSTM.

[0040] Furthermore, the training and optimization process of the meteorological time-series prediction model includes:

[0041] The meteorological time series prediction model is trained and optimized based on the true meteorological characteristics of the samples and the mean square error loss function of the predicted meteorological characteristics.

[0042] The present invention also provides a low-altitude aviation meteorological service system based on multi-source data fusion, wherein the low-altitude aviation meteorological service system is equipped with the aforementioned low-altitude aviation meteorological service method based on multi-source data fusion.

[0043] Furthermore, this includes a low-altitude meteorological service platform built on WebGIS, a data storage server, a data processing server, and a geographic information server;

[0044] The site wind field data is the detection data of the site's lidar stored on the data storage server;

[0045] The meteorological time-series forecasting model and the meteorological physical guidance model are mounted on the data processing server;

[0046] The elevation geographic image is stored in the geographic information server;

[0047] The wind speed map for the entire low-altitude airspace is displayed on the low-altitude meteorological service platform.

[0048] In particular, by dividing STGNN nodes and edges onto the elevation geographic image, the continuous geographic space is transformed into graph-structured data. This allows the model to explicitly model the meteorological data correlations between nodes in different regions through a graph convolutional network (GCN), overcoming the neglect of spatial correlations by traditional grid models and improving the spatial resolution of data fusion. The design of the gating coefficients directly embeds core meteorological principles (such as the relationship between pressure gradient force and wind field direction, and the driving effect of temperature difference on airflow), enabling LSTM to follow physical laws when extracting historical time-series features.

[0049] Compared with existing technologies, the beneficial effects of this invention are that, by using STGNN and LSTM, it can capture long-term temporal dependencies in meteorological data and model spatial correlations. By embedding wavelet decomposition, the principle of temperature difference buoyancy, and the laws of turbulence generation into the neural network, it avoids the physical contradictions that are prone to occur in pure data-driven models, and more accurately depicts the distribution law of turbulence intensity in terrain, providing a basis for visualized low-altitude aviation meteorological services.

[0050] In particular, this invention decomposes station wind field data using learnable wavelet transform, breaking down complex wind fields into low-frequency wind direction components reflecting macroscopic airflow trends and high-frequency turbulence components capturing microscale disturbances. Turbulence modeling based on Navier-Stokes equations, through the collaborative calculation of drift and diffusion terms, overcomes the limitations of empirical formulas in roughly estimating turbulence intensity, ensuring a high degree of consistency between turbulence predictions and fluid dynamics laws. By combining Archimedes' principle of buoyancy based on temperature difference with topographic parameters from elevation geographic images, the influence of thermal factors on the wind speed field is quantified, making the predicted wind speed field more closely reflect actual observation data in its diurnal variations. Furthermore, through data decomposition, physical modeling, and dynamic fusion, this invention effectively solves the problems of physical fragmentation and poor scene adaptability in wind field prediction for low-altitude aviation meteorological services.

[0051] In particular, this invention divides STGNN nodes and edges onto elevation geographic images, transforming continuous geographic space into graph-structured data. This allows the model to explicitly model the meteorological data correlations between nodes in different regions through a graph convolutional network (GCN), overcoming the neglect of spatial correlations by traditional grid models and improving the spatial resolution of data fusion. The design of the gating coefficients directly embeds core meteorological principles (such as the relationship between pressure gradient force and wind field direction, and the driving effect of temperature difference on airflow), enabling LSTM to follow physical laws when extracting historical time-series features. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating the low-altitude aviation meteorological service method based on multi-source data fusion according to an embodiment of the present invention.

[0053] Figure 2 This is a flowchart illustrating the meteorological physics-guided model of the low-altitude aviation meteorological service method based on multi-source data fusion, according to an embodiment of the present invention.

[0054] Figure 3 This is a schematic diagram of the training and optimization process of the meteorological physics-guided model of the low-altitude aviation meteorological service method based on multi-source data fusion according to an embodiment of the present invention.

[0055] Figure 4 This is a flowchart illustrating the meteorological time-series prediction model of the low-altitude aviation meteorological service method based on multi-source data fusion, according to an embodiment of the present invention.

[0056] Figure 5 This is a schematic diagram showing the display effect of the low-altitude airspace wind speed annotation platform of the low-altitude aviation meteorological service method and system based on multi-source data fusion according to an embodiment of the present invention.

[0057] Figure 6 This is a schematic diagram illustrating the effect of using a full low-altitude airspace wind speed annotation map of the low-altitude aviation meteorological service method and system based on multi-source data fusion for displaying wind speeds along low-altitude routes, as shown in this embodiment of the invention.

[0058] Figure 7 This is a schematic diagram illustrating the station wind field data acquisition effect of the low-altitude aviation meteorological service method and system based on multi-source data fusion according to an embodiment of the present invention.

[0059] Figure 8 This is a schematic diagram of the architecture of the low-altitude meteorological service platform for fusing and processing multi-source data in an embodiment of the low-altitude aviation meteorological service method and system based on multi-source data fusion according to an embodiment of the present invention.

[0060] Figure 9 This is a schematic diagram of the low-altitude meteorological service platform of the low-altitude aviation meteorological service method and system based on multi-source data fusion, according to an embodiment of the present invention. Detailed Implementation

[0061] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0062] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0063] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate the direction or positional relationship, are based on the direction or positional relationship shown in the drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0064] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0065] This invention provides a method and system for low-altitude aviation meteorological services based on multi-source data fusion. By using STGNN and LSTM, it can capture long-term temporal dependencies in meteorological data and model spatial correlations. By embedding wavelet decomposition, the principle of temperature difference buoyancy, and turbulence generation laws into the neural network, it avoids the physical contradictions that are prone to occur in pure data-driven models, and more accurately depicts the distribution law of turbulence intensity in terrain, providing a basis for visualized low-altitude aviation meteorological services.

[0066] like Figure 1 As shown, this embodiment proposes a low-altitude aviation meteorological service method based on multi-source data fusion, including:

[0067] The stored meteorological data is used to generate predicted meteorological features by performing meteorological prediction with embedded pressure gradient gating through a meteorological time series prediction model based on STGNN and LSTM network architecture.

[0068] The stored station wind field data, elevation geographic images, and the predicted meteorological features are used to perform wind field meteorological decomposition, meteorological dynamic calculation, and turbulence modeling through a meteorological physics guided model to generate a predicted wind speed field. The wind field meteorological decomposition is used to perform wavelet decomposition on the station wind field data, meteorological dynamic calculation is performed based on the principle of buoyancy due to temperature difference, and turbulence modeling is constructed based on the principle of neural network modeling based on the Navier-Stokes fluid equations.

[0069] The predicted meteorological features and the predicted wind speed field are dynamically weighted and fused through an attention mechanism to generate the wind speed prediction value of the STGNN node, and the wind speed prediction value is displayed on the elevation geographic image to generate a wind speed annotation map of the entire low-altitude airspace.

[0070] like Figure 2 As shown, in this embodiment, the process of generating the predicted wind speed field includes:

[0071] The wind field data from the aforementioned stations are subjected to wind field meteorological decomposition based on learnable wavelet transform to generate low-frequency wind direction components and high-frequency turbulence components.

[0072] The turbulence components and the predicted meteorological features are fitted to form a turbulence model to generate turbulence prediction values;

[0073] The daily thermal buoyancy prediction value is calculated based on the Archimedes' principle of temperature difference buoyancy using temperature data.

[0074] The predicted thermal buoyancy value is generated by weighting the predicted value of the thermal buoyancy with the valley height gradient, coastline length and urban heat island gradient values ​​obtained from the elevation geographic image.

[0075] The predicted turbulence value, the wind direction component, and the predicted hot wind value are added together to generate the predicted wind speed field;

[0076] The meteorological data includes the temperature data.

[0077] Specifically, the process of performing wind field meteorological decomposition based on learnable wavelet transform can be represented as:

[0078] C l =W l *V

[0079]

[0080] In the formula, C l W represents the turbulent component of the l-th layer. l This represents the wave filter bank of the meteorological wavelet decomposition (MetWavelet), where V represents the original station wind field data. Higher layer numbers correspond to smaller-scale, higher-frequency turbulence components. weather C represents the wind direction component, C1 represents the larger-scale turbulence component at the cloud scale, and C L These represent the smallest-scale turbulent components of local gusts, which are crucial for refined forecasting.

[0081] Specifically, the process of calculating the predicted value of thermal buoyancy can be expressed as:

[0082]

[0083] In the formula, F thermal This represents the predicted value of thermal buoyancy, and β represents the coefficient of thermal expansion, preferably 3.4 × 10⁻⁶. -3 K -1 t0 represents the maximum heating time predicted for the STGNN node, which is 14 hours for fields and 14 x 0.3 = 4.2 hours for water bodies. g represents the acceleration due to gravity, and ΔT t T represents the temperature difference between the Earth's surface and the surrounding air, where t represents the current time. diel This represents a 24-hour daily cycle, therefore, through βgΔT t The Archimedes force can be calculated by measuring temperature differences and then... By performing diurnal cycle modulation, the predicted value of thermo-buoyancy that varies with the diurnal cycle can be calculated.

[0084] Specifically, the process of calculating and predicting the wind speed field can be expressed as:

[0085] V out =V weather +V t +γ·F thermal ·H grad

[0086] In the formula, Vout V represents the predicted wind speed field. weather V represents the wind direction component. t Let F represent the turbulence prediction value, γ represent the learnable terrain coupling coefficient, and F be optimized through the total loss function. thermal H represents the predicted value of thermal buoyancy. grad This represents the valley height gradient, coastline length, or urban heat island gradient value derived from elevation geographic images. Wherein, γ·F thermal ·H grad This represents the predicted hot wind value, with the valley height gradient being... That is, the gradient values ​​of the elevation data in the x and y directions of the elevation geographic image, the coastline length is the coastline normal vector, and the gradient value of the urban heat island is... Where ΔT urban-rural This represents the recorded urban-rural temperature difference data, which are definitive values ​​derived from local historical data. It represents the urban density gradient, which is determined by analyzing the distribution range and density changes of urban built-up areas, such as commercial areas and residential areas.

[0087] In this embodiment, the process of generating turbulence prediction values ​​includes:

[0088] The gradient value based on the air pressure data and the summation based on the turbulence components are used to generate the drift term;

[0089] The predicted meteorological features and learnable parameters are fitted and calculated to generate diffusion values, and the diffusion values ​​are multiplied by turbulent kinetic energy to generate diffusion terms.

[0090] The drift term and the diffusion term are substituted into the turbulence modeling model to generate the turbulence prediction value;

[0091] The meteorological data also includes the air pressure data.

[0092] Specifically, the calculation process of the diffusion term can be expressed as follows:

[0093] f diff =σ turb ·tanh(WC l )

[0094] In the formula, f diff Let σ represent the diffusion term. turb Represents turbulent kinetic energy. Where k is the turbulent kinetic energy, which is determined based on the turbulent energy spectrum obtained from the station's wind field data, tanh represents the tanh activation function, W represents the learnable parameter, and C... l This represents the turbulent component.

[0095] Specifically, the turbulence modeling model can be expressed as:

[0096]

[0097] In the formula, V t v′ t+Δt f represents the predicted value of turbulence. drift f represents the drift term. drift Let v′ represent the diffusion term. t The turbulent fluctuation velocity is represented by the wind field data from the station closest to the STGNN node to be predicted, and Δt represents the prediction time step. t This represents a random value belonging to (0,1). This is a random term used to ensure that the predicted turbulence values ​​do not overfit.

[0098] Understandably, the aforementioned turbulence modeling model is determined by discretizing the differential equations of turbulence energy cascade modeling using the Euler-Maruyama discretization scheme. These differential equations are: The turbulence modeling model employs a discrete form of neural stochastic differential equations, connecting continuous-time theory with actual numerical simulations to fit the drift and diffusion terms, where v′ t +f drift Δt is used to replace the pressure term in the differential equation. and viscous terms The diffusion term σdW used to replace the differential equation t The above-mentioned v is the viscous dissipation coefficient of the wind fluid.

[0099] In this embodiment, the process of generating drift terms includes:

[0100] Based on the product of the gradient value and the learnable energy transfer, a predicted pressure gradient value is generated.

[0101] The turbulent dissipation prediction is generated based on the product of the turbulent components and the learnable dissipation coefficient.

[0102] The predicted pressure gradient and the predicted turbulent dissipation are summed to generate the drift term.

[0103] Specifically, the calculation process for the drift term can be expressed as follows:

[0104]

[0105] In the formula, f drift Let ρ represent the drift term and ρ represent the learnable energy transfer. This represents the gradient value of the air pressure data, which is the partial derivative of the air pressure data with respect to the wind direction. Used to describe the fluid acceleration caused by a pressure gradient as it flows from high pressure to low pressure, ν represents the learnable dissipation coefficient, and C l h represents the turbulent component. t It indicates the predicted meteorological characteristics.

[0106] like Figure 3 As shown, in this embodiment, the training and optimization process of the meteorological physics-guided model includes:

[0107] A decomposition loss term is constructed based on the differences between the meteorological characteristics of the sample and the wind direction component and the turbulence component;

[0108] A wind speed fitting loss term is constructed based on the difference between the actual wind speed field of the sample and the predicted wind speed field.

[0109] A total loss function is constructed based on the weighted sum of the decomposition loss term and the wind speed fitting loss term, and the total loss function is used for the training and optimization of the meteorological physical guided model.

[0110] Specifically, the total loss function can be expressed as:

[0111]

[0112] In the formula, Loss represents the total loss function, T represents the total learning period, and V out Indicates the predicted wind speed field. Represents the true wind speed field of the sample, || ||2 represents the L2 norm, V represents the decomposition of the loss term. t V represents the meteorological characteristics of the sample. weather Indicates wind direction component, Let λ represent the turbulent component of the total gradient L, and λ1 be the weighting coefficient, preferably 0.6, because wavelet decomposition has a large number of learnable parameters. Therefore, the convergence speed of the model can be accelerated by decomposing the loss term.

[0113] In particular, by decomposing station wind field data using learnable wavelet transform, the complex wind field is broken down into a low-frequency wind direction component reflecting macroscopic airflow trends and a high-frequency turbulence component capturing microscale disturbances. Turbulence modeling based on the Navier-Stokes equations, through the coordinated calculation of drift and diffusion terms, overcomes the limitations of empirical formulas in roughly estimating turbulence intensity, resulting in turbulence predictions that closely match fluid dynamics laws. Furthermore, by combining Archimedes' principle of buoyancy based on temperature difference with topographic parameters from elevation geographic images, the influence of thermal factors on the wind speed field is quantified, making the predicted wind speed field more closely reflect actual observation data in its diurnal variations. Finally, through data decomposition, physical modeling, and dynamic fusion, the problems of physical fragmentation and poor scene adaptability in low-altitude aviation meteorological services for wind field prediction are effectively solved.

[0114] like Figure 4As shown, in this embodiment, the process of generating predicted meteorological features includes:

[0115] Divide the elevation geographic image into STGNN nodes and STGNN edges;

[0116] The meteorological data is processed through a graph convolutional network based on the STGNN edges to perform data fusion between STGNN nodes, thereby generating node-fused meteorological data.

[0117] The node fusion meteorological data is processed by using an LSTM with embedded pressure gradient gating to extract the node's historical meteorological features in order to generate the predicted meteorological features.

[0118] Specifically, the process of generating node-fused meteorological data through a graph convolutional network can be represented as follows:

[0119]

[0120] In the formula, This indicates that STGNN node i has access to the meteorological data h of its neighbor STGNN node j. j The nodes integrate meteorological data. Let σ represent the set of all neighboring STGNN nodes of STGNN node i, and let α represent the Sigmoid loss function. ij W represents the attention weight of STGNN node i to its neighbor STGNN node j. (l) This represents the learnable weights.

[0121] like Figure 4 As shown, in this embodiment, the process of generating predicted meteorological features using an LSTM with embedded pressure gradient gating includes:

[0122] The gating coefficient is calculated based on the learnable weight matrix, the magnitude of the pressure gradient, the angle between the pressure data and the station wind field data, and the difference in temperature data.

[0123] The predicted meteorological features are generated by element-wise multiplication of the gating coefficients and the temporary meteorological features output by the LSTM.

[0124] Specifically, the features of the node-fused meteorological data within the [tk, t-1] time window are input into the LSTM, and the LSTM outputs temporary meteorological features.

[0125] Specifically, the process of generating predicted meteorological features can be represented as:

[0126]

[0127] In the formula, h t Indicates predicted meteorological characteristics, g physLet σ represent the gating coefficients, σ ​​represent the Sigmoid loss function, and U represent the learnable weight matrix. Cosθ represents the magnitude of the pressure gradient. t ΔT represents the cosine of the angle between the air pressure data and the station wind field data. t The temperature data difference is represented by the multiplication of the modulus of the pressure gradient, the angle between the pressure data and the station wind field data, and the temperature data difference with a learnable weight matrix, followed by a weighted sum to generate the gating coefficient. This represents the temporary meteorological characteristics of the LSTM output, and ⊙ represents element-wise multiplication.

[0128] In this embodiment, the training and optimization process of the meteorological time series prediction model includes:

[0129] The meteorological time series prediction model is trained and optimized based on the true meteorological characteristics of the samples and the mean squared error loss function (MSE) of the predicted meteorological characteristics.

[0130] Specifically, the hyperparameters of the meteorological time-series prediction model and the meteorological physical guidance model include: STGNN node count (sites and prediction interpolation grid points) 200, graph update frequency per hour, graph convolution type Directional GAT, number of heads 8, hidden layer dimension (node ​​feature dimension) 128, number of layers 3, activation function GELU, Dropout 0.2, LSTM: input sequence length (24-hour historical data) 24, hidden layer size 128, number of layers 2, Dropout 0.3, gating fusion parameters: gating dimension 128, gating layer number 2, wind speed weight coefficient 0.5, terrain gradient coefficient 0.3, and meteorological coefficient 0.7.

[0131] In particular, by dividing STGNN nodes and edges onto the elevation geographic image, the continuous geographic space is transformed into graph-structured data. This allows the model to explicitly model the meteorological data correlations between nodes in different regions through a graph convolutional network (GCN), overcoming the neglect of spatial correlations by traditional grid models and improving the spatial resolution of data fusion. The design of the gating coefficients directly embeds core meteorological principles (such as the relationship between pressure gradient force and wind field direction, and the driving effect of temperature difference on airflow), enabling LSTM to follow physical laws when extracting historical time-series features.

[0132] This embodiment also provides a low-altitude aviation meteorological service system based on multi-source data fusion, wherein the low-altitude aviation meteorological service system is equipped with the aforementioned low-altitude aviation meteorological service method based on multi-source data fusion.

[0133] In this embodiment, the meteorological service system includes a low-altitude meteorological service platform built on WebGIS, a data storage server, a data processing server, and a geographic information server;

[0134] The site wind field data is the detection data of the site's lidar stored on the data storage server;

[0135] The meteorological time-series forecasting model and the meteorological physical guidance model are mounted on the data processing server;

[0136] The elevation geographic image is stored in the geographic information server;

[0137] The wind speed map for the entire low-altitude airspace is displayed on the low-altitude meteorological service platform.

[0138] Specifically, such as Figure 8 and 9 The low-air meteorological service platform shown, based on a WebGIS platform, enables the querying and display of various airport and town real-time conditions and forecasts, radar mosaics, satellite cloud images, tropical cyclones, and European medium- and long-term refined numerical model forecast products on mobile tablet devices, and provides customized route and flight mission area functions. Specifically, the meteorological time-series forecasting model and meteorological physical guidance model described in this embodiment are encapsulated in a... Figure 8 The system's manual analysis and early warning product demonstrates how it integrates and predicts wind fields based on multi-source data, including station wind field data, elevation geographic images, and meteorological data stored in the system. The resulting full low-altitude airspace wind speed map is displayed as follows. Figure 5 and 6 As shown, in Figure 5 In the aforementioned low-altitude airspace wind speed mapping, portions of the predicted wind speed field exceeding the safety threshold are marked in red to provide a visual wind speed warning. Figure 6 The system can display wind speed and direction in the area traversed by the flight path. The effect of collecting station wind field data using station lidar in this embodiment is as follows: Figure 7 As shown, it only covers the area around the station and cannot provide accurate wind speed and direction for areas not covered by the station.

[0139] In this embodiment, STGNN and LSTM are used to capture both the long-term temporal dependencies of meteorological data and model spatial correlations. Wavelet decomposition, the principle of buoyancy due to temperature difference, and the laws of turbulence generation are embedded into the neural network, avoiding the physical contradictions that are prone to occur in purely data-driven models. This more accurately depicts the distribution of turbulence intensity in the terrain, providing a basis for visualized low-altitude aviation meteorological services. The station wind field data is decomposed using learnable wavelet transform, breaking down the complex wind field into a low-frequency wind direction component reflecting macroscopic airflow trends and a high-frequency turbulence component capturing microscale disturbances. Turbulence modeling based on the Navier-Stokes equations, through the collaborative calculation of drift and diffusion terms, overcomes the limitations of empirical formulas in roughly estimating turbulence intensity, making the predicted turbulence values ​​highly consistent with fluid dynamics laws. Combining the Archimedes principle of buoyancy due to temperature difference with topographic parameters from elevation geographic images, the influence of thermal factors on the wind speed field is quantified, making the predicted wind speed field more closely resemble actual observation data in its diurnal variations. Furthermore, through data decomposition, physical modeling, and dynamic fusion, the problems of physical fragmentation and poor scene adaptability in wind field prediction in low-altitude aviation meteorological services are effectively solved. STGNN nodes and edges are divided on the elevation geographic image, transforming continuous geographic space into graph-structured data. This allows the model to explicitly model the meteorological data correlations between nodes in different regions through a graph convolutional network (GCN), overcoming the neglect of spatial correlations by traditional grid models and improving the spatial resolution of data fusion. The design of the gating coefficients directly embeds core meteorological principles (such as the relationship between pressure gradient force and wind field direction, and the driving effect of temperature difference on airflow), enabling LSTM to follow physical laws when extracting historical time-series features.

[0140] Those skilled in the art will recognize that the models and algorithmic steps of the examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0141] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0142] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A low-altitude aviation meteorological service method based on multi-source data fusion, characterized in that, The method comprises the following steps: The stored meteorological data is input into a meteorological time series prediction model based on an STGNN and an LSTM network architecture to perform meteorological prediction with pressure gradient gating to generate predicted meteorological features; The stored site wind field data, elevation geographical image and the predicted meteorological features are input into a meteorological physical guidance model to perform wind field meteorological decomposition, meteorological dynamic calculation and turbulence modeling to generate a predicted wind speed field, wherein the wind field meteorological decomposition is used to perform wavelet decomposition on the site wind field data, the meteorological dynamic calculation is based on the air temperature difference buoyancy principle, and the turbulence modeling is based on the neural network modeling principle constructed according to the Navier-Stokes fluid equation; The predicted wind speed field is displayed on the elevation geographical image to generate a full low-altitude airspace wind speed labeling image; The process of generating the predicted wind speed field comprises the following steps: The site wind field data is subjected to wind field meteorological decomposition based on a learnable wavelet transform to generate a low-frequency wind direction component and a high-frequency turbulence component; The turbulence component and the predicted meteorological features are subjected to fitting turbulence modeling to generate a turbulence prediction value; The air temperature data is used to calculate a daily thermal buoyancy prediction value based on the Archimedes air temperature difference buoyancy principle; The thermal buoyancy prediction value and the mountain valley height gradient, coastline length and urban heat island gradient value obtained based on the elevation geographical image are subjected to weighted calculation to generate a predicted thermal wind value; The turbulence prediction value, the wind direction component and the predicted thermal wind value are added to generate the predicted wind speed field; The meteorological data comprises the air temperature data. 2.The method of claim 1, wherein, The process of generating the turbulence prediction value comprises the following steps: A gradient value based on air pressure data and a sum based on the turbulence component are used to generate a drift term; The predicted meteorological features and a learnable parameter are used for fitting calculation to generate a diffusion value, and the diffusion value and turbulence kinetic energy are multiplied to generate a diffusion term; The drift term and the diffusion term are substituted into a turbulence modeling model to generate the turbulence prediction value; The meteorological data further comprises the air pressure data. 3.The method of claim 2, wherein, The process of generating the drift term comprises the following steps: A product of the gradient value and a learnable energy transfer is used to generate an air pressure gradient prediction value; A product of the turbulence component and a learnable dissipation coefficient is used to generate a turbulence dissipation prediction value; The air pressure gradient prediction value and the turbulence dissipation prediction value are summed to generate the drift term. 4.The method of claim 1, wherein, The training and optimization process of the meteorological physical guidance model comprises the following steps: A difference between a sample's meteorological features and the wind direction component and the turbulence component is used to construct a decomposition loss term; A difference between a sample's real wind speed field and the predicted wind speed field is used to construct a wind speed fitting loss term; A weighted sum of the decomposition loss term and the wind speed fitting loss term is used to construct a total loss function, and the total loss function is used for training and optimization of the meteorological physical guidance model. 5.The method for low-altitude aviation weather service based on multi-source data fusion according to claim 1, characterized in that, The process of generating the predicted meteorological features comprises the following steps: STGNN nodes and STGNN edges are divided on the elevation geographical image; The meteorological data is fused between the STGNN nodes through a graph convolution network according to the STGNN edges to generate node fused meteorological data; The node fusion meteorological data is extracted by an LSTM embedded with a barometric pressure gradient gate to generate the predicted meteorological feature.

6. The method for low-altitude aviation meteorological service based on multi-source data fusion according to claim 5, characterized in that, The process of generating the predicted meteorological feature by the LSTM embedded with the barometric pressure gradient gate comprises: calculating a gate coefficient based on a learnable weight matrix, a modulus of the barometric pressure gradient, an included angle of the barometric pressure data and the site wind field data, and a difference of the air temperature data; generating the predicted meteorological feature based on an element-wise multiplication of the gate coefficient and a temporary meteorological feature output by the LSTM. 7.The low-altitude aviation weather service method based on multi-source data fusion according to claim 1, wherein, The training optimization process of the meteorological time series prediction model comprises: training and optimizing the meteorological time series prediction model based on a mean square error loss function of the real meteorological feature and the predicted meteorological feature of a sample.

8. A low-altitude aviation meteorological service system based on multi-source data fusion, characterized in that, The low-altitude aviation meteorological service system carries the low-altitude aviation meteorological service method based on multi-source data fusion according to any one of claims 1 to 7.

9. The low-altitude aviation meteorological service system based on multi-source data fusion according to claim 8, characterized in that, The low-altitude meteorological service platform based on WebGIS, a data storage server, a data processing server, and a geographic information server are included. The site wind field data is detection data of a site laser radar stored in the data storage server. The meteorological time series prediction model and the meteorological physical guidance model are carried in the data processing server. The high-elevation geographic image is stored in the geographic information server. The full low-altitude airspace wind speed label map is displayed on the low-altitude meteorological service platform.

Images