Search

CN-120337729-B - AttUnet-based ECMWF mode element deviation correction method

CN120337729BCN 120337729 BCN120337729 BCN 120337729BCN-120337729-B

Abstract

The invention provides an ECMWF mode element deviation correcting method based on AttUnet, and belongs to the field of weather forecast. The method is characterized in that a deep convolutional neural network is used for outputting the surface 4 factors of 2m temperature, surface air pressure, 2m specific humidity and 10m wind in an ECMWF mode, and deviation correction is carried out on an ECMWF mode output result with the resolution of 0.125 degrees. According to the method, firstly, a feature library is constructed according to ECMWF mode output elements, on the basis, a XgBoost tool is used for carrying out feature analysis on a sample library and sequencing the importance of the feature library, and meanwhile, factors are further screened by combining with manual experience and terrain information is considered. And then, predicting and correcting deviation by using a deep convolutional neural network to obtain a more accurate 0.125-degree grid product. Meanwhile, in the model debugging stage, the model has good forecasting performance on extreme disastrous weather by optimizing the learning rate and the loss function.

Inventors

  • DONG JUN
  • HUANG JUN
  • WU GUANGSHENG
  • PAN QIXIN
  • CUI SHUANGSHUANG
  • CHEN SHAOWEI

Assignees

  • 广州市粤港澳大湾区气象智能装备研究中心

Dates

Publication Date
20260508
Application Date
20250328

Claims (10)

  1. 1. An ECMWF mode element bias correction method based on AttUnet, comprising: the method comprises the steps of obtaining historical mode forecast data, label data and geographical static data, wherein the historical mode forecast data is ECMWF mode forecast data of 0-72 hours, the label data is HRCLDAS data, the label data comprises weather four elements including a two-meter temperature element, a two-meter specific humidity element, surface air pressure and a decameter wind element, and the geographical static data is DEM data; The method comprises the steps of determining relevant characteristic factors according to ECMWF mode data preliminarily, and constructing a preselected characteristic factor library, wherein the characteristic factors comprise vertical layer height, vertical layer weft wind, vertical layer vertical velocity, vertical layer divergence, vertical layer vorticity, vertical layer specific humidity, vertical layer temperature, two-meter specific humidity, two-meter dew point temperature, decameter u wind component, decameter v wind component, surface air pressure, ground temperature, surface albedo, atmospheric column water content, total cloud amount and precipitation amount; using XgBoost tools to sort the importance of the feature factors of the meteorological four elements according to the importance analysis scoring requirement, and screening to obtain features with scoring meeting the requirement; Combining the physical law with the screening operation input by the forecaster, and further screening the characteristics meeting the requirements of the scores to obtain the element correction characteristics of the four meteorological elements; Processing ECMWF mode data and HRCLDAS data into a required data set according to the element correction characteristics, screening data samples of the data set for each time, removing sample data with data quality not meeting requirements to obtain removal result data, and carrying out normalization processing on the removal result data to obtain a standardized result; Training a depth learning model based on AttUnet architecture by utilizing the standardized result to obtain a target deviation correction model; Performing deviation correction on the meteorological four elements output by the mode by using the target deviation correction model; The two meter specific humidity factor employs the following loss function: ; Where loss 1 represents the loss function of a two meter specific humidity element, y i represents the true value of the ith sample of a two meter specific humidity element, A predicted value representing an ith sample of the two meter specific humidity element; the decameter wind element adopts the following loss function: ; where loss 2 represents the loss function of a decade wind element, Representing the true value of the ith sample of the decade wind element, Representing the predicted value of the ith sample of the decade wind element.
  2. 2. The method for correcting the deviation of the ECMWF mode element based on AttUnet as set forth in claim 1, wherein the four meteorological elements are characterized by: The characteristic factors corresponding to the two-meter temperature elements are ground temperature, earth surface albedo, total water content of an atmosphere column, two-meter temperature characteristic factors and DEM; The characteristic factors corresponding to the two-meter specific humidity element are 850hPa specific humidity, two-meter specific humidity characteristic factors and DEM; The characteristic factors corresponding to the surface air pressure elements are 500hPa potential height, 850hPa potential height, two-meter temperature, surface air pressure characteristic factors and DEM; The characteristic factors corresponding to the decameter wind elements are 850hPa warp wind, 850hPa weft wind, a surface air pressure characteristic factor, a decameter u wind component characteristic factor, a decameter v wind component characteristic factor and a DEM.
  3. 3. The method for correcting the deviation of the ECMWF mode element based on AttUnet as set forth in claim 1, wherein the normalizing the culling result data is specifically: the data set is normalized according to the following formula: x a =（x i -min(x))/(max(x)-min(x)); Where min (x) represents the minimum value in the pre-normalization dataset, max (x) represents the maximum value in the pre-normalization dataset, x i represents the i-th element in the pre-normalization dataset, and x a represents the post-normalization value.
  4. 4. The method for correcting the deviation of the ECMWF model element based on AttUnet as set forth in claim 1, wherein the training of the deep learning model based on AttUnet architecture is performed to obtain a target deviation correction model, specifically: When training a deep learning model, optimizing model training parameters by adopting an early-stop method and a simulated annealing strategy, evaluating the output of the deep learning model by adopting RMSE, designing different loss functions according to different element characteristics, evaluating extreme weather of the output of the deep learning model by using POD, CSI and FAR, and adjusting and optimizing the loss functions and learning rate of the deep learning model according to various evaluation index results to obtain a target deviation correction model.
  5. 5. The ECMWF pattern element bias correction method based on AttUnet as defined in claim 4, further comprising, prior to optimizing the deep learning model using a loss function, an early stop method, and a simulated annealing strategy: The standardized results are divided into a training set, a verification set and a test set according to the proportion of 7:1.5:1.5.
  6. 6. The ECMWF pattern element bias correction method based on AttUnet of claim 1, further comprising, prior to obtaining the historical pattern forecast data, the tag data, and the geo-static data: And averaging HRCLDAS data and DEM data to a 12.5km area grid through resolution processing.
  7. 7. The method for correcting element bias of ECMWF pattern based on AttUnet as set forth in claim 6, wherein a calculation formula for performing the resolution processing on the HRCLDAS data is specifically: ; Wherein T (i,j) is the initial data of the HRCLDAS data, T (i,j) is the HRCLDAS data subjected to the resolution processing, i represents the row number index of the grid, and j represents the column number index of the grid.
  8. 8. The method for correcting element bias of ECMWF mode based on AttUnet as claimed in claim 6, wherein the calculation formula for performing the resolution processing on the DEM data is specifically: ; wherein T (i,j) is the initial DEM data, the resolution is 0.001 degrees, T (i,j) is the DEM data processed by the resolution, i represents the row number subscript of the grid, and j represents the column number subscript of the grid.
  9. 9. The ECMWF pattern element bias correction method based on AttUnet as defined in claim 6, wherein the screening of the data samples of each time of the data set is specifically: removing HRCLDAS data with the correlation coefficient smaller than 0.9 by calculating the correlation coefficient between HRCLDAS data of the previous and subsequent times; The calculation formula of the correlation coefficient comprises: ; Wherein R is a correlation coefficient calculation result, X and Y respectively represent front and back time data, Representing the average value of a certain data set, Representing the average of the other data set, i represents the number of samples, and N is the total number of samples used to calculate the correlation coefficient.
  10. 10. An ECMWF pattern element bias correction method based on AttUnet as claimed in claim 1, characterized in that data outside a preset historical extremum are rejected before or after normalization of the dataset.

Description

AttUnet-based ECMWF mode element deviation correction method Technical Field The invention relates to the field of meteorological data processing, in particular to an ECMWF mode element deviation correcting method based on AttUnet. Background The errors of the weather pattern and the initial value errors are all important factors causing the deviation of weather forecast, and the errors have different performances in different areas, seasons and weather processes. Therefore, at least one error correction is required before the actual forecast is made. The traditional error correction method has various defects, such as complex flow, low calculation efficiency, unsatisfactory correction effect and the like. In addition, the accuracy of the correction is limited by personal knowledge and experience based on subjective correction methods. In order to improve the accuracy of pattern release and reduce the workload of subjective correction, developing an objective better pattern release algorithm has become a urgent issue for the current business. The existing mode release methods can be roughly classified into a complete prediction method (perfect prognostic, abbreviated as PP) method or a mode output statistics (model output statistics, abbreviated as MOS) method for predicting meteorological element values. However, these methods are based on a multiple linear regression model, and cannot reasonably describe nonlinear relationships among meteorological elements, geographic elements and predicted objects, so that accuracy of meteorological element prediction is poor. The former performs the correction by building a linear or simple nonlinear statistical model between the observed values and the mode forecast values, and the latter performs the correction by building a linear or simple nonlinear statistical model between the observed values and a set of associated atmospheric variable numerical forecast estimates. Most of the release methods are multiple linear regression models, nonlinear relations among meteorological elements, geographic elements and forecast objects cannot be reasonably described, self-adaption capability is poor, and element forecast accuracy is required to be improved. Along with the continuous development of numerical weather forecast and interpretation application technology thereof, multidimensional and multielement information in weather big data is excavated by utilizing a deep learning technology, a release method of an artificial intelligence technology is developed, the fine forecast level of weather elements is further improved, and the method is a development trend of the application of numerical forecast products in recent years. Disclosure of Invention The invention provides an ECMWF mode element deviation correcting method based on AttUnet, which aims to solve the technical problem of how to improve accuracy of meteorological element deviation correction. In order to solve the above technical problems, the present invention provides a method for correcting an ECMWF mode element bias based on AttUnet, including: The method comprises the steps of obtaining historical mode forecast data, label data and geographical static data, wherein the historical mode forecast data is ECMWF mode forecast data of 0-72 hours, the label data is HRCLDAS data, the label data comprises weather four elements including two-meter temperature elements, two-meter specific humidity elements, surface air pressure and ten-meter wind elements, and the geographical static data is DEM data; The method comprises the steps of determining relevant characteristic factors according to ECMWF mode data preliminarily, and constructing a preselected characteristic factor library, wherein the characteristic factors comprise vertical layer height, vertical layer latitudinal wind, vertical layer vertical velocity, vertical layer divergence, vertical layer vorticity, vertical layer specific humidity, vertical layer temperature, 2m specific humidity, 2m dew point temperature, decameter u wind component, decameter v wind component, surface air pressure, ground temperature, surface albedo, atmospheric column water content, total cloud amount and precipitation amount; using XgBoost tools to sort the importance of the feature factors of the meteorological four elements according to the importance analysis scoring requirement, and screening to obtain features with scoring meeting the requirement; Combining the physical law with the screening operation input by the forecaster, and further screening the characteristics meeting the requirements of the scores to obtain the element correction characteristics of the four meteorological elements; Processing ECMWF mode data and HRCLDAS data into a required data set according to the element correction characteristics, screening data samples of the data set for each time, removing sample data with data quality not meeting requirements to obtain removal result data, and carrying out norm