Low-altitude visibility prediction method, system, device and medium based on wind field driving

Abstract

The application discloses a low-altitude visibility prediction method and system based on wind field driving, equipment and a medium. The method comprises the following steps: inputting a obtained multivariate time series matrix into a wind door control graph attention network to obtain an advection space feature matrix; fusing the advection space feature matrix and the multivariate time series matrix to obtain a fused multivariate time series matrix; splitting the fused multivariate time series matrix into multiple independent channels to obtain a one-dimensional time series corresponding to each independent channel; processing the one-dimensional time series through a Transformer encoder to determine a target time-varying burst feature vector; determining a dry aerosol extinction coefficient estimate value and a hygroscopic growth parameter estimate value based on the target time-varying burst feature vector; and performing visibility prediction based on the dry aerosol extinction coefficient estimate value and the hygroscopic growth parameter estimate value to obtain a visibility prediction result. The application can improve the accuracy of low-altitude visibility prediction.

Description

Technical Field

[0001] This application relates to the field of meteorological forecasting technology, and in particular to a method, system, device and medium for low-altitude visibility prediction based on wind field drive. Background Technology

[0002] The core airspace for low-altitude visibility forecasting is typically located within 1,000 meters above the ground (especially in the 0 to 300 meter range). Aircraft in this airspace (such as industrial drones and electric vertical takeoff and landing (eVTOL) aircraft) are highly sensitive to low-altitude micro-meteorological elements (especially wind shear, strong turbulence, and sudden low visibility) due to their size, payload, and power system limitations.

[0003] Sudden drops in visibility (such as sudden advection fog or radiation fog) are a major meteorological disaster factor causing low-altitude aircraft to veer off course, collide with obstacles, or fail to make emergency landings. Traditional aviation weather forecasting relies heavily on numerical weather prediction and a sparse network of conventional ground-based meteorological stations. However, numerical prediction models are limited by grid resolution (usually at the kilometer level) and computation time, making it difficult to provide high-frequency warnings at the "hundred-meter level" or "minute level" for low-altitude flights. Meanwhile, ground-based station data only reflects local conditions in a two-dimensional plane and cannot accurately characterize the three-dimensional evolution of aerosols and water vapor within the low-altitude boundary layer. Existing methods for visibility prediction using deep learning technology, while capable of combining multi-source meteorological data for short-term visibility prediction, suffer from problems such as ignoring dynamic changes in instantaneous wind speed and direction, the model being prone to getting trapped in local optima, and introducing spurious high-frequency noise, resulting in relatively low accuracy in low-altitude visibility prediction. Summary of the Invention

[0004] This application aims to propose a wind-driven low-altitude visibility prediction method, system, device, and medium that can improve the accuracy of low-altitude visibility prediction.

[0005] In a first aspect, embodiments of this application provide a low-altitude visibility prediction method based on wind field driving, the method comprising:

[0006] The acquired multi-source heterogeneous meteorological data are preprocessed to obtain a multivariate time series matrix. The multi-source heterogeneous meteorological data includes ground observation sequences of low-altitude flight areas and three-dimensional real-time wind field data. The ground observation sequences include temperature, relative humidity, air pressure, horizontal visibility, and aerosol mass concentration. The three-dimensional real-time wind field data includes horizontal zonal wind, horizontal meridional wind, and vertical wind.

[0007] The multivariate time series matrix is ​​input into the damper control graph attention network to obtain the advection spatial feature matrix. The damper control graph attention network includes a linear mapping layer, a graph attention mechanism, and an activation function.

[0008] The advection spatial feature matrix and the multivariate time series matrix are fused to obtain the fused multivariate time series matrix;

[0009] The fused multivariate time series matrix is ​​split into multiple independent channels to obtain a one-dimensional time series corresponding to each independent channel; the one-dimensional time series is processed by a Transformer encoder to determine the target time-varying burst feature vector.

[0010] Based on the target time-varying burst feature vector, the estimated values ​​of the dry aerosol extinction coefficient and the moisture absorption growth parameter are determined; based on the estimated values ​​of the dry aerosol extinction coefficient and the moisture absorption growth parameter, visibility is predicted to obtain the visibility prediction result.

[0011] In some embodiments, the preprocessing of the acquired multi-source heterogeneous meteorological data to obtain a multivariate time series matrix includes:

[0012] After mapping the three-dimensional real-time wind field data to the geographic coordinates of each automatic weather station using the Kriging interpolation method, the three-dimensional real-time wind field data and the ground observation sequences obtained by each automatic weather station are concatenated column by column at the same time step to obtain a multivariate time series matrix.

[0013] In some implementations, the step of inputting the multivariate time series matrix into the damper control map attention network to obtain the advection spatial feature matrix includes:

[0014] All automatic weather stations used to acquire ground observation sequences are constructed as a graph neural network, where one node in the graph neural network corresponds to one automatic weather station;

[0015] Obtain the 3D wind field vector for each node and the unit direction vector between two nodes;

[0016] The multivariate time series matrix is ​​reduced in dimensionality through the linear mapping layer, and the initial hidden layer features of each node are calculated.

[0017] The graph attention mechanism is used to calculate the basic similarity between the initial hidden layer features of two nodes;

[0018] Calculate the inner product between the three-dimensional wind field vector and the unit direction vector corresponding to the current node, and input the inner product result into the activation function to obtain the gating factor;

[0019] Multiply the basic similarity and the gating factor of the two nodes to obtain the edge weights of the two nodes;

[0020] A spatial adjacency matrix is ​​constructed using the edge weights, and the initial hidden features of the current node and its neighboring nodes are weighted and summed using the spatial adjacency matrix to obtain the advection spatial feature matrix.

[0021] In some implementations, splitting the fused multivariate time series matrix into multiple independent channels to obtain a one-dimensional time series corresponding to each independent channel includes:

[0022] The fused multivariate time series matrix is ​​split into multiple independent channels according to physical characteristics to obtain a one-dimensional time series corresponding to each independent channel, and the one-dimensional time series corresponds to a physical characteristic.

[0023] In some implementations, the step of processing the one-dimensional time series using a Transformer encoder to determine the target time-varying burst feature vector includes:

[0024] The one-dimensional time series is truncated into multiple overlapping sequence blocks, and each overlapping sequence block is input into the Transformer encoder to obtain the local time-varying burst feature vector corresponding to each overlapping sequence block;

[0025] Flatten and concatenate the local time-varying burst feature vectors corresponding to all overlapping sequence blocks of the one-dimensional time series to obtain a one-dimensional channel feature vector; concatenate the one-dimensional channel feature vectors corresponding to each channel to obtain the target time-varying burst feature vector.

[0026] In some implementations, determining the estimated values ​​of the dry aerosol extinction coefficient and the hygroscopic growth parameter based on the target time-varying burst feature vector includes:

[0027] The target time-varying burst feature vector is input into the fully connected layer to obtain the estimated values ​​of the dry aerosol extinction coefficient and the hygroscopic growth parameter.

[0028] In some embodiments, the visibility prediction based on the estimated dry aerosol extinction coefficient and the estimated hygroscopic growth parameter, to obtain the visibility prediction result, includes:

[0029] ;

[0030] in, This indicates the visibility prediction result. This represents the estimated value of the extinction coefficient of dry aerosols. This represents the predicted relative humidity value. This represents the estimated value of the moisture absorption growth parameter.

[0031] Secondly, embodiments of this application also provide a low-altitude visibility prediction system based on wind field driving, the system comprising:

[0032] The data preprocessing module is used to preprocess the acquired multi-source heterogeneous meteorological data to obtain a multivariate time series matrix. The multi-source heterogeneous meteorological data includes ground observation sequences of low-altitude flight areas and three-dimensional real-time wind field data. The ground observation sequences include temperature, relative humidity, air pressure, horizontal visibility, and aerosol mass concentration. The three-dimensional real-time wind field data includes horizontal zonal wind, horizontal meridional wind, and vertical wind.

[0033] The feature extraction module is used to input the multivariate time series matrix into the damper control graph attention network to obtain the advection space feature matrix. The damper control graph attention network includes a linear mapping layer, a graph attention mechanism, and an activation function.

[0034] The data fusion module is used to fuse the advection spatial feature matrix and the multivariate time series matrix to obtain the fused multivariate time series matrix;

[0035] The data processing module is used to split the fused multivariate time series matrix into multiple independent channels to obtain a one-dimensional time series corresponding to each independent channel; and to process the one-dimensional time series through a Transformer encoder to determine the target time-varying burst feature vector.

[0036] The visibility prediction module is used to determine the estimated values ​​of the dry aerosol extinction coefficient and the moisture absorption growth parameter based on the target time-varying burst feature vector; and to perform visibility prediction based on the estimated values ​​of the dry aerosol extinction coefficient and the moisture absorption growth parameter to obtain the visibility prediction result.

[0037] Thirdly, embodiments of this application also provide an electronic device, including at least one control processor and a memory for communicatively connecting to the at least one control processor; the memory stores instructions executable by the at least one control processor, which, when executed by the at least one control processor, enable the at least one control processor to perform a wind-driven low-altitude visibility prediction method as described above.

[0038] Fourthly, embodiments of this application also provide a computer-readable storage medium storing computer-executable instructions for causing a computer to execute a wind-driven low-altitude visibility prediction method as described above.

[0039] Compared with the prior art, this application has the following beneficial effects:

[0040] This application preprocesses acquired multi-source heterogeneous meteorological data to obtain a multivariate time series matrix. The multi-source heterogeneous meteorological data includes ground observation sequences from low-altitude flight areas and three-dimensional real-time wind field data. The multivariate time series matrix is ​​then input into a wind gate control map attention network to obtain an advection spatial feature matrix. By comprehensively considering the three-dimensional real-time wind field data and the ground observation sequences, the advection spatial feature matrix is ​​extracted, addressing the problem of neglecting the dynamic changes in instantaneous wind speed and direction. This allows for a realistic simulation of the wind-driven advection transport process of water vapor and aerosols. The advection spatial feature matrix and the multivariate time series matrix are fused to obtain a fused multivariate time series matrix. This fused multivariate time series matrix is ​​then split into multiple independent channels, obtaining a one-dimensional time series for each channel. A Transformer encoder is used to process the one-dimensional time series to determine the target time-varying sudden feature vector. This solves the problem of introducing spurious high-frequency noise and easily missing sudden fog signals, significantly improving the detection rate of extreme events. Based on the target time-varying burst feature vector, the estimated values ​​of the dry aerosol extinction coefficient and the hygroscopic growth parameter are determined. Visibility is then predicted based on these estimates, resulting in a visibility prediction. This solves the problem of the model easily getting trapped in local optima and can eliminate distorted predictions that violate the common sense of atmospheric physics, thereby improving the accuracy of low-altitude visibility prediction. Attached Figure Description

[0041] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0042] Figure 1 This is a flowchart illustrating an embodiment of the wind-driven low-altitude visibility prediction method provided in this application.

[0043] Figure 2 This is a schematic diagram of the internal logic of the wind gate control map attention network in the best embodiment of the wind field-driven low-altitude visibility prediction method provided in this application.

[0044] Figure 3 This is a flowchart illustrating the differentiable physical computation layer in the best embodiment of the wind-driven low-altitude visibility prediction method provided in this application.

[0045] Figure 4 This is a schematic diagram of an embodiment of the wind-driven low-altitude visibility prediction system provided in this application;

[0046] Figure 5 This is a schematic diagram of the structure of an embodiment of the electronic device provided in this application. Detailed Implementation

[0047] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0048] In the description of this application, the use of terms such as "first," "second," etc., is for the purpose of distinguishing technical features only and should not be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated or the order of the technical features indicated.

[0049] In the description of this application, it should be understood that the orientation descriptions, such as up, down, etc., are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application.

[0050] In the description of this application, it should be noted that, unless otherwise explicitly defined, terms such as "setup," "installation," and "connection" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this application in conjunction with the specific content of the technical solution.

[0051] To address the issue of low accuracy in low-altitude visibility prediction in related technologies, this application proposes a wind-driven low-altitude visibility prediction method, system, device, and medium.

[0052] Reference Figure 1 This application provides a schematic flowchart of a wind-driven low-altitude visibility prediction method. This wind-driven low-altitude visibility prediction method is applied to an electronic device, which may be a server or a mobile terminal, etc. Figure 1 As shown, the wind-driven low-altitude visibility prediction method may include, but is not limited to, the following steps S101 to S105.

[0053] Step S101: Preprocess the acquired multi-source heterogeneous meteorological data to obtain a multivariate time series matrix. The multi-source heterogeneous meteorological data includes ground observation sequences of low-altitude flight areas and three-dimensional real-time wind field data. The ground observation sequences include temperature, relative humidity, air pressure, horizontal visibility, and aerosol mass concentration. The three-dimensional real-time wind field data includes horizontal zonal wind, horizontal meridional wind, and vertical wind.

[0054] In this step, interpolation is used to map the 3D real-time wind field data to the geographic coordinates of the automatic weather station used to obtain the ground observation sequence. The 3D real-time wind field data and the ground observation sequence are then concatenated column-by-column at the same time step to form a multivariate time series matrix. The multi-source heterogeneous meteorological data includes ground observation sequences from the low-altitude flight area and 3D real-time wind field data. The ground observation sequences include temperature, relative humidity, air pressure, horizontal visibility, and aerosol mass concentration. The 3D real-time wind field data includes horizontal zonal winds, horizontal meridional winds, and vertical winds.

[0055] By comprehensively considering three-dimensional real-time wind field data and ground observation sequences, a solid data foundation can be laid for the subsequent extraction of the advection spatial feature matrix.

[0056] The interpolation method described above can be Kriging interpolation or other interpolation methods known to those skilled in the art. This embodiment does not specifically limit the method.

[0057] The aforementioned acquisition of multi-source heterogeneous meteorological data can be achieved through equipment such as automatic weather stations, radar, and satellites.

[0058] Step S102: Input the multivariate time series matrix into the damper control graph attention network to obtain the advection spatial feature matrix. The damper control graph attention network includes a linear mapping layer, a graph attention mechanism, and an activation function.

[0059] In this step, the multivariate time series matrix is ​​input into the wind gate control graph attention network. First, the automatic weather station is used as a node to construct a graph neural network, and the multivariate time series matrix is ​​reduced in dimensionality through a linear mapping layer to calculate the initial hidden features of each node. Then, the basic similarity between the initial hidden features of two nodes is calculated through the graph attention mechanism. Next, the three-dimensional wind field vector and unit direction vector corresponding to the current node are obtained through activation function processing to obtain the gate factor. Finally, the basic similarity and the gate factor are multiplied to obtain the edge weight between the two nodes, and a dynamic spatial adjacency matrix is ​​constructed based on the edge weight. The initial hidden features of the current node and its surrounding nodes (i.e., neighboring nodes) are weighted and summed according to the dynamic spatial adjacency matrix to obtain the advection spatial feature matrix.

[0060] By considering the multivariate time series matrix of three-dimensional real-time wind field data, the advection spatial feature matrix is ​​extracted, overcoming the shortcomings of traditional static graph neural networks (GNNs) that rely on fixed edge weights based on Euclidean distance. This approach can realistically simulate the advection transport process of water vapor and aerosols dominated by the wind field, laying a solid data foundation for accurate low-altitude visibility prediction in the later stages.

[0061] Step S103: Merge the advection spatial feature matrix and the multivariate time series matrix to obtain the fused multivariate time series matrix.

[0062] In this step, the advection spatial feature matrix and the multivariate time series matrix are concatenated and fused to obtain the fused multivariate time series matrix. This data fusion enriches the data information and lays a solid data foundation for extracting accurate target time-varying burst feature vectors in subsequent steps.

[0063] Step S104: Split the fused multivariate time series matrix into multiple independent channels to obtain a one-dimensional time series corresponding to each independent channel; process the one-dimensional time series using a Transformer encoder to determine the target time-varying burst feature vector.

[0064] In this step, the fused multivariate time series matrix is ​​split into multiple independent channels according to physical characteristics, obtaining a one-dimensional time series for each independent channel. Then, the one-dimensional time series is truncated into multiple overlapping sequence blocks. A Transformer encoder is used to process each overlapping sequence block to obtain a local time-varying burst feature vector corresponding to each overlapping sequence block. These local time-varying burst feature vectors are flattened and concatenated to obtain a one-dimensional channel feature vector. Finally, the one-dimensional channel feature vectors corresponding to all channels are concatenated to obtain the target time-varying burst feature vector.

[0065] By processing each channel separately and then concatenating the one-dimensional channel feature vectors corresponding to each channel, the target time-varying burst feature vector is obtained. This avoids the numerical pollution of low-frequency humidity evolution features by high-frequency wind speed noise, solves the problem of introducing false high-frequency noise and easily missing sudden fog signals, and can significantly improve the detection rate of extreme events.

[0066] Step S105: Based on the target time-varying burst feature vector, determine the estimated value of the dry aerosol extinction coefficient and the estimated value of the hygroscopic growth parameter; based on the estimated value of the dry aerosol extinction coefficient and the estimated value of the hygroscopic growth parameter, perform visibility prediction to obtain the visibility prediction result.

[0067] In this step, by inputting the target time-varying burst feature vector into the fully connected layer, the estimated values ​​of the dry aerosol extinction coefficient and the moisture absorption growth parameter can be predicted. Then, based on these estimates, visibility is predicted, yielding the following visibility prediction results:

[0068] ;

[0069] in, This indicates the visibility prediction result. This represents the estimated value of the extinction coefficient of dry aerosols. This represents the predicted relative humidity value. This represents the estimated value of the moisture absorption growth parameter.

[0070] By first predicting the estimated values ​​of the dry aerosol extinction coefficient and the hygroscopic growth parameter through a fully connected layer, and then performing visibility prediction, the semantic gap between the physical equation and the deep learning computation graph is bridged. For the first time, the static atmospheric extinction physical equation is reconstructed into a "differentiable deterministic operator" inside the deep neural network, which solves the problem that the model is prone to getting trapped in local optima. It can eliminate distorted predictions that violate the common sense of atmospheric physics, thereby improving the accuracy of low-altitude visibility prediction.

[0071] In some implementations, step S101 above, which involves preprocessing the acquired multi-source heterogeneous meteorological data to obtain a multivariate time series matrix, may include step S201 as shown below.

[0072] Step S201: After mapping the three-dimensional real-time wind field data to the geographic coordinates of each automatic weather station using the Kriging interpolation method, the three-dimensional real-time wind field data and the ground observation sequences obtained by each automatic weather station are concatenated column by column at the same time step to obtain a multivariate time series matrix.

[0073] In this step, surface observation sequences from various automatic weather stations (AWS) within the target area are acquired, including parameters such as temperature, relative humidity, air pressure, horizontal visibility, and aerosol mass concentration. Simultaneously, hourly updated, 1-kilometer resolution low-altitude three-dimensional real-time wind field data (including horizontal zonal wind, horizontal meridional wind, and vertical wind) from numerical weather prediction are obtained. Kriging interpolation is used to map the three-dimensional real-time wind field data to the geographic coordinates of each AWS, and the three-dimensional real-time wind field data and the surface observation sequences acquired from each AWS are concatenated column-wise at the same time step to form a multivariate time series matrix.

[0074] The aforementioned Kriging interpolation method, also known as the spatial autocovariance optimal interpolation method, is a statistical interpolation method used to estimate the attribute values ​​of unknown sample points. It uses regionalized variables as a basis and leverages the variogram to perform linear, unbiased, and optimal estimation of the interpolation points.

[0075] In some implementations, step S102 above, which involves inputting the multivariate time series matrix into the damper control map attention network to obtain the advection spatial feature matrix, may include steps S301 to S307 as shown below.

[0076] Step S301: Construct a graph neural network for all automatic weather stations used to obtain ground observation sequences, where one node in the graph neural network corresponds to one automatic weather station.

[0077] In this step, all automatic weather stations used to acquire ground observation sequences are constructed as a graph neural network, where each node in the graph neural network corresponds to an automatic weather station. If there are... An automatic weather station, a graph neural network has Each node.

[0078] Step S302: Obtain the three-dimensional wind field vector of each node and the unit direction vector between two nodes.

[0079] In this step, the current time is obtained. node Three-dimensional wind field vector at the location and obtaining nodes Pointing to node unit direction vector .

[0080] Step S303: Reduce the dimensionality of the multivariate time series matrix through a linear mapping layer and calculate the initial hidden layer features of each node.

[0081] In this step, the multivariate time series matrix is... Feature dimensionality reduction is performed using a linear embedding layer to calculate the time of each node. Initial hidden layer feature representation .

[0082] Step S304: Calculate the basic similarity between the initial hidden features of two nodes using a graph attention mechanism.

[0083] In this step, the traditional Graph Attention (GAT) mechanism is used to compute nodes. With nodes Each of their initial hidden layer features and basic similarity between .

[0084] Step S305: Calculate the inner product between the three-dimensional wind field vector and the unit direction vector corresponding to the current node, and input the inner product result into the activation function to obtain the gating factor.

[0085] In this step, the inner product between the 3D wind field vector and the unit direction vector corresponding to the current node is calculated, and the inner product result is input into the activation function. To obtain the gating factor for: .

[0086] Step S306: Multiply the basic similarity and gating factor of the two nodes to obtain the edge weights of the two nodes.

[0087] In this step, the basic similarity and gating factor corresponding to the two nodes are multiplied together to obtain the edge weights of the two nodes. for: .

[0088] Step S307: Construct a spatial adjacency matrix using edge weights, and then use the spatial adjacency matrix to perform a weighted summation of the initial hidden features of the current node and its neighboring nodes to obtain the advection spatial feature matrix.

[0089] In this step, the initial hidden features of the current node and its surrounding nodes (i.e., neighboring nodes) are weighted and summed using the spatial adjacency matrix, which serves as the new feature of the current node. Each node contains the initial hidden features of its surrounding nodes, and the final output is a spatial feature matrix containing upstream air mass advection information. (i.e., the advection space characteristic matrix).

[0090] In some implementations, step S104 above, which involves splitting the fused multivariate time series matrix into multiple independent channels to obtain a one-dimensional time series corresponding to each independent channel, may include step S401 as shown below.

[0091] Step S401: The fused multivariate time series matrix is ​​split into multiple independent channels according to physical characteristics to obtain a one-dimensional time series corresponding to each independent channel. The one-dimensional time series corresponds to a physical characteristic.

[0092] In this step, a channel-independent mechanism is employed to split the fused multivariate time series matrix into independent one-dimensional channels (i.e., one-dimensional time series corresponding to each independent channel) based on physical characteristics. For example, sequences containing only the humidity channel or sequences containing only the wind speed channel. This avoids numerical contamination of low-frequency humidity evolution characteristics by high-frequency wind speed noise.

[0093] In some implementations, step S104 above, which involves processing the one-dimensional time series using a Transformer encoder to determine the target time-varying burst feature vector, may include steps S501 to S502 as shown below.

[0094] Step S501: Truncate the one-dimensional time series into multiple overlapping sequence blocks, and input each overlapping sequence block into the Transformer encoder to obtain the local time-varying burst feature vector corresponding to each overlapping sequence block.

[0095] In this step, the length is The one-dimensional time series is truncated into overlapping sequence blocks (Patch) of length P by sliding with step size S. Each sequence block is mapped to an independent token and fed into the Transformer encoder for self-attention calculation. The Transformer encoder outputs a local time-varying burst feature vector corresponding to each overlapping sequence block.

[0096] Step S502: Flatten and stitch together the local time-varying burst feature vectors corresponding to all overlapping sequence blocks of the one-dimensional time series to obtain the one-dimensional channel feature vector; stitch together the one-dimensional channel feature vectors corresponding to each channel to obtain the target time-varying burst feature vector.

[0097] In this step, the local time-varying burst feature vectors corresponding to multiple overlapping sequence blocks belonging to the same channel are flattened and spliced ​​together to form a one-dimensional channel feature vector representing that channel. Then, the one-dimensional channel feature vectors of all channels are spliced ​​and fused to output the final time-varying burst feature vector (i.e., the target time-varying burst feature vector).

[0098] By truncating a one-dimensional time series into multiple overlapping sequence blocks, it naturally plays a role in local smoothing and resisting high-frequency noise within the neural network, replacing external VMD decomposition. Moreover, when encountering events such as sudden drops in visibility caused by advection fog, it can encode the entire "sudden drop waveform" as a complete semantic unit, significantly improving the detection rate of extreme events.

[0099] In some implementations, step S105 above, which determines the estimated value of the dry aerosol extinction coefficient and the estimated value of the hygroscopic growth parameter based on the target time-varying burst feature vector, may include step S601 as shown below.

[0100] Step S601: Input the target time-varying burst feature vector into the fully connected layer to obtain the estimated values ​​of the dry aerosol extinction coefficient and the hygroscopic growth parameter.

[0101] In this step, instead of directly outputting visibility prediction results, the estimated values ​​of the dry aerosol extinction coefficient and hygroscopic growth parameter are first predicted through a fully connected layer. This bridges the semantic gap between the physical equation and the deep learning computation graph, and for the first time reconstructs the static atmospheric extinction physical equation into a "differentiable deterministic operator" within the deep neural network. By transforming Koschmieder's law into a rigid computation layer that does not contain any learnable weight parameters but allows continuous gradient propagation, and directly concatenating it at the end of the feature output of the temporal encoder, the gradient of the prediction error can be stably propagated back through this deterministic operator following the chain rule during the backpropagation of model training.

[0102] In some implementations, visibility prediction is performed based on estimates of the dry aerosol extinction coefficient and the hygroscopic growth parameter to obtain visibility prediction results, which may include:

[0103] ;

[0104] in, This indicates the visibility prediction result. This represents the estimated value of the extinction coefficient of dry aerosols. This represents the predicted relative humidity value. This represents the estimated value of the moisture absorption growth parameter.

[0105] In this embodiment, the above formula design forcibly restricts the model in... At extremely high altitudes, low visibility must be output, thus eliminating distorted predictions that violate the common sense of atmospheric physics.

[0106] To facilitate understanding by those skilled in the art, a set of preferred embodiments is provided below:

[0107] Existing deep learning-based meteorological sequence forecasting schemes have the following three main substantive defects in actual aviation meteorological support operations:

[0108] (1) Feature extraction of non-stationary abrupt change signals suffers from endpoint distortion and loss of anomalous signals.

[0109] On the one hand, existing signal decomposition methods such as VMD inevitably introduce endpoint effects (i.e., Gibbs artifacts) during global integration operations when processing real-time streaming truncated data, introducing false high-frequency noise. Furthermore, the iterative solution process incurs significant computational delays, making it difficult to meet the high-frequency real-time simulation requirements for low-altitude flight. On the other hand, the sparse attention mechanism employed by conventional Informer models to reduce computational overhead tends to filter out seemingly irregular "outliers" in the time series. However, in micrometeorology, these outliers are often key precursors to sudden drops in local visibility (such as advection fog intrusion). Misjudgment by this mechanism leads to a persistently high rate of missed detections for sudden severe weather events.

[0110] (2) Static spatial maps cannot characterize material transport processes dominated by fluid mechanics.

[0111] Most existing spatial network models rely on the absolute geographical distances of meteorological stations to construct static adjacency matrices. However, the formation and evolution of low-level advection fog is essentially a dynamic transport process of water vapor and aerosols driven by a three-dimensional wind field. The static graph structure solidifies the information transmission paths between nodes, completely ignoring the dynamic changes in instantaneous wind speed and direction. Under non-prevailing wind conditions, static models are prone to aggregating a large amount of invalid downstream or crosswind node data, causing spatial feature extraction to completely deviate from the actual weather situation.

[0112] (3) Physical information networks based on soft constraints cannot guarantee the rationality of predictions under extreme conditions.

[0113] Existing physical fusion schemes generally employ the approach of adding a penalty term to the physical equations (i.e., soft constraints) to the loss function. This heuristic method has inherent contradictions in multi-objective parameter optimization, and the model is prone to getting trapped in local optima. When extreme weather combinations that rarely occur in historical samples appear in the test set (such as high humidity, low pressure, extremely weak winds, and high aerosol concentration), the soft constraints often fail, leading to physical distortions in the model's final output, such as "high visibility even under near-saturated high humidity conditions," which is unacceptable in the field of aviation safety.

[0114] To address the aforementioned shortcomings of existing technologies, this embodiment provides a low-altitude visibility prediction method based on wind field-driven dynamic graphs and physical constraints. This embodiment aims to solve the problem of missing physical laws in weather forecasts by purely data-driven models, and the insufficient representation of advection transport processes by traditional graph networks. Specifically:

[0115] (1) Provides a temporal feature encoding mechanism based on local slices and channel independence.

[0116] By abandoning global decomposition algorithms and probabilistic sparse attention mechanisms, local sequence slicing and channel-independent mapping are performed on multivariate time-series data. While eliminating endpoint effects and iterative computation delays, the semantics of local fluctuations in low-visibility emergencies are fully preserved, significantly reducing the underreporting rate of extreme weather.

[0117] (2) A dynamic spatial map topology method based on instantaneous wind field is provided.

[0118] By using the three-dimensional real-time wind field vector as a dynamic gating factor for the edge weights of the graph network, the information transmission path between two nodes is activated and strengthened only when the wind field vector satisfies the advection condition of wind blowing from the upstream station to the local station. This method realistically simulates the dynamic advection trajectory of water vapor and aerosols, overcoming the blindness of static graph networks in predicting air mass movement patterns.

[0119] (3) Provides a differentiable hard constraint calculation framework with embedded atmospheric physics laws.

[0120] By converting Koschmieder's law and related extinction evolution relationships in atmospheric optics into differentiable deterministic operators, and then using these operators as rigid computational layers connected in series at the final output of a neural network (i.e., a Transformer encoder), the neural network is only responsible for predicting hidden physical parameters (such as the extinction coefficient of dry aerosols). The final visibility prediction result is rigorously derived from the physical operators, thus completely eliminating distorted predictions that violate the common sense of atmospheric physics in the algorithm mechanism.

[0121] The technical solution of this embodiment specifically includes the following contents:

[0122] Step S11: Acquire multi-source heterogeneous meteorological data of the low-altitude flight area and perform spatiotemporal assimilation preprocessing. Provide the model with standardized input including three-dimensional dynamic features and local micrometeorological features. First, access the meteorological observation network of various meteorological observation stations (referring to the ground observation network composed of automatic weather stations (AWS) distributed in different geographical locations) through the meteorological dedicated network to obtain the ground observation sequences of each AWS in the target area. The elements include: temperature (T), relative humidity (RH), air pressure (P), horizontal visibility (V), and aerosol mass concentration (PM2.5 / PM10). At the same time, acquire the 1-kilometer resolution, hourly updated low-altitude three-dimensional real-time wind field data (including horizontal zonal wind U, horizontal meridional wind V, and vertical wind W) issued by numerical weather prediction. Use Kriging interpolation to map the three-dimensional real-time wind field data to the geographical coordinates of each automatic weather station, and concatenate the three-dimensional real-time wind field data with the original ground observation sequences of each automatic weather station at the same time step to form a multivariate time series matrix. :

[0123] ;

[0124] in, This refers to the number of weather stations (i.e., the number of automatic weather stations). For historical time steps, The total number of meteorological characteristic channels after integrating wind field and ground observations.

[0125] Step S12: Input the preprocessed multivariate time series matrix from Step S11 into the windward control graph attention network to dynamically calculate the spatial adjacency matrix and extract the advection spatial feature matrix between nodes through feature aggregation. This overcomes the shortcomings of traditional static graph neural networks (GNNs) that fix edge weights based on Euclidean distance, and realistically simulates the wind-driven advection transport process of water vapor and aerosols. (Refer to...) Figure 2 The logic is as follows. The specific calculation method includes the following parts:

[0126] All automatic weather stations are constructed as a spatial graph neural network (GNN), with one node representing one automatic weather station.

[0127] Hidden layer feature extraction of nodes: extracting features from multivariate time series matrices. Feature dimensionality reduction is performed using a linear embedding layer to calculate the dimensionality of each node at time step [time value missing]. Initial hidden layer feature representation .

[0128] Basic attention coefficient calculation: Nodes are calculated using the traditional Graph Attention (GAT) mechanism. With nodes Each of their initial hidden layer features and basic similarity between .

[0129] Wind field physical gating calculation: obtaining time node Three-dimensional wind field vector at the location and nodes Pointing to node unit direction vector Calculate the inner product (dot product) of the two and activate it using the Sigmoid function. Generate gating factor .

[0130] Dynamic weight generation and advection feature aggregation: computation nodes With nodes Edge weights between The edge weight is an element in the dynamic spatial adjacency matrix. Then, this spatial adjacency matrix is ​​used to perform a weighted sum of the initial hidden features of the current node and its surrounding nodes (i.e., neighboring nodes), which serves as the new feature of the current node j. Each node contains the initial hidden features of its surrounding nodes, ultimately outputting a spatial feature matrix containing upstream air mass advection information. (i.e., the spatial characteristic matrix of advection), spatial characteristic matrix Each element in can be represented as .

[0131] If and only if the wind direction changes from node Blow towards the node Time (inner product is positive), gating factor When the gating factor approaches 1, the system enables the transmission of features from upstream to downstream; when the wind direction is opposite (the inner product is negative), the gating factor approaches 0, forcibly cutting off pseudo-correlation edges.

[0132] Step S13: The spatial feature matrix obtained in step S12 After being concatenated and fused with the original time-series data (i.e., the multivariate time series matrix), the data is input into a time-series encoder (PatchTST) based on a local slicing and channel independence mechanism to extract local time-varying burst features. This step aims to address the right-side boundary truncation distortion present in conventional variational mode decomposition (VMD) and the shortcomings of the Informer model in sparse sampling, which easily misses burst fogging signals.

[0133] The specific processing procedure of the time-series encoder is as follows: First, a channel-independent mechanism is adopted to split the fused multivariate time series matrix into independent one-dimensional channels (i.e., one-dimensional time series, such as a sequence containing only a humidity channel or a sequence containing only a wind speed channel) according to physical characteristics. Each channel is independently fed into the Transformer encoder. This avoids the numerical contamination of low-frequency humidity evolution characteristics by high-frequency wind speed noise. Within each independent one-dimensional channel, point-by-point attention calculation is not performed on the entire sequence; instead, a length of [missing information] is used. One-dimensional time series with step size Sliding cutoff with length of The overlapping sequence blocks (Patch) are mapped to an independent token and fed into the Transformer encoder for self-attention computation. The Transformer encoder outputs a local time-varying burst feature vector corresponding to each overlapping sequence block. Subsequently, the local time-varying burst feature vectors corresponding to multiple overlapping sequence blocks belonging to the same channel are flattened and concatenated to form a one-dimensional channel feature vector representing each meteorological variable (one channel corresponds to one meteorological variable). Finally, the one-dimensional channel feature vectors of all channels are concatenated and fused to form the final time-varying burst feature vector (i.e., the target time-varying burst feature vector) and output to the fully connected layer.

[0134] Patching operations (i.e., sliding truncation operations) naturally play a role in local smoothing and resisting high-frequency noise within neural networks, replacing external VMD decomposition. Moreover, when encountering events such as sudden drops in visibility caused by advection fog, it can encode the entire "sudden drop waveform" as a complete semantic unit, significantly improving the detection rate of extreme events.

[0135] Step S14: The target time-varying burst feature vector output from Step S13 is connected to the differentiable physics computation layer. Based on the atmospheric extinction equation, hard physical constraints are executed to deduce the final visibility prediction value. This step aims to abandon the existing method of "adding a soft penalty term to the loss function" and achieve hard constraints on physical rules from the fundamental network graph structure. (Refer to...) Figure 3 The logical flow, as shown in the diagram Specifically, it includes:

[0136] Hidden physical parameter mapping: The fully connected layer (connected to the output of the Transformer encoder) receives the target time-varying burst feature vector output in step S13. Instead of directly outputting the visibility prediction result, it predicts two intermediate parameters with physical meaning: the estimated value of the dry aerosol extinction coefficient. and estimated values ​​of moisture absorption growth parameters The extinction coefficient of dry aerosols can be considered as the ability of aerosol particles (i.e., dry aerosols) to attenuate light after moisture has been removed.

[0137] Differentiable physical constraint layer derivation: A non-parametric, fully differentiable Koschmieder computation node is cascaded at the end of the network. This node receives the two hidden parameters mentioned above, as well as the relative humidity forecast provided by an external numerical weather prediction (NWP) product. Perform the following deterministic physical equations:

[0138] ;

[0139] in, This represents the final visibility prediction result output by the network. This design imposes strict limitations on the model's performance. At extremely high altitudes, low visibility must be output, thus eliminating distorted predictions that violate the common sense of atmospheric physics from the algorithm mechanism.

[0140] The prediction network model in this embodiment includes a damper control map attention network, a Transformer encoder, and fully connected layers. During the model training phase, mean squared error (MSE) is used as the loss function for the prediction network model in this embodiment. Specifically, the true horizontal visibility labels at the target time are obtained from historical samples. The predicted visibility output of the differentiable physics computation layer The error is calculated, and the loss function formula is as follows:

[0141] ;

[0142] in, This represents the number of samples in the training batch. Indicates the first One visibility prediction result, Indicates the first A true horizontal visibility label value. During backpropagation, since the Koschmieder physical computation nodes are continuously differentiable operators, the loss function... The error gradient can directly affect and The derivative is calculated and then propagated back to the fully connected layer, the Transformer encoder, and the damper control attention network according to the chain rule. This allows the weight parameters of the entire deep neural network to be jointly updated using real visibility data, thus completing end-to-end closed-loop training.

[0143] During the initialization phase of deep neural network training, the windward control map attention network, Transformer encoder, and fully connected layers can be trained from scratch using either a random initialization strategy or by using pre-trained model weights based on meteorological big data. When using pre-trained weights for initialization, the error gradient calculated using the real visibility labels is used to globally fine-tune the pre-trained weights during end-to-end training combined with differentiable Koschmieder computation layers. This leverages the prior knowledge of general atmospheric evolution laws learned by the pre-trained model, significantly accelerating the convergence speed of the prediction network model in this embodiment at specific airports or low-altitude airspace, while ensuring that the final model parameters still strictly adhere to the dynamic hard constraints of the physical computation layer.

[0144] Compared with the prior art, the method of this embodiment has the following beneficial effects:

[0145] (1) At the level of spatiotemporal feature analysis: The method in this embodiment substantially overcomes the blindness of spatial feature extraction caused by the reliance on static geographical distance in traditional graph networks. By introducing a dynamic gating mechanism driven by three-dimensional wind field vectors, the network can reconstruct the advection transport trajectory of water vapor and aerosols in real time, just like a fluid dynamics model, effectively blocking the pseudo-correlation interference of crosswind or downwind nodes, and significantly improving the accuracy of advection fog source tracing. At the same time, the method in this embodiment replaces the existing VMD decomposition and Informer architecture with channel independence and sequence slicing mechanism, which not only avoids the endpoint fly-out effect and high computational delay caused by VMD when processing real-time truncated data, but also solves the defect that the Informer sparse attention mechanism is prone to filtering out sudden fog precursors as noise. The method in this embodiment greatly reduces the false negative rate of low-altitude extreme weather while ensuring millisecond-level inference timeliness.

[0146] (2) At the level of physical reliability of the prediction results:

[0147] Although Koschmieder's law and related extinction evolution relationships in atmospheric optics have been widely proposed and applied in existing physical and meteorological measurement techniques, existing technologies only treat them as static post-processing formulas independent of prediction models, which cannot solve the underlying defect of pure data-driven deep learning models outputting physically distorted results under extreme weather conditions.

[0148] This embodiment overcomes the optimization limitations of conventional Physical Information Neural Networks (PINNs), which heavily rely on the "soft constraints" of the loss function. It bridges the semantic gap between physical equations and deep learning computational graphs, reconstructing the static atmospheric extinction physical equations into a "differentiable deterministic operator" within the deep neural network for the first time. Specifically, this embodiment transforms Koschmieder's law into a rigid computational layer that does not contain any learnable weight parameters but allows for continuous gradient propagation, and directly concatenates it to the end of the feature output of the temporal encoder. During the backpropagation of model training, the gradient of the prediction error can be stably propagated back through this deterministic operator following a chain rule.

[0149] This architectural innovation forces physical laws to substantially intervene in the forward inference process. When faced with unseen extreme weather samples (such as extremely high humidity and calm winds), the model output is constrained by the underlying physical operator mechanism of the network, fundamentally eliminating the predictive paradoxes of "high humidity and high visibility" that may arise from simply relying on data fitting, which violate common sense in atmospheric science. This is not a conventional application of existing physical formulas, but a substantial improvement to the underlying network architecture in the interdisciplinary field of meteorological artificial intelligence, achieving unexpected technical results and providing a meteorological foundation with extremely high robustness and absolute physical consistency for low-altitude flight safety.

[0150] Reference Figure 4 This application also provides a wind-driven low-altitude visibility prediction system, which includes:

[0151] The data preprocessing module 100 is used to preprocess the acquired multi-source heterogeneous meteorological data to obtain a multivariate time series matrix. The multi-source heterogeneous meteorological data includes ground observation sequences of low-altitude flight areas and three-dimensional real-time wind field data. The ground observation sequences include temperature, relative humidity, air pressure, horizontal visibility, and aerosol mass concentration. The three-dimensional real-time wind field data includes horizontal zonal wind, horizontal meridional wind, and vertical wind.

[0152] The feature extraction module 200 is used to input the multivariate time series matrix into the damper control graph attention network to obtain the advection space feature matrix. The damper control graph attention network includes a linear mapping layer, a graph attention mechanism, and an activation function.

[0153] The data fusion module 300 is used to fuse the advection spatial feature matrix and the multivariate time series matrix to obtain the fused multivariate time series matrix.

[0154] The data processing module 400 is used to split the fused multivariate time series matrix into multiple independent channels to obtain a one-dimensional time series corresponding to each independent channel; and to process the one-dimensional time series through a Transformer encoder to determine the target time-varying burst feature vector.

[0155] The visibility prediction module 500 is used to determine the estimated values ​​of the dry aerosol extinction coefficient and the moisture absorption growth parameter based on the target time-varying burst feature vector; and to predict visibility based on the estimated values ​​of the dry aerosol extinction coefficient and the moisture absorption growth parameter to obtain the visibility prediction result.

[0156] It should be noted that since the low-altitude visibility prediction system based on wind field drive in this embodiment is based on the same inventive concept as the low-altitude visibility prediction method based on wind field drive described above, the corresponding content in the method embodiment is also applicable to this system embodiment, and will not be described in detail here.

[0157] Reference Figure 5 This application also provides an electronic device, which includes:

[0158] At least one memory;

[0159] At least one processor;

[0160] At least one program;

[0161] The program is stored in memory, and the processor executes at least one program to implement the wind-driven low-altitude visibility prediction method described above in this disclosure.

[0162] This electronic device can be any smart terminal, including mobile phones, tablets, personal digital assistants (PDAs), and in-vehicle computers.

[0163] The electronic devices according to embodiments of this application will now be described in detail.

[0164] The processor 1600 can be implemented using a general-purpose central processing unit (CPU), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this disclosure.

[0165] The memory 1700 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 1700 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1700 and is called and executed by the processor 1600 to execute the wind farm-driven low-altitude visibility prediction method of the embodiments of this disclosure.

[0166] The input / output interface 1800 is used to implement information input and output.

[0167] The communication interface 1900 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0168] Bus 2000 transmits information between various components of the device (e.g., processor 1600, memory 1700, input / output interface 1800, and communication interface 1900);

[0169] The processor 1600, memory 1700, input / output interface 1800 and communication interface 1900 are connected to each other within the device via bus 2000.

[0170] This disclosure also provides a storage medium, which is a computer-readable storage medium storing computer-executable instructions for causing a computer to execute the above-described wind-driven low-altitude visibility prediction method.

[0171] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0172] The embodiments described in this disclosure are for the purpose of more clearly illustrating the technical solutions of this disclosure and do not constitute a limitation on the technical solutions provided by this disclosure. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by this disclosure are also applicable to similar technical problems.

[0173] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this disclosure, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0174] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0175] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0176] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0177] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0178] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0179] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0180] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0181] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks. The embodiments of this application have been described in detail above with reference to the accompanying drawings, but this application is not limited to the above embodiments. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of this application.

[0182] The embodiments of this application have been described in detail above with reference to the accompanying drawings. However, this application is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of this application.

Claims

1. A low-altitude visibility prediction method based on wind field drive, characterized in that, The method includes: The acquired multi-source heterogeneous meteorological data are preprocessed to obtain a multivariate time series matrix. The multi-source heterogeneous meteorological data includes ground observation sequences of low-altitude flight areas and three-dimensional real-time wind field data. The ground observation sequences include temperature, relative humidity, air pressure, horizontal visibility, and aerosol mass concentration. The three-dimensional real-time wind field data includes horizontal zonal wind, horizontal meridional wind, and vertical wind. The multivariate time series matrix is ​​input into the damper-controlled graph attention network to obtain the advection spatial feature matrix. The damper-controlled graph attention network includes a linear mapping layer, a graph attention mechanism, and an activation function, including: All automatic weather stations used to acquire ground observation sequences are constructed as a graph neural network, where one node in the graph neural network corresponds to one automatic weather station; Obtain the 3D wind field vector for each node and the unit direction vector between two nodes; The multivariate time series matrix is ​​reduced in dimensionality through the linear mapping layer, and the initial hidden layer features of each node are calculated. The graph attention mechanism is used to calculate the basic similarity between the initial hidden layer features of two nodes; Calculate the inner product between the three-dimensional wind field vector and the unit direction vector corresponding to the current node, and input the inner product result into the activation function to obtain the gating factor; Multiply the basic similarity and the gating factor of the two nodes to obtain the edge weights of the two nodes; A spatial adjacency matrix is ​​constructed using the edge weights, and the initial hidden features of the current node and its neighboring nodes are weighted and summed using the spatial adjacency matrix to obtain the advection spatial feature matrix. The advection spatial feature matrix and the multivariate time series matrix are fused to obtain the fused multivariate time series matrix; The fused multivariate time series matrix is ​​split into multiple independent channels to obtain a one-dimensional time series corresponding to each independent channel; the one-dimensional time series is processed by a Transformer encoder to determine the target time-varying burst feature vector. Based on the target time-varying burst feature vector, the estimated values ​​of the dry aerosol extinction coefficient and the moisture absorption growth parameter are determined; based on the estimated values ​​of the dry aerosol extinction coefficient and the moisture absorption growth parameter, visibility is predicted to obtain the visibility prediction result.

2. The low-altitude visibility prediction method based on wind field drive according to claim 1, characterized in that, The preprocessing of the acquired multi-source heterogeneous meteorological data to obtain a multivariate time series matrix includes: After mapping the three-dimensional real-time wind field data to the geographic coordinates of each automatic weather station using the Kriging interpolation method, the three-dimensional real-time wind field data and the ground observation sequences obtained by each automatic weather station are concatenated column by column at the same time step to obtain a multivariate time series matrix.

3. The low-altitude visibility prediction method based on wind field drive according to claim 1, characterized in that, The step of splitting the fused multivariate time series matrix into multiple independent channels to obtain a one-dimensional time series corresponding to each independent channel includes: The fused multivariate time series matrix is ​​split into multiple independent channels according to physical characteristics to obtain a one-dimensional time series corresponding to each independent channel, and the one-dimensional time series corresponds to a physical characteristic.

4. The low-altitude visibility prediction method based on wind field drive according to claim 1, characterized in that, The step of processing the one-dimensional time series using a Transformer encoder to determine the target time-varying burst feature vector includes: The one-dimensional time series is truncated into multiple overlapping sequence blocks, and each overlapping sequence block is input into the Transformer encoder to obtain the local time-varying burst feature vector corresponding to each overlapping sequence block; Flatten and concatenate the local time-varying burst feature vectors corresponding to all overlapping sequence blocks of the one-dimensional time series to obtain a one-dimensional channel feature vector; concatenate the one-dimensional channel feature vectors corresponding to each channel to obtain the target time-varying burst feature vector.

5. The low-altitude visibility prediction method based on wind field drive according to claim 1, characterized in that, The step of determining the estimated values ​​of the dry aerosol extinction coefficient and the estimated values ​​of the hygroscopic growth parameter based on the target time-varying burst feature vector includes: The target time-varying burst feature vector is input into the fully connected layer to obtain the estimated values ​​of the dry aerosol extinction coefficient and the hygroscopic growth parameter.

6. The low-altitude visibility prediction method based on wind field drive according to claim 1, characterized in that, The visibility prediction based on the estimated dry aerosol extinction coefficient and the estimated hygroscopic growth parameter yields the following results: ； in, This indicates the visibility prediction result. This represents the estimated value of the extinction coefficient of dry aerosols. This represents the predicted relative humidity value provided by external numerical weather prediction products. This represents the estimated value of the moisture absorption growth parameter.

7. A low-altitude visibility prediction system based on wind field drive, characterized in that, The system includes: The data preprocessing module is used to preprocess the acquired multi-source heterogeneous meteorological data to obtain a multivariate time series matrix. The multi-source heterogeneous meteorological data includes ground observation sequences of low-altitude flight areas and three-dimensional real-time wind field data. The ground observation sequences include temperature, relative humidity, air pressure, horizontal visibility, and aerosol mass concentration. The three-dimensional real-time wind field data includes horizontal zonal wind, horizontal meridional wind, and vertical wind. The feature extraction module is used to input the multivariate time series matrix into the damper-controlled graph attention network to obtain the advection spatial feature matrix. The damper-controlled graph attention network includes a linear mapping layer, a graph attention mechanism, and an activation function, including: All automatic weather stations used to acquire ground observation sequences are constructed as a graph neural network, where one node in the graph neural network corresponds to one automatic weather station; Obtain the 3D wind field vector for each node and the unit direction vector between two nodes; The multivariate time series matrix is ​​reduced in dimensionality through the linear mapping layer, and the initial hidden layer features of each node are calculated. The graph attention mechanism is used to calculate the basic similarity between the initial hidden layer features of two nodes; Calculate the inner product between the three-dimensional wind field vector and the unit direction vector corresponding to the current node, and input the inner product result into the activation function to obtain the gating factor; Multiply the basic similarity and the gating factor of the two nodes to obtain the edge weights of the two nodes; A spatial adjacency matrix is ​​constructed using the edge weights, and the initial hidden features of the current node and its neighboring nodes are weighted and summed using the spatial adjacency matrix to obtain the advection spatial feature matrix. The data fusion module is used to fuse the advection spatial feature matrix and the multivariate time series matrix to obtain the fused multivariate time series matrix; The data processing module is used to split the fused multivariate time series matrix into multiple independent channels to obtain a one-dimensional time series corresponding to each independent channel; and to process the one-dimensional time series through a Transformer encoder to determine the target time-varying burst feature vector. The visibility prediction module is used to determine the estimated values ​​of the dry aerosol extinction coefficient and the moisture absorption growth parameter based on the target time-varying burst feature vector; and to perform visibility prediction based on the estimated values ​​of the dry aerosol extinction coefficient and the moisture absorption growth parameter to obtain the visibility prediction result.

8. An electronic device, characterized in that, It includes at least one control processor and a memory for communicatively connecting to the at least one control processor; the memory stores instructions executable by the at least one control processor, which, when executed by the at least one control processor, enable the at least one control processor to perform the wind-driven low-altitude visibility prediction method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions for causing a computer to perform the wind-driven low-altitude visibility prediction method as described in any one of claims 1 to 6.

Images