Intelligent prediction method and system of wave height based on global and local information fusion

Abstract

The present invention belongs to the technical field of sea wave height prediction, and discloses a sea wave height intelligent prediction method and system based on global and local information fusion. The method first uses sea surface radial wind field data and latitudinal wind field data as input, extracts and fuses spatial features of the global region, and obtains radial wind field fusion features and latitudinal wind field fusion features; then the obtained radial wind field fusion features and latitudinal wind field fusion features are respectively divided into different sub-ocean wave height prediction fields, and then all sub-ocean wave height prediction fields are subjected to feature aggregation to obtain aggregate features; finally, the aggregate features are used as input to perform local sea area refined sea wave height prediction, and obtain global regional prediction results. The present invention improves the wave height prediction accuracy, and solves the problem of insufficient prediction accuracy of traditional methods.

Description

Technical Field

[0001] The invention belongs to the technical field of ocean wave height prediction, and in particular relates to an ocean wave height intelligent prediction method and system based on global and local information fusion. Background Art

[0002] Refined prediction of wave height in local sea areas is of great significance for protecting the safety of coastal areas. Accurate prediction of wave height can help people understand the sea conditions in time, take preventive measures, reduce disaster risks, and ensure the safety of marine activities. Considering the high cost and spatial sparsity of professional equipment, traditional field measurement methods cannot achieve refined prediction in some spaces, resulting in data loss. Therefore, the advanced wave height prediction method is to design a data-driven neural network model to predict wave height.

[0003] For example, the existing method uses a multivariate 3-layer LSTM architecture based on matrix-to-matrix mapping to establish the potential spatiotemporal relationship between wind components and effective wave height fields. In order to reduce computational overhead, a parallel modeling strategy is adopted to flexibly divide the prediction domain into a finite number of windows and train them independently as sub-models. Since the change of wave height is mainly affected by the wind field, this method takes wind field data at multiple past moments as input, uses a long short-term memory network to extract the spatiotemporal characteristics of the physical field, and performs matrix transformation on the output features through a fully connected neural network, ultimately realizing the prediction of wave height, solving the problem that some spaces cannot be predicted by traditional methods.

[0004] However, this method has the following problems: (1) The wind field plays an extremely important role in the change of wave height, and local wind field information is the key to causing extreme changes in wave height. The existing method realizes global spatial feature extraction by dimensional transformation of the wind field data matrix, but does not take into account the impact of local wind field information on wave height and the fusion of local and global information, so it is impossible to accurately predict extreme changes in wave height. (2) Due to the large area of ​​the global ocean and the strong regional differences, the prediction effect of global wave height using a single model is poor. The existing model divides the global ocean into multiple local oceans and models them separately to achieve refined prediction. However, since some surrounding spatial information is lost in the boundary part of the local ocean after segmentation, the overall model prediction accuracy is reduced. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides an intelligent prediction method and system for wave height based on global and local information fusion. First, the global region pre-trained boundary missing information completion module is used to extract and fuse the spatial features of the global region, so as to ensure that the sub-region boundary is integrated with the surrounding information; then, the multi-scale physical space global-local information fusion local sea area refined prediction module is used to extract global and local information from the spatiotemporal features in the sub-region under multiple scale perspectives, and the high-precision solution of the sub-region wave height is achieved. Finally, the prediction results of multiple sub-regions are spliced ​​to obtain the prediction results of the large global region.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0007] The method for intelligent prediction of wave height based on global and local information fusion includes the following steps:

[0008] Step S1: Extract and fuse spatial features of the global region:

[0009] The radial wind field data U and the latitudinal wind field data V of the sea surface are used as input. For each radial wind field data and latitudinal wind field data, different convolutional layers are used to process the spatial features at multiple scales. The local information and global information of the spatial features at each scale are extracted through the CNN module and the Transformer module respectively. Then, the local information and the global information are fused to obtain the fused features. Finally, the fused features at different scales are added together to obtain the radial wind field fused feature H and the latitudinal wind field fused feature HH.

[0010] Step S2: Feature aggregation:

[0011] The H and HH obtained in step S1 are respectively divided into different sub-ocean wave height prediction fields, and then all sub-ocean wave height prediction fields are subjected to feature aggregation to obtain the aggregated feature K;

[0012] Step S3: Refined prediction of wave height in local sea area:

[0013] Taking the aggregate feature K as input, we first extract and fuse the global and local information of the spatiotemporal features in the sub-region from multiple scale perspectives; then we perform a high-precision solution of the sub-region wave height; finally, we splice the prediction results of multiple sub-regions to obtain the global regional prediction result.

[0014] Further, in step S1, the radial wind field data and the latitudinal wind field data of the sea surface at n moments are named , As input; the specific calculation process is as follows: construct convolution layers based on convolution kernels of different sizes, process the radial wind field data U and the latitudinal wind field data V using different convolution layers respectively, and output spatial features at multiple scales. Taking U as an example, after being processed by different convolution layers conv0, conv1, and conv2, the spatial features F0, F1, and F2 at multiple scales are output;

[0015] Then, the local and global information of F0, F1, and F2 are extracted through the CNN module and the Transformer module respectively. The specific calculation process is as follows:

[0016] In the Transformer module, taking F0 as an example, different weights are assigned to F0 based on the self-attention mechanism and the weight matrix is ​​calculated. Then, the weight matrix is ​​used to assign different attentions to different features, and finally a feature M1 containing the relationship between features is obtained. Similarly, F1 and F2 are processed by the Transformer module to obtain M2 and M3.

[0017] In the CNN module, convolution operations are performed on F0, F1, and F2 to obtain F3, F4, and F5 respectively;

[0018] The features F3, F4, and F5 obtained by the CNN module are local information at different scales, and the features M1, M2, and M3 obtained by the Transformer module are global information at different scales.

[0019] Furthermore, the features obtained by the CNN module and the Transformer module are added and fused to obtain the fused feature FU. Specifically, F3 and M1 are added, F4 and M2 are added, and F5 and M3 are added. The fused features are represented as FU0, FU1, and FU2, respectively. Then, convolution operations are performed to obtain BG0, BG1, and BG2, respectively. BG0, BG1, and BG2 are added to obtain the radial wind field fusion feature H. Similarly, for the latitudinal wind field data V, the latitudinal wind field fusion feature HH is obtained.

[0020] Furthermore, in step S2, H and HH are divided into ζ sub-ocean wave height prediction fields respectively. and , and then perform feature aggregation to obtain the aggregate feature K, as shown in the following formula:

[0021] ;

[0022] Among them, AK is the ocean wave height prediction field of the radial wind field, BK is the ocean wave height prediction field of the zonal wind field, and concat represents the feature aggregation operation. and They are ζ sub-ocean radial wave height prediction fields and ζ sub-ocean zonal wave height prediction fields respectively.

[0023] Furthermore, step S3 is specifically: taking the aggregated feature K as input, firstly using different convolutional layers to process to obtain the large-scale spatial feature R0, the medium-scale spatial feature R1, and the small-scale spatial feature R2, respectively, the spatial features at each scale are respectively extracted local information and global information by CNN and Transformer, and then the local information and the global information are fused to obtain fused features FUUU0, FUUU1, and FUUU2, respectively, the fused features at different scales are restored to the original size through the convolution operation, and the large-scale spatial feature BL0, the medium-scale spatial feature BL1, and the small-scale spatial feature BL2 are respectively obtained, and then BL0, BL1, and BL2 are respectively flattened to G0, G1, and G2, and then the feature matrices G0G1, G0G2, and G1G2 are obtained through the cross-attention operation, and then the cross-fusion results are mapped back to the same shape as the original features to obtain G0', G1', and G2', and then G0', G1', and G2' are fused to obtain the final global area prediction result P.

[0024] Furthermore, the specific calculation process of step S3 is as follows: construct convolutional layers based on convolutional kernels of different sizes, process the aggregated features K using different convolutional layers, output spatial features R0, R1, and R2 at multiple scales, and then extract global and local information from R0, R1, and R2 through CNN modules and Transformer modules respectively.

[0025] The specific calculation process is as follows: In the Transformer module, taking R0 as an example, different weights are assigned to R0 based on the self-attention mechanism and the weight matrix is ​​calculated. Then, the weight matrix is ​​used to assign different attentions to different features, and finally a feature MM1 containing the relationship between features is obtained. Similarly, R1 and R2 are processed by the Transformer module to obtain MM2 and MM3.

[0026] In the CNN module, convolution operations are performed on R0, R1, and R2 to obtain R3, R4, and R5 respectively;

[0027] Then, the features obtained by the CNN module and the Transformer module are added and fused to obtain the fused feature FUUU. Specifically, R3 and MM1 are added, R4 and MM2 are added, and R5 and MM3 are added. The fused features are represented as FUUU0, FUUU1, and FUUU2 respectively. Then, convolution operation is performed to obtain BL0, BL1, and BL2 respectively.

[0028] Then BL0, BL1, and BL2 are flattened into G0, G1, and G2 respectively, and then the feature matrices G0G1, G0G2, and G1G2 are obtained through cross-attention operation. Then the cross-fusion results are mapped back to the same shape as the original features to obtain G0', G1', and G2'. G0', G1', and G2' are added and fused to obtain the final global area prediction result P;

[0029] Finally, the output features are placed at their original geographic locations and spliced ​​with other sub-ocean wave height prediction fields to obtain the predicted wave height P all .

[0030] The present invention also provides an intelligent prediction system for wave height based on global and local information fusion, which is used to implement the intelligent prediction method for wave height based on global and local information fusion as described above. The system includes a global area pre-trained boundary missing information completion module, a multi-scale physical space global-local information fusion local sea area refined prediction module,

[0031] The global region pre-trained boundary missing information completion module extracts and fuses spatial features of the global region to ensure that the sub-region boundary fuses surrounding information;

[0032] The multi-scale physical space global-local information fusion local sea area refined prediction module realizes global and local information extraction of spatiotemporal features in sub-areas from multiple scale perspectives and outputs predicted wave heights.

[0033] Furthermore, the global area pre-trained boundary missing information completion module includes different convolutional layers, CNN modules, Transformer modules and information aggregation modules. The input sea surface radial wind field data U and latitudinal wind field data V are processed by different convolutional layers to obtain spatial features at multiple scales. The spatial features at each scale are respectively extracted by CNN modules and Transformer modules for local information and global information. Then, the local information and the global information are fused to obtain fused features. Finally, the fused features at different scales are added to obtain radial wind field fusion features H and latitudinal wind field fusion features HH respectively. In the information aggregation module, H and HH are respectively divided into different sub-ocean wave height prediction fields, and all sub-ocean wave height prediction fields are feature aggregated to obtain aggregated features K.

[0034] Furthermore, the multi-scale physical space global-local information fusion local sea area refined prediction module includes different convolutional layers, CNN modules, Transformer modules and cross-fusion modules. The processing process of the convolutional layer, CNN module and Transformer module of the multi-scale physical space global-local information fusion local sea area refined prediction module is the same as that of the global area pre-trained boundary missing information completion module. The obtained fusion features are input into the cross-fusion module, and the cross-attention operation is first performed, and the result is mapped back to the same shape as the original feature, and then the fusion operation is performed to obtain the final global area prediction result P. Finally, the output feature is placed in its original geographical location and spliced ​​with other sub-ocean wave height prediction fields to finally obtain the predicted wave height P. all .

[0035] Compared with the prior art, the present invention has the advantages of:

[0036] (1) A global regional pre-training boundary missing information completion module was designed. Global features were extracted through Transformer and local features were extracted through CNN. The global features were added to the sub-model to achieve the fusion of global and local features, ensuring that the sub-region boundary was integrated with surrounding information, thereby improving the accuracy of accurate prediction of extreme wave height changes.

[0037] (2) A multi-scale physical space global-local information fusion local sea area refined prediction module was designed. The global ocean was divided into multiple local sea areas for modeling, and boundary information was added to complete the model. The spatiotemporal characteristics in the sub-areas were extracted from global and local information at multiple scales. The local sea area refined prediction module was then used to achieve high-precision solution of sub-area wave heights. Finally, the prediction results of multiple sub-areas were spliced ​​to obtain the global area prediction result, which solved the problem of insufficient prediction accuracy of traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0039] Figure 1 It is a schematic diagram of the system structure of the present invention;

[0040] Figure 2 This is a schematic diagram of the structure of the global area pre-training boundary missing information completion module of the present invention;

[0041] Figure 3 It is a schematic diagram of the structure of the multi-scale physical space global-local information fusion local sea area refined prediction module of the present invention. DETAILED DESCRIPTION

[0042] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0043] This embodiment provides an intelligent prediction method for wave height based on global and local information fusion. Figure 1 The system implementation shown includes the following steps:

[0044] Step S1: Extract and fuse spatial features of the global region:

[0045] The radial wind field data U and the latitudinal wind field data V on the sea surface are taken as input. For each radial wind field data and latitudinal wind field data, different convolutional layers are first used to obtain spatial features at multiple scales. The spatial features at each scale are respectively extracted local information and global information through the CNN module and the Transformer module. Then the local information and the global information are fused to obtain the fused features (for the radial wind field data U, the fused features FU are obtained, and for the latitudinal wind field data V, the fused features FUU are obtained). Finally, the fused features at different scales are added together to obtain the radial wind field fused features H and the latitudinal wind field fused features HH.

[0046] Combination Figure 2 As shown, step S1 is implemented by the global area pre-trained boundary missing information completion module, and the sea surface radial wind field data and zonal wind field data with dimensions of M*N*T at n moments are named , As input; the specific calculation process is as follows: construct convolution layers based on convolution kernels of different sizes, process the radial wind field data U and the latitudinal wind field data V using different convolution layers respectively, and output spatial features at multiple scales. Taking U as an example, after being processed by different convolution layers conv0, conv1, and conv2, the spatial features F0, F1, and F2 at multiple scales are output;

[0047] ;

[0048] ;

[0049] ;

[0050] F0, F1, and F2 are all spatial features at multiple scales. Then, the local and global information of F0, F1, and F2 are extracted through the CNN module and the Transformer module respectively. The specific calculation process is as follows:

[0051] In the Transformer module, taking F0 as an example, different weights are assigned to F0 based on the self-attention mechanism and the weight matrix is ​​calculated. Then, the weight matrix is ​​used to assign different attention levels to different features, and finally a feature M1 containing the relationship between features is obtained.

[0052] As an example, specifically, setting the weight , respectively Multiply the result :

[0053] ;

[0054] ;

[0055] ;

[0056] Where V represents the matrix element multiplication;

[0057] Then calculate the weight matrix :

[0058] = ;

[0059] Final use The relationship weights between the features in the calculation are used to calculate M1 and allocate attention to different features:

[0060] ;

[0061] The final feature M1 is the feature matrix of mutual guidance of spatial features at this spatial scale.

[0062] Similarly, F1 and F2 are processed by the Transformer module to obtain M2 and M3.

[0063] In the CNN module, the convolution operation conv3 is performed on F0, and the formula is as follows:

[0064] ;

[0065] Similarly, convolution operations conv4 and conv5 are performed on F1 and F2 to obtain F4 and F5. The formula is as follows:

[0066] ;

[0067] .

[0068] The features F3, F4, and F5 obtained by the CNN module are local information at different scales, and the features M1, M2, and M3 obtained by the Transformer module are global information at different scales.

[0069] Next, the features obtained by the CNN module and the Transformer module are added and fused to obtain the fused feature FU. Specifically, F3 and M1 are added, F4 and M2 are added, and F5 and M3 are added. The fused features are represented as FU0, FU1, and FU2, respectively. Then, convolution operations are performed to obtain BG0, BG1, and BG2, respectively. BG0, BG1, and BG2 are added to obtain the radial wind field fusion feature H.

[0070] As an example, for F0, the fused feature is expressed as FU0, and the formula is as follows:

[0071] ;

[0072] Then perform the convolution operation conv6 to restore the image to its original size and get BG0. The formula is as follows:

[0073] ;

[0074] Finally, a large-scale spatial feature BG0 with a dimension of M×N×T is obtained.

[0075] Similarly, the mesoscale and microscale spatial features BG1 and BG2 can be obtained, and finally the radial wind field fusion feature H is obtained, and the formula is as follows:

[0076] ;

[0077] Similarly, if Figure 2 As shown in the figure, for the latitudinal wind field data V, after the same processing, the latitudinal wind field fusion feature HH can be obtained, and the formula is as follows:

[0078] .

[0079] in, Figure 2 In the figure, the input V is processed by different convolutional layers, and the spatial features at multiple scales are output as FF0, FF1, and FF2; the features after processing and fusion by the CNN module and the Transformer module are represented as FUU0, FUU1, and FUU2; then the convolution operation is performed to obtain the large-scale, medium-scale, and small-scale spatial features BGG0, BGG1, and BGG2, respectively, and BGG0, BGG1, and BGG2 are added together to obtain the latitudinal wind field fusion feature HH.

[0080] Finally, the real data values ​​of M×N spatial points are The loss calculation is performed as shown in the following formula. After all sample predictions and loss calculations in a batch are completed, the final loss value of the radial wind field is obtained by the following formula , perform back-propagation operation to update model parameters;

[0081] ;

[0082] Where m is the batch size and i is the training batch number. Data values ​​output for radial wind field calculations. is the radial true data value.

[0083] Similarly, the final loss value of the zonal wind field is for:

[0084] ;

[0085] Where m is the batch size and i is the training batch number. The data value output for the zonal wind field calculation. is the true data value in the latitudinal direction.

[0086] Step S2: Feature aggregation:

[0087] The H and HH obtained in step S1 are respectively divided into different sub-ocean wave height prediction fields, and then all sub-ocean wave height prediction fields are subjected to feature aggregation to obtain the aggregated feature K.

[0088] In step S2, H and HH are divided into ζ sub-ocean wave height prediction fields respectively. and , dimension is X T, and then perform feature aggregation to obtain the aggregate feature K, as shown in the following formula:

[0089] ;

[0090] Among them, AK is the ocean wave height prediction field of the radial wind field, BK is the ocean wave height prediction field of the zonal wind field, and They are ζ sub-ocean radial wave height prediction fields and ζ sub-ocean zonal wave height prediction fields respectively.

[0091] Step S3: Refined prediction of wave height in local sea area:

[0092] Taking the aggregate feature K as input, we first extract and fuse the global and local information of the spatiotemporal features in the sub-region from multiple scale perspectives; then we perform a high-precision solution of the sub-region wave height; finally, we splice the prediction results of multiple sub-regions to obtain the global regional prediction result.

[0093] Combination Figure 3 As shown, step S3 is implemented by a multi-scale physical space global-local information fusion local sea area refined prediction module, specifically: taking the aggregated feature K as input, firstly, using different convolution layers to process to obtain the large-scale spatial feature R0, the medium-scale spatial feature R1, and the small-scale spatial feature R2, the spatial features at each scale are respectively extracted local information and global information by CNN and Transformer, and then the local information and the global information are fused to obtain fused features FUUU0, FUUU1, and FUUU2, respectively, the fused features at different scales are restored to the original size by convolution operation, and the large-scale spatial feature BL0, the medium-scale spatial feature BL1, and the small-scale spatial feature BL2 are respectively obtained, and then BL0, BL1, and BL2 are flattened to G0, G1, and G2, respectively, and then the feature matrices G0G1, G0G2, and G1G2 are obtained by cross-attention operation, and then the cross-fusion results are mapped back to the same shape as the original features to obtain G0', G1', and G2', and then G0', G1', and G2' are fused to obtain the final global area prediction result P.

[0094] Step S3, set the dimension to The aggregated feature K is used as input, and then the wind field features are calculated , wind field characteristics is the east-west wind speed component and north-south wind speed components The square root of the sum of the squares of The specific calculation process is as follows: construct convolutional layers conv7, conv8, and conv9 based on convolutional kernels of different sizes, process K using different convolutional layers, and output spatial features R0, R1, and R2 at multiple scales.

[0095] ;

[0096] ;

[0097] ;

[0098] Then, the global and local information of R0, R1 and R2 are extracted through the CNN module and Transformer module respectively.

[0099] The specific calculation process is as follows: In the Transformer module, taking R0 as an example, different weights are assigned to R0 based on the self-attention mechanism and the weight matrix is ​​calculated. Then, the weight matrix is ​​used to assign different attentions to different features, and finally a feature MM1 containing the relationship between features is obtained. Similarly, R1 and R2 are processed by the Transformer module to obtain MM2 and MM3.

[0100] In detail, in the Transformer module, only feature matrices are accepted And set the weight, weight Respectively Multiply the result :

[0101] ;

[0102] ;

[0103] ;

[0104] Then calculate the weight matrix of the relationship between features :

[0105] ;

[0106] Final use The relationship weights between the features in the calculation are used to calculate MM1 and allocate attention to different features:

[0107] ;

[0108] The final feature matrix MM1 is the feature matrix of mutual guidance of spatial features at this spatial scale.

[0109] In the CNN module, convolution operations are performed on R0, R1, and R2 to obtain R3, R4, and R5 respectively. Taking the convolution operation on R0 as an example, the formula is as follows:

[0110] ;

[0111] Similarly, convolution operation is performed on R1 and R2. The formula is as follows:

[0112] ;

[0113] .

[0114] Then, the features obtained by the CNN module and the Transformer module are added and fused to obtain the fused feature FUUU. Specifically, R3 and MM1 are added, R4 and MM2 are added, and R5 and MM3 are added. The fused features are represented as FUUU0, FUUU1, and FUUU2 respectively. Then, convolution operation is performed to obtain BL0, BL1, and BL2 respectively.

[0115] Then flatten BL0, BL1, and BL2 into dimensions of G0, G1, G2, and then the feature matrices G0G1, G0G2, G1G2 are obtained through cross attention operation. Then the cross fusion results are mapped back to the same shape as the original features to obtain G0', G1', G2'. G0', G1', G2' are added and fused to obtain the final global area prediction result P.

[0116] Finally, the output features are placed in their original geographical locations and spliced ​​with other sub-ocean wave height prediction fields, resulting in a final dimension of The predicted wave height P all .

[0117] The features obtained by the large-scale CNN module And the features obtained by the Transformer module After fusion, the large-scale features after fusion are expressed as , the formula is as follows:

[0118] ;

[0119] Similarly, the features obtained by the mesoscale CNN module And the features obtained by the Transformer module After fusion, the fused mesoscale features are expressed as The features obtained by the small-scale CNN module And the features obtained by the Transformer module After fusion, the fused small-scale features are expressed as , the formula is as follows:

[0120] ;

[0121] .

[0122] After that, a convolution operation is performed to restore the image to its original size. The formula is as follows:

[0123] ;

[0124] ;

[0125] .

[0126] The final dimension is Large-scale spatial features BL0, medium-scale spatial features BL1, and small-scale spatial features BL2;

[0127] Then, the cross attention operation is performed, that is, the large-scale spatial features BL0, medium-scale spatial features BL1, and small-scale spatial features BL2 are flattened into dimensions of The two-dimensional features G0, G1, G2 are then interacted to obtain the feature matrix G0 ’ 、G1 ’ , G2 ’ ; Finally, G0 ’ 、G1 ’ , G2 ’ Fusion, to obtain the final global regional prediction result P;

[0128] ;

[0129] ;

[0130] ;

[0131] Finally, we get P;

[0132] .

[0133] Finally, the output features are placed in their original geographical locations and spliced ​​with other sub-ocean wave height prediction fields, resulting in a final dimension of The predicted wave height P all .

[0134] Finally, the real data values ​​of M×N spatial points are Calculate the loss, where is the predicted value of the physical field of wave height at the next moment in the jth sample, is the true value of the wave height physical field at the next moment in the jth sample. As shown in the following formula; after completing the prediction and loss calculation of all samples in a batch (batch size is defined as m), the final loss value Loss is obtained through the following formula, and the back propagation operation is performed to update the model parameters.

[0135] ;

[0136] The present invention also provides an intelligent prediction system for wave height based on global and local information fusion, which is used to implement the intelligent prediction method for wave height based on global and local information fusion as described above. The system includes a global area pre-trained boundary missing information completion module and a multi-scale physical space global-local information fusion local sea area refined prediction module.

[0137] The global area pre-trained boundary missing information completion module extracts and fuses spatial features of the global area to ensure that the sub-area boundary is integrated with surrounding information; the global area pre-trained boundary missing information completion module includes different convolutional layers, CNN modules, Transformer modules and information aggregation modules. The input sea surface radial wind field data U and latitudinal wind field data V are processed by different convolutional layers to obtain spatial features at multiple scales. The spatial features at each scale are respectively extracted from local information and global information by the CNN module and the Transformer module, and then the local information and the global information are fused to obtain fused features. Finally, the fused features at different scales are added to obtain radial wind field fusion features H and latitudinal wind field fusion features HH respectively; in the information aggregation module, H and HH are respectively divided into different sub-ocean wave height prediction fields, and all sub-ocean wave height prediction fields are subjected to feature aggregation to obtain aggregated features K.

[0138] The multi-scale physical space global-local information fusion local sea area refined prediction module realizes global and local information extraction from the spatiotemporal features in the sub-area under multiple scale perspectives, including different convolutional layers, CNN modules, Transformer modules and cross-fusion modules. The processing process of the convolutional layer, CNN module and Transformer module of the multi-scale physical space global-local information fusion local sea area refined prediction module is the same as that of the global area pre-trained boundary missing information completion module. The obtained fusion feature is input into the cross-fusion module, and a cross-attention operation is first performed, and the result is mapped back to the same shape as the original feature, and then a fusion operation is performed to obtain the final global area prediction result P. Finally, the output feature is placed in its original geographical location and spliced ​​with other sub-ocean wave height prediction fields to finally obtain the predicted wave height P. all .

[0139] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Any changes, modifications, additions or substitutions made by ordinary technicians in this technical field within the essential scope of the present invention should fall within the protection scope of the present invention.

Claims

1. An intelligent prediction method for wave height based on global and local information fusion, characterized in that: The following steps are involved: Step S1: Extract and fuse spatial features of the global region: The radial wind field data U and the latitudinal wind field data V of the sea surface are used as input. For each radial wind field data and latitudinal wind field data, different convolutional layers are used to process the spatial features at multiple scales. The local information and global information of the spatial features at each scale are extracted through the CNN module and the Transformer module respectively. Then, the local information and the global information are fused to obtain the fused features. Finally, the fused features at different scales are added together to obtain the radial wind field fused feature H and the latitudinal wind field fused feature HH. Specifically, step S1 names the radial wind field data and the latitudinal wind field data of the sea surface at n moments as U0, ..., U n-1 , V0, ..., V n-1 As input; the specific calculation process is as follows: construct convolution layers based on convolution kernels of different sizes, process the radial wind field data U and the latitudinal wind field data V using different convolution layers respectively, and output spatial features at multiple scales. For U, after being processed by different convolution layers conv0, conv1, and conv2, the spatial features F0, F1, and F2 at multiple scales are output; Then, the local and global information of F0, F1, and F2 are extracted through the CNN module and the Transformer module respectively. The specific calculation process is as follows: In the Transformer module, different weights are assigned to F0 based on the self-attention mechanism and the weight matrix is ​​calculated. Then, the weight matrix is ​​used to assign different attentions to different features, and finally a feature M1 containing the relationship between features is obtained. Similarly, F1 and F2 are processed by the Transformer module to obtain M2 and M3. In the CNN module, convolution operations are performed on F0, F1, and F2 to obtain F3, F4, and F5 respectively; The features F3, F4, and F5 obtained by the CNN module are local information at different scales, and the features M1, M2, and M3 obtained by the Transformer module are global information at different scales; The features obtained by the CNN module and the Transformer module are added and fused to obtain the fused feature FU. Specifically, F3 and M1 are added, F4 and M2 are added, and F5 and M3 are added. The fused features are represented as FU0, FU1, and FU2 respectively. Then, a convolution operation is performed to obtain BG0, BG1, and BG2 respectively. BG0, BG1, and BG2 are added to obtain the radial wind field fusion feature H. Similarly, for the latitudinal wind field data V, the latitudinal wind field fusion feature HH is obtained. Step S2: Feature aggregation: The H and HH obtained in step S1 are respectively divided into different sub-ocean wave height prediction fields, and then all sub-ocean wave height prediction fields are subjected to feature aggregation to obtain the aggregated feature K; Step S3: Refined prediction of wave height in local sea area: Taking the aggregate feature K as input, firstly, the spatiotemporal features in the sub-region are extracted and fused globally and locally at multiple scales; then the sub-region wave height is solved with high precision; finally, the prediction results of multiple sub-regions are spliced ​​to obtain the global regional prediction result; Step S3 is specifically as follows: taking the aggregated feature K as input, firstly using different convolutional layers to process respectively to obtain the spatial feature R0 at a large scale, the spatial feature R1 at a medium scale, and the spatial feature R2 at a small scale, the spatial features at each scale are respectively extracted local information and global information by CNN and Transformer, and then the local information and the global information are fused to obtain fused features FUUU0, FUUU1, and FUUU2, respectively, the fused features at different scales are restored to the original size through the convolution operation, and the large-scale spatial feature BL0, the medium-scale spatial feature BL1, and the small-scale spatial feature BL2 are respectively obtained, and then BL0, BL1, and BL2 are flattened to G0, G1, and G2, and then the feature matrices G0G1, G0G2, and G1G2 are obtained through the cross attention operation, and then the cross fusion result is mapped back to the same shape as the original feature to obtain G0', G1', and G2', and then G0', G1', and G2' are fused to obtain the final global area prediction result P.

2. The method for intelligent prediction of wave height based on global and local information fusion according to claim 1 is characterized in that: In step S2, H and HH are divided into ζ sub-ocean wave height prediction fields AK0, ..., AK ζ-1 and BK0, ..., BK ζ-1 , and then perform feature aggregation to obtain the aggregated feature K, K = concat(AK, BK); Among them, AK is the ocean wave height prediction field of the radial wind field, BK is the ocean wave height prediction field of the zonal wind field, concat represents the feature aggregation operation, AK0,...,AK ζ-1 and BK0,...,BK ζ-1 They are ζ sub-ocean radial wave height prediction fields and ζ sub-ocean zonal wave height prediction fields respectively.

3. The method for intelligent prediction of wave height based on global and local information fusion according to claim 1 is characterized in that: The specific calculation process of step S3 is as follows: construct convolution layers based on convolution kernels of different sizes, process the aggregated features K using different convolution layers, output spatial features R0, R1, and R2 at multiple scales, and then extract global and local information from R0, R1, and R2 through CNN modules and Transformer modules respectively. The specific calculation process is as follows: In the Transformer module, for R0, different weights are assigned to R0 based on the self-attention mechanism and the weight matrix is ​​calculated. Then, the weight matrix is ​​used to assign different attentions to different features, and finally a feature MM1 containing the relationship between features is obtained. Similarly, R1 and R2 are processed by the Transformer module to obtain MM2 and MM3. In the CNN module, convolution operations are performed on R0, R1, and R2 to obtain R3, R4, and R5 respectively; Then, the features obtained by the CNN module and the Transformer module are added and fused to obtain the fused feature FUUU. Specifically, R3 and MM1 are added, R4 and MM2 are added, and R5 and MM3 are added. The fused features are represented as FUUU0, FUUU1, and FUUU2 respectively. Then, convolution operation is performed to obtain BL0, BL1, and BL2 respectively. Then BL0, BL1, and BL2 are flattened into G0, G1, and G2 respectively, and then the feature matrices G0G1, G0G2, and G1G2 are obtained through cross-attention operation. Then the cross-fusion results are mapped back to the same shape as the original features to obtain G0', G1', and G2'. G0', G1', and G2' are added and fused to obtain the final global area prediction result P; Finally, the output features are placed at their original geographic locations and spliced ​​with other sub-ocean wave height prediction fields to obtain the predicted wave height P all .

4. The intelligent prediction system of wave height based on global and local information fusion is characterized by: Used to implement the wave height prediction method based on global and local information fusion as described in any one of claims 1-3, the system includes a global area pre-trained boundary missing information completion module, a multi-scale physical space global-local information fusion local sea area refined prediction module, the global area pre-trained boundary missing information completion module extracts and fuses spatial features of the global area to ensure that the sub-area boundary is integrated with surrounding information; The multi-scale physical space global-local information fusion local sea area refined prediction module realizes global and local information extraction of spatiotemporal features in sub-areas from multiple scale perspectives and outputs predicted wave heights.

5. The intelligent prediction system for wave height based on global and local information fusion according to claim 4 is characterized in that: The global area pre-trained boundary missing information completion module includes different convolutional layers, CNN modules, Transformer modules and information aggregation modules. The input sea surface radial wind field data U and latitudinal wind field data V are processed by different convolutional layers to obtain spatial features at multiple scales. The spatial features at each scale are respectively extracted from local information and global information by CNN modules and Transformer modules. Then the local information and the global information are fused to obtain fused features. Finally, the fused features at different scales are added to obtain radial wind field fusion features H and latitudinal wind field fusion features HH. In the information aggregation module, H and HH are respectively divided into different sub-ocean wave height prediction fields, and all sub-ocean wave height prediction fields are feature aggregated to obtain aggregated features K.

6. The intelligent prediction system for wave height based on global and local information fusion according to claim 4 is characterized in that: The multi-scale physical space global-local information fusion local sea area refined prediction module includes different convolutional layers, CNN modules, Transformer modules and cross-fusion modules. The processing process of the convolutional layers, CNN modules and Transformer modules of the multi-scale physical space global-local information fusion local sea area refined prediction module is the same as that of the global area pre-trained boundary missing information completion module. The obtained fusion features are input into the cross-fusion module, and the cross-attention operation is first performed, and the results are mapped back to the same shape as the original features, and then the fusion operation is performed to obtain the final global area prediction result P. Finally, the output features are placed in their original geographical locations and spliced ​​with other sub-ocean wave height prediction fields to finally obtain the predicted wave height P. all .

Images