CN-120275926-B - Multimode radar echo extrapolation method based on Doppler dual-polarization weather radar

CN120275926BCN 120275926 BCN120275926 BCN 120275926BCN-120275926-B

Abstract

The invention discloses a multimode radar echo extrapolation method based on Doppler dual-polarization weather radar, which belongs to the technical field of weather, and specifically comprises the steps of firstly collecting Doppler dual-polarization weather base data, converting the Doppler dual-polarization weather base data into multimode physical parameter data, then constructing a multimode radar echo extrapolation model comprising a multimode encoder, a depth feature extraction module and a decoder, wherein the encoder is composed of a dual-branch structure, extracting initial features of original multimode information, inputting the initial features into the depth feature extraction module, further extracting depth space-time features through a self-attention mechanism, finally finishing layer-by-layer decoding of the depth space-time features through the decoder, introducing the initial features to strengthen the transmission of shallow feature information, and inputting the new observed radar information into the trained echo extrapolation model to predict radar echo images at future moments. The invention provides powerful technical support for real-time weather monitoring, rainfall forecasting and extreme weather early warning.

Inventors

GAO FENG
YE YUANKANG
LIU CHANG
SONG LIANCHENG
SUN JIANAN

Assignees

哈尔滨工程大学

Dates

Publication Date: 20260505
Application Date: 20250325

Claims (5)

1. A multimode radar echo extrapolation method based on Doppler dual-polarization weather radar is characterized by comprising the following steps: Firstly, acquiring base data based on Doppler dual-polarization weather radar, unifying data processing through an inversion algorithm, and converting the data into multi-mode physical parameter data; The multi-modal physical parameter data includes horizontal polarization radar reflectivity Differential radar reflectivity Specific differential phase Correlation coefficient of common bias And velocity spectrum width ; Then, constructing a multi-mode radar echo extrapolation model of the Doppler dual-polarization weather radar, and training by utilizing multi-mode data; converting the radar physical information observed in real time into multi-mode physical parameter data, and inputting the multi-mode physical parameter data into a trained multi-mode radar echo extrapolation model to obtain a radar echo prediction image at a future moment; the multi-mode radar echo extrapolation model comprises a multi-mode encoder, a depth feature extraction module and a decoder; The multi-mode encoder consists of a dual-branch structure, namely an inter-space feature extraction branch and an intra-space feature extraction branch, wherein the inter-space feature extraction branch and the intra-space feature extraction branch are used for carrying out primary feature extraction on original multi-mode physical parameter data, and the inter-factor feature RFF extracted by the inter-space feature extraction branch and the intra-factor feature AFF extracted by the intra-space feature extraction branch are overlapped in a channel dimension to generate shallow aggregation features FF and input the shallow aggregation features FF to a depth feature extraction module Finally, finishing the layer-by-layer decoding of the depth space-time features by a decoder, and introducing preliminary features in the last layer of the decoder To strengthen the transmission of the shallow characteristic information, and finally obtain the extrapolation result of the radar echo image.
2. A multi-modal radar echo extrapolation method as claimed in claim 1, wherein the spatial feature extraction branches are formulated as follows, Represents the first The first mode is at The characteristics of the layers are such that, ; Represents the first The information of the individual original modalities is provided, SiLU represents a non-linear activation function, Representing a two-dimensional convolution operation, Representing the step length of the convolution structure if Odd number of Taking out 1 of the mixture, Is even if 2, Taking; Representing the first layer to be passed to the decoder RFF represents the factor-to-factor characteristics extracted by the space-to-space characteristic extraction branch in the multi-mode encoder; Layer 4 features representing a first modality, concat representing feature stitching in a channel dimension; The formula of the feature extraction branch in the space is as follows: representing the information obtained by stacking all the modal information in the channel dimension, and finally At the time of the maximum value of 4, Branches representing feature extraction in space are extracted to feature AFF in factor.
3. The method for multi-modal radar echo extrapolation based on Doppler dual polarized weather radar as claimed in claim 1, wherein the depth feature extraction module is formulated as follows, When (when) The initial value is 1, and the initial value is 1, For the shallow aggregation feature FF extracted from the multi-mode encoder, L layers are co-stacked; ; Represents the formula of the attention of the person, Represents a Point-wise convolution with a convolution kernel size of 1, Representing a gated self-attention unit, Representing a self-regularized non-monotonic neural activation function, Representing the formula of a multi-layer perceptron, Representing a depth-separable convolution, Represents the first Depth spatiotemporal characteristics of the layers, Represents BatchNorm functions; Finally, the output deep space-time characteristics of the depth characteristic extraction module are as follows 。
4. A method of multi-modal radar echo extrapolation based on doppler dual polarized weather radar as claimed in claim 3 wherein the gated self-attention unit The formula is as follows: Wherein, the Representing a dilation depth convolution, The characteristics of the input are represented by, The representative of the force of interest is sought, Representing the characteristics of the output.
5. A multi-modal radar echo extrapolation method based on doppler dual polarization weather radar as claimed in claim 2 wherein the decoder data flow formula is as follows: Wherein, the And Representing the features after the first and second upsampling, Representing the convolved feature after the first upsampling, Representing the inverse convolution of the data with the data, And finally, predicting the radar echo data.

Description

Multimode radar echo extrapolation method based on Doppler dual-polarization weather radar Technical Field The invention belongs to the technical field of weather, and particularly relates to a multimode radar echo extrapolation method based on a Doppler dual-polarization weather radar. Background Doppler weather radar is an advanced active remote sensing detection device and is widely applied to the fields of atmospheric environment monitoring and weather forecast. The basic principle is based on Doppler effect, and by transmitting electromagnetic wave and receiving reflected signals, the movement speed of the scatterer relative to the radar in cloud, rain and other weather is measured in real time. Under certain conditions, the radar can invert the atmospheric wind field distribution, the vertical velocity change of the air flow and the local turbulence state by utilizing the phase and frequency change of the echo signals, and scientific basis is provided for monitoring and early warning strong convection weather. The traditional radar echo extrapolation method mainly comprises a cross correlation method, a single centroid method and an optical flow method, wherein the methods predict the distribution of future echoes by relying on the correlation or motion characteristics among signals, but the prediction accuracy and the applicability of the algorithm are difficult to meet the actual demands due to the limitation of the algorithm under complex meteorological conditions. In recent years, with the remarkable improvement of the performance of computer hardware (such as GPU and TPU), the deep learning technology is rapidly developed, and the performance superior to that of the traditional method is shown in a plurality of fields such as image processing, space-time sequence prediction and the like. In this context, data-driven based radar echo extrapolation models are one of the hot spots and optimal solutions for current research. The existing research generally reduces the radar echo extrapolation problem to a space-time sequence predictor problem of deep learning, and focuses on modeling and characterization of the self-motion characteristics of radar echoes. Although such models have made a significant breakthrough in short-term prediction accuracy, the evolution of the radar echo is not only affected by the internal dynamic process, but also is closely related to other physical parameters. Most current models fail to adequately fuse these additional multi-modal physical parameter information, thus lacking predictive capability in the face of complex weather systems. Through extrapolation prediction of unobserved areas, global monitoring of a large range of weather conditions can be achieved, so that evolution of a weather system can be known more accurately, precipitation distribution and wind field change can be monitored, and early warning capacity of extreme weather can be improved. By constructing a model based on deep learning and fully excavating and fusing historical multiple physical parameter information, the prediction accuracy of the future radar echo condition is expected to be improved, and more accurate data support is provided for real-time monitoring and prediction of ground rainfall information. However, no method for effectively utilizing the information of the multi-mode physical parameters of the weather radar to optimize the radar echo extrapolation effect exists in the prior art. Therefore, under the support of a deep learning algorithm and a multi-source data fusion technology, a novel radar echo extrapolation method and system are developed, and the method and system have important theoretical significance and wide application prospect. Disclosure of Invention Aiming at the problems that in the prior art, only single radar echo information is relied on for extrapolation prediction, so that prediction precision is insufficient and multiple physical parameter information cannot be fully fused, the invention discloses a multi-mode radar echo extrapolation method based on a Doppler dual-polarization weather radar, which fully integrates various physical parameter information acquired by a radar system, effectively captures complex dynamics characteristics in the radar echo evolution process, remarkably improves prediction accuracy and system robustness, and provides powerful data support for monitoring and extreme weather early warning of a large-scale weather system. The multi-mode radar echo extrapolation method based on the Doppler dual-polarization weather radar comprises the following steps: step one, acquiring basic data based on Doppler dual-polarization weather radar, unifying data processing through an inversion algorithm, and converting the basic data into multi-mode physical parameter data; The multi-mode physical parameter data comprise a horizontal polarization radar reflectivity Z H, a differential radar reflectivity Z DR, a specific differential phase K DP, a common b