A typhoon gale prediction method and system based on wind-terrain multi-scale spatial correlation consideration

Abstract

The application relates to a typhoon gale prediction method and system based on consideration of wind-terrain multi-scale space correlation, a data preprocessing and prediction target establishment module acquires typhoon wind field and terrain data in the future 1 hour, and performs standardization processing; current time meteorological observation station data is interpolated into grid data, and a difference value of the grid data and the data in the future 1 hour is calculated and taken as a prediction target; a multi-scale wind-terrain space correlation extraction module inputs the standardized data into a network, and through feature high-dimensional projection, multi-scale fusion and full connection prediction, a refined prediction result of a typhoon wind field grid is output; and a model optimization and parameter adjustment module enhances the gale area weight through a weighted loss function, and avoids wind speed distribution compression; the problems that the existing typhoon wind field prediction method is insufficient in wind field prediction accuracy and reliability under complex terrain and extreme typhoon path conditions, and is defective in stability, adaptability and refined simulation are solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of typhoon wind field reconstruction and simulation, and relates to a typhoon gale prediction method and system based on wind-terrain multi-scale spatial correlation consideration, in particular to a typhoon gale fine prediction method and system based on wind-terrain multi-scale spatial correlation consideration under the influence of mountainous terrain. BACKGROUND

[0002] In recent years, with the development of meteorological numerical prediction technology, remote sensing observation means and artificial intelligence methods, the prediction method of typhoon wind field has been continuously enriched, and the accuracy and calculation efficiency of wind field simulation have been improved. Traditional wind field modeling methods based on physical parameterization model, empirical formula or numerical simulation gradually evolve towards the direction of fusion of multi-source data, statistical learning and deep neural network driving, which greatly promotes the progress of typhoon wind field prediction technology. However, the existing methods still have certain limitations in dealing with complex terrain and extreme typhoon weather conditions. On the one hand, the traditional model based on empirical formula or physical parameterization is prone to systematic deviation in complex underlying surface area, and it is difficult to accurately depict the local wind field variation characteristics. On the other hand, although the wind field modeling method based on deep learning improves the overall fitting ability, there is a problem of underestimating extreme wind speed. In addition, some methods rely on a large amount of numerical simulation data or high computational load in the modeling process, resulting in slow reasoning speed and limited real-time application. Although statistical learning methods can achieve fast correction in specific scenarios, they have limited ability to capture the complex nonlinear relationship between terrain and wind field, and lack adaptability.

[0003] Among them, the Chinese invention patent CN116861766A discloses a typhoon wind field prediction method and system based on a generative adversarial network. The method is based on the idea of multiple nested generative adversarial network, and adds a typhoon wind field prediction model outside the typhoon cloud system prediction model. The typhoon wind field prediction model based on the generative adversarial network can efficiently generate simulated typhoon wind field results.

[0004] The Chinese invention patent CN111598301A discloses a typhoon wind field correction method and device combining multiple algorithms and a readable storage medium. The method combines the traditional linear correction method with the nonlinear method by using the selected influence factors, and on the basis of eliminating systematic errors, it further corrects random errors, so as to obtain more accurate typhoon wind field prediction results.

[0005] The Chinese invention patent CN118779639A discloses a typhoon wind field generation method based on marine satellite observation, belonging to the field of meteorological science and numerical weather prediction. When predicting the typhoon field through the Holland theoretical model, the invention uses the BP neural network algorithm to predict the key parameter B of the pressure profile affecting the intensity of the typhoon. By combining the typhoon field obtained by the Holland theoretical model and the ERA5 wind field data, a more refined typhoon field is obtained.

[0006] Although the existing typhoon wind field prediction and correction techniques have made some progress, there are still many deficiencies. For example, the typhoon wind field prediction method based on generative adversarial network (CN116861766A) can efficiently generate simulated wind fields, but due to the instability of generative adversarial network training and pattern collapse, it may lead to insufficient diversity and accuracy of wind field prediction under extreme weather conditions, especially in the process of large-scale typhoon events. The correction method combining multiple algorithms (CN111598301A) can eliminate systematic errors and random errors, but it is less adaptable when dealing with complex terrain and typhoon path changes. The typhoon field generation method based on model superposition (CN118779639A) has made breakthroughs in improving the refinement of wind fields, but due to its dependence on Holland theory and ERA5 data, it is still difficult to accurately capture complex terrain or local wind field changes. In addition, due to the data lag of ERA5 data, it cannot be applied to real-time application scenarios. Overall, the existing technology still has limitations in stability, adaptability and refined simulation, especially in complex terrain and extreme typhoon path conditions, the accuracy and reliability of wind field prediction need to be improved. SUMMARY

[0007] Therefore, the present invention provides a typhoon gale prediction method and system based on wind-terrain multi-scale spatial correlation consideration to solve the problems of insufficient accuracy and reliability of wind field prediction in complex terrain and extreme typhoon path conditions, and defects in stability, adaptability and refined simulation of existing typhoon wind field prediction methods.

[0008] To achieve the above-mentioned purpose, the present invention provides the following technical solutions:

[0009] A typhoon gale prediction method based on wind-terrain multi-scale spatial correlation consideration, comprising the following steps:

[0010] S1, establish a prediction target; obtain future 1-hour tropical cyclone path prediction data, including engineering typhoon wind field data W t+1 =(u t+1 ,v t+1 ) and terrain data T t+1The standardized wind field data W Norm_t+1 and the standardized terrain data T Norm_t+1 are obtained by standardizing the data t+1 The difference between the grid data obtained by interpolating the standardized current-time meteorological observation station data and the future 1-hour meteorological observation station data is taken as the model prediction target

[0011] S2, predicting the typhoon wind field grid; the standardized wind field, terrain data, and grid observation are input into a multi-scale wind-terrain spatial correlation extraction network to predict the typhoon wind field grid, wherein the multi-scale wind-terrain spatial correlation extraction network is composed of a feature high-dimensional projection module, a multi-scale high-dimensional feature fusion module, and a full-connection prediction module

[0012] S3, optimizing the model parameters; by designing a weighted loss function, the weight of the strong wind area is enhanced, the compression effect of the model on the wind speed distribution in the training process is avoided, and thus the prediction accuracy of the typhoon strong wind is improved, and the extreme wind speed change is better captured.

[0013] Further, step S1 is specifically:

[0014] S11, obtaining the future 1-hour tropical cyclone path forecast data, and obtaining the current-time meteorological observation station data and the future 1-hour engineering typhoon wind field data W t+1 =(u t+1 ,v t+1 ) and terrain data T t+1 according to the path forecast data

[0015] S12, preprocessing the obtained engineering typhoon model wind field data and terrain data, respectively standardizing different data to generate standardized data; the standardized wind field data and the standardized terrain data are respectively represented as W Norm_t+1 and T Norm_t+1 ; the standardization process is as follows:

[0016] W Norm_t+1 = Norm(W t+1 )

[0017] T Norm_t+1 = Norm(T t+1 )

[0018] wherein Norm(*) represents the standardization operation, W t+1 represents the input engineering typhoon wind field data, T t+1 represents the input terrain data, W Norm_t+1 represents the standardized future 1-hour typhoon wind field data, and T Norm_t+1represent the standardized terrain data; the current time weather observation site data is interpolated into a grid using a bicubic interpolation method, and the grid resolution is the same as the W Norm_t+1 and T Norm_t+1 , denoted as W obs_t ;

[0019] S13, the difference between the standardized current time weather observation site data and the future 1 hour weather observation site data is taken as the model prediction target, and the prediction target is calculated as:

[0020]

[0021] wherein W target_t+1 is the future 1 hour predicted wind field, represents a deep learning network.

[0022] Further, the tropical cyclone path prediction data in step S11 is usually the China Meteorological Administration prediction data provided by the Central Meteorological Observatory.

[0023] Further, step S2 is specifically:

[0024] S21, the standardized wind field and terrain data are input into a feature high-dimensional projection module PM, and the feature high-dimensional projection module PM outputs a total high-dimensional feature F total of the wind field and terrain, and the calculation formula of the feature high-dimensional projection module PM is as follows:

[0025] F W = PM(W Norm_t+1 , W obs_t+1 )

[0026] F T = PM(T Norm_t+1 )

[0027] F total = Concatenate(F W , F T )

[0028] wherein PM(*) represents a feature high-dimensional projection module, F W represents a high-dimensional typhoon wind field feature processed by the feature high-dimensional projection module, F T represents a high-dimensional terrain feature processed by the feature high-dimensional projection module, Concatenate(*) represents a splicing operation, and F total represents a total high-dimensional feature of the typhoon wind field and terrain;

[0029] Each feature high-dimensional projection module PM is composed of three convolution modules, two Maxpooling pooling layers, and an attention layer, each convolution module includes a convolution layer and a Silu activation layer, and the calculation process of each feature high-dimensional projection module PM is:

[0030] F1 = MaxPooling(BN(SiLU(Conv(Input)))

[0031] F2 = MaxPooling(BN(SiLU(Conv(F1)))

[0032] F3 = BN(SiLU(Conv(F2)))

[0033] F output = Attention(SiLU(Conv(F3)))

[0034] wherein, Input represents the input of the feature extraction module PM, Conv(*) is a convolution operation, SiLU(*) is a SiLU activation operation, BN(*) is a Batch Normalization operation, MaxPooling(*) is a maximum pooling operation, Attention(*) represents an attention module processing operation, F output is the output of the feature high-dimensional projection module PM;

[0035] S22, the typhoon wind field and the terrain total high-dimensional feature F total is input into the multi-scale high-dimensional feature fusion module FM, and a multi-scale wind-terrain high-dimensional fusion feature F fused is output; the multi-scale high-dimensional feature fusion module takes the high-dimensional feature F total as input, fuses the high-dimensional wind field feature and the high-dimensional terrain feature through three multi-scale fusion (MSF) layers and three attention modules, captures the correlation between the wind field and the terrain, and outputs a multi-scale high-dimensional fusion feature F fused as an input of the full connection prediction module; the calculation formula of the multi-scale high-dimensional feature fusion module FM is as follows:

[0036] T1 = MSF(F total )

[0037] T 1_Att = Attention(T1)

[0038] T2 = MSF(T 1_Att )

[0039] T 2_Att = Attention(T2)

[0040] T3 = MSF(T 2_Att )

[0041] F fused = Attention(T3)

[0042] where F total is the input high-dimensional feature, MSF(*) is the feature fusion operation of the MSF module, T1 is the feature fused by the first MSF layer, T 1_Att is the feature after weight adjustment of the first attention module, T2 is the feature fused by the second MSF layer, T 2_Att is the feature after weight adjustment of the second attention module, T3 is the feature fused by the third MSF layer, and T 3_Att is the feature after weight adjustment of the third attention module. After the feature fusion work of the three MSF layers and the three attention modules, the fused feature F fused is obtained.

[0043] S23, the multi-scale wind-terrain high-dimensional fusion feature F fused is used by the full-connection prediction module to perform prediction, and the typhoon wind field grid prediction W pred_t+1 is output.

[0044] Further, the attention module in step S21 includes a wind residual spatial attention mechanism WR-SA and a channel attention mechanism CA; the wind residual spatial attention mechanism WR-SA (Wind Residual-Spatial Attention) aims to amplify the weight of key spatial positions, and the formula is as follows:

[0045] A S1 = Input W -Center(W Norm_t+1 )

[0046] A S2 = Sigmoid(Conv(A S1 ))

[0047] A S3 = MeanPooling(A S2 ) + Maxpooling(A S2 )

[0048] F out = A S3 ⊙ Input

[0049] where Input represents the input grid data, W Norm represents the input standardized typhoon wind field grid data, Center(*) represents taking the center value of the grid and copying it as the corresponding dimension, Conv(*) is the convolution operation, Sigmoid(*) is the Sigmoid activation operation, ⊙ represents the dot product operation, A S1 represents the difference between each grid point of the input wind field grid and the center grid point of the input wind field grid, and A S2A S3 A S2 F out A

[0050] The wind residual spatial attention mechanism uses the wind field grid and the residual of its center grid point as input, dynamically allocates the weight of each spatial position, and effectively improves the performance of the network.

[0051] The channel attention mechanism CA aims to amplify the key channel weight, and the formula is as follows:

[0052] A C1 = MeanPooling(Input)

[0053] A C2 = SiLU(FC(A1))

[0054] A C3 = Sigmoid(FC(A2))

[0055] F out = A C3 ⊙Input

[0056] Where, Input represents input grid data, MeanPooling(*) is an average pooling operation, which is used to average each channel variable into a channel response value, FC(*) is a linear layer conversion, SiLU(*) is a SiLU activation operation, Sigmoid(*) is a Sigmoid activation operation, and ⊙ represents a dot product operation, A C1 represents the channel response value corresponding to each channel, A C2 represents the high-dimensional channel response feature processed by the linear layer and SiLU activation operation, A C3 represents the weight corresponding to each channel, F out represents the output processed by the channel attention mechanism.

[0057] This mechanism can dynamically allocate channel weights according to the importance of the channel, so that the model pays more attention to important channels and improves the network performance.

[0058] Further, each feature fusion module MSF in step S22 is composed of three dilated convolution layers with different dilated rates, three SiLU activation layers, a convolution fusion module, and an attention module. Each convolution module includes a convolution layer, and the network formula is as follows:

[0059] S 1.1= BN (SiLU (D_Conv d＝1 (Input))

[0060] S 1.2 = BN (SiLU (D_Conv d＝2 (Input))

[0061] S 1.3 = BN (SiLU (D_Conv d＝3 (Input))

[0062] S out = Conv (Concatenate (S 1.1 , S 1.2 , S 1.3 ))

[0063] where D_Conv d＝1 represents a dilated convolution with a dilation rate of 1, D_Conv d＝2 represents a dilated convolution with a dilation rate of 2, D_Conv d＝3 represents a dilated convolution with a dilation rate of 3, S 1.1 is a local scale fusion feature processed by a dilated convolution module with a dilation rate of 1, S 1.2 is a medium scale fusion feature processed by a dilated convolution module with a dilation rate of 2, S 1.3 is a large scale fusion feature processed by a dilated convolution module with a dilation rate of 3, S2 is a multi-scale high-dimensional fusion feature fused by a convolution fusion module, and S out is a multi-scale high-dimensional fusion feature output by the MSF layer.

[0064] Further, the form of the step S3 weighted loss function is as follows:

[0065]

[0066] where N represents the number of training samples, u pred_t+1,i and v pred_t+1,i represent the future 1-hour wind speed components (east-west and north-south) predicted by the i-th sample model, u obs_t+1,i and v obs_t+1,i represent the future 1-hour observed wind speed components of the i-th sample, u obs_t,i and v obs_t,i represent the current observed wind speed components of the i-th sample. α is a control factor of the gale weighting strength, which is usually positive and can be adjusted.

[0067] The weighted loss function is used as a target function in model training, and the model parameters are iteratively updated by an optimization algorithm (such as Adam or SGD). It gives greater weight to high wind speed areas, preferentially reduces the prediction error in these areas, improves the fitting accuracy in high wind areas, maintains the consistency of the overall wind field structure, avoids the compression of wind speed to the average level, and realizes the fine capture and accurate warning of extreme wind field characteristics.

[0068] The typhoon gale prediction system based on the consideration of wind-terrain multi-scale spatial correlation comprises:

[0069] The data preprocessing and prediction target establishment module is responsible for obtaining 1-hour tropical cyclone path prediction data, including engineering typhoon wind field data and terrain data, and standardizing these data; the standardized current time meteorological observation station data is interpolated into grid data, and the difference between the grid data and the future 1-hour meteorological observation station data is calculated as the prediction target of the model;

[0070] The multi-scale wind-terrain spatial correlation extraction module inputs the standardized wind field, terrain data and grid observation data into the multi-scale wind-terrain spatial correlation extraction network, and outputs the prediction result of the typhoon wind field grid through the feature high-dimensional projection, multi-scale high-dimensional feature fusion and full connection prediction sub-modules, so as to realize the fine prediction of the typhoon wind field;

[0071] The model optimization and parameter adjustment module enhances the weight of the gale area by designing a weighted loss function, avoids the compression effect of the model on the wind speed distribution in the training process, iteratively updates the model parameters through an optimization algorithm (such as Adam or SGD), thereby improves the prediction accuracy of the typhoon gale, better captures the extreme wind speed change, and realizes the optimization of the model performance.

[0072] An electronic device comprises a memory for storing a computer program and a processor for executing the computer program to implement the steps of the typhoon gale prediction method based on the consideration of wind-terrain multi-scale spatial correlation.

[0073] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the typhoon gale prediction method based on the consideration of wind-terrain multi-scale spatial correlation.

[0074] The typhoon gale prediction method based on the consideration of wind-terrain multi-scale spatial correlation has the following advantages:

[0075] 1. The typhoon gale prediction method based on wind-terrain multi-scale spatial correlation consideration disclosed in the present application, by introducing wind-terrain multi-scale spatial correlation, a wind field prediction model capable of self-adapting to the multi-scale influence of complex terrain and typhoon path change is constructed. By considering the interactive influence of multi-scale spatial variation of wind field and terrain effect, the accuracy of typhoon gale prediction can be improved, especially in complex terrain area and extreme weather conditions, the local change of typhoon wind field can be more accurately simulated. Compared with the prior art, the present application solves the problems of instability and poor adaptability of the existing method in simulating wind field, improves the accuracy and reliability of wind field simulation, especially in high complexity terrain and extreme weather conditions.

[0076] 2. The typhoon gale prediction method based on wind-terrain multi-scale spatial correlation consideration disclosed in the present application, by introducing wind-terrain multi-scale spatial correlation, the interactive effect of multi-scale variation of wind field and terrain effect can be accurately considered, so that the accuracy of typhoon gale prediction under complex terrain conditions can be significantly improved. Compared with the prior art, the method of the present application can better adapt to complex terrain and typhoon path change, and improve the reliability of typhoon wind field prediction results. In the case of extreme weather and complex terrain, the present application can effectively handle the interaction between complex variables, provide high-precision prediction results, and has important significance for disaster warning and weather forecast.

[0077] Other advantages, objects, and features of the present application will be apparent to those skilled in the art from the following specification, and will be learned from practice of the present application. The objects and other advantages of the present application can be achieved and obtained by the following specification. BRIEF DESCRIPTION OF DRAWINGS

[0078] In order to make the objects, technical solutions and advantages of the present application clearer, the preferred detailed description of the present application will be combined with the drawings as follows, wherein:

[0079] Figure 1 The flow chart of the typhoon gale prediction method based on wind-terrain multi-scale spatial correlation consideration of the present application. DETAILED DESCRIPTION

[0080] The embodiments of the present application are described below through specific examples, and those skilled in the art can easily understand other advantages and effects of the present application from the disclosure of the present specification. The present application can also be implemented or applied through other different embodiments, and the details in the specification can be modified or changed based on different views and applications without departing from the spirit of the present application.

[0081] As Figure 1The typhoon gale prediction method based on the wind-terrain multi-scale spatial correlation consideration shown includes the following steps:

[0082] S1, establishing a prediction target; obtaining future 1-hour tropical cyclone path prediction data, including engineering typhoon wind field data W t+1 =(u t+1 ,v t+1 ) and terrain data T t+1 , standardizing these data to obtain standardized wind field data W Norm_t+1 and standardized terrain data T Norm_t+1 ; after interpolating the current time meteorological observation station data into grid data, the difference between the future 1-hour meteorological observation station data is used as the model prediction target.

[0083] S11, obtaining 1-hour future tropical cyclone path prediction data (usually provided by the Central Meteorological Observatory of the China Meteorological Administration), extracting the prediction center position according to the path prediction data to obtain the current time meteorological observation station data, future 1-hour engineering typhoon wind field data W t+1 =(u t+1 ,v t+1 ) and terrain data T t+1 .

[0084] S12, preprocessing the obtained engineering typhoon model wind field data and terrain data, standardizing different data respectively to generate standardized data. The standardized wind field data and standardized terrain data are represented as W Norm_t+1 and T Norm_t+1 respectively; the standardization process is as follows:

[0085] W Norm_t+1 = Norm(W t+1 )

[0086] T Norm_t+1 = Norm(T t+1 )

[0087] Wherein, Norm(*) represents the standardization operation, W t+1 represents the input engineering typhoon wind field data, T t+1 represents the input terrain data, W Norm_t+1 represents the standardized future 1-hour typhoon wind field data, and T Norm_t+1 represents the standardized terrain data. The current time meteorological observation station data is interpolated into a grid using a bicubic interpolation method, and the grid resolution is the same as W Norm_t+1 and T Norm_t+1 , denoted as W obs_t .

[0088] S13, the difference between the standardized current time meteorological observation station data and the future 1 hour meteorological observation station data is taken as a model prediction target. The prediction target is calculated as:

[0089]

[0090] wherein W target_t+1 is a future 1 hour predicted wind field, represents a deep learning network.

[0091] S2, a typhoon wind field grid is predicted. The standardized wind field, terrain data, and grid observation are input into a multi-scale wind-terrain spatial correlation extraction network to perform prediction, and a typhoon wind field grid prediction is output, wherein the multi-scale wind-terrain spatial correlation extraction network is composed of a feature high-dimensional projection module, a multi-scale high-dimensional feature fusion module, and a full connection prediction module.

[0092] S21, the standardized wind field and terrain data are input into the feature high-dimensional projection module PM, and the feature high-dimensional projection module PM outputs the total high-dimensional features F total of the wind field and terrain. The calculation formula of the feature high-dimensional projection module PM is as follows:

[0093] F W =PM(W Norm_t+1 ,W obs_t+1 )

[0094] F T =PM(T Norm_t+1 )

[0095] F total =Concatenate(F W ,F T )

[0096] wherein PM(*) represents the feature high-dimensional projection module, F W represents the high-dimensional typhoon wind field features processed by the feature high-dimensional projection module, F T represents the high-dimensional terrain features processed by the feature high-dimensional projection module, Concatenate(*) represents a splicing operation, and F total represents the total high-dimensional features of the typhoon wind field and terrain. Each feature high-dimensional projection module PM is composed of three convolution modules, two Maxpooling pooling layers, and an attention layer, each convolution module includes a convolution layer and a Silu activation layer, and the calculation process of the feature high-dimensional projection module PM is as follows:

[0097] F1=MaxPooling(BN(SiLU(Conv(Input))))

[0098] F2 = MaxPooling(BN(SiLU(Conv(F1)))

[0099] F3 = BN(SiLU(Conv(F2)))

[0100] F output = Attention(SiLU(Conv(F3)))

[0101] wherein Input represents the input of the feature extraction module PM, Conv(*) is a convolution operation, SiLU(*) is a SiLU activation operation, BN(*) is a Batch Normalization operation, MaxPooling(*) is a maximum pooling operation, Attention(*) represents an attention module processing operation, and F output is the output of the feature high-dimensional projection module PM.

[0102] The attention module includes a wind residual spatial attention mechanism WR-SA and a channel attention mechanism CA. The wind residual spatial attention mechanism WR-SA (WindResidual-SpatialAttention) aims to amplify the weight of key spatial positions, and the formula is as follows:

[0103] A S1 = Input W -Center(W Norm_t+1 )

[0104] A S2 = Sigmoid(Conv(A S1 ))

[0105] A S3 = MeanPooling(A S2 ) + Maxpooling(A S2 )

[0106] F out = A S3 ⊙ Input

[0107] wherein Input represents the input grid data, W Norm represents the input normalized typhoon wind field grid data, Center(*) represents taking the grid center value and copying it as the corresponding dimension, Conv(*) is a convolution operation, Sigmoid(*) is a Sigmoid activation operation, and represents a dot product operation, A S1 represents the difference between each grid point of the input wind field grid and the center grid point of the input wind field grid, A S2 represents the weight of each grid position, and A S3 represents the output of the wind residual spatial attention mechanism WR-SA.A represents the grid position weight processed by two pooling operations, used to match A S2 dimension and Input dimension, F out A represents the output processed by the spatial attention mechanism. The spatial attention mechanism uses the residual of the wind field grid and its center grid point as input to dynamically allocate the weight of each spatial position, effectively improving the performance of the network.

[0108] Channel attention mechanism CA(ChannelAttention) aims to amplify the weight of key channels, and the formula is as follows:

[0109] A C1 = MeanPooling(Input)

[0110] A C2 = SiLU(FC(A1))

[0111] A C3 = Sigmoid(FC(A2))

[0112] F out = A C3 ⊙Input

[0113] Where, Input represents the input grid data, MeanPooling(*) is the average pooling operation, which is used to average the variables of each channel into a channel response value, FC(*) is the Linear linear layer conversion, SiLU(*) is the SiLU activation operation, Sigmoid(*) is the Sigmoid activation operation, ⊙ represents the dot product operation, A C1 represents the channel response value corresponding to each channel, A C2 represents the high-dimensional channel response feature processed by the linear layer and SiLU activation operation, A C3 represents the weight corresponding to each channel, F out represents the output processed by the channel attention mechanism. This mechanism can dynamically allocate channel weights according to the importance of channels, so that the model pays more attention to important channels and improves the performance of the network.

[0114] S22, the typhoon wind field and the total high-dimensional terrain feature F total input multi-scale high-dimensional feature fusion module FM, output multi-scale wind-terrain high-dimensional fusion feature F fused The multi-scale high-dimensional feature fusion module takes high-dimensional feature F total as input, and fuses high-dimensional wind field features and high-dimensional terrain features through three multi-scale fusion layers (Multi-scale Fusion, MSF) and three attention modules to capture the correlation between wind field and terrain, and outputs multi-scale high-dimensional fusion feature F fused, as the input of the full connection prediction module. The multi-scale high-dimensional feature fusion module calculation formula is as follows:

[0115] T1=MSF(F total )

[0116] T 1_Att =Attention(T1)

[0117] T2=MSF(T 1_Att )

[0118] T 2_Att =Attention(T2)

[0119] T3=MSF(T 2_Att )

[0120] F fused =Attention(T3)

[0121] Where F total is the input high-dimensional feature, MSF(*) is the feature fusion operation of the MSF module, T1 is the feature fused by the first MSF layer, T 1_Att is the feature adjusted by the first attention module weight, T2 is the feature fused by the second MSF layer, T 2_Att is the feature adjusted by the second attention module weight, T3 is the feature fused by the third MSF layer, and T 3_Att is the feature adjusted by the third attention module weight. After three MSF layers and three attention modules complete the feature fusion work, the fused feature F fused is obtained. Each feature fusion module MSF is composed of three dilated convolution layers with different expansion rates, three SiLU activation layers, a convolution fusion module, and an attention module. Each convolution module contains a convolution layer. The network formula is as follows:

[0122] S 1.1 =BN(SiLU(D_Conv d＝1 (Input)))

[0123] S 1.2 =BN(SiLU(D_Conv d＝2 (Input)))

[0124] S 1.3 =BN(SiLU(D_Conv d＝3 (Input)))

[0125] S out =Conv(Concatenate(S 1.1 ,S 1.2 ,S 1.3))

[0126] where D_Conv d＝1 represents a dilated convolution with a dilation rate of 1, D_Conv d＝2 represents a dilated convolution with a dilation rate of 2, D_Conv d＝3 represents a dilated convolution with a dilation rate of 3, S 1.1 is the local scale fusion feature processed by the dilated convolution module with a dilation rate of 1, S 1.2 is the medium scale fusion feature processed by the dilated convolution module with a dilation rate of 2, S 1.3 is the large scale fusion feature processed by the dilated convolution module with a dilation rate of 3, S2 is the multi-scale high-dimensional fusion feature fused by the convolution fusion module, S out is the multi-scale high-dimensional fusion feature output by the MSF layer.

[0127] S23, the fully connected prediction module utilizes the multi-scale wind-terrain high-dimensional fusion feature F fused to perform prediction and output the typhoon wind field grid prediction W pred_t+1 .

[0128] S3, optimize the model parameters. In order to avoid the compression effect of the model on the wind speed distribution in the training process in the typhoon gale prediction, the application designs a weighted loss function, which aims to improve the prediction accuracy of the model on the typhoon gale by giving appropriate weight to the gale area. The weighted loss function enhances the weight of the gale area, avoids the loss of the area with larger wind speed being excessively compressed or ignored in the training process, so that the model can better capture the extreme wind speed change of the typhoon. Specifically, the form of the weighted loss function is as follows:

[0129]

[0130] where N represents the number of training samples, u pred_t+1,i and v pred_t+1,i represent the future 1-hour wind speed components (east-west and north-south) predicted by the model of the i-th sample, u obs_t+1,i and v obs_t+1,i represent the future 1-hour observed wind speed components of the i-th sample, u obs_t,i and v obs_t,iThe current observation wind speed component representing the i-th sample. a is a control factor of the strong wind weighting intensity, usually positive, which can be adjusted. The weighted loss function is used as the objective function in the model training process to participate in gradient calculation, and the model parameters are iteratively updated through optimization algorithms such as Adam or SGD. Since the high wind speed area has a larger weight, the model will preferentially reduce the prediction error in these areas, thereby improving the fitting accuracy in the strong wind area. After the model is trained and optimized, it can significantly improve the prediction accuracy in the strong wind area while maintaining the consistency of the overall wind field structure, avoiding the phenomenon of compressing the wind speed value to the average level, and is conducive to the fine capture and accurate warning of extreme wind field characteristics.

[0131] Finally, it should be pointed out that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the purpose and scope of the technical solutions, which should be covered in the scope of the claims of the present application.

Claims

1. A typhoon gale prediction method based on wind-terrain multi-scale spatial correlation consideration, characterized in that, The method comprises the following steps: S1, establishing a prediction target; obtaining 1-hour tropical cyclone path prediction data, including engineering typhoon wind field data and terrain data , standardizing the data to obtain standardized wind field data and standardized terrain data ; after interpolating the standardized current-time meteorological observation station data into grid data, the difference between the grid data and 1-hour future meteorological observation station data is taken as the model prediction target; Step S1 is specifically: S11, acquire tropical cyclone path forecast data for the next 1 hour, extract the forecast center position according to the path forecast data, acquire current time meteorological observation station data, and future 1 hour typhoon wind field data and terrain data ; S12, preprocessing the obtained engineering typhoon model wind field data and terrain data, respectively standardizing different data to generate standardized data; the standardized wind field data and the standardized terrain data are respectively represented as and ; the standardization process is that: wherein, represents the standardized operation, represents the inputted engineering typhoon wind field data, represents the inputted terrain data, represents the standardized future 1-hour typhoon wind field data, represents the standardized terrain data; the current time meteorological observation site data is interpolated into a grid by using a bi-cubic interpolation method, and the grid resolution is the same as and are the same, and are denoted as ; S13, the difference between the standardized current time meteorological observation station data and the future 1-hour meteorological observation station data is taken as the model prediction target, and the prediction target is calculated as: wherein forecast the wind field for the next 1 hour, represents a deep learning network; S2, predict the typhoon wind field grid; input the standardized wind field, terrain data and grid observation into the multi-scale wind-terrain spatial correlation extraction network for prediction, and output the typhoon wind field grid prediction; the multi-scale wind-terrain spatial correlation extraction network is composed of a feature high-dimensional projection module, a multi-scale high-dimensional feature fusion module and a full connection prediction module; Step S2 is specifically: S21, standardized wind field, terrain data input feature high-dimensional projection module feature high-dimensional projection module output wind field and terrain total high-dimensional feature feature high-dimensional projection module The calculation formula is as follows: wherein, a representative feature high-dimensional projection module, a representative high-dimensional typhoon wind field feature processed by the feature high-dimensional projection module, a representative high-dimensional terrain feature processed by the feature high-dimensional projection module, a representative splicing operation, a representative total high-dimensional feature of the typhoon wind field and the terrain. Each feature high-dimensional projection module Each feature high-dimensional projection module is composed of three convolution modules and two Maxpooling pooling layers and an attention layer, each convolution module contains a convolution layer and a Silu activation layer The calculation process is as follows: wherein, representing a feature extraction module input to, is a convolution operation, is a SiLU activation operation, is a Batch Normalization operation, is a max-pooling operation, representing an attention module processing operation, is an output of a feature high-dimensional projection module . S22. Combine the typhoon wind field and terrain high-dimensional features. Input the multi-scale high-dimensional feature fusion module FM, output multi-scale wind-topography high-dimensional fusion features. The multi-scale high-dimensional feature fusion module uses high-dimensional features... As input, the high-dimensional wind field features and high-dimensional terrain features are fused through three multi-scale fusion (MSF) layers and three attention modules to capture the correlation between wind field and terrain, and output multi-scale high-dimensional fused features. The input to the fully connected prediction module is used as the input; the calculation formula for the multi-scale high-dimensional feature fusion module is as follows: wherein is the input high-dimensional feature, is the feature fusion operation of the module, is the fused feature of the first MSF layer, is the feature after weight adjustment of the first attention module, is the fused feature of the second MSF layer, is the feature after weight adjustment of the second attention module, is the fused feature of the third MSF layer is the feature after weight adjustment of the third attention module, after the feature fusion work of three MSF layers and three attention modules, the fused feature is obtained. S23, the full connection prediction module utilizes the multi-scale wind-terrain high-dimensional fusion features make a prediction, output a typhoon wind field grid prediction ; S3, optimize the model parameters; by designing a weighted loss function, the weight of the strong wind area is enhanced, the compression effect of the model on the wind speed distribution in the training process is avoided, and the prediction accuracy of the typhoon strong wind is improved, so that the extreme wind speed change can be better captured.

2. The typhoon-gale prediction method according to claim 1, wherein Step S11, the tropical cyclone path prediction data is usually the prediction data provided by the China Meteorological Administration of the Central Meteorological Observatory.

3. The typhoon-gale prediction method according to claim 1, wherein The attention module in step S21 includes a wind residual spatial attention mechanism WR-SA and a channel attention mechanism CA; the wind residual spatial attention mechanism WR-SA (Wind Residual-Spatial Attention) aims to amplify the weight of the key spatial position, and the formula is as follows: wherein represents input grid data, represents input normalized typhoon wind field grid data, represents taking the grid center value and copying it as corresponding dimensions, is a convolution operation, is an activation operation, represents a dot product operation, represents the difference between each grid point of the input wind field grid and the center grid point of the input wind field grid, represents the weight of each grid position, represents the grid position weight after two kinds of pooling operations, used to match dimensions and dimensions, represents the output after the spatial attention mechanism processing; The channel attention mechanism CA (Channel Attention) aims to amplify the weight of the key channel, and the formula is as follows: wherein, represents input grid data, is an average pooling operation for averaging variables of each channel into a channel response value, is a Linear linear layer transformation, is a SiLU activation operation, is an activation operation, represents a point multiplication operation, represents a channel response value corresponding to each channel, represents a high-dimensional channel response feature processed by a linear layer and an activation operation, represents a weight corresponding to each channel, represents an output processed by a channel attention mechanism.

4. The typhoon-gale prediction method according to claim 3, characterized by, In step S22, each feature fusion module MSF is composed of three hollow convolution layers with different expansion rates, three SiLU activation layers, a convolution fusion module and an attention module; each convolution module includes a convolution layer, and the network formula is as follows: wherein represents a dilated convolution with dilation rate of 1, represents a dilated convolution with dilation rate of 2, represents a dilated convolution with dilation rate of 3, is a local scale fusion feature processed by a dilated convolution module with dilation rate of 1, is a medium scale fusion feature processed by a dilated convolution module with dilation rate of 2, is a large scale fusion feature processed by a dilated convolution module with dilation rate of 3, is a multi-scale high-dimensional fusion feature fused by the convolution fusion module, is a multi-scale high-dimensional fusion feature output by the MSF layer.

5. The typhoon-gale prediction method according to claim 4, characterized by, The form of the weighted loss function in step S3 is as follows: wherein represents the number of training samples, and represents the future 1-hour wind speed component (east-west and north-south) predicted by the i-th sample model, and represents the future 1-hour observed wind speed component of the i-th sample, and represents the current observed wind speed component of the i-th sample; is a control factor for the weighted intensity of the gale, usually positive and adjustable.

6. A system using a typhoon gale prediction method based on wind-topography multiscale spatial correlation consideration as claimed in claim 1, characterized in that, It comprises: A data preprocessing and prediction target establishment module, which is responsible for acquiring future 1-hour tropical cyclone path prediction data, including engineering typhoon wind field data and terrain data, and standardizing these data; The standardized current time meteorological observation station data is interpolated into grid data, and the difference between the grid data and the future 1-hour meteorological observation station data is calculated as the prediction target of the model; A multi-scale wind-terrain spatial correlation extraction module, which inputs the standardized wind field, terrain data and grid observation data into the multi-scale wind-terrain spatial correlation extraction network, and outputs the prediction result of the typhoon wind field grid through the sub-modules of feature high-dimensional projection, multi-scale high-dimensional feature fusion and full connection prediction, so as to realize the fine prediction of the typhoon wind field; A model optimization and parameter adjustment module, which enhances the weight of the strong wind area by designing a weighted loss function, and avoids the compression effect of the model on the wind speed distribution in the training process; The model parameters are iteratively updated through the optimization algorithm, so as to improve the prediction accuracy of the typhoon strong wind, better capture the extreme wind speed change, and realize the optimization of the model performance.

7. An electronic device, comprising: It comprises: A memory for storing a computer program; A processor for executing the computer program to realize the steps of the typhoon strong wind prediction method based on the consideration of wind-terrain multi-scale spatial correlation in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is used for realizing the steps of the typhoon gale prediction method based on the wind-terrain multi-scale spatial correlation consideration according to any one of claims 1-5 when executed by the processor.

Images