Search

CN-121995541-A - Typhoon rainfall intelligent prediction method based on multi-scale space-time coupling

CN121995541ACN 121995541 ACN121995541 ACN 121995541ACN-121995541-A

Abstract

The invention discloses a typhoon rainfall intelligent prediction method based on multi-scale space-time coupling, which comprises the following steps of 1) collecting typhoon rainfall data, wherein the data comprises satellite rainfall data, numerical weather forecast data and typhoon path information, 2) preprocessing the data, comprising data standardization and space-time clipping, 3) data coding, 4) spatial feature extraction, wherein the spatial feature extraction is performed by adopting a double-branch spatial coding module and is divided into a local convolution branch LC and a global Fourier transform branch GF, 5) multi-scale coupling, and hidden representation output by the spatial coding module is used as input to obtain hidden features of the next time step, and 6) decoding is performed by using a decoder to output a typhoon rainfall field prediction result. According to the typhoon rainfall prediction model constructed by the invention, the modeling capacity and the prediction precision of the typhoon rainfall process are obviously improved by constructing the self-adaptive space coding and combining a multi-scale space-time characteristic extraction structure, a flexible decoder and a region loss function.

Inventors

  • QIN MENGJIAO
  • Lin Wuxia
  • Zhao Bufan
  • CHEN XIJIANG

Assignees

  • 武汉理工大学

Dates

Publication Date
20260508
Application Date
20251224

Claims (10)

  1. 1. The intelligent typhoon rainfall prediction method based on multi-scale space-time coupling is characterized by comprising the following steps of: 1) Collecting typhoon rainfall data, wherein the data comprise satellite rainfall data, numerical weather forecast data and typhoon path information; The satellite rainfall data are spatially continuous distribution images of a typhoon rainfall field, and the numerical weather forecast data comprise the wind speed and the potential height of typhoons; 2) Data preprocessing, including data standardization and space-time clipping; 3) Encoding data; after all the preprocessed data are aligned in the time dimension, uniformly processing the data into a five-dimensional tensor format, wherein X= (B, T, C, H and W), B represents the batch size of the data, T represents the time step, C represents the channel number, H represents the space height of the data, and W represents the space width of the data; 4) Extracting spatial features; the space feature extraction is carried out by adopting a double-branch space coding module and is divided into a local convolution branch LC and a global Fourier transform branch GF; The local convolution branches are based on a shallow convolution neural network, and feature graphs of high-frequency detail features are extracted from an input image through multi-layer convolution and downsampling operations; the global frequency domain branch is used for capturing large-scale space dependence and global association in a rainfall field according to an input image; The two branches realize the fusion of the double-branch characteristics through repeated iterative interaction; 5) Taking the fused double-branch characteristic output by the space coding module as input to obtain the hidden characteristic of the next time step: 6) And decoding by using a decoder, and outputting a typhoon rainfall field prediction result.
  2. 2. The intelligent typhoon rainfall prediction method based on multi-scale space-time coupling according to claim 1, wherein in the step 2), space-time clipping is implemented by adopting a dynamic clipping strategy with a typhoon center as an anchor point, and a region with a fixed space size is clipped from an original image corresponding to satellite rainfall data and a weather forecast variable field with the typhoon center of each time step as the anchor point, and the region covers typhoon cores and surrounding regions thereof, so that the spatial physical alignment of the rainfall data and weather variables is ensured.
  3. 3. The intelligent typhoon rainfall prediction method based on multi-scale space-time coupling according to claim 1, wherein in the step 3), the data coding is specifically as follows; 3.1 Satellite rainfall data processing, namely stacking the cut multi-frame images in the time dimension to form an image tensor with a time sequence structure; 3.2 Numerical meteorological variable processing, namely, the numerical meteorological variable subjected to spatial clipping and layer-by-layer normalization processing is consistent with a rainfall image in a spatial dimension and is stacked into tensors after being aligned in a time dimension; 3.3 Processing typhoon path information, namely imaging discrete point data of typhoon path information to form a channel consistent with a satellite rainfall data image structure; 3.4 Multi-mode tensor fusion indicates that all preprocessed data are aligned in the time dimension and then are uniformly processed into a five-dimensional tensor format, wherein X= (B, T, C, H and W), B indicates the batch size of the data, T indicates the time step, C indicates the channel number, H indicates the space height of the data, and W indicates the space width of the data.
  4. 4. The intelligent typhoon rainfall prediction method based on multi-scale space-time coupling according to claim 1, wherein in the step 4), the local convolution branches are formed by The stacked ConvNormReLU units are formed by the following core structural forms: ; ; Wherein GNorm denotes the group normalization, leakyReLU is the leak ReLU activation function, Representing the output of the ith LC block; Representing a two-dimensional convolution layer, and D representing the number of characteristic channels.
  5. 5. The intelligent typhoon rainfall prediction method based on multi-scale space-time coupling according to claim 1, wherein in the step 4), the global fourier transform branch processing procedure includes: Step 4.1) dividing an input image into non-overlapping image blocks, and linearly mapping the non-overlapping image blocks into embedded vectors; Step 4.2) embedding vectors for the obtained image blocks Converting the spatial domain features into frequency domain features through two-dimensional fast Fourier transform; step 4.3) performing linear transformation on the frequency domain characteristics by using a multi-layer feedforward neural network (MLP) to map the frequency domain characteristics into a linear space; And 4.4) converting the processed frequency domain characteristics back to a space domain through inverse Fourier transform (IFFT) to realize the combination of frequency domain information and space details.
  6. 6. The intelligent prediction method of typhoon rainfall based on multi-scale space-time coupling according to claim 1, wherein in the step 5), the decoder comprises a spatial decoder and a temporal decoder; Wherein, the space decoder consists of The decoding blocks are used for gradually recovering the spatial resolution of the input coded features through a deconvolution layer, mapping the low-dimensional space-time features back to the original rainfall field size, and guaranteeing the high-resolution output of the predicted image; ; 。
  7. 7. The intelligent prediction method of typhoon rainfall based on multi-scale space-time coupling according to claim 6, wherein in the step 5), a time decoder adopts ConvNormReLU units to expand a time channel; Features derived from spatial decoders The time dimension of (2) is spliced with the channel dimension to form a tensor with the dimension of T×C, and the tensor is mapped to K×C, wherein K is the target prediction length, and then the obtained feature map is subjected to dimension transformation to obtain the target prediction dimension ; The calculation process is as follows: 。
  8. 8. the intelligent typhoon rainfall prediction method based on multi-scale space-time coupling according to claim 1, wherein loss functions calculate typhoon kernel areas respectively Peripheral region Integral area Is a precipitation prediction error of (1); Is provided with Representing the predicted and actual precipitation fields respectively, Is a binary mask for the cyclone core region, Is a mask for the peripheral region. The regional MSE loss function is defined as follows: ; Total loss function: ; Wherein, the Indicating the actual precipitation of the b-th sample at time t, Is a binary mask for the kernel region, A mask that is a peripheral region; , And Is a weight coefficient.
  9. 9. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the method of any of claims 1 to 8 when the computer program is executed.
  10. 10. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the method of any of claims 1 to 8.

Description

Typhoon rainfall intelligent prediction method based on multi-scale space-time coupling Technical Field The invention relates to a weather disaster prediction technology, in particular to a typhoon rainfall intelligent prediction method based on multi-scale space-time coupling. Background Typhoons and storm are main expression forms of typhoons, have far-reaching influence range and complex and changeable properties, often cause secondary geological disasters such as flood, landslide, debris flow and the like, and cause huge impact to human society. Therefore, the method accurately predicts the spatial distribution and evolution trend of rainfall in the typhoon process, and has important significance for disaster prevention and reduction and public early warning response. Numerical weather forecast (NWP) models such as WRF and ECMWF widely used at present perform well in terms of large-scale weather system simulation, but are difficult to realize high-frequency update and high-resolution output due to strong initial condition sensitivity and high calculation resource consumption. In addition, single deep learning models have limited predictive power in the face of highly varying nonlinear systems such as typhoons. Aiming at the problems, a typhoon rainfall prediction model with local details, global structures and long-short-term evolution processes is required to be constructed, and the typhoon rainfall prediction model has the advantages of strong generalization capability, high-precision expression and flexible output capability so as to meet the actual requirements of current intelligent meteorological service and disaster prevention and reduction business. Disclosure of Invention Aiming at the defects in the prior art, the technical problem to be solved by the invention is to provide the typhoon rainfall intelligent prediction method based on multi-scale space-time coupling. The technical scheme adopted for solving the technical problems is that the typhoon rainfall intelligent prediction method based on multi-scale space-time coupling comprises the following steps: 1) Collecting typhoon rainfall data, wherein the data comprise satellite rainfall data, numerical weather forecast data and typhoon path information; The satellite rainfall data are spatially continuous distribution images of a typhoon rainfall field, and the numerical weather forecast data comprise the wind speed and the potential height of typhoons; 2) Data preprocessing, including data standardization and space-time clipping; 3) Encoding data; after all the preprocessed data are aligned in the time dimension, uniformly processing the data into a five-dimensional tensor format, wherein X= (B, T, C, H and W), B represents the batch size of the data, T represents the time step, C represents the channel number, H represents the space height of the data, and W represents the space width of the data; 4) Extracting spatial features; the space feature extraction is carried out by adopting a double-branch space coding module and is divided into a local convolution branch LC and a global Fourier transform branch GF; The local convolution branches are based on a shallow convolution neural network, and feature graphs of high-frequency detail features are extracted from an input image through multi-layer convolution and downsampling operations; the global frequency domain branch is used for capturing large-scale space dependence and global association in a rainfall field according to an input image; The two branches realize the fusion of the double-branch characteristics through repeated iterative interaction; 5) Taking the fused double-branch characteristic output by the space coding module as input to obtain the hidden characteristic of the next time step: 6) And decoding by using a decoder, and outputting a typhoon rainfall field prediction result. According to the above scheme, in the step 2), the space-time clipping adopts a dynamic clipping strategy with the typhoon center as an anchor point, and the typhoon center of each time step is used as an anchor point, so that a region with a fixed space size is clipped from the original image corresponding to the satellite rainfall data and the weather forecast variable field, and the region covers the typhoon core and the surrounding region thereof, thereby ensuring the space physical alignment of the rainfall data and the weather variables. According to the above scheme, in the step 3), the data coding is specifically as follows; 3.1 Satellite rainfall data processing, namely stacking the cut multi-frame images in the time dimension to form an image tensor with a time sequence structure; 3.2 Numerical meteorological variable processing, namely, the numerical meteorological variable subjected to spatial clipping and layer-by-layer normalization processing is consistent with a rainfall image in a spatial dimension and is stacked into tensors after being aligned in a time dimension; 3.3 Processing typho