Correction method suitable for sea-gas coupling mode typhoon generation season prediction

Abstract

The invention provides a correction method suitable for sea-gas coupling mode typhoon generation season prediction. The correction method comprises the following steps: determining basic parameters; establishing a key meteorological element error prediction basic model, generating an error prediction model of each key meteorological element at each grid point through training, and combining to form a model cluster for predicting typhoon errors in a correction region; correcting the predicted value of the sea-gas coupling mode by adopting the model cluster; and performing statistical analysis on the prediction correction value of each key meteorological element of each grid point in the prediction correction area to obtain a correction result of typhoon generation season prediction. According to the correction method suitable for the sea-gas coupling mode typhoon generation season prediction, the accuracy of the sea-gas coupling mode typhoon generation season prediction can be improved, application of machine learning in typhoon numerical prediction is facilitated, and scientific and technological support can be provided for typhoon disaster prevention.

Description

Technical Field

[0001] The present invention belongs to the technical field of typhoon generation prediction and correction, and particularly relates to a correction method applicable to seasonal prediction of typhoon generation in a sea-air coupled model. Background Art

[0002] Typhoons (the present invention generally refers to tropical cyclones) often cause serious disasters, and typhoon prediction is an urgent need for disaster prevention and reduction. Numerical prediction based on a sea-air coupled model is an important method for typhoon prediction. This method predicts typhoon activities by running a sea-air coupled model to obtain prediction data of key meteorological elements. However, there are errors in typhoon prediction based on a sea-air coupled model, especially large errors in seasonal-scale typhoon generation prediction. How to reduce the prediction error of key meteorological elements at the seasonal scale of the sea-air coupled model to improve the accuracy of seasonal typhoon generation prediction is a problem to be solved at present. Summary of the Invention

[0003] In view of the problems existing in the prior art, the present invention provides a correction method applicable to seasonal prediction of typhoon generation in a sea-air coupled model, which helps to solve the above problems.

[0004] The technical solution adopted by the present invention is as follows:

[0005] The present invention provides a correction method applicable to seasonal prediction of typhoon generation in a sea-air coupled model, including the following steps:

[0006] Step S1, determining basic parameters:

[0007] Determine the typhoon generation season and the prediction correction area, and grid the prediction correction area into N grid points; determine M key meteorological elements f 1 , f 2 ,..., f M ; the M key meteorological elements f 1 , f 2 ,..., f M are combined into a key meteorological feature F = f 1 , f 2 ,..., f M ; determine the number m of consecutive historical years for each typhoon generation season prediction;

[0008] Step S2, establishing a basic model for predicting the error of key meteorological elements;

[0009] Step S3, let i = 1;

[0010] Step S4, for grid point i, i = 1, 2,..., N, obtain a sample set S i = {s i1 , si2 ,..., s iQ}; where, Q represents the number of samples in the sample set S i ; for each sample s ik , k = 1, 2,..., Q, is the prediction error sequence of the key meteorological feature F in the consecutive m + 1 historical years at grid point i, expressed as: s ik = {(ΔF i,y , ΔF i,y+1 ,..., ΔF i,y+m-1 ), ΔF i,y+m}; where, ΔF i,Y represents the prediction error of the key meteorological feature F at grid point i in the Y-th historical year, Y = y, y + 1,..., y + m, including M key meteorological elements f 1 , f 2 ,..., f M at grid point i; for each key meteorological element f j , j = 1, 2,..., M, its prediction error Δf i,j,Y in the Y-th historical year at grid point i is the difference between the predicted value of the key meteorological element f j predicted by the ocean - atmosphere coupled model in the Y-th historical year at grid point i and the standard value of the key meteorological element f j in the reanalysis grid data in the Y-th historical year at grid point i;

[0011] Step S5, let j = 1;

[0012] Step S6, use the sample set S i = {s i1 , s i2 ,..., s iQ} to train the key meteorological element error prediction basic model, and obtain the key meteorological element error prediction model Model j of the key meteorological element f j,i at grid point i;

[0013] Step S7, determine whether j is equal to M; if equal, execute Step S8; if not equal, let j = j + 1, and return to Step S6, and loop continuously like this;

[0014] Step S8, determine whether i is equal to N; if equal, execute Step S9; if not equal, let i = i + 1, and return to Step S4, and loop continuously like this;

[0015] Step S9, thus obtain M * N key meteorological element error prediction models, and combine them to form a model cluster for typhoon error prediction in the prediction correction area;

[0016] Step S10. When making actual predictions, each key meteorological element error prediction model in the model cluster is used to predict the corresponding key meteorological element error in the next year of the corresponding grid point based on the sample values in the most recent m historical years, and the error prediction value is obtained, expressed as: ER j,i , representing the error prediction value of the key meteorological element f j at grid point i;

[0017] Step S11. The error prediction value ER j,i is used to correct the predicted value of the key meteorological element f j at grid point i in the next year of the most recent m historical years predicted by the sea-air coupling model, and the corrected predicted value of the key meteorological element f j at grid point i is obtained;

[0018] Among them: The correction method is: the corrected predicted value of the key meteorological element f j at grid point i = the predicted value of the key meteorological element f j at grid point i predicted by the sea-air coupling model - the error prediction value ER j,i ;

[0019] Step S12. Statistical analysis is performed on the corrected predicted values of each key meteorological element at each grid point in the prediction correction area to obtain the correction result of typhoon generation season prediction.

[0020] Preferably, the typhoon generation season is specifically from June to August in summer every year.

[0021] Preferably, the predicted value of the key meteorological element f j at grid point i in the Yth historical year predicted by the sea-air coupling model is specifically the seasonal average value of the prediction result of the key meteorological element f j at grid point i in the typhoon generation season in the Yth historical year predicted by the sea-air coupling model;

[0022] The standard value of the key meteorological element f j at grid point i in the Yth historical year in the reanalysis grid point data is specifically the seasonal average value of the standard value of the key meteorological element f j at grid point i in the typhoon generation season in the Yth historical year in the reanalysis grid point data.

[0023] Preferably, in step S4, for grid point i, i = 1, 2,..., N, the sample set S i ={s i1 , s i2 ,..., s iQ} is obtained specifically as follows:

[0024] Step S4.1: Determine the continuous historical year range as the 1st historical year to the (Q + m)th historical year; among them, the 1st historical year to the (Q + m)th historical year are arranged in ascending order of time.

[0025] Step S4.2: For each key meteorological element f j , obtain the predicted values from the 1st historical year to the (Q + m)th historical year at grid point i predicted by the sea - air coupling model, and form a sea - air coupling model prediction sequence, denoted as: pf i,j,1 , pf i,j,2 ,..., pf i,j,Q+m ;

[0026] For each key meteorological element f j , obtain the standard values of the re - analyzed grid data from the 1st historical year to the (Q + m)th historical year at grid point i, and form a re - analyzed grid data standard value sequence, denoted as: rf i,j,1 , rf i,j,2 ,..., rf i,j,Q+m ;

[0027] Step S4.3: Calculate the difference between the data of each same historical year in the sea - air coupling model prediction sequence pf i,j,1 , pf i,j,2 ,..., pf i,j,Q+m and the re - analyzed grid data standard value sequence rf i,j,1 , rf i,j,2 ,..., rf i,j,Q+m to obtain the prediction error sequence of the key meteorological element f j at grid point i, denoted as: Δf i,j,1 , Δf i,j,2 ,..., Δf i,j,Q+m ;

[0028] Step S4.4: Therefore, at grid point i, for each key meteorological element f j , a prediction error sequence Δf i,j,1 , Δf i,j,2 ,..., Δf i,j,Q+m is obtained;

[0029] Combine the prediction errors of the M key meteorological elements in the same historical year to form the prediction error of the key meteorological feature F at grid point i; specifically, for the hth historical year, h = 1, 2,..., Q + m, the prediction error ΔF i,h of the key meteorological feature F at grid point i = {Δf i,1,h , Δf i,2,h ,..., Δf i,M,h};

[0030] Since there are a total of Q + m historical years, the prediction error sequence of the key meteorological feature F at grid point i is obtained: ΔF i,1 , ΔF i,2 ,..., ΔF i,Q+m ;

[0031] Step S4.5, for the prediction error sequence ΔF of the key meteorological feature F at grid point i i,1 , ΔF i,2 ,..., ΔF i,Q+m , perform window shifting with a step size of 1 and a window length of m + 1, thus obtaining Q samples, specifically:

[0032] The first sample s i1 : ΔF i,1 , ΔF i,2 ,..., ΔF i,1+m ;

[0033] The second sample s i2 : ΔF i,2 , ΔF i,3 ,..., ΔF i,2+m ;

[0034] The third sample s i3 : ΔF i,3 , ΔF i,4 ,..., ΔF i,3+m ;

[0035] And so on

[0036] The Qth sample s iQ : ΔF i,Q , ΔF i,Q+1 ,..., ΔF i,Q+m

[0037] Thus, the sample set S i = {s i1 , s i2 ,..., s iQ}

[0038] Preferably, in step S6, taking the key meteorological element f j as the prediction target variable, using the sample set S i = {s i1 , s i2 ,..., s iQ} to train the key meteorological element error prediction basic model, and obtaining the key meteorological element error prediction model Model j of the key meteorological element f j,i at the grid point i position, specifically:

[0039] Step S6.1, the key meteorological element error prediction basic model is pre-set with an error threshold EE(d) corresponding to each training time d;

[0040] Step S6.2, set the initial value of the training time d to 1;

[0041] Step S6.3, input the samples s i1 , s i2 ,..., s iQ into the key meteorological element error prediction basic model at the same time;

[0042] Among them: for the sample s ik ={(ΔF i,y , ΔF i,y+1 ,..., ΔF i,y+m-1 ), ΔF i,y+m}, k = 1, 2,..., Q, the key meteorological element error prediction basic model uses ΔF i,y , ΔF i,y+1 ,..., ΔF i,y+m-1 as the input sequence, and predicts the predicted value Model_Δf j of the sea-air coupling model prediction error of the key meteorological element f i,j,y+m at the grid point i in the (y + m)-th historical year; then extract the prediction error Δf i,y+m of the key meteorological element f j from the prediction error ΔF i,j,y+m of the key meteorological feature F at the grid point i in the (y + m)-th historical year; use the formula |Model_Δf i,j,y+m -Δf i,j,y+m | to obtain the absolute error E ik of the sample s i,j，k ;

[0043] Therefore, for the samples s i1 , s i2 ,..., s iQ , the absolute errors E i,j，1 , E i,j，2 ,..., E i,j，Q are obtained respectively; take the average value of E i,j，1 , E i,j，2 ,..., E i,j，Q to obtain the mean absolute error of this training

[0044] Step S6.4, judge whether the mean absolute error of this training is less than the error threshold EE(d); if it is less, output the model obtained at this time of training, which is the key meteorological element f jKey meteorological element error prediction model at grid point i, Model j,i ; If it is not less than, adjust the model parameters obtained by training at this time, set d = d + 1, and return to step S6.3.

[0045] Preferably, the error threshold EE(d) corresponding to each training time d is set as follows: as the training time d increases, the error threshold EE(d) first gradually increases, and after increasing to a set value, it remains unchanged.

[0046] Preferably, when using the sample set S i ={s i1 , s i2 ,..., s iQ} to train the key meteorological element error prediction basic model, divide the sample set S i ={s i1 , s i2 ,..., s iQ} into a training set and a test set according to a set ratio;

[0047] First, use the training set to train the key meteorological element error prediction basic model, and then use the test set to test and detect the key meteorological element error prediction basic model obtained by training.

[0048] Preferably, the M key meteorological elements f 1 , f 2 ,..., f M are specifically 7 key meteorological elements;

[0049] The key meteorological elements f 1 , f 2 ,..., f 7 are respectively: relative humidity at the 600-hPa isobaric surface, typhoon potential intensity, zonal wind at the 850-hPa isobaric surface, zonal wind at the 200-hPa isobaric surface, meridional wind at the 850-hPa isobaric surface, meridional wind at the 200-hPa isobaric surface, absolute vorticity at the 850-hPa isobaric surface.

[0050] Preferably, step S12 is specifically:

[0051] For the key meteorological elements f 1 , f 2 ,..., f 7 , their prediction correction values at grid point i are respectively expressed as: g i,1 , g i,2 ,..., g i,7 ;

[0052] Use the following formula to obtain the typhoon potential generation index I at grid point ii :[[]]

[0053] I i =(g i,1 / 50)[[]] 3 *(g i,2 / 70)[[]] 3 *{1 + 0.1*[(g i,3 -g i,4 )[[]] 2 +(g i,5 -g i,6 )[[]] 2 )[[]] 1 / 2}[[]] -2 *|10 5 *g i,7 |[[]] 3 / 2

[0054] Thus, the typhoon potential generation index for each grid point in the prediction correction area is obtained, and the typhoon generation season is predicted through the typhoon potential generation index.

[0055] The correction method for typhoon generation season prediction applicable to the ocean - atmosphere coupled model provided by the present invention has the following advantages:

[0056] The present invention provides a correction method for typhoon generation season prediction applicable to the ocean - atmosphere coupled model. This method can improve the accuracy of typhoon generation season prediction of the ocean - atmosphere coupled model. The present invention not only helps the application of machine learning in typhoon numerical prediction, but also provides scientific and technological support for preventing typhoon disasters. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 is a flowchart of a correction method for typhoon generation season prediction applicable to the ocean - atmosphere coupled model provided by the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0058] In order to make the technical problems, technical solutions and beneficial effects solved by the present invention more clear, the present invention will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0059] Referring to Figure 1 , the present invention provides a correction method for typhoon generation season prediction applicable to the ocean - atmosphere coupled model, including the following steps:

[0060] Step S1, determining basic parameters:

[0061] Determine the typhoon generation season and the prediction correction area, and the prediction correction area is gridded into N grid points; determine M key meteorological elements f 1 , f2 ,..., f M ; M key meteorological elements f 1 , f 2 ,..., f M Combined into the key meteorological feature F = f 1 , f 2 ,..., f M ; Determine the number of consecutive historical years m for each typhoon genesis season prediction;

[0062] In this step, the typhoon genesis season is determined according to the climate characteristics of the prediction correction area. For example, it can be from June to August in summer every year.

[0063] The M key meteorological elements f 1 , f 2 ,..., f M , determined according to the climate characteristics of the prediction correction area. For example, it can be 7 key meteorological elements f 1 , f 2 ,..., f 7 , which are respectively: the relative humidity of the 600-hPa isobaric surface, the potential intensity of the typhoon, the zonal wind of the 850-hPa isobaric surface, the zonal wind of the 200-hPa isobaric surface, the meridional wind of the 850-hPa isobaric surface, the meridional wind of the 200-hPa isobaric surface, and the absolute vorticity of the 850-hPa isobaric surface. The present invention does not limit the types and quantities of the key meteorological elements.

[0064] Step S2, establish a basic model for predicting the errors of key meteorological elements;

[0065] The basic model for predicting the errors of key meteorological elements can be an artificial neural network prediction model. The present invention does not limit the specific structure of the artificial neural network prediction model. As an example, the following structure can be adopted: Establish a Transformer neural network: Set parameters such as an encoder and a decoder. For example, it includes 2 encoder layers, 4 decoder layers, 8 attention heads, and the number of input features is 512; the fully connected hidden layer has 2048 neurons and the output layer has 1 neuron; the activation function is ReLU; the dropout method is used, and its random dropout rate is set to 0.1.

[0066] Step S3, let i = 1;

[0067] Step S4, for grid point i, i = 1, 2,..., N, obtain the sample set S i ={s i1 , s i2 ,..., s iQ}; where, Q represents the sample set S iThe number of samples in the medium; for each sample s ik , where k = 1, 2,..., Q, is the prediction error sequence of the key meteorological feature F in the consecutive m + 1 historical years at grid point i, expressed as: s ik = {(ΔF i,y , ΔF i,y+1 ,..., ΔF i,y+m-1 ), ΔF i,y+m}; where ΔF i,Y represents the prediction error of the key meteorological feature F at grid point i in the Y-th historical year, Y = y, y + 1,..., y + m, and includes the prediction errors of M key meteorological elements f 1 , f 2 ,..., f M at grid point i; for each key meteorological element f j , j = 1, 2,..., M, its prediction error Δf i,j,Y at grid point i in the Y-th historical year is the difference between the predicted value of the key meteorological element f j at grid point i predicted by the ocean - atmosphere coupled model in the Y-th historical year and the standard value of the key meteorological element f j at grid point i in the reanalysis grid data in the Y-th historical year;

[0068] In this step, the predicted value of the key meteorological element f j at grid point i predicted by the ocean - atmosphere coupled model in the Y-th historical year is specifically the seasonal average of the prediction results of the key meteorological element f j at grid point i in the typhoon - generating season predicted by the ocean - atmosphere coupled model in the Y-th historical year; for example, for the key meteorological element of relative humidity at the 600 - hPa isobaric surface, the seasonal average of the relative humidity at the 600 - hPa isobaric surface in June - August of summer in the Y-th historical year predicted by the ocean - atmosphere coupled model is used as the predicted value of the relative humidity at the 600 - hPa isobaric surface.

[0069] The standard value of the key meteorological element f j at grid point i in the reanalysis grid data in the Y-th historical year is specifically the seasonal average of the standard values of the key meteorological element f j at grid point i in the typhoon - generating season in the reanalysis grid data in the Y-th historical year.

[0070] As a specific implementation, for grid point i, i = 1, 2,..., N, obtain the sample set S i = {s i1 , s i2 ,..., s iQ}, specifically:

[0071] Step S4.1: Determine the range of consecutive historical years as from the 1st historical year to the (Q + m)th historical year; among them, the 1st historical year to the (Q + m)th historical year are arranged in ascending order of time.

[0072] Step S4.2: For each key meteorological element f j , obtain the predicted values of the sea - air coupled model at grid point i from the 1st historical year to the (Q + m)th historical year, and form a sea - air coupled model prediction sequence, denoted as: pf i,j,1 , pf i,j,2 ,..., pf i,j,Q+m ;

[0073] For each key meteorological element f j , obtain the standard values of the re - analysis grid data at grid point i from the 1st historical year to the (Q + m)th historical year, and form a re - analysis grid data standard value sequence, denoted as: rf i,j,1 , rf i,j,2 ,..., rf i,j,Q+m ;

[0074] Step S4.3: Calculate the difference between the data of each same historical year in the sea - air coupled model prediction sequence pf i,j,1 , pf i,j,2 ,..., pf i,j,Q+m and the re - analysis grid data standard value sequence rf i,j,1 , rf i,j,2 ,..., rf i,j,Q+m to obtain the prediction error sequence of the key meteorological element f j at grid point i, denoted as: Δf i,j,1 , Δf i,j,2 ,..., Δf i,j,Q+m ;

[0075] Step S4.4: Therefore, at grid point i, for each key meteorological element f j , a prediction error sequence Δf i,j,1 , Δf i,j,2 ,..., Δf i,j,Q+m is obtained;

[0076] Combine the prediction errors of the M key meteorological elements in the same historical year to form the prediction error of the key meteorological feature F at grid point i; specifically, for the hth historical year, h = 1, 2,..., Q + m, the prediction error ΔF i,h of the key meteorological feature F at grid point i = {Δf i,1,h , Δf i,2,h ,..., Δf i,M,h};

[0077] Since there are Q + m historical years in total, the prediction error sequence of the key meteorological feature F at grid point i is obtained: ΔF i,1 , ΔF i,2 ,..., ΔF i,Q+m ;

[0078] Step S4.5, for the prediction error sequence ΔF of the key meteorological feature F at grid point i i,1 , ΔF i,2 ,..., ΔF i,Q+m , with a step size of 1 and a window length of m + 1 for window shifting, Q samples are obtained, specifically:

[0079] The first sample s i1 : ΔF i,1 , ΔF i,2 ,..., ΔF i,1+m ;

[0080] The second sample s i2 : ΔF i,2 , ΔF i,3 ,..., ΔF i,2+m ;

[0081] The third sample s i3 : ΔF i,3 , ΔF i,4 ,..., ΔF i,3+m ;

[0082] And so on

[0083] The Qth sample s iQ : ΔF i,Q , ΔF i,Q+1 ,..., ΔF i,Q+m

[0084] Thus, the sample set S i = {s i1 , s i2 ,..., s iQ} is obtained.

[0085] Step S5, let j = 1;

[0086] Step S6, use the sample set S i = {s i1 , s i2 ,..., s iQ} to train the key meteorological element error prediction basic model, and obtain the key meteorological element f j at the grid point i position, the key meteorological element error prediction model Model j,i ;

[0087] Specifically, this step is as follows:

[0088] Step S6.1, the error threshold EE(d) corresponding to each training times d is preset in the basic model for predicting key meteorological element errors;

[0089] Step S6.2, let the initial value of the training times d be 1;

[0090] Step S6.3, input the samples s i1 , s i2 ,..., s iQ into the basic model for predicting key meteorological element errors at the same time;

[0091] Where: for the sample s ik ={(ΔF i,y , ΔF i,y+1 ,..., ΔF i,y+m-1 ), ΔF i,y+m}, k = 1, 2,..., Q, the basic model for predicting key meteorological element errors takes ΔF i,y , ΔF i,y+1 ,..., ΔF i,y+m-1 as the input sequence, and predicts the predicted value Model_Δf j of the sea-air coupling model prediction error of the key meteorological element f i,j,y+m at the grid point i in the (y + m)-th historical year; then, from the prediction error ΔF i,y+m of the key meteorological feature F at the grid point i in the (y + m)-th historical year, extract the prediction error Δf j of the key meteorological element f i,j,y+m ; use the formula |Model_Δf i,j,y+m - Δf i,j,y+m | to obtain the absolute error E ik of the sample s i,j，k ;

[0092] Therefore, for the samples s i1 , s i2 ,..., s iQ , the absolute errors E i,j，1 , E i,j，2 ,..., E i,j，Q are obtained respectively; take the average value of E i,j，1 , E i,j，2 ,..., E i,j，Q to obtain the mean absolute error of this training

[0093] Step S6.4, judge whether the mean absolute error of this training is less than the error threshold EE(d); if it is less than, output the model obtained by this training, which is the key meteorological element fj Key meteorological element error prediction model Model at grid point i j,i ; If it is not less than, adjust the model parameters obtained by training at this time, let d = d + 1, and return to step S6.3.

[0094] As a preferred method, the error threshold EE(d) corresponding to each training number d is set as follows: as the training number d increases, the error threshold EE(d) first gradually increases, and after increasing to the set value, it remains unchanged.

[0095] Furthermore, when using the sample set S i ={s i1 , s i2 ,..., s iQ} to train the key meteorological element error prediction basic model, divide the sample set S i ={s i1 , s i2 ,..., s iQ} into a training set and a test set according to a set ratio; first use the training set to train the key meteorological element error prediction basic model, and then use the test set to test and detect the key meteorological element error prediction basic model obtained by training.

[0096] Step S7, determine whether j is equal to M; if it is equal, execute step S8; if it is not equal, let j = j + 1, and return to step S6, and loop continuously like this;

[0097] Step S8, determine whether i is equal to N; if it is equal, execute step S9; if it is not equal, let i = i + 1, and return to step S4, and loop continuously like this;

[0098] Step S9, thus obtain M * N key meteorological element error prediction models, and combine them to form a model cluster for typhoon error prediction in the prediction correction area;

[0099] Step S10, when making an actual prediction, use each key meteorological element error prediction model in the model cluster, based on the sample values of the nearest m historical years, to predict the corresponding key meteorological element error in the next year of the corresponding grid point in the nearest m historical years, and obtain an error prediction value, denoted as: ER j,i , representing the error prediction value of the key meteorological element f j at grid point i;

[0100] Step S11, use the error prediction value ER j,i , to correct the predicted value of the key meteorological element f j at grid point i in the next year of the nearest m historical years predicted by the sea - air coupling model, and obtain the key meteorological element f at grid point ij Predicted correction value;

[0101] Wherein: the correction method is: for the key meteorological element f at grid point i j Predicted correction value = Predicted value of the key meteorological element f at grid point i predicted by the ocean - atmosphere coupling model j - Error predicted value ER j,i ;

[0102] Step S12, perform statistical analysis on the predicted correction values of each key meteorological element at each grid point in the predicted correction area to obtain the correction result of typhoon generation season prediction.

[0103] Specifically, when there are M key meteorological elements f 1 , f 2 ,..., f M , specifically 7 key meteorological elements, and the key meteorological elements f 1 , f 2 ,..., f 7 are respectively: relative humidity at the 600 - hPa isobaric surface, typhoon potential intensity, zonal wind at the 850 - hPa isobaric surface, zonal wind at the 200 - hPa isobaric surface, meridional wind at the 850 - hPa isobaric surface, meridional wind at the 200 - hPa isobaric surface, absolute vorticity at the 850 - hPa isobaric surface, for the key meteorological elements f 1 , f 2 ,..., f 7 , their predicted correction values at grid point i are respectively expressed as: g i,1 , g i,2 ,..., g i,7 ;

[0104] Then: use the following formula to obtain the typhoon potential generation index I at grid point i i :

[0105] I i =(g i,1 / 50) 3 *(g i,2 / 70) 3 *{1 + 0.1*[(g i,3 - g i,4 ) 2 +(g i,5 - g i,6 ) 2 1 / 2} -2

[0106] *|10 5 * g i,7 | 3 / 2

[0107] ​The typhoon potential generation index of each grid point in the prediction correction area is obtained, and the typhoon generation season is predicted through the typhoon potential generation index.

[0108] To facilitate the understanding of the present invention, an embodiment is introduced below:

[0109] Step S1, determine the basic parameters:

[0110] In this embodiment, the geographical coordinate range of the prediction correction area is: (0°-71.25°N, 99.375°E-180°), with 39*44 = 1716 grid points;

[0111] The typhoon generation season is from June to August every year; there are 7 key meteorological elements, and the 7 key meteorological elements f 1 , f 2 ,..., f 7 are respectively: the relative humidity of the 600-hPa isobaric surface, the typhoon potential intensity, the zonal wind of the 850-hPa isobaric surface, the zonal wind of the 200-hPa isobaric surface, the meridional wind of the 850-hPa isobaric surface, the meridional wind of the 200-hPa isobaric surface, and the absolute vorticity of the 850-hPa isobaric surface. The number of consecutive historical years m for each typhoon generation season prediction is set to 3.

[0112] Step S2, establish a basic model for predicting the errors of key meteorological elements;

[0113] The basic model for predicting the errors of key meteorological elements is: establish a Transformer neural network, including 2 encoder layers, 4 decoder layers, 8 attention heads, and the number of input features is 512; the fully connected hidden layer has 2048 neurons and the output layer has 1 neuron; the activation function is ReLU; the dropout method is used, and its random dropout rate is set to 0.1.

[0114] Step S3, generate sample data:

[0115] In this embodiment, for each grid point, a total of 80 samples are obtained, that is: Q = 80, and the acquisition method is:

[0116] (1) Determine the sample acquisition time range as: the 1st historical year to the 83rd historical year;

[0117] (2) Download and re-analyze gridded data (such as Gaussian grids or fixed grids, etc.). In this embodiment, it is ERA5 re-analysis gridded data, so as to obtain the standard values of each key meteorological element at each grid point in each historical year. The standard value is specifically the average value of the standard values in June - August of summer; therefore, for each key meteorological element, each grid point obtains a standard value in each historical year; since there are 83 historical years in total, an interannual standard value time series is obtained;

[0118] Obtain the predicted values of the ocean - atmosphere coupled model. Convert the prediction results of the atmospheric component model and the ocean component model in the ocean - atmosphere coupled model into the grid type of re - analysis gridded data. In this embodiment, it is the gridded data of the prediction results of the ICM global ocean - atmosphere coupled model, so as to obtain the predicted values of each key meteorological element at each grid point in each historical year, which is also the average value of the predicted values in June - August of summer; since there are 83 historical years in total, an interannual predicted value time series is obtained;

[0119] Therefore, for each key meteorological element at each grid point, subtract its interannual predicted value from the interannual standard value of the corresponding year to obtain the prediction error of each historical year. Since there are 83 historical years in total, a prediction error time series formed by arranging 83 prediction errors is obtained.

[0120] For example, when the key meteorological element is the relative humidity of the 600 - hPa isobaric surface, obtain the interannual change time series of the seasonal average value of the relative humidity of the 600 - hPa isobaric surface at each grid point in the selected geographical area predicted by the ocean - atmosphere coupled model, and the interannual change time series of the corresponding seasonal average value of the relative humidity of the 600 - hPa isobaric surface at the corresponding grid points in the corresponding geographical area in the re - analysis data. Subtract the relative humidity values of the 600 - hPa isobaric surface in the corresponding years of the two time series to obtain the relative humidity prediction error time series of the 600 - hPa isobaric surface.

[0121] (3) For the same grid point in the same historical year, obtain the prediction errors of 7 key meteorological elements, which are respectively expressed as: Δf 1 , Δf 2 ,..., Δf 7 . Therefore, the overall prediction error ΔF of the same grid point in the same historical year = {Δf 1 , Δf 2 ,..., Δf 7};

[0122] So, for the same grid point, obtain the overall prediction error sequence from the 1st historical year to the 83rd historical year, which is expressed as: ΔF 1 , ΔF 2 ,..., ΔF 83 ;

[0123] Since the number m of consecutive historical years is set to 3, that is, the historical data of 3 consecutive years are required to predict the prediction error of the key meteorological elements in the next year; therefore, in the present invention, in ΔF 1 , ΔF 2 ,..., ΔF 83 , a sample is formed by using the prediction errors of every four consecutive years, that is: ΔF 1 , ΔF 2 , ΔF 3 , ΔF 4 form a sample, ΔF 2 , ΔF 3 , ΔF 4 , ΔF 5 form a sample, and so on, until ΔF 80 , ΔF 81 , ΔF 82 , ΔF 83 form a sample, thus forming 80 samples. For each sample, the training method of its basic model for predicting the error of key meteorological elements will be introduced in the subsequent steps.

[0124] In addition, in practical applications, in the overall prediction error sequence ΔF 1 , ΔF 2 ,..., ΔF 83 of the present invention described above, it is the overall prediction error sequence after normalization, and the normalization method is:

[0125] For each key meteorological element, obtain its prediction errors at all grid points in all historical years, with a total of 1716 * 83 values, and re - represent them as: Δf 1 ', Δf' 2 ,..., Δf' 1716*83 ; take the maximum absolute value in Δf 1 ', Δf' 2 ,..., Δf' 1716*83 , denoted as: Δf max . For each prediction error Δf' x , x = 1, 2,..., 1716 * 83, the following formula is used for normalization to obtain the normalized prediction error Δf x :

[0126]

[0127] where: int() is the function of taking the integer part.

[0128] Through the above normalization method, the prediction errors of each key meteorological element at all grid points in all historical years are scaled down to the interval (-1, 1).

[0129] When the above normalization method is adopted, for the key meteorological element error prediction model of each key meteorological element, the predicted error output by it needs to be denormalized:

[0130] Assume that the predicted error directly output by the key meteorological element error prediction model is Δf * , then the following formula is used for denormalization to obtain the denormalized predicted error Δf ** :

[0131]

[0132] where: Δf' e is the maximum value of the absolute values in Δf 1 ', Δf' 2 ,..., Δf' 1716*83 .

[0133] In practical applications, other methods can also be used for normalization and denormalization, and the present invention does not limit this.

[0134] Step S4. For each key meteorological element at each grid point, 80 samples corresponding to the key meteorological element at the corresponding grid point are used to train and test the key meteorological element error prediction basic model, and the trained key meteorological element error prediction model is obtained.

[0135] (1) Divide the 80 samples into a training set and a test set. The earlier and more continuous groups of samples are used as the training set, and the later and less continuous groups of samples are used as the test set. For example, the training set has 75 samples and the test set has 5 samples.

[0136] (2) Set the training parameters, including: the learning rate is 0.0001, and the number of training rounds and the number of iteration rounds are set. Among them, the number of training rounds is 10, and 50,000 rounds of iterative operations are performed in each training round. The batch_size of the iterative operation is set to 75, the loss function is the mean square error, and the optimizer uses Adam.

[0137] (3) Use the training set to train the key meteorological element error prediction basic model for multiple training rounds and iteration rounds, and use GPU to accelerate the training.

[0138] where: Assume that it is currently necessary to train and generate an error prediction model for the relative humidity (hereinafter referred to as relative humidity) of the 600-hPa isobaric surface at a certain grid point, then:

[0139] In each training round, 75 samples are input. For each sample, with the sample ΔF 1 , ΔF 2 , ΔF 3 , ΔF 4For example, the basic model for predicting the errors of key meteorological elements is based on the prediction errors ΔF of all key meteorological elements in the previous three years. 1 , ΔF 2 , ΔF 3 , the prediction error of the relative humidity in the fourth year is predicted to obtain the predicted value Model_Δf of the prediction error of the sea-air coupling model for relative humidity. Then, the relative humidity prediction error Δf is extracted from the prediction error ΔF in the fourth year, and the sample ΔF is obtained using the formula |Model_Δf - Δf|. 4 , ΔF 1 , ΔF 2 , ΔF 3 , ΔF 4 The absolute error E of;

[0140] A total of 75 samples are input accordingly. Therefore, 75 absolute errors E are obtained, and then the average value of the 75 absolute errors E is taken to obtain the mean absolute error of this training. If the mean absolute error is less than the error threshold EE, the model at this time is output, which is the relative humidity error prediction model at the grid point i position; if it is not less than, the model parameters obtained by training at this time are adjusted and retrained.

[0141] Therefore, for each grid point, 7 key meteorological element error prediction models are finally generated; since there are 1716 grid points in total, 1716 * 7 key meteorological element error prediction models are finally obtained.

[0142] In the present invention, for the same key meteorological element at the same grid point, different error thresholds EE are used in each training round when training the corresponding prediction model.

[0143] As a way, the setting method of the error threshold EE is:

[0144] For example, for a certain grid point, for the relative humidity key meteorological element at the 600-hPa isobaric surface, its annual sequence of prediction errors is: Δf 1 ', Δf' 2 ,..., Δf' 83 ; take the average value of the absolute values of Δf 1 ', Δf' 2 ,..., Δf' 83 as Therefore, the E error threshold EE for the first training round is The threshold of E for the second training round is The thresholds of E for the third to tenth training rounds are

[0145] (4) Use the test set to test the trained prediction model;

[0146] The test set consists of 5 samples. When using the test set to test the trained prediction model, the test method is basically the same as the training method, with the difference that the error threshold EE for each training round is set smaller. For example, the error threshold EE for the first training round is The threshold of E for the second training round is The thresholds of E for the third to the tenth training rounds are

[0147] Step S5, use the key meteorological element error prediction model for actual prediction and correction

[0148] For each key meteorological element at each grid point, load the corresponding key meteorological element error prediction model for the corresponding grid point, input the prediction errors of all key meteorological elements at this grid point in the recent three years, and output the prediction error for the next year of all key meteorological elements at this grid point in the recent three years. Use this prediction error to correct the predicted value of the key meteorological element at this grid point predicted by the ocean-atmosphere coupled model.

[0149] Thus, the correction of the predicted values of the ocean-atmosphere coupled model for all key meteorological elements at all grid points in the prediction correction area is realized.

[0150] Statistically analyze the prediction correction values of each key meteorological element at each grid point in the prediction correction area to obtain the correction results of typhoon generation season prediction.

[0151] It should be emphasized that the present invention is not limited to the specific content mentioned in the previous five steps, and those of ordinary skill in the art can make simple changes or substitutions to it. For example:

[0152] (1) The parameters of the Transformer neural network can be changed, depending on whether it can effectively improve the prediction skill.

[0153] (2) The number of training rounds and the number of iteration rounds can be changed, depending on whether it can effectively improve the prediction skill.

[0154] The present invention provides a correction method applicable to typhoon generation season prediction of the ocean-atmosphere coupled model. This method can improve the accuracy of typhoon generation season prediction of the ocean-atmosphere coupled model. The present invention not only helps the application of machine learning in typhoon numerical prediction, but also provides scientific and technological support for preventing typhoon disasters.

[0155] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A correction method for seasonal prediction of typhoon formation in sea-air coupled model, characterized in that: The following steps are involved: Step S1, determine basic parameters: Determine the typhoon generation season and the forecast correction area, the forecast correction area is gridded into N grid points; determine M key meteorological elements f1, f2, ..., f that affect typhoon generation M ; M key meteorological elements f1,f2,...,f M Combination of key meteorological features F = f1, f2, ..., f M ; Determine the number of consecutive historical years m for each typhoon generation season forecast; Step S2, establishing a basic model for error prediction of key meteorological elements; Step S3, let i=1; Step S4, for grid point i, i = 1, 2, ..., N, obtain sample set S i ={s i1 ,s i2 ,...,s iQ }; where Q represents the sample set S i The number of samples in; for each sample s ik , k = 1, 2, ..., Q, is the forecast error sequence of the key meteorological feature F at grid point i for m+1 consecutive historical years, expressed as: s ik ={(ΔF i,y ,ΔF i,y+1 ,...,ΔF i,y+m-1 ),ΔF i,y+m }; where ΔF i,Y represents the prediction error of the key meteorological feature F at grid point i in the Yth historical year, Y = y, y + 1, ..., y + m, including M key meteorological elements f1, f2, ..., f M Forecast error at grid point i; for each key meteorological element f j , j = 1, 2, ..., M, its prediction error Δf at grid point i in the Yth historical year i,j,Y , is the key meteorological element f at grid point i in the Yth historical year predicted by the ocean-atmosphere coupled model j The predicted value of the key meteorological element f at grid point i in the Yth historical year in the reanalysis grid data j The difference between the standard values ​​of Step S5, set j=1; Step S6, using sample set S i ={s i1 ,s i2 ,...,s iQ } Train the key meteorological element error prediction basic model to obtain the key meteorological element f j Error prediction model of key meteorological elements at grid point i j,i ; Step S7, determine whether j is equal to M; if so, execute step S8; if not, set j=j+1, return to step S6, and repeat the cycle; Step S8, determine whether i is equal to N; if so, execute step S9; if not, set i=i+1, return to step S4, and repeat the cycle; Step S9, thereby obtaining M*N key meteorological element error prediction models, and combining them to form a model cluster for typhoon error prediction in the prediction correction area; Step S10, when performing actual prediction, each key meteorological element error prediction model in the model cluster is used to predict the corresponding key meteorological element error of the corresponding grid point in the next year of the latest m historical years based on the sample values ​​of the latest m historical years, and obtain the error prediction value, which is expressed as: ER j,i , represents the key meteorological element f at grid point i j The error prediction value of Step S11, using the error prediction value ER j,i , the key meteorological elements f of the next year predicted by the ocean-atmosphere coupled model for the grid point i in the most recent m historical years j The predicted value is corrected to obtain the key meteorological element f of grid point i j The forecast revised value of Among them: The correction method is: the key meteorological element f of grid point i j The forecast correction value of = the key meteorological element f of grid point i predicted by the ocean-air coupled model j The predicted value - the error predicted value ER j,i ; Step S12, statistically analyzing the forecast correction value of each key meteorological element at each grid point in the forecast correction area to obtain the correction result of the typhoon formation season forecast.

2. A correction method for seasonal prediction of typhoon formation in sea-air coupling model according to claim 1, characterized in that: The typhoon generation season is specifically the summer period from June to August every year.

3. The correction method for seasonal prediction of typhoon formation in sea-air coupling model according to claim 1, characterized in that: The key meteorological element f predicted by the ocean-atmosphere coupled model at grid point i in the Yth historical year j The predicted value is specifically the key meteorological element f at grid point i in the typhoon generation season of the Yth historical year predicted by the sea-air coupling model. j The seasonal average of the forecast results; The key meteorological element f at grid point i in the Yth historical year in the reanalysis grid data j The standard value of is specifically the key meteorological element f at grid point i in the typhoon generation season of the Yth historical year in the reanalysis grid data. j The seasonal average of the standard values.

4. The correction method for seasonal prediction of typhoon formation in sea-air coupling model according to claim 1, characterized in that: Step S4, for grid point i, i = 1, 2, ..., N, obtain sample set S i ={s i1 ,s i2 ,...,s iQ }, specifically: Step S4.1, determining the range of consecutive historical years as: the 1st historical year to the Q+mth historical year; wherein the 1st historical year to the Q+mth historical year are arranged in ascending order of time; Step S4.2, for each key meteorological element f j , obtain the predicted values ​​of the sea-air coupling model at grid point i from the 1st historical year to the Q+mth historical year, and form the sea-air coupling model prediction sequence, expressed as: pf i,j,1 ,pf i,j,2 ,...,pf i,j,Q+m ; For each key meteorological element f j , obtain the standard values ​​of the reanalysis grid data at grid point i from the 1st historical year to the Q+mth historical year, and form a standard value sequence of the reanalysis grid data, expressed as: rf i,j,1 ,rf i,j,2 ,...,rf i,j,Q+m ; Step S4.3, predict the sequence pf for the air-sea coupled model i,j,1 ,pf i,j,2 ,...,pf i,j,Q+m and the standard value series rf of the reanalysis grid data i,j,1 ,rf i,j,2 ,...,rf i,j,Q+m The data of each same historical year are subtracted to obtain the key meteorological elements f j The prediction error sequence at grid point i is expressed as: Δf i,j,1 ,Δf i,j,2 ,...,Δf i,j,Q+m ; Step S4.4, therefore, at grid point i, for each key meteorological element f j , we obtain the prediction error sequence Δf i,j,1 ,Δf i,j,2 ,...,Δf i,j,Q+m ; The prediction errors of the M key meteorological elements in the same historical year are combined to form the prediction error of the key meteorological feature F at grid point i; specifically, for the hth historical year, h = 1, 2, ..., Q + m, the prediction error of the key meteorological feature F at grid point i is ΔF i,h ={Δf i,1,h ,Δf i,2,h ,...,Δf i,M,h }; Since there are Q+m historical years in total, the prediction error sequence of the key meteorological feature F at grid point i is obtained: ΔF i,1 ,ΔF i,2 ,...,ΔF i,Q+m ; Step S4.5: For the prediction error sequence ΔF of the key meteorological feature F at grid point i i,1 ,ΔF i,2 ,...,ΔF i,Q+m , with 1 as the step size and m+1 as the window length, the window is moved to obtain Q samples, specifically: The first sample s i1 :ΔF i,1 ,ΔF i,2 ,...,ΔF i,1+m ; The second sample i2 :ΔF i,2 ,ΔF i,3 ,...,ΔF i,2+m ; The third sample i3 :ΔF i,3 ,ΔF i,4 ,...,ΔF i,3+m ; And so on The Qth sample s iQ :ΔF i,Q ,ΔF i,Q+1 ,...,ΔF i,Q+m Thus we get the sample set S i ={s i1 ,s i2 ,...,s iQ }.

5. The correction method for seasonal prediction of typhoon formation in sea-air coupling model according to claim 1, characterized in that: Step S6, using key meteorological elements f j To predict the target variable, the sample set S is used i ={s i1 ,s i2 ,...,s iQ } Train the key meteorological element error prediction basic model to obtain the key meteorological element f j Error prediction model of key meteorological elements at grid point i j,i , specifically: Step S6.1, the error prediction basic model of key meteorological elements is pre-set with an error threshold EE(d) corresponding to each training number d; Step S6.2, set the initial value of the number of training times d to 1; Step S6.3, sample s i1 ,s i2 ,...,s iQ At the same time, it is input into the basic model for error prediction of key meteorological elements; Where: For sample s ik ={(ΔF i,y ,ΔF i,y+1 ,...,ΔF i,y+m-1 ),ΔF i,y+m }, k = 1, 2, ..., Q, the error prediction basic model of key meteorological elements is based on ΔF i,y ,ΔF i,y+1 ,...,ΔF i,y+m-1 The input sequence is used to predict the key meteorological factor f at grid point i in the historical year y+m. j The predicted value of the prediction error of the sea-air coupling model Model_Δf i,j,y+m ; Then, the prediction error ΔF of the key meteorological feature F at the grid point i in the y+mth historical year is i,y+m In the paper, we extract the key meteorological elements f j The prediction error Δf i,j,y+m ; Using the formula Model_Δf i,j,y+m -Δf i,j,y+m |, get the sample s ik The absolute error E i,j，k ; Therefore, for sample s i1 ,s i2 ,...,s iQ , and the absolute error E i,j，1 ,E i,j，2 ,...,E i,j，Q ; for E i,j，1 ,E i,j，2 ,...,E i,j，Q Take the average value to get the mean absolute error of this training Step S6.4, determine the mean absolute error of this training Is it less than the error threshold EE(d)? If it is, then the model trained at this time is output, which is the key meteorological element f j Error prediction model of key meteorological elements at grid point i j,i ; If it is not less than, adjust the model parameters obtained by training at this time, set d=d+1, and return to step S6.

3.

6. A correction method for seasonal prediction of typhoon formation in sea-air coupling model according to claim 5, characterized in that: The error threshold EE(d) corresponding to each training number d is set as follows: as the training number d increases, the error threshold EE(d) first increases gradually, and after increasing to a set value, remains unchanged.

7. The correction method for seasonal prediction of typhoon formation in sea-air coupling model according to claim 1, characterized in that: When using the sample set S i ={s i1 ,s i2 ,...,s iQ When training the basic model for error prediction of key meteorological elements, the sample set S i ={s i1 ,s i2 ,...,s iQ } Divide into training set and test set according to the set ratio; First, the training set is used to train the key meteorological element error prediction basic model, and then the test set is used to test the trained key meteorological element error prediction basic model.

8. The correction method for seasonal prediction of typhoon formation in sea-air coupling model according to claim 1, characterized in that: The M key meteorological elements f1, f2, ..., f M , specifically 7 key meteorological elements; The key meteorological elements f1, f2, ..., f7 are: relative humidity of 600-hPa isobaric surface, typhoon potential intensity, zonal wind of 850-hPa isobaric surface, zonal wind of 200-hPa isobaric surface, meridional wind of 850-hPa isobaric surface, meridional wind of 200-hPa isobaric surface, and absolute vorticity of 850-hPa isobaric surface.

9. A correction method for seasonal prediction of typhoon formation in sea-air coupling model according to claim 8, characterized in that: Step S12 is specifically as follows: For the key meteorological elements f1, f2, ..., f7, their forecast correction values ​​at grid point i are expressed as: g i,1 ,g i,2 ,...,g i,7 ; The typhoon potential generation index I of grid point i is obtained by the following formula: i : I i =(g i,1 / 50) 3 *(g i,2 / 70) 3 *{1+0.1*[(g i,3 -g i,4 ) 2 +(g i,5 -g i,6 ) 2 ] 1 / 2 } -2 *|10 5 *g i,7 | 3 / 2 The typhoon potential generation index of each grid point in the forecast correction area is obtained, and the typhoon generation season forecast is made based on the typhoon potential generation index.

Images