CN-121479664-B - Sand and dust weather assimilation data generation method and device based on AI algorithm fusion and re-optimization

Abstract

The application provides a sand and dust weather assimilation data generation method and device based on AI algorithm fusion and re-optimization, and relates to the technical field of computer science and atmospheric science fusion. The method comprises the steps of obtaining multichannel observation data provided by satellites, introducing a Bayesian optimization process to enable background errors and observation errors to be mutually independent to obtain assimilated analysis field data, constructing a combined optimization framework, responding to the input analysis field data, generating a preliminary optimized data field by using CNN as CNN global features, obtaining SVR local correction features through SVR, and balancing the CNN global features and the SVR local correction features through dynamic weighted fusion. The application obviously improves the reconstruction precision of the large-range sand and dust data by combining and comparing the cooperative processing among the multi-satellite channel data and the actual sand and dust weather process, and provides more accurate support for the construction of the whole data after the current sand and dust assimilation.

Inventors

• CHEN BIN
• WANG HUANZHI
• GUAN XIAODAN

Assignees

• 兰州大学

Dates

Publication Date

20260508

Application Date

20251110

Claims (5)

1. 1. A method of generating sand weather assimilation data, the method comprising: Acquiring multichannel observation data provided by a satellite, wherein the multichannel observation data comprises background field data and observation field data; taking a 4DEnVar method combining four-dimensional variation and set Kalman filtering as a basic assimilation model, introducing a Bayesian optimization process to enable a background error and an observation error to be mutually independent, inputting the multichannel observation data into the basic assimilation model for assimilation calculation, and outputting analysis field data after assimilation; Constructing a joint optimization framework, wherein the joint optimization framework fuses a convolutional neural network CNN and a support vector regression SVR, the joint optimization framework responds to input analysis field data, captures the spatial relevance and multi-scale characteristics of multi-time window data by utilizing a CNN multi-layer convolution structure, generates a preliminary post-optimization data field as CNN global characteristics, carries out secondary modeling on a high-frequency residual through an epsilon-insensitive loss function of the SVR to obtain SVR local correction characteristics, and realizes balance of the CNN global characteristics and the SVR local correction characteristics through dynamic weighted fusion, wherein the high-frequency residual is the difference value between the preliminary post-optimization data field and real observation field data, and takes the fused data as final sand weather assimilation data; capturing the spatial correlation and multi-scale characteristics of multi-time window data by using a CNN multi-layer convolution structure, and generating a preliminary post-optimization data field as CNN global characteristics, wherein the method comprises the following steps: Introducing a dynamic weight function aiming at space self-adaption based on CNN, wherein the dynamic weight function promotes the contribution of assimilation data in an extreme value mutation region through gradient weighting, and simultaneously inhibits oscillation noise introduced in the assimilation process of a stable region based on multi-time window background field similarity; constructing an architecture containing residual connection in the joint optimization framework, and determining a loss function of the architecture; Generating optimization data which simultaneously meets extreme value fidelity and stable region stability as CNN global characteristics by combining the dynamic weight function and the loss function through a cyclic iteration mechanism; the dynamic weight function is expressed as: in the formula, As a function of the dynamic weights, As a background field uncertainty parameter, In order to analyze the gradient of the field data, For background field data, max is the maximum value, exp is an exponential function based on a natural constant, and ISDI is a dust intensity index for quantifying the dust intensity in dust weather; the loss function is expressed as: Where L is the loss function, Analysis field data re-optimized for CNN algorithm, In order to observe the field data, Analyzing field data, wherein TV is the total variation for measuring the spatial fluctuation degree; generating optimized data simultaneously meeting extreme value fidelity and stability region stability as analysis field data after CNN algorithm re-optimization through the following formula: wherein R (x) is a regularization constraint term based on a CNN framework, The loss function of the data is analyzed for background fields and extreme regions, For the loss function of the background field and the gentle region analysis data, X b is background field data, y is observation field data, X is extreme region analysis data, arg min is a variable value which enables a subsequent target function to obtain a minimum value; The dynamic weighting fusion is used for constructing a self-adaptive weight function based on gradient sensitivity to realize dynamic balance of CNN global features and SVR local correction features, and the expression of the self-adaptive weight function is as follows: wherein alpha and beta are contribution weights for controlling the gradient sensitive area and the stable area respectively, The gradient of the CNN global characteristic, namely the observed data gradient after the CNN algorithm is optimized, For SVR local correction characteristics, namely the assimilation data obtained by SVR algorithm correction, Is CNN global characteristic, namely assimilated data corrected by CNN algorithm, W DeepSVR (x) is self-adaptive weight, exp is exponential function based on natural constant, max is maximum value, And locally correcting the characteristic scale parameters of the difference between the SVR and the CNN global characteristic.
2. 2. A method of generating sand weather assimilation data according to claim 1, wherein said satellite comprises an FY4B satellite and/or a Himawari-9 satellite.
3. 3. A sand weather assimilation data generating apparatus for realizing the sand weather assimilation data generating method according to any one of claims 1-2, characterized in that the apparatus comprises: the system comprises a data acquisition module, a data processing module and a data processing module, wherein the data acquisition module is configured to acquire multichannel observation data provided by satellites, and the multichannel observation data comprises background field data and observation field data; The data processing module is configured to take a 4DEnVar method combining four-dimensional variation and set Kalman filtering as a basic assimilation model, introduce a Bayesian optimization process to enable a background error and an observation error to be mutually independent, input the multichannel observation data into the basic assimilation model for assimilation calculation, and output analysis field data after assimilation; The combined optimization module is configured to construct a combined optimization framework, the combined optimization framework fuses a convolutional neural network CNN and support vector regression SVR, the combined optimization framework responds to input analysis field data, a CNN multi-layer convolution structure is utilized to capture spatial correlation and multi-scale characteristics of multi-time window data, a preliminary post-optimization data field is generated to serve as CNN global characteristics, a high-frequency residual is subjected to secondary modeling through an epsilon-insensitive loss function of the SVR to obtain SVR local correction characteristics, the high-frequency residual is a difference value between the preliminary post-optimization data field and real observation field data, balance of the CNN global characteristics and the SVR local correction characteristics is achieved through dynamic weighting fusion, and fused data serves as final sand weather assimilation data.
4. 4. An electronic device comprising a processor and a memory communicatively coupled to the processor; The memory stores computer-executable instructions; The processor executes computer-executable instructions stored by the memory to implement the sand weather assimilation data generation method of any of claims 1-2.
5. 5. A computer-readable storage medium, wherein computer-executable instructions are stored in the computer-readable storage medium, which when executed by a processor, are adapted to implement the method for generating sand weather assimilation data according to any one of claims 1-2.

Description

Sand and dust weather assimilation data generation method and device based on AI algorithm fusion and re-optimization Technical Field The application relates to the technical field of computer science and atmospheric science fusion, in particular to a method and a device for generating assimilation data of sand and dust weather in the crossing field of data assimilation, deep learning and statistical learning. Background The data assimilation technology is widely applied in the field of atmospheric science, and the data assimilation can effectively improve the data precision by combining the observed data with a numerical model. However, for the deep analysis of the environment and climate effect of the dust aerosol, since the existing research is mostly based on the collaborative research in two aspects of multi-source observation and traditional numerical simulation, the data fluctuation between the initial background field and the boundary field further accumulates the error between the simulation data and the live data. For the data assimilation field, the most widely used methods at present include ensemble kalman filtering (EnKF) and four-dimensional variation assimilation (4 DVar). The method is suitable for processing uncertainty propagation of nonlinear system data by constructing an error covariance dynamic structure through multiple disturbance simulation and combining current data, but has lower efficiency and possibly faces numerical instability aiming at a high-dimensional atmospheric model of a multi-time window, and the method is suitable for application requirements of high resolution, multiple time nodes and strong constraint by optimizing a minimum objective cost function of background field observation and analysis field simulation in a certain time window according to an optimization theory. The proposed four-dimensional variation in this context is an improved method in combination with a set kalman filter (4 DEnVar). The statistical characteristics of the system are estimated by introducing an EnKF set assimilation method, and the initial conditions are optimized by combining a variation method, so that the calculation efficiency is improved while the high precision is ensured. In order to optimize the problem of unstable fluctuation of data errors before and after assimilation, an artificial intelligent algorithm, particularly a deep learning method, is also gradually an important supplement for improving the characterization and observation assimilation performance of a complex system, has strong nonlinear fitting capacity and characteristic extraction capacity in the fields of remote sensing inversion, pollution source identification, meteorological variable reconstruction and the like, and greatly expands the application range and effect of assimilation technology. The main technologies of error re-optimization for data assimilation based on artificial intelligence algorithm currently comprise classical neural networks, supervised learning, generation of countermeasure networks and other methods. These techniques have advantages and disadvantages in the context of separate use, with requirements for application scenarios and computing resources. Therefore, when the data assimilation and artificial intelligence algorithm error re-optimization-based sand and dust weather assimilation data construction method is designed, the method is more focused on improving the computing efficiency and the precision and the capability of processing a nonlinear and high-dimensional system, and has important application potential in the fields of extreme sand and dust prediction, sand and dust multi-time window change prediction and the like. Disclosure of Invention The application provides a method and a device for generating dust and sand weather assimilation data, and aims to create a 4DEnVar_deep SVR data assimilation scheme for developing error re-optimization collaborative analysis based on a Python constructed more concise artificial intelligent algorithm. Specifically, the deep learning neural network and the statistical learning algorithm are combined, the post-optimization is performed aiming at errors generated by assimilation of the variation sets, and the final model algorithm can be applied to reconstruction of sand and dust weather process data. The reconstruction accuracy of the large-range sand and dust data is remarkably improved through the combination comparison of the cooperative processing among the multi-satellite channel data and the actual sand and dust weather process, and more accurate support is provided for the integral construction of the current sand and dust data. Specifically, the application develops data error re-optimization based on an artificial intelligence algorithm, and combines deep learning, statistical learning and variation set assimilation algorithm to overcome the systematic deviation of the current data assimilation process in the aspects of initial f