A method for identifying surface weather patterns based on sequence models
By constructing radial scanning feature extraction in a polar coordinate system, the two-dimensional pressure field is converted into a one-dimensional sequence. This is coupled with the pressure gradient field and azimuth relationship, which solves the problem of insufficient surface pressure field feature extraction in existing technologies and achieves efficient weather pattern recognition and improved model interpretability.
Patent Information
- Application Number
- CN202611330810.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-31
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies struggle to effectively extract key features of the ground pressure field, resulting in insufficient accuracy in weather pattern identification and high computational complexity, making it difficult to meet real-time and resource efficiency requirements.
A sequence model-based method for identifying surface weather patterns is adopted. By constructing radial scanning feature extraction in a polar coordinate system, the two-dimensional pressure field is converted into a one-dimensional sequence. The pressure gradient field and azimuth relationship are coupled together, and a sequence encoder is used for feature extraction and classification.
It improves the accuracy and physical interpretability of weather pattern identification, reduces computational complexity, and enhances the model's generalization ability and physical rationality across different regions.
Smart Images

Figure CN123069078A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of meteorological data processing technology, and in particular to a method for identifying surface weather patterns based on sequence models. Background Technology
[0002] Surface weather pattern identification is a fundamental task in weather forecasting. It involves identifying the spatial distribution patterns of surface pressure fields to determine the type of current weather system, such as the leading edge of a high-pressure system, the trailing edge of a high-pressure system, the leading edge of a low-pressure system, the trailing edge of a low-pressure system, and isobaric fields. Traditional weather pattern identification methods mainly rely on manual analysis of weather maps and isobaric lines, which is highly subjective and inefficient. With the development of deep learning technology, image recognition methods based on convolutional neural networks have been applied to weather pattern identification, processing pressure fields as two-dimensional images. While this improves automation, it suffers from high computational complexity and large memory consumption, making it difficult to meet the real-time and resource efficiency requirements of operational applications.
[0003] Compared to upper-level pressure fields, surface pressure fields are more susceptible to the influence of topography and local circulation, resulting in a more complex pressure distribution. Traditional processing methods based on overall pressure field images struggle to efficiently extract key features. Furthermore, existing convolutional neural network-based methods, when processing large-scale pressure field data, require dense sampling and convolution operations across the entire two-dimensional pressure field. This leads to a large number of model parameters, slow inference speed, and a lack of targeted modeling of the physical properties of the pressure field, affecting the model's interpretability and generalization ability across different regions.
[0004] Chinese patent CN114594532A discloses a method for predicting cold wave weather. This method acquires temperature data from multiple observation stations within a target area at multiple preset time points, generates a regional average temperature based on the temperature data, and then generates a time-temperature feature curve, which is used to identify cold wave weather. However, it still uses a standard two-dimensional convolutional network to process meteorological features, failing to specifically model the spatial structure characteristics and physical laws of the pressure field. When dealing with complex weather conditions, feature extraction is insufficient, affecting the recognition accuracy and the interpretability of the model. Summary of the Invention
[0005] In view of this, the present invention provides a ground weather pattern identification method based on a sequence model, which solves the technical problems of the prior art that fails to effectively extract features based on the spatial structure characteristics of the pressure field and has insufficient accuracy in identifying complex weather patterns. By extracting features through radial scanning, the two-dimensional pressure field is converted into a one-dimensional sequence, which improves the accuracy and physical interpretability of weather pattern identification, while reducing computational complexity.
[0006] The technical solution of this invention is implemented as follows: This invention provides a method for identifying ground weather patterns based on sequence models, comprising the following steps: S1. Determine the center position of the target area, establish a polar coordinate system with the center position of the target area as the origin, and set up multiple radial layers and multiple angular directions in the polar coordinate system to form a sampling position array; S2. Obtain meteorological grid data of the target area, interpolate the air pressure values of each sampling location in the sampling location array to obtain the air pressure values of each sampling location, and form an air pressure feature matrix. S3. Organize the pressure feature matrix into sequential data according to the angular direction, where each position in the sequential data corresponds to each angular direction, and the feature of each position is composed of multiple radial layer pressure values in that angular direction; by mapping the features of each position to the feature space and adding position encoding to the mapped features, an encoded sequence is obtained. S4. Input the encoded sequence into the sequence encoder for feature extraction to obtain encoded features; the sequence encoder includes multiple coding layers, and each coding layer establishes the correlation between different positions in the sequence through an attention mechanism. S5. Aggregate the encoded features to obtain a global feature representation, and perform classification prediction based on the global feature representation to obtain the weather situation category.
[0007] Based on the above technical solutions, preferably, the angle of the polar coordinate system increases clockwise from zero degrees in the due north direction to 360 degrees, and the radius of the polar coordinate system represents the geographical distance from the center position.
[0008] Based on the above technical solutions, preferably, step S2 specifically includes: Acquire meteorological grid data for the target area, and use the finite difference method to calculate the pressure gradient modulus of each meteorological grid point in the meteorological grid data. The polar coordinates of each sampling location in the sampling location array are converted to geographic coordinates, and multiple meteorological grid points around the sampling location are selected from the meteorological grid data as interpolation grid points. Calculate the spherical distance and the azimuth angle of the line connecting each sampling position and each interpolation grid point; calculate the directional deviation angle between the radial direction of each sampling position and the azimuth angle of the line connecting the points. The interpolation weights of each interpolation grid point are constructed based on the spherical distance, the direction deviation angle, and the pressure gradient modulus. The pressure value at each sampling location is calculated using a weighted average method. The air pressure values at each sampling location are organized into an air pressure feature matrix according to radial layers and angular directions.
[0009] Based on the above technical solutions, preferably, the formula for calculating the interpolation weight is as follows: ; ; in, For the first Interpolation weights for each interpolation grid point For sampling position and the first spherical distance between interpolation grid points This is the directional deviation angle between the radial direction of the sampling location and the azimuth angle of the connecting line. This is the orientation alignment factor. For the first Pressure gradient modulus at each interpolation grid point The maximum pressure gradient modulus across all interpolation grid points. It is a tiny constant. The directional modulation intensity coefficient, For sampling position to the th The azimuth angle of the line connecting the interpolation grid points The radial angle of the sampling position.
[0010] Based on the above technical solutions, preferably, step S3 specifically includes: The pressure feature matrix is transposed into sequence data with angular direction as the sequence position, and the radial pressure features of each sequence position are mapped from the radial layer dimension to the hidden feature dimension. Calculate the radial energy center based on the air pressure values of each radial layer in the angular direction corresponding to the sequence position; The rotational position encoding frequency parameters of the sequence positions are adaptively adjusted based on the radial energy center of each sequence position to obtain the adjusted frequency parameters. The feature vectors at each sequence position are rotated based on the adjusted frequency parameters to obtain the encoded sequence containing positional information.
[0011] Based on the above technical solution, preferably, the radial energy center represents the radial distribution location of the air pressure anomaly in that angular direction, and the calculation formula is: ; in, For sequence position radial energy center, The radial layer number, For the first The radius of each radial layer For the first The radial layers are located in the sequence position. The normalized pressure value at that location.
[0012] Based on the above technical solutions, preferably, the formula for calculating the adjusted frequency parameter is as follows: ; ; in, For sequence position The adjusted frequency parameters, For standard frequency parameters, As the base, Indexed by feature dimensions, To hide the feature dimension, Radial modulation intensity coefficient, and These are the minimum and maximum radii for the radial scan, respectively.
[0013] Based on the above technical solutions, preferably, the application of rotation transformation based on the adjusted frequency parameters to the feature vectors at each sequence position specifically includes: Pair the feature vectors at each sequence position along multiple dimensions, and construct a rotation matrix for each pair of dimensions. The calculation formula is as follows: ; in, For sequence position The eigenvector after rotation transformation One portion, For sequence position The eigenvector after rotation transformation One portion, For sequence position The angle value is converted to radians. For sequence position The adjusted frequency parameters, and The first feature vector is the first feature vector. and One component; Construct rotation matrices for all dimensions to obtain the rotated encoded feature vectors.
[0014] Based on the above technical solutions, preferably, step S4 specifically includes: The radial average gradient intensity at the sequence position is calculated based on the pressure difference between adjacent radial layers in the angular direction corresponding to the sequence position. An attention bias matrix is constructed based on the radial average gradient intensity at each sequence position and the angular interval between sequence positions; In each coding layer of the sequence encoder, the input features of the coding layer are mapped to a query matrix, a key matrix, and a value matrix through linear transformation. The attention score matrix is obtained by calculating the matrix product of the query matrix and the key matrix and then scaling it. The scaled attention score matrix is added to the corresponding bias value in the attention bias matrix and then normalized to obtain the attention weight matrix. The attention output is obtained by weighted aggregation of the value matrix based on the attention weight matrix. The attention outputs of multiple attention heads are concatenated and then linearly transformed to obtain multi-head attention output. The multi-head attention output is then residually connected with the input features of the coding layer and layer normalized to obtain normalized features. The normalized features are nonlinearly transformed through a feedforward neural network to obtain the output of the feedforward network. The output of the feedforward network is then residually connected with the normalized features and layer normalized to obtain the output features of the coding layer.
[0015] More preferably, the formula for calculating the corresponding bias value in the attention bias matrix is: ; ; in, For sequence position and sequence position Attention bias between The bias intensity coefficient, and Sequence position and sequence position The radial average gradient intensity, The geometric mean of the positional gradient intensities of the two sequences is given. It is the minimum angular difference between the positions of the two sequences. and Sequence position and sequence position The corresponding angle and direction is the standard deviation parameter for angular proximity.
[0016] The present invention has the following advantages over the prior art: (1) By establishing a polar coordinate system with the center of the target area as the origin, and forming a sampling position array through multiple radial layers and angular directions, the two-dimensional pressure field is converted into a one-dimensional sequence representation, which effectively reduces the data dimension and computational complexity, improves the accuracy and physical interpretability of weather pattern identification, and reduces computational complexity while ensuring identification accuracy. (2) By coupling the pressure gradient field with the orientation relationship, the interpolation weights can be adaptively adjusted according to the physical structure of the actual pressure field, which improves the accuracy and physical rationality of feature extraction. (3) By automatically adjusting the frequency parameters of the location coding according to the radial energy distribution of each sequence location, the physical characteristics of the pressure field are integrated into the location coding mechanism, which enhances the model's ability to express weather systems at different scales. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 This is a flowchart of a ground weather pattern identification method based on a sequence model according to the present invention; Figure 2 This is a schematic diagram of polar coordinate sampling for a ground weather pattern identification method based on a sequence model according to the present invention; Figure 3 This is a flowchart of the sequence encoding process for a ground weather pattern identification method based on a sequence model according to the present invention. Figure 4 This is a block diagram of the sequence encoder structure of a ground weather pattern recognition method based on a sequence model according to the present invention. Detailed Implementation
[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0020] like Figure 1 and Figure 2 As shown, this invention provides a method for identifying ground weather patterns based on sequence models, comprising the following steps: S1. Determine the center position of the target area, establish a polar coordinate system with the center position of the target area as the origin, and set multiple radial layers and multiple angular directions in the polar coordinate system to form a sampling position array; wherein, the angle of the polar coordinate system is 0 degrees north and increases clockwise to 360 degrees, and the radius of the polar coordinate system represents the geographical distance from the center position.
[0021] Understandably, the angle values in polar coordinates range from 0 degrees to 360 degrees.
[0022] In one embodiment of the present invention, the angular resolution is 1 degree, that is, the number of angular sampling points. The angle sequence is Degree, among which .
[0023] This invention employs a polar coordinate system established with the center of the target region as the origin. By constructing a sampling position array through multiple radial layers and angular directions, the two-dimensional pressure field is converted into a one-dimensional sequence representation. This effectively reduces the data dimensionality and computational complexity, improves the accuracy and physical interpretability of weather pattern recognition, and reduces computational complexity while ensuring recognition accuracy. Compared with traditional convolutional neural networks that directly process the entire two-dimensional grid field, this invention significantly reduces the number of model parameters and improves inference speed.
[0024] S2. Obtain meteorological grid data of the target area, interpolate the air pressure values of each sampling location in the sampling location array to obtain the air pressure values of each sampling location, and form an air pressure feature matrix.
[0025] Specifically, step S2 includes: Acquire meteorological grid data for the target area, and use the finite difference method to calculate the pressure gradient modulus of each meteorological grid point in the meteorological grid data. The polar coordinates of each sampling location in the sampling location array are converted to geographic coordinates, and multiple meteorological grid points around the sampling location are selected from the meteorological grid data as interpolation grid points. Calculate the spherical distance and the azimuth angle of the line connecting each sampling position and each interpolation grid point; calculate the directional deviation angle between the radial direction of each sampling position and the azimuth angle of the line connecting the points. The interpolation weights of each interpolation grid point are constructed based on the spherical distance, the direction deviation angle, and the pressure gradient modulus. The pressure value at each sampling location is calculated using a weighted average method. The air pressure values at each sampling location are organized into an air pressure feature matrix according to radial layers and angular directions.
[0026] In one embodiment of the present invention, FNL global reanalysis meteorological data at a target time is acquired. This data provides a global sea level pressure field with a spatial resolution of 1 degree or 0.25 degrees. For the target area, pressure grid data within a range of at least 10 degrees around the coverage center is extracted to ensure that all sampling locations within the subsequent radial scan area can obtain pressure values through interpolation from the FNL grid. For locations in the FNL grid field... Grid points Its pressure gradient modulus The approximation calculated using the finite difference method is as follows:
[0027] in This represents the longitude grid spacing of the FNL meteorological grid data. This represents the latitudinal grid spacing of the FNL meteorological grid data. This represents the air pressure value at the grid point adjacent to grid point k to the east. This represents the air pressure value at the grid point adjacent to grid point k to the west. This represents the air pressure value at the grid point adjacent to grid point k to the north. This represents the air pressure value at the grid point adjacent to grid point k to the south.
[0028] For each sampling location in the sampling location array, its polar coordinates are converted to geographic coordinates. Multiple meteorological grid points around the sampling location are selected from the meteorological grid data as interpolation grid points. For each radial scan sampling location... It is necessary to convert its polar coordinates to geographic coordinates. The conversion formula is:
[0029]
[0030] in The kilometer per degree is the degree conversion factor corresponding to the Earth's average radius. and Sampling positions The corresponding latitude and longitude, This represents the radius of the j-th radial layer, i.e., the geographical distance of that radial layer from the center of the target area. This represents the angle value in the i-th angular direction, increasing clockwise from true north (zero degrees). For each sampling position requiring interpolation... The four nearest surrounding grid points are selected from the FNL grid field as interpolation grid points, denoted as . .
[0031] Calculate the spherical distance and azimuth angle of the line connecting each sampling location to each interpolation grid point. Calculate the directional deviation angle between the radial direction of each sampling location and the azimuth angle of the line connecting them. Calculate the spherical distance from each interpolation grid point to the interpolation point. The Haversine formula is used:
[0032] in Kilometers represent the average radius of the Earth. Calculate the azimuth angle of the line connecting the interpolation point to each grid point. This azimuth angle represents the angle from the interpolation point to the grid point. The geographical direction is calculated clockwise with true north as zero degrees:
[0033] Direction deviation angle Indicates the radial direction of the sampling position and the direction of the line The angle between them.
[0034] Furthermore, the formula for calculating the interpolation weights is: ; ; in, For the first Interpolation weights for each interpolation grid point For sampling position and the first spherical distance between interpolation grid points This is the directional deviation angle between the radial direction of the sampling location and the azimuth angle of the connecting line. This is the orientation alignment factor. For the first Pressure gradient modulus at each interpolation grid point The maximum pressure gradient modulus across all interpolation grid points. It is a tiny constant. The directional modulation intensity coefficient, For sampling position to the th The azimuth angle of the line connecting the interpolation grid points The radial angle of the sampling position.
[0035] This invention couples the pressure gradient field with the orientation relationship, enabling the interpolation weights to be adaptively adjusted according to the physical structure of the actual pressure field. This strengthens the weights of interpolation grid points located in the radial direction of the sampling location with larger pressure gradients, while weakening the weights of interpolation grid points located in the tangential direction or with smaller pressure gradients, thereby improving the accuracy and physical rationality of feature extraction.
[0036] In one embodiment of the present invention, the directional modulation intensity coefficient The value is 0.3, a tiny constant. The value is set to 0.01 hPa per degree. The four nearest surrounding grid points are selected from the FNL grid field as interpolation grid points, denoted as . The final interpolated air pressure value is:
[0037] in For the first The air pressure value at each interpolation grid point.
[0038] The air pressure values at each sampling location were organized into an air pressure feature matrix according to the radial layer and angular direction, and the air pressure feature matrix was normalized.
[0039] S3. Organize the pressure feature matrix into sequential data according to the angular direction, where each position in the sequential data corresponds to each angular direction, and the feature of each position is composed of multiple radial layer pressure values in that angular direction; by mapping the features of each position to the feature space and adding position encoding to the mapped features, an encoded sequence is obtained. like Figure 3As shown, specifically, step S3 includes: The pressure feature matrix is transposed into sequence data with angular direction as the sequence position, and the radial pressure features of each sequence position are mapped from the radial layer dimension to the hidden feature dimension. The radial energy center is calculated based on the air pressure values of each radial layer along the corresponding angular direction of the sequence position; where the radial energy center represents the radial distribution location of the air pressure anomaly along that angular direction, and the calculation formula is: ; in, For sequence position radial energy center, The radial layer number, For the first The radius of each radial layer For the first The radial layers are located in the sequence position. The normalized atmospheric pressure value at the location; if the near-field atmospheric pressure at a certain sequence location is abnormally large in the angular direction, then A smaller value indicates that the main influence in that direction comes from the near-field system; conversely, if the far-field pressure anomaly is large, then... A relatively large value indicates that the system is primarily affected by remote systems. The rotational position encoding frequency parameters of the sequence positions are adaptively adjusted based on the radial energy center of each sequence position to obtain the adjusted frequency parameters: ; ; in, For sequence position The adjusted frequency parameters, For standard frequency parameters, As the base, Indexed by feature dimensions, To hide the feature dimension, Radial modulation intensity coefficient, and These are the minimum and maximum radii of the radial scan, respectively; The feature vectors at each sequence position are rotated based on the adjusted frequency parameters to obtain the encoded sequence containing positional information.
[0040] In one embodiment of the present invention, a rotation transformation based on the adjusted frequency parameters is applied to the feature vectors at each sequence position, specifically including: Pair the feature vectors at each sequence position along multiple dimensions, and construct a rotation matrix for each pair of dimensions. The calculation formula is as follows: ; in, For sequence position The eigenvector after rotation transformation One portion, For sequence position The eigenvector after rotation transformation One portion, For sequence position The angle value is converted to radians. For sequence position The adjusted frequency parameters, and The first feature vector is the first feature vector. and One component; Construct rotation matrices for all dimensions to obtain the rotated encoded feature vectors.
[0041] This invention automatically adjusts the frequency parameters of the position encoding based on the radial energy distribution of each sequence position, so that the sequence positions with the radial energy center in the near field adopt a higher rotation frequency to capture local rapid changes. The physical characteristics of the pressure field are integrated into the position encoding mechanism, which enhances the model's ability to express weather systems at different scales. At the same time, the periodicity of the rotation position encoding automatically satisfies the periodic boundary conditions of the polar coordinate system, avoiding the need for manual boundary setting.
[0042] S4. Input the encoded sequence into the sequence encoder for feature extraction to obtain encoded features; the sequence encoder includes multiple coding layers, and each coding layer establishes the correlation between different positions in the sequence through an attention mechanism. like Figure 4 As shown, specifically, step S4 includes: The radial average gradient intensity at the sequence position is calculated based on the pressure difference between adjacent radial layers in the angular direction corresponding to the sequence position. An attention bias matrix is constructed based on the radial average gradient intensity at each sequence position and the angular interval between sequence positions; In each coding layer of the sequence encoder, the input features of the coding layer are mapped to a query matrix Q, a key matrix K, and a value matrix V through linear transformation. The attention score matrix is obtained by calculating the matrix product of the query matrix and the key matrix and then scaling it. The scaled attention score matrix is added to the corresponding bias value in the attention bias matrix and then normalized to obtain the attention weight matrix. The attention output is obtained by weighted aggregation of the value matrix based on the attention weight matrix. The attention outputs of multiple attention heads are concatenated and then linearly transformed to obtain multi-head attention output. The multi-head attention output is then residually connected with the input features of the coding layer and layer normalized to obtain normalized features. The normalized features are nonlinearly transformed through a feedforward neural network to obtain the output of the feedforward network. The output of the feedforward network is then residually connected with the normalized features and layer normalized to obtain the output features of the coding layer.
[0043] Furthermore, the formula for calculating the corresponding bias value in the attention bias matrix is as follows: ; ; in, For sequence position and sequence position Attention bias between The bias intensity coefficient, and Sequence position and sequence position The radial average gradient intensity, The geometric mean of the positional gradient intensities of the two sequences is given. It is the minimum angular difference between the positions of the two sequences. and Sequence position and sequence position The corresponding angle and direction is the standard deviation parameter for angular proximity.
[0044] In one embodiment of the present invention, the bias intensity coefficient γ is taken as 0.2, and the standard deviation parameter of the angle proximity is... A value of 45 degrees enhances the correlation between sequence positions with larger radial average gradient strength, while also enhancing the correlation between sequence positions with smaller angular intervals.
[0045] This invention calculates the radial average gradient intensity of each sequence position and constructs an attention bias matrix based on the gradient intensity and angular interval. It incorporates the physical characteristics of the pressure field into the attention calculation, thereby enhancing the attention association between sequence positions with larger radial gradients and between sequence positions with smaller angular intervals. This improves the physical interpretability and feature learning ability of the model.
[0046] S5. Aggregate the encoded features to obtain a global feature representation, and perform classification prediction based on the global feature representation to obtain the weather situation category.
[0047] Specifically, step S5 includes: Global average pooling is performed on the output of the last layer of the encoder, aggregating the features of all sequence positions into a single global representation vector. ; By inputting the global feature representation into a linear classifier, the predicted probability distribution of weather condition categories is obtained:
[0048]
[0049] in and These are the learnable parameters of the classifier. For unnormalized classification scores, For the predicted probabilities of each class, satisfying and The five weather pattern categories correspond to: uniform pressure field, front of high pressure, rear of high pressure, front of low pressure, and rear of low pressure.
[0050] The predicted weather pattern category is the one with the highest probability:
[0051] in, Indicates the predicted weather pattern category.
[0052] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for identifying surface weather patterns based on sequence models, characterized in that, Includes the following steps: S1. Determine the center position of the target area, establish a polar coordinate system with the center position of the target area as the origin, and set up multiple radial layers and multiple angular directions in the polar coordinate system to form a sampling position array; S2. Obtain meteorological grid data of the target area, interpolate the air pressure values of each sampling location in the sampling location array to obtain the air pressure values of each sampling location, and form an air pressure feature matrix. S3. Organize the pressure feature matrix into sequential data according to the angular direction, where each position in the sequential data corresponds to each angular direction, and the feature of each position is composed of multiple radial layer pressure values in that angular direction; by mapping the features of each position to the feature space and adding position encoding to the mapped features, an encoded sequence is obtained. S4. Input the encoded sequence into the sequence encoder for feature extraction to obtain the encoded features; The sequence encoder consists of multiple coding layers, and each coding layer establishes the correlation between different positions in the sequence through an attention mechanism; S5. Aggregate the encoded features to obtain a global feature representation, and perform classification prediction based on the global feature representation to obtain the weather situation category.
2. The method for identifying ground weather patterns based on a sequence model as described in claim 1, characterized in that, The angles of the polar coordinate system increase clockwise from zero degrees in the due north direction to 360 degrees, and the radius of the polar coordinate system represents the geographical distance from the center position.
3. The method for identifying ground weather patterns based on a sequence model as described in claim 1, characterized in that, Step S2 specifically includes: Acquire meteorological grid data for the target area, and use the finite difference method to calculate the pressure gradient modulus of each meteorological grid point in the meteorological grid data. The polar coordinates of each sampling location in the sampling location array are converted to geographic coordinates, and multiple meteorological grid points around the sampling location are selected from the meteorological grid data as interpolation grid points. Calculate the spherical distance and the azimuth angle of the line connecting each sampling position and each interpolation grid point; calculate the directional deviation angle between the radial direction of each sampling position and the azimuth angle of the line connecting the points. The interpolation weights of each interpolation grid point are constructed based on the spherical distance, the direction deviation angle, and the pressure gradient modulus. The pressure value at each sampling location is calculated using a weighted average method. The air pressure values at each sampling location are organized into an air pressure feature matrix according to radial layers and angular directions.
4. The method for identifying ground weather patterns based on a sequence model as described in claim 3, characterized in that, The formula for calculating the interpolation weights is: ; ; in, For the first Interpolation weights for each interpolation grid point For sampling position and the first spherical distance between interpolation grid points This is the directional deviation angle between the radial direction of the sampling location and the azimuth angle of the connecting line. This is the orientation alignment factor. For the first Pressure gradient modulus at each interpolation grid point The maximum pressure gradient modulus across all interpolation grid points. It is a tiny constant. The directional modulation intensity coefficient, For sampling position to the th The azimuth angle of the line connecting the interpolation grid points The radial angle of the sampling position.
5. The method for identifying ground weather patterns based on a sequence model as described in claim 1, characterized in that, Step S3 specifically includes: The pressure feature matrix is transposed into sequence data with angular direction as the sequence position, and the radial pressure features of each sequence position are mapped from the radial layer dimension to the hidden feature dimension. Calculate the radial energy center based on the air pressure values of each radial layer in the angular direction corresponding to the sequence position; The rotational position encoding frequency parameters of the sequence positions are adaptively adjusted based on the radial energy center of each sequence position to obtain the adjusted frequency parameters. The feature vectors at each sequence position are rotated based on the adjusted frequency parameters to obtain the encoded sequence containing positional information.
6. The method for identifying ground weather patterns based on a sequence model as described in claim 5, characterized in that, The radial energy center represents the radial distribution location of the air pressure anomaly in that angular direction, and the calculation formula is: ; in, For sequence position radial energy center, The radial layer number, For the first The radius of each radial layer For the first The radial layers are located in the sequence position. The normalized pressure value at that location.
7. The method for identifying ground weather patterns based on a sequence model as described in claim 6, characterized in that, The formula for calculating the adjusted frequency parameters is as follows: ; ; in, For sequence position The adjusted frequency parameters, For standard frequency parameters, As the base, Indexed by feature dimensions, To hide the feature dimension, Radial modulation intensity coefficient, and These are the minimum and maximum radii for the radial scan, respectively.
8. The method for identifying ground weather patterns based on a sequence model as described in claim 7, characterized in that, The application of a rotation transformation based on the adjusted frequency parameters to the feature vectors at each sequence position specifically includes: Pair the feature vectors at each sequence position along multiple dimensions, and construct a rotation matrix for each pair of dimensions. The calculation formula is as follows: ; in, For sequence position The eigenvector after rotation transformation One portion, For sequence position The eigenvector after rotation transformation One portion, For sequence position The corresponding angle value converted to radians. For sequence position The adjusted frequency parameters, and The first feature vector is the first feature vector. and One component; A rotation matrix is constructed for all dimensions to obtain the rotated encoded feature vector.
9. The method for identifying ground weather patterns based on a sequence model as described in claim 1, characterized in that, Step S4 specifically includes: The radial average gradient intensity at the sequence position is calculated based on the pressure difference between adjacent radial layers in the angular direction corresponding to the sequence position. An attention bias matrix is constructed based on the radial average gradient intensity at each sequence position and the angular interval between sequence positions; In each coding layer of the sequence encoder, the input features of the coding layer are mapped to a query matrix, a key matrix, and a value matrix through linear transformation. The attention score matrix is obtained by calculating the matrix product of the query matrix and the key matrix and then scaling it. The scaled attention score matrix is added to the corresponding bias value in the attention bias matrix and then normalized to obtain the attention weight matrix. The attention output is obtained by weighted aggregation of the value matrix based on the attention weight matrix. The attention outputs of multiple attention heads are concatenated and then linearly transformed to obtain multi-head attention output. The multi-head attention output is then residually connected with the input features of the coding layer and layer normalized to obtain normalized features. The normalized features are nonlinearly transformed through a feedforward neural network to obtain the output of the feedforward network. The output of the feedforward network is then residually connected with the normalized features and layer normalized to obtain the output features of the coding layer.
10. The method for identifying ground weather patterns based on a sequence model as described in claim 9, characterized in that, The formula for calculating the corresponding bias value in the attention bias matrix is as follows: ; ; in, For sequence position and sequence position Attention bias between The bias intensity coefficient, and Sequence position and sequence position The radial average gradient intensity, The geometric mean of the positional gradient intensities of the two sequences is given. It is the minimum angular difference between the positions of the two sequences. and Sequence position and sequence position The corresponding angle and direction is the standard deviation parameter for angular proximity.
Citation Information
Patent Citations
Cold-wave weather prediction method and device, electronic equipment and computer readable medium
CN114594532A
