Search

CN-121995544-A - Method and device for forecasting medium-small scale extreme precipitation

CN121995544ACN 121995544 ACN121995544 ACN 121995544ACN-121995544-A

Abstract

The invention discloses a medium-small scale extreme precipitation prediction method and device, which comprise the steps of S1, obtaining an atmospheric precipitation GNSS PWV image sequence and a weather radar reflectivity image sequence according to multi-mode weather observation data of a region to be detected, S2, obtaining a water vapor characteristic map and a radar characteristic map which contain long-range dependence according to the atmospheric precipitation GNSS PWV image sequence and the weather radar reflectivity image sequence, S3, carrying out deformable cross attention fusion on the water vapor characteristic map and the radar characteristic map to generate fusion characteristics after time-space alignment, S4, decoding and detail enhancing the fusion characteristics to obtain multi-scale characteristics, and S5, mapping the multi-scale characteristics into a precipitation intensity distribution map of a future continuous period. By adopting the technical scheme of the invention, the accuracy and the image fineness of extreme precipitation prediction are obviously improved.

Inventors

  • WANG JIALE
  • TIAN JINYU
  • SHI CHUANG

Assignees

  • 北京航空航天大学

Dates

Publication Date
20260508
Application Date
20260123

Claims (8)

  1. 1. The method for forecasting the extreme precipitation of the medium and small scale is characterized by comprising the following steps of Step S1, obtaining an atmospheric precipitation GNSS PWV image sequence and a weather radar reflectivity image sequence according to multi-mode weather observation data of a region to be detected; S2, obtaining a water vapor characteristic map and a radar characteristic map which contain long-range dependence according to an atmospheric precipitation GNSS PWV image sequence and a weather radar reflectivity image sequence; s3, carrying out deformable cross attention fusion on the water vapor feature map and the radar feature map to generate fusion features after space-time alignment; S4, decoding and detail enhancing are carried out on the fusion characteristics to obtain multi-scale characteristics; And S5, mapping the multi-scale characteristics into a rainfall intensity distribution map of a future continuous period.
  2. 2. The method for forecasting medium-small-scale extreme precipitation according to claim 1, wherein in step S2, the atmospheric precipitation GNSS PWV image sequence and the weather radar reflectivity image sequence are input into two independent Bi-Mamba branches, and space-time feature coding of linear complexity is performed by using a bidirectional state model SSM, so as to obtain a water vapor feature map and a radar feature map which contain long-range dependence.
  3. 3. The method for forecasting medium and small-scale extreme precipitation according to claim 2, wherein the Bi-Mamba branch comprises a stack of Patch Embedding layers and a plurality of Bi-Mamba modules, and the Bi-Mamba modules capture global space-time dependency in long-sequence meteorological data through forward and backward scanning mechanisms.
  4. 4. The method for forecasting the medium-and small-scale extreme precipitation according to claim 3, wherein in the step S3, PWV features are used as queries (Query), radar features are used as keys (Key) and values (Value), sampling offset is generated through an offset prediction network, radar features are dynamically sampled based on the sampling offset, interaction weights of multi-mode features are calculated through a deformable attention mechanism, and fusion features after time-space alignment are generated.
  5. 5. The method for forecasting medium-small-scale extreme precipitation according to claim 4, wherein in the step S3, the offset forecasting network receives the PWV feature and the radar feature simultaneously, learns the spatial misalignment information between the features through a convolution layer, and outputs the two-dimensional spatial offset for each sampling point.
  6. 6. The method for forecasting medium-small-scale extreme precipitation according to claim 5, wherein in step S3, non-grid-aligned dynamic interpolation sampling is performed on the radar feature map by using the generated spatial offset, so that the sparse-distribution GNSS PWV features achieve adaptive aggregation of effective context information in the high-resolution radar echo.
  7. 7. The method for forecasting medium and small-scale extreme precipitation according to claim 6, wherein in step S4, the fused features are reshaped and upsampled, and meanwhile, local detail features of different receptive fields are captured through a multi-scale convolutional neural network CNN.
  8. 8. A medium and small scale extreme precipitation forecasting device, comprising: the first processing module is used for obtaining an atmospheric precipitation GNSS PWV image sequence and a weather radar reflectivity image sequence according to the multi-mode weather observation data of the region to be detected; the second processing module is used for obtaining a water vapor characteristic map and a radar characteristic map which contain long-range dependence according to the atmospheric precipitation GNSS PWV image sequence and the weather radar reflectivity image sequence; the third processing module is used for carrying out deformable cross attention fusion on the water vapor characteristic diagram and the radar characteristic diagram to generate fusion characteristics after space-time alignment; the fourth processing module is used for decoding and detail enhancing the fusion characteristics to obtain multi-scale characteristics; a fifth processing module for mapping the multi-scale features to a precipitation intensity profile for a future continuous period.

Description

Method and device for forecasting medium-small scale extreme precipitation Technical Field The invention belongs to the technical field of weather disaster forecasting, and particularly relates to a method and a device for forecasting medium-small scale extreme precipitation. Background In recent years, global warming causes frequent extreme weather events, especially disasters such as urban waterlogging and landslide caused by short-time strong rainfall, and the like, so that the public safety of society is a serious challenge. The traditional extreme rainfall prediction mainly depends on a Numerical Weather Prediction (NWP) mode, and is limited by the approximation of a physical parameterization scheme, the uncertainty of an initial field and extremely high calculation resource consumption, and the problems of long cranking time, insufficient space-time resolution, low fixed-point quantitative prediction precision and the like exist in the presence of strong convection weather of the NWP in which local burstiness and rapid evolution are faced, so that the timeliness requirement of short-term prediction (0-6 hours) is difficult to meet. Along with the perfection of a modern meteorological observation system, the multisource remote sensing data provides a rich view angle for capturing a precipitation process. The Global Navigation Satellite System (GNSS) can monitor the water vapor conveying and accumulating process all-weather through inverting the atmospheric Precipitation (PWV) and is regarded as a sensitive expiration early warning signal of strong precipitation, and the weather radar can accurately describe the three-dimensional space structure and strength of a precipitation cloud cluster through echo reflectivity. However, these two types of data have significant multi-modal heterogeneity, GNSS data is typically sparse site data or interpolated low frequency fields, focusing on "water vapor environments", and radar data is a high resolution grid image, focusing on "precipitation results". How to efficiently integrate two data sources with different physical meanings and different spatial resolutions is a key for improving the forecasting precision. In recent years, deep learning has made remarkable progress in weather short-term prediction. Early Convolutional Neural Networks (CNNs) had difficulty capturing time-evolution features, while recurrent neural networks (e.g., LSTM, GRU) while being adept at processing time-series, employing serial computing mode, training was inefficient and had difficulty capturing the historical dependence of very long sequences. Although the following transducer architecture introduces a global attention mechanism and improves the long-distance modeling capability, the quadratic computational complexity of the architecture leads to huge memory consumption when processing high-resolution meteorological image sequences, and efficient reasoning is difficult to realize under limited resources. In addition, in the multi-modal fusion aspect, the existing method mostly adopts simple channel splicing (Concat) or element-by-element addition. However, there is often a spatial misalignment (SPATIAL MISALIGNMENT) of the atmospheric water vapor field and precipitation cloud (e.g., wind shear causes the water vapor center to be misaligned with the precipitation center). The traditional rigid fusion mode cannot adaptively correct the geometric deformation, so that the characteristic fusion is insufficient, and the accurate positioning of a polar precipitation landing zone is limited. Disclosure of Invention In order to solve the problems in the prior art, the invention provides a method and a device for forecasting the extreme precipitation of small and medium dimensions, which are used for solving the problems of low long-sequence modeling efficiency and multi-source data space isomerism in the traditional method by deeply fusing GNSS atmospheric vapor field and weather radar echo data and utilizing a state space model with linear complexity and a self-adaptive feature alignment mechanism, thereby realizing high-efficiency and high-precision short-term forecasting of the extreme precipitation event. In order to achieve the above object, the present invention provides the following solutions: a method for forecasting extreme precipitation of small and medium scale includes Step S1, obtaining an atmospheric precipitation GNSS PWV image sequence and a weather radar reflectivity image sequence according to multi-mode weather observation data of a region to be detected; S2, obtaining a water vapor characteristic map and a radar characteristic map which contain long-range dependence according to an atmospheric precipitation GNSS PWV image sequence and a weather radar reflectivity image sequence; s3, carrying out deformable cross attention fusion on the water vapor feature map and the radar feature map to generate fusion features after space-time alignment; S4, decoding and detail enhancing