Radar echo data processing method and system based on multi-elevation subspace coding

Abstract

The invention discloses a radar echo data processing method and system based on multi-elevation subspace coding, and relates to the technical field of meteorological data processing, and the method comprises the steps: carrying out the preprocessing of multi-elevation radar echo data; independently coding the radar echo reflectivity of each elevation angle into the corresponding subspace to obtain the subspace feature representation of each elevation angle; the subspace feature representations corresponding to the elevation angles are integrated to form a complete implicit vector space; integrating the implicit vector space with the spatial position embedded information and the time period embedded information, inputting the information into a pre-trained time sequence model, capturing the dependency and evolution rule of a convective weather system in time and space, and forecasting the multi-elevation radar echo reflectivity in future 2 hours; the time sequence model is constructed based on a deep learning network with a self-attention mechanism. According to the method, the radar echo reflectivity of different elevation angles is fully utilized, the complex evolution process of a meteorological system is captured, and the accuracy of meteorological forecast is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological data processing, and in particular to a radar echo data processing method and system based on multi-elevation subspace coding. Background Art

[0002] Meteorological radar echo data plays an important role in weather forecasting and disaster monitoring. Doppler meteorological radar usually performs multi-elevation scanning in VCP21, VCP11 and other modes to obtain radar echo reflectivity, radial velocity, spectrum width and other data at each elevation angle (VCP21 is: 0.5°, 1.5°, 2.4°, 3.4°, 4.3°, 6.0°, 9.9°, 14.6°, 19.5°). These data can provide high-resolution spatiotemporal structure of the occurrence and development of convective weather systems, and provide important support for short-term weather forecasts. Existing radar data processing methods usually process multi-elevation data in a unified manner, and generate radar echo combination reflectivity, multi-layer CAPPI radar echo reflectivity and other data through projection transformation, interpolation and other methods, resulting in partial information loss and introduction of interpolation errors. At the same time, due to the complexity of the meteorological system and the limitation of algorithm computing power resources, traditional methods cannot achieve high spatiotemporal resolution (0-2 hours, 6 minutes, library length 500 meters, 9 elevation angles) radar echo prediction. Summary of the invention

[0003] The present invention provides a radar echo data processing method and system based on multi-elevation subspace coding, which can efficiently utilize radar data at different elevation angles and combine geographical location and time period information to achieve high temporal and spatial resolution multi-elevation radar echo reflectivity prediction.

[0004] In order to solve the above problems, the technical solution adopted by the present invention is as follows:

[0005] In a first aspect, the present invention provides a radar echo data processing method based on multi-elevation subspace coding, comprising the following steps:

[0006] S1. Preprocessing the multi-elevation-angle weather radar base data;

[0007] S2. For the preprocessed meteorological radar base data, the radar echo reflectivity at each elevation angle is independently encoded into the corresponding subspace to obtain the subspace feature representation of each elevation angle;

[0008] S3, integrating the subspace feature representations corresponding to each elevation angle to form a complete latent vector space;

[0009] S4. Integrate the latent vector space with the spatial position embedding information and the time period embedding information as the input vector, and input it into the pre-trained time series model to capture the spatiotemporal dependency and evolution law of the convective weather system, and predict the multi-elevation radar echo reflectivity within the next 2 hours; the time series model is constructed based on a deep learning network with a self-attention mechanism.

[0010] As a further description of the above technical solution: In step S1, the method for preprocessing the meteorological radar base data includes decoding the FMT format base data, reading the multi-elevation radar echo reflectivity (ZH) data, and selecting the elevation layer of VCP21: 0.5°, 1.5°, 2.4°, 3.4°, 4.3°, 6.0°, 9.9°, 14.6°, 19.5° (9 layers in total). Each layer takes the original library length (S-band radar: 500 meters), with a maximum of 300 library lengths, to obtain the radar echo reflectivity data of the elevation layer. Because the starting azimuth of each elevation angle does not necessarily start at 0°, it is necessary to further align and interpolate the data of each elevation layer to normalize the azimuth of 0°-359°, with an equal interval of 1°. Finally, after preprocessing, each meteorological radar base data generates standard multi-elevation radar echo reflectivity data with the dimensions of elevation (9) × storage length (300) × azimuth (360); the high temporal and spatial resolution multi-elevation radar echo reflectivity data for 2 consecutive hours is: forecast time (20) × elevation (9) × storage length (300) × azimuth (360).

[0011] As a further description of the above technical solution: in step S2, the radar echo reflectivity (library length × azimuth) of each elevation angle corresponds to an independent subspace encoder, and the subspace encoder is a convolutional encoder, including a convolution layer, a pooling layer and an activation function layer.

[0012] As a further description of the above technical solution: in step S3, the subspace feature representation corresponding to each elevation angle is weighted and integrated through the attention mechanism.

[0013] As a further description of the above technical solution: In step S4, the spatial position embedded information is information obtained by encoding the geographical location data of the radar station, which is used to assist the spatial positioning of the time series model. The geographical location data of the radar station includes longitude, latitude and altitude.

[0014] As a further description of the above technical solution: encoding the geographic location data of the radar station is to embed the geographic location data of the radar station into independent vector spaces respectively, and then integrate them into a comprehensive spatial location representation.

[0015] As a further description of the above technical solution: the time period embedded information is generated by sine and cosine functions.

[0016] As a further description of the above technical solution: the time series model is built based on the Transformer network, including several self-attention layers and feedforward neural network layers, each layer captures the spatiotemporal dependency of radar echo data through the self-attention mechanism.

[0017] In a second aspect, the present invention provides a radar echo data processing system, comprising:

[0018] A data receiving module is used to receive and pre-process multi-elevation-angle radar volume scanning data;

[0019] A subspace encoding module is used to independently encode the radar echo reflectivity at different elevation angles into the corresponding subspace;

[0020] The feature integration module is used to integrate the subspace feature representations of each elevation angle into a complete latent vector space;

[0021] Position encoding module, used to generate spatial position embedding information;

[0022] A time period encoding module, used for generating time period embedding information;

[0023] The spatiotemporal series prediction module integrates the latent vector space, spatial position embedding information and time period embedding information, and performs fusion prediction based on the time series model to output the multi-elevation radar echo reflectivity every 6 minutes in the next 2 hours.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] The present invention can respectively encode the data features of each elevation angle in the multi-elevation-angle radar echo reflectivity data processing, and improve the high-temporal-resolution radar echo prediction capability through spatial position, time periodicity information and spatiotemporal sequence modeling; the method of the present invention makes full use of the original library data of different elevation angles and normalized azimuth angles, maximizes the use of data information, captures the complex spatiotemporal evolution process of the meteorological system, and significantly improves the disaster weather monitoring and early warning capabilities.

[0026] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the embodiments of the present invention are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0028] Figure 1 is a flow chart of multi-elevation radar echo data processing in an embodiment;

[0029] Figure 2 2 is a schematic diagram of the structure of a subspace encoder in an embodiment;

[0030] Figure 3 is a schematic diagram of spatial position information encoding in an embodiment;

[0031] Figure 4 is a schematic diagram of time period information encoding in an embodiment;

[0032] Figure 5 4 is a network structure diagram of spatiotemporal sequence modeling in an embodiment. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0034] Please refer to Figure 1 The embodiment of the present invention provides a radar echo data processing method based on multi-elevation subspace coding, which is as follows:

[0035] 1. Preprocess the weather radar basic data.

[0036] Decode the FMT format base data, read the multi-elevation radar echo reflectivity (ZH) data, and select the elevation layer of VCP21: 0.5°, 1.5°, 2.4°, 3.4°, 4.3°, 6.0°, 9.9°, 14.6°, 19.5° (9 layers in total). Take the original library length (S-band radar: 500 meters) for each layer, and take a maximum of 300 library lengths to obtain the radar echo reflectivity data of the elevation layer. Because the starting azimuth of each elevation angle does not necessarily start at 0°, further alignment and interpolation are required to normalize the data of each elevation layer to 0°-359° azimuth, with an equal interval of 1°. Finally, after preprocessing, each meteorological radar base data generates standard multi-elevation radar echo reflectivity data with the dimensions of elevation (9) × storage length (300) × azimuth (360); the high temporal and spatial resolution multi-elevation radar echo reflectivity data for 2 consecutive hours is: forecast time (20) × elevation (9) × storage length (300) × azimuth (360).

[0037] 2. Encode the radar echo data of each elevation angle independently into the corresponding subspace to obtain the subspace feature representation of each elevation angle.

[0038] The radar echo data at each elevation angle corresponds to an independent subspace encoder, which is tailored for each elevation angle and uses structures such as convolutional neural networks (CNN) to extract meteorological features unique to the elevation angle. The features output by the subspace encoder are represented as high-dimensional vectors, capturing the independent features of each elevation angle data. Figure 2 The subspace coding structure of a single elevation angle data is shown, with an emphasis on the encoding process of the convolutional neural network.

[0039] Subspace encoder design:

[0040] Suppose we choose an elevation angle of 3.4°.

[0041] (1) Convolutional layer design

[0042] 1) Kernel size: Depending on the spatial resolution and feature scale of the 3.4° elevation data, we may choose kernels of different sizes, such as 3×3, stride 2, and padding 2. For example, if the data contains large-scale meteorological features (such as large clouds), we may choose a larger kernel to capture these features.

[0043] 2) Number of convolution kernels: Depending on the complexity of the features, we may need multiple convolution kernels to capture features of different scales.

[0044] For example, we might use 16 convolution kernels to capture basic features and 32 convolution kernels to capture finer features.

[0045] (2) Pooling layer design

[0046] Pooling strategy: In order to reduce the feature dimension and extract key information, we may use maximum pooling or average pooling after the convolution layer. The size of the pooling window can be selected according to the spatial resolution of the data and the required feature abstraction level.

[0047] (3) Activation function layer

[0048] Activation function selection: To introduce nonlinearity, we may choose ReLU or Leaky ReLU as the activation function. These functions can help the model learn more complex feature representations.

[0049] (4) Adjustment and optimization

[0050] 1) Number of layers and depth: Depending on the complexity of the elevation data, we may need to adjust the depth of the network, increasing or decreasing the number of layers to optimize performance.

[0051] 2) Regularization and Normalization: To prevent overfitting, we may add a Dropout layer or a LayerNormalization layer to the network.

[0052] In this embodiment, the configured subspace encoder is a convolution encoder to extract features from radar echo data. The subspace encoder structure includes a convolution layer, a pooling layer, and an activation function layer, which are used to extract radar feature information at corresponding elevation angles. The convolution kernel size, the number of layers, and the activation function can be adjusted according to the different features of each elevation angle data to enhance the adaptability of the model to the features of each elevation angle.

[0053] 3. Integrate the subspace feature representations corresponding to each elevation angle to form a complete latent vector space.

[0054] The complete latent vector space contains feature information at multiple elevation angles, which can characterize meteorological information at different altitudes, fully express the atmospheric structure in the vertical direction, and increase the ability to capture the vertical structure of the atmosphere.

[0055] In this embodiment, the subspace feature representation corresponding to each elevation angle is weighted and integrated through a multi-head self-attention mechanism. The multi-head self-attention mechanism allows the model to weight data of different elevation angles according to their importance, so as to ensure that data of important elevation angles has a higher weight in the joint representation.

[0056] Multi-head self-attention mechanism:

[0057] The input of the multi-head attention mechanism MHA (Multi-Head Attention) includes three vectors: query vector (query), key vector (key) and value vector (value). For a given query vector, MHA will perform a weighted sum of the key vector, where the weight is calculated by the similarity between the query vector and the key vector, and then multiply the weighted sum by the value vector for output. The calculation formula of MHA is shown in Formula 1:

[0058] MultiHead(Q,K,V)=Concat(head1,…,head h )W O (1)

[0059] Where Q, K, V represent the query vector, key vector, and value vector respectively, h represents the number of heads, headi represents the output of the i-th head, and WO is the output transformation matrix. The output headi of each head can be expressed as Formula 2:

[0060]

[0061] in are the query, key, and value transformation matrices of the i-th head, respectively, and Attention is the attention calculation function. In MHA, the self-attention mechanism is generally used to calculate attention.

[0062] The calculation of the self-attention mechanism can be expressed as the following formula 3:

[0063]

[0064] where d k is the dimension of the key vector, the softmax similarity is normalized, the weight of each key vector is calculated, and then the weight is multiplied by the value vector, and finally the weighted sum is performed to get the attention output.

[0065] The MHA process can be summarized into the following steps:

[0066] 1) Divide the input sequence data into multiple headers;

[0067] 2) Perform independent query, key, and value linear transformations on each header;

[0068] 3) Perform self-attention calculation on each head to obtain the output of the head;

[0069] 4) Concatenate the outputs of all heads together and perform a linear transformation on the outputs.

[0070] 4. Combination Figure 3 As shown, the geographic location data of the radar station is encoded to obtain spatial location embedded information.

[0071] The spatial location embedding information is used to assist the spatial positioning of the model. The geographic location data of the radar station includes longitude, latitude and altitude.

[0072] The geographic location data of the radar stations are embedded into independent vector spaces, which can be achieved based on embedding layers or sine and cosine functions, and then these vector spaces are integrated into a comprehensive spatial position representation to enhance the spatial characteristic expression of the space-time sequence model and ensure the precise positioning of the model in space.

[0073] 5. Combination Figure 4 As shown, the time period information is encoded to obtain the time period embedded information.

[0074] In this embodiment, the time period information includes months and hours. The time period information is periodically encoded, that is, the months and hours are converted into time period embedding information through sine and cosine functions, and the time period information is converted into a vector representation to represent the periodic characteristics of the months and hours; the generated inter-period embedding information can help the model understand periodic weather phenomena such as seasonality and diurnal changes, and provide support for long-term prediction of weather systems.

[0075] The specific formula for generating time period embedded information includes:

[0076] Month code: Use and Perform periodic representation;

[0077] Hour coding: Use and Perform periodic representation.

[0078] 6. Integrate the latent vector space with the spatial position embedding information and the time period embedding information as the input vector and input it into the pre-trained time series model to capture the temporal and spatial dependence and evolution law of the convective weather system and predict the multi-elevation radar echo reflectivity within the next 2 hours.

[0079] In this embodiment, if Figure 5 As shown in the figure, the time series model is built based on the Transformer network, which includes several self-attention layers and feedforward neural network layers. Each layer captures the spatiotemporal dependencies of radar echo data through the self-attention mechanism, and decodes and generates the multi-elevation radar echo reflectivity within the next 2 hours.

[0080] The embodiments of the present invention realize high-temporal and spatial resolution radar echo reflectivity forecasting through subspace coding, feature integration, and the assistance of spatial position and time period information, combined with spatiotemporal sequence modeling, and maximize the application of radar detection data information, laying a solid foundation for research on severe convective weather system identification, tracking, and early warning based on multi-elevation radar echo reflectivity, and have significant innovation and application prospects.

[0081] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A radar echo data processing method based on multi-elevation subspace coding, characterized in that: The following steps are involved: S1. Preprocessing the multi-elevation-angle weather radar base data; S2. For the preprocessed meteorological radar base data, the radar echo reflectivity at each elevation angle is independently encoded into the corresponding subspace to obtain the subspace feature representation of each elevation angle; S3, integrating the subspace feature representations corresponding to each elevation angle to form a complete latent vector space; S4. Integrate the latent vector space with the spatial position embedding information and the time period embedding information as the input vector, and input it into the pre-trained time series model to capture the temporal and spatial dependence and evolution law of the convective weather system, and predict the multi-elevation angle radar echo reflectivity within the next 2 hours; the time series model is constructed based on a deep learning network with a self-attention mechanism.

2. The method according to claim 1, characterized in that In step S1, the method for preprocessing the multi-elevation-angle meteorological radar base data includes radar base data decoding and elevation angle data normalization processing.

3. The method according to claim 1, characterized in that In step S2, the radar echo reflectivity at each elevation angle corresponds to an independent subspace encoder, and the subspace encoder is a convolutional encoder including a convolutional layer, a pooling layer and an activation function layer.

4. The method according to claim 1, characterized in that: In step S3, the subspace feature representations corresponding to each elevation angle are weighted and integrated through the attention mechanism.

5. The method according to claim 1, characterized in that In step S4, the spatial position embedded information is information obtained by encoding the geographical location data of the radar station, which is used to assist the spatial positioning of the time series model. The geographical location data of the radar station includes longitude, latitude and altitude.

6. The method according to claim 5, characterized in that The encoding of the radar station's geographic location data is to embed the radar station's geographic location data into independent vector spaces respectively, and then integrate them into a comprehensive spatial location representation.

7. The method according to claim 5, characterized in that The time period embedded information is generated by sine and cosine functions.

8. The method according to claim 5, characterized in that The time series model is built based on the Transformer network, including several self-attention layers and feedforward neural network layers, each of which captures the temporal dependency of radar echo data through the self-attention mechanism.

9. The method according to claim 5, characterized in that The multi-elevation angle radar echo reflectivity within the next 2 hours has a time resolution of 6 minutes and a spatial resolution of the radar echo library length. The multiple elevation angles specifically include: 0.5°, 1.5°, 2.4°, 3.4°, 4.3°, 6.0°, 9.9°, 14.6° and 19.5°.

10. A radar echo data processing system, characterized in that: include: A data receiving module is used to receive and pre-process multi-elevation-angle radar volume scanning data; A subspace encoding module is used to independently encode the radar echo reflectivity at different elevation angles into the corresponding subspace; The feature integration module is used to integrate the subspace feature representations of each elevation angle into a complete latent vector space; Position encoding module, used to generate spatial position embedding information; A time period encoding module, used for generating time period embedding information; The spatiotemporal series prediction module integrates the latent vector space, spatial position embedding information and time period embedding information, and performs fusion prediction based on the time series model to output the multi-elevation radar echo reflectivity every 6 minutes in the next 2 hours.

Images