CN-122019921-A - Land-gas coupling data mixed assimilation method based on LSTM introduced space-time hysteresis effect

CN122019921ACN 122019921 ACN122019921 ACN 122019921ACN-122019921-A

Abstract

The invention discloses a land-gas coupling data mixed assimilation method based on LSTM introduced space-time hysteresis effect, which comprises the following steps of generating set members by utilizing RANDOMCV, obtaining a set of set prediction fields by utilizing short-time set prediction, calculating expanded land variable set disturbance to obtain flow-dependent background error information of land information, constructing a long-short memory method model, training a weight coefficient of a time-space hysteresis term by taking a preprocessed satellite and conventional data of a past period as input, adding the trained space-time hysteresis term into an analysis increment based on the analysis increment of mixed assimilation, and carrying out satellite or conventional data assimilation by utilizing a mixed assimilation method of coupling land information control variables. According to the invention, the influence of land information control variables is introduced, and the space-time hysteresis effect caused by the difference between land and gas is realized by utilizing the LSTM model training, so that the model variable information is updated and optimized, and a better analysis field is obtained.

Inventors

SHEN FEIFEI
LIU ZIXIAN
YUAN XIAOLIN
XU DONGMEI
SUN MEI
WANG YI
GUO YAKAI
Shu Aiqing
ZHAO RUONAN

Assignees

南京信息工程大学

Dates

Publication Date: 20260512
Application Date: 20260128
Priority Date: 20250320

Claims (10)

1. A land-gas coupling data mixed assimilation method based on LSTM to introduce space-time lag effect is characterized by comprising the following steps: s1, generating set members by utilizing RANDOMCV random disturbance, and adding land information of a set disturbance part into a mixed assimilation system as an expansion control variable to obtain an expanded set disturbance variable; s2, obtaining a group of integrated forecasting fields by utilizing short-time integrated forecasting, and calculating the integrated average value and standard deviation of each variable in the integrated members so as to calculate the expanded disturbance of the land variable integrated to obtain the flow-dependent background error information of the land information; s3, constructing an LSTM model, wherein the LSTM model comprises an input sequence, a weight training module, a standard gating equation and a time space convolution module; S4, constructing a mixed assimilation analysis increment considering a space-time hysteresis effect based on the introduced land information control variable, so that the influence of the land space-time state difference and the flow dependence background error information of the land information can be fully considered, and performing data assimilation of satellites or conventional data based on a newly constructed coupled land information control variable and a mixed assimilation method of the space-time hysteresis effect.
2. The method for mixed assimilation of land-gas coupling data based on LSTM induced space-time lag effect according to claim 1, wherein in said S1 land information of aggregate disturbance part is added in the mixed assimilation system, control variable of aggregate flow dependent part is expanded, and the expanded aggregate disturbance variable The method comprises the following steps: (1) disturbance is set for atmospheric variables; disturbance is set for the extended land variable.
3. The land-gas coupling data mixed assimilation method based on LSTM introduced time-space hysteresis effect according to claim 1 is characterized in that in S2, a group of integrated forecast fields is obtained by utilizing short-time integrated forecast on the basis of random disturbance generation of integrated members, integrated disturbance information of variables is obtained by calculating integrated mean value and standard deviation of all integrated member variables in the integrated forecast fields, then expanded land variable integrated disturbance information is calculated, discrete degree of land variables in the integrated members is quantitatively represented by calculating mean value and standard deviation of all variables among the members and used for directly quantifying surface variable uncertainty of different spatial positions and moments, stream dependent background error information is obtained based on calculation results of the mean value and the standard deviation, time-space variation characteristics of land variable errors along with weather system evolution can be dynamically reflected, and formulas of land information mean value and standard deviation are as follows: (2) (3) (4) Wherein the method comprises the steps of Is the average value of the land-based variables, Land variable representing kth member, K being the number of collection members; For a disturbance of the set of land-based variables, Representing land variable set perturbation generated by a kth set member; stdv is the standard deviation of land variable.
4. The method for mixed assimilation of land-gas coupling data based on LSTM induced space-time hysteresis effect according to claim 1, wherein the LSTM model is used for training the weight and bias coefficient of the space-time hysteresis effect, wherein the space-time hysteresis effect is considered in the construction of the loss function, and after the space-time hysteresis term is added to the land variable, the space-time hysteresis term is subjected to mixed assimilation and the indirect constraint on the time weight and bias is carried out according to the prediction error of the short-time prediction; the training steps include (1) giving initial values of coefficients by using an Xavier/Glorot initialization method, wherein the initial values of bias coefficients are set to 0, (2) inputting past time series data and the initialized coefficients into a standard gating equation to calculate and obtain hidden states h (t), (3) substituting the hidden states h (t) into time and space weight formulas to obtain two results, (4) substituting time and space weights into time and space hysteresis expressions to obtain final time and space hysteresis terms, (5) using past assimilation results with time and space hysteresis terms after training to conduct short-time prediction and comparing the prediction results with observation data with corresponding time to construct a loss function S with time and space hysteresis effect, (6) checking the change condition of the loss function, skipping a training circulation process when the loss function reduces the amplitude to meet the requirement and taking the final generated results as optimal weights and biases, (7) calculating gradients of the weights and the bias coefficients based on the loss function, (8) updating the weights and the bias coefficients by using SGD after calculating the gradients and substituting the gradients into the LSTM model, and (5) executing the steps (2) - (8).
5. The LSTM-introduced spatio-temporal hysteresis based hybrid assimilation method of land-gas coupled data according to claim 1, wherein in said S3, the LSTM model first needs to train weights and bias coefficients of time and space lags using assimilation data of a past period of time as an input sequence before calculating a spatio-temporal hysteresis term, the input sequence expression is as follows: (5) For the input sequence, t is the target assimilation moment, And The analysis increment of the atmosphere and the land surface are shown, and M represents the assimilation of the past M times.
6. The method for mixed assimilation of land-gas coupling data based on LSTM introduced time-space hysteresis effect according to claim 1, wherein the weight training module utilizes forward propagation during training, firstly utilizes Xavier/Glorot to initialize and give out first weight and bias coefficient, then substitutes past time series data and makes forward propagation, adds hysteresis term into land variable analysis increment after calculating first time-space hysteresis term to make mixed assimilation and make short-time forecast, then uses result with hysteresis term in the past to make short-time forecast, compares forecast result with actual observation and calculates loss, then adopts BPTT to calculate gradient according to loss and utilizes SGD to update and optimize weight and bias coefficient, training of weight is completed in advance before calculating time hysteresis term of current assimilation time, after training of weight and bias coefficient, land gas data of window area of current assimilation time is put into input sequence, then input sequence is transferred into standard gate time-space control equation to make calculation, finally substitutes convolution module after calculation, convolution module is as follows: (6) (7) (8) (9) (10) Wherein the method comprises the steps of 、 , 、 , 、 The weights and bias coefficients are trained during the training process, For time weights, h (t) is the hidden state in the standard gating equation, To the extent to which different assimilation moments have contributed to the time-lag effect in the past, For the time lag term, L is the total time step length, Is a lattice point At the position of The land increment of the moment in time, As the weight of the spatial offset, For an offset propagation delay, As a term of the spatial lag, ) Is a position deviation Delay of Land delta of time, N is a spatial neighborhood.
7. The LSTM-introduced spatio-temporal hysteresis-based hybrid assimilation method of land-gas coupled data according to claim 4, wherein the loss function expression under consideration of the spatio-temporal hysteresis effect in the training process is as follows: (11) Wherein n represents the total number of samples, For short forecast with t as initial time The variation of the land surface of the rear part, To correspond to The actual observed land variable of the moment.
8. The LSTM-introduced spatio-temporal hysteresis based hybrid assimilation method of land-gas coupled data according to claim 1, wherein the analysis increment defining land-surface variables in the hybrid assimilation system is: (12) For land information increment brought by 3DVar static background error covariance, since land information is not introduced into 3DVar as control variable, land information increment of static part Setting to 0, K is the number of the collection members; Delta of land information variable for aggregate provided stream dependent error background information, wherein A control variable vector representing a K-th set member, the control variables including an atmospheric variable and a land information variable; The term of time-space lag is time lag term And a spatial lag term And, it takes into account the hysteresis effect of land with respect to the atmosphere; The analytical increment for defining the atmospheric variables in the mixed assimilation system is: (13) Wherein the method comprises the steps of Is the increase in atmospheric information from the 3DVar static background error covariance, Represents the atmospheric information increment brought by the collection, Representing the set perturbation of the atmospheric variables generated by the kth set member.
9. Computer equipment, characterized in that it comprises a memory, a processor and a computer program stored in the memory and running on the processor, the processor executing the steps of the method for realizing the hybrid assimilation of land-gas coupled data based on LSTM induced space-time hysteresis effect according to any one of claims 1-8.
10. A computer-readable storage medium, wherein the computer-readable storage medium stores a computer program for performing the LSTM-introduced spatio-temporal hysteresis-based land-gas coupled data hybrid assimilation method according to any of claims 1 to 8.

Description

Land-gas coupling data mixed assimilation method based on LSTM introduced space-time hysteresis effect Technical Field The invention relates to a land-gas coupling data mixed assimilation method based on LSTM introduced space-time hysteresis effect, belonging to the technical field of numerical simulation and data assimilation. Background In modern numerical weather forecast and climate prediction, land information is used as a key element of land-gas interaction, and directly influences landing energy, moisture exchange and the evolution of boundary layer processes. The land surface humidity, temperature, soil humidity and other surface information play an important role in the formation and development of weather systems. For example, land temperature affects the intensity of surface evaporation and latent heat flux, which are key factors for convection and precipitation triggering, and soil humidity and temperature affect surface sensible heat flux and local atmospheric stability, further regulating the development of atmospheric flows and weather systems. In addition, land information is affected by various environmental factors, such as vegetation coverage, soil state, season difference, etc., and is a slower process relative to the change of the atmosphere, so that the data observed by satellites and the like does not actually reflect the real state of the land information, and space-time hysteresis effects exist between the land and the air. Conventional control variable settings (e.g., temperature, humidity, and wind farm) typically do not take into account interactions between land information and atmospheric information, weakening the impact of land information in an assimilation system. Meanwhile, the traditional assimilation system has insufficient consideration of land information characteristics in the assimilation of satellite and other data, and in most cases, the space-time hysteresis of land gas is ignored, so that real information of land cannot be fed back correctly in the assimilation process, and the final result and the actual observation are large in phase diameter. Moreover, in dealing with the space-time hysteresis effect, the conventional method often depends on a fixed hysteresis window or only on static background error covariance (such as 3D-Var), and cannot be adjusted according to the actual situation when facing seasonal changes or surface situation changes. This can deviate the adjustment of the space-time lag of the land from the actual lag situation, making it difficult to positively improve the forecast and possibly even negatively. Disclosure of Invention Aiming at the problems and the defects existing in the prior art, the invention provides a mixed assimilation method which is used for coupling land information control variables and taking the space-time hysteresis characteristic of land gas into consideration based on long short memory (LSTM) model training. The land information control variable is expanded by utilizing the aggregation/variation mixing and assimilation method, and the land information is considered to be slower than the atmospheric variation, so that the space-time hysteresis effect is added into the land analysis increment and is integrated into the assimilation system, the coupling between the land information and the atmospheric information is enhanced, the influence is correctly fed back, the analysis and the forecast of a weather system are improved, and the accuracy of numerical weather forecast and weather forecast is improved. The technical scheme is that the land-air coupling data mixed assimilation method based on LSTM introduced space-time hysteresis effect combines expanded land information control variables and considers the space-time hysteresis effect under LSTM model training, fully considers the space-time difference of land-air data and the interaction of land-air, improves the analysis and forecast of the land-air information and the atmosphere information in the mode, and comprises the following contents: s1, generating set members by utilizing RANDOMCV random disturbance, and adding land information of a set disturbance part into a mixed assimilation system as an expansion control variable to obtain an expanded set disturbance variable; s2, obtaining a group of integrated forecasting fields by utilizing short-time integrated forecasting, and calculating the integrated average value and standard deviation of each variable in the integrated members so as to calculate the expanded disturbance of the land variable integrated to obtain the flow-dependent background error information of the land information; s3, constructing an LSTM model, wherein the LSTM model comprises an input sequence, a weight training module, a standard gating equation and a time space convolution module; S4, constructing a mixed assimilation analysis increment considering a space-time hysteresis effect based on the introduced land information control variable,