A multi-threshold processing hierarchical precipitation revision method, system, device and medium

By extracting features from the native grid and constructing a multi-expert network, and employing dynamic adaptive weights and multi-threshold processing, the systematic bias in precipitation forecasting in tropical regions was resolved, resulting in higher accuracy in precipitation correction and improved product quality.

CN122132704APending Publication Date: 2026-06-02湖南省气象信息中心
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
湖南省气象信息中心
Filing Date
2026-04-30
Publication Date
2026-06-02

AI Technical Summary

Technical Problem

Existing precipitation forecasting technologies suffer from significant systematic biases in tropical regions, including interpolation distortion, underestimation of extreme values, probability distortion, gradient explosion, and abrupt changes in the space physics field, resulting in low accuracy of precipitation corrections and poor quality of gridded products.

Method used

In the native grid coordinate system of the global forecasting system, neighborhood statistical features, hydrodynamic extended features, and temporal evolution features are extracted to construct a multi-dimensional feature matrix. This matrix is ​​trained through a multi-expert network and subjected to graded precipitation correction using dynamic adaptive weights and multi-threshold processing to eliminate spatial abrupt changes and probabilistic distortions.

Benefits of technology

It improves the accuracy of precipitation correction, solves the problems of interpolation distortion, underestimation of extreme values ​​and abrupt changes in the spatial physical field, and enhances the quality of precipitation grid products.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122132704A_ABST
    Figure CN122132704A_ABST
Patent Text Reader

Abstract

This application discloses a multi-threshold processing method, system, device, and medium for graded precipitation correction. The method involves: constructing a sample set; extracting samples from different precipitation level intervals within the sample set to construct a training dataset; determining a dynamic adaptive weight vector; training a first expert network (including a binary classification deep neural network) and a second expert network (including a baseline regression model and a high-level regression model) based on the training dataset and the dynamic adaptive weight vector; determining a probability neighborhood interval; inputting the feature data to be predicted into the trained first and second expert networks to obtain the second probability, the output values ​​of the baseline regression model, and the output values ​​of the high-level regression model; and performing graded precipitation correction using the probability neighborhood interval, the second probability, the output values ​​of the baseline regression model, and the output values ​​of the high-level regression model. This application can improve the accuracy of precipitation correction, thereby improving the quality of precipitation gridded products.
Need to check novelty before this filing date? Find Prior Art