An extended-scale prediction method for ocean climate based on artificial intelligence

Abstract

The present invention discloses an artificial intelligence-based extended-scale prediction method for ocean climate, which specifically relates to the field of ocean climate prediction, including fusing and standardizing multi-source spatiotemporal data, extracting multi-scale features, generating extreme event samples, and then conducting joint training to generate adaptively optimized prediction results. An artificial intelligence-based extended-scale prediction method for ocean climate effectively reduces the prediction error amplification caused by insufficient cross-scale interaction through a cascaded spatiotemporal attention mechanism; based on a conditional variational autoencoder, it supplements historical sparse samples through physical encoding and residual correction mechanisms, reduces the model's dependence on training set coverage, and improves the generalization ability of extreme event features; by dynamically calculating the feature correlation matrix of atmospheric and ocean modes using a gated recurrent unit, and introducing a conservation law loss function to constrain the feature evolution direction, the correlation and physical consistency of cross-scale features are strengthened.

Description

Technical Field

[0001] The present invention relates to the technical field of ocean climate prediction, and more specifically, to an artificial intelligence-based ocean climate extended-scale prediction method. Background Art

[0002] With the deep integration of ocean observation technology and numerical simulation, the demand for accuracy and timeliness of extended-scale ocean climate prediction is particularly urgent. Traditional technology consists of three parts: atmosphere-ocean coupling model, assimilation system and parameterization scheme. It constructs a gridded prediction system by solving a set of equations and relies on observation data assimilation and parameter optimization to achieve climate prediction. However, due to the computational complexity of solving the discretized physical equations, when extended to longer scales, there is a phenomenon of accumulated errors in the parameterization scheme leading to prediction bias.

[0003] In order to break through the limitations of computational efficiency and error control, existing technologies use deep learning algorithms. By constructing a spatiotemporal convolutional neural network, it uses an end-to-end training mechanism to directly learn the mapping relationship between multi-source observation data and model outputs, and by introducing an attention mechanism, the mean square error of sea temperature prediction on a monthly scale is reduced.

[0004] However, it still has some shortcomings in actual use. For example, the separated design leads to insufficient extraction of cross-scale interaction features, causing the amplification of the prediction error of sub-seasonal vortex kinetic energy transfer; data-driven strongly relies on the coverage of the training set. When applied to extreme events such as El Niño, the standard deviation of the prediction results increases due to the sparsity of historical samples. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides an artificial intelligence-based ocean climate extended-scale prediction method, which solves the problems raised in the above-mentioned background technology through the following scheme.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] An artificial intelligence-based extended-scale ocean climate prediction method, comprising:

[0008] S1: Obtain the first ocean scale including ocean satellite remote sensing and buoy observations, perform spatiotemporal alignment reconstruction through a physical constraint adversarial network, and output standardized grid data as the first key scale;

[0009] S2: Deploy the spatiotemporal attention mechanism and extract first-scale features including vortex and circulation modes based on the first key scale;

[0010] S3: Based on the first-scale features, the conditional variational autoencoder is driven to generate extreme event samples as the second-scale features;

[0011] S4: Deploy a cross-modal feature interaction gate to dynamically calculate the feature correlation matrix of the first-scale features and the second-scale features, and construct a hybrid optimization objective that includes a conservation law loss function and a data-driven loss function to generate third-scale features;

[0012] S5: Based on the third scale feature, the cross-modal coupling coefficient is calculated using a differentiable verification mechanism embedded in the vorticity pseudo-energy conservation equation, and the weight distribution of the cross-modal feature interaction gate is dynamically adjusted;

[0013] S6: Dynamically adjust the loss function proportional coefficient of the hybrid optimization objective and the computing power distribution of the spatiotemporal attention mechanism according to the prediction error distribution output by the differentiable verification mechanism.

[0014] Preferably, the deployment of the spatiotemporal attention mechanism in S2 specifically includes:

[0015] A cascaded architecture using spatiotemporal graph neural networks and 3D convolutional attention;

[0016] The spatiotemporal graph neural network uses the standardized spatiotemporal grid corresponding to the first key scale as nodes, and the nodes contain four-dimensional features of sea surface temperature, sea surface height, latitudinal velocity, and meridional velocity. Based on the sea temperature gradient and velocity shear between adjacent nodes, a Gaussian kernel function is used to calculate the dynamic adjacency matrix as the edge weight;

[0017] The three-dimensional convolutional attention performs maximum pooling along the time dimension and average pooling along the spatial dimension, uses a two-dimensional convolution kernel to generate a spatial weight map and generates an attention matrix by fusing weights across time slices.

[0018] Preferably, the step S2 of obtaining the first scale feature specifically includes:

[0019] The Gaussian kernel function is used to calculate the dynamic adjacency matrix to capture the local interaction characteristic map of the kilometer-scale vortex, which is specifically expressed as:

[0020] .

[0021] in, Represented as a node and nodes The edge weights between Represented as a node The sea temperature gradient at Represented as a node The sea temperature gradient at Expressed as the temperature gradient bandwidth coefficient, Expressed as the velocity correlation weight, Represented as a node The velocity vector at Represented as a node The velocity vector at .

[0022] Preferably, the step S3, generating the second scale feature, specifically includes:

[0023] Physically encoding the thermodynamic characteristics corresponding to the first scale characteristics;

[0024] splicing the development stages of extreme events with the characteristics of the physical code;

[0025] Feature reconstruction is performed through multi-layer 3D deconvolution of the decoder in the conditional variational autoencoder, and physical constraints are introduced before each layer to calculate the thermodynamic residual of the generated sample, and the output is corrected through residual connections.

[0026] Preferably, the S3, physical coding content as the physical information driven condition includes the tropical Pacific thermocline depth, the spatial gradient of the zonal wind stress anomaly and the ocean mixed layer heat content anomaly extracted from the first scale feature.

[0027] Preferably, the step S4 of calculating a feature correlation matrix of the first scale feature and the second scale feature specifically includes:

[0028] Quantify atmospheric and ocean characteristics at each spatial location ( ) by calculating the atmospheric eigenvector With the ocean feature vector The cosine similarity between Measure, specifically expressed as:

[0029] .

[0030] Quantify the global correlation between atmospheric and oceanic characteristics by and ocean features Perform average pooling operation in the spatial dimension and calculate the mutual information based on the distribution of target area features , specifically expressed as:

[0031] .

[0032] in, Expressed as the joint probability distribution of atmospheric and oceanic characteristics, and Represented as marginal probability distributions of atmospheric and oceanic features, respectively.

[0033] Preferably, the step S4 of calculating a feature correlation matrix of the first scale feature and the second scale feature further includes:

[0034] The characteristic correlation matrix between the atmospheric and ocean modes is updated using the exponential moving average , specifically expressed as:

[0035] .

[0036] in, Expressed as the updated feature correlation matrix, Expressed as the feature correlation matrix of the previous time step, Represented as the feature correlation matrix of the current time step, Expressed as a time decay factor;

[0037] The time decay factor decays exponentially with the increase of the time step, which can be specifically expressed as:

[0038] .

[0039] in, Expressed as the time since the last update, Expressed as the decay constant.

[0040] Preferably, the S4, hybrid optimization objective includes a conservation law loss function that constrains the feature evolution direction by embedding the vorticity pseudo-energy conservation equation, and a data-driven loss function that uses an improved Wasserstein distance to evaluate the matching degree of the predicted field distribution.

[0041] The technical effects and advantages of the present invention are as follows:

[0042] 1. This invention uses a cascaded spatiotemporal attention mechanism to construct a dynamic adjacency matrix based on sea temperature gradients and velocity shear, capturing the multi-scale feature tensors of local interactions between kilometer-scale vortices and hundred-kilometer-scale circulation modes, achieving coupled extraction of cross-scale features and effectively reducing the amplification of prediction errors caused by insufficient cross-scale interactions.

[0043] 2. Based on a conditional variational autoencoder, this invention generates extreme event samples based on physical information and thermodynamic constraints. Physical coding and residual correction mechanisms are used to supplement historical sparse samples, reducing the model's dependence on training set coverage, reducing the standard deviation of extreme event predictions, and improving the generalization ability of extreme event characteristics.

[0044] 3. The present invention deploys cross-modal feature interaction gates, uses gated recurrent units to dynamically calculate the feature correlation matrices of atmospheric and oceanic modes, generates physically interpretable interaction features through tensor fusion, and introduces conservation law loss functions to constrain the feature evolution direction, thereby strengthening the correlation and physical consistency of cross-scale features and avoiding kinetic energy transfer prediction deviations caused by the separate design. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a flowchart of the steps of an artificial intelligence-based ocean climate extended-scale prediction method provided according to an embodiment of the present application.

[0046] Figure 2 This is a block diagram of a spatiotemporal attention mechanism in an artificial intelligence-based ocean climate extended-scale prediction system provided according to an embodiment of the present application. DETAILED DESCRIPTION

[0047] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0048] The terms used in the following examples of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the specification of the present application, the singular expressions "a", "an", "said", "above", "the", and "this" are intended to include plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used in the present application refers to and includes any or all possible combinations of one or more of the listed items.

[0049] In the following, the terms "first," "second," and "third" are used for descriptive purposes only and should not be understood to imply or suggest relative importance or implicitly indicate the number of the technical features indicated. Therefore, a feature defined as "first," "second," and "third" may explicitly or implicitly include one or more of the features. In the description of the embodiments of this application, unless otherwise specified, "plurality" means two or more.

[0050] As attached Figure 1 The AI-based extended-scale ocean climate prediction method shown in the figure includes fusing and standardizing multi-source spatiotemporal data, extracting multi-scale features, generating extreme event samples, and then conducting joint training to generate adaptively optimized prediction results. Specifically, it includes:

[0051] S1: Obtain the first ocean scale including ocean satellite remote sensing and buoy observations, perform spatiotemporal alignment reconstruction through a physical constraint adversarial network, and output standardized grid data as the first key scale;

[0052] S2: Deploy the spatiotemporal attention mechanism and extract first-scale features including vortex and circulation modes based on the first key scale;

[0053] S3: Based on the first-scale features, the conditional variational autoencoder is driven to generate extreme event samples as the second-scale features;

[0054] S4: Deploy a cross-modal feature interaction gate to dynamically calculate the feature correlation matrix of the first-scale features and the second-scale features, and construct a hybrid optimization objective that includes a conservation law loss function and a data-driven loss function to generate third-scale features;

[0055] S5: Based on the third scale feature, the cross-modal coupling coefficient is calculated using a differentiable verification mechanism embedded in the vorticity pseudo-energy conservation equation, and the weight distribution of the cross-modal feature interaction gate is dynamically adjusted;

[0056] S6: Dynamically adjust the loss function proportional coefficient of the hybrid optimization objective and the computing power distribution of the spatiotemporal attention mechanism according to the prediction error distribution output by the differentiable verification mechanism.

[0057] Specifically, in S1, the first ocean scale is constructed by acquiring multimodal data of ocean satellite remote sensing, buoy array observations and reanalysis data, and a physical constrained adversarial network is used to perform spatiotemporal alignment and missing value reconstruction to output standardized spatiotemporal grid data, namely the first key scale.

[0058] In this embodiment, the ocean satellite remote sensing data in the first ocean scale is obtained by accessing the L2 level products of the ocean dynamic satellite, including but not limited to sea surface height, sea surface temperature, sea surface wind field data, etc., and the spatial resolution is set to 0.25°×0.25°, and the time sampling interval is 6 hours; the data observed by the buoy array are obtained by integrating the temperature and salinity profile data of the Argo buoy array and the sea-air flux data through the TRITON moored buoy, and are converted into 1°×1° grid data by using the Kriging interpolation method, and are time-aligned to the timestamp of the ocean satellite remote sensing data; the reanalysis data calls the ocean mixed layer depth and subsurface flow field data in the ECMWF ERA5 reanalysis data set, and are unified to a spatial grid of 0.5°×0.5° by bilinear interpolation.

[0059] It should be noted that the generator architecture of the physical constraint adversarial network receives the multimodal data as input, and extracts deep spatiotemporal correlation features through a spatiotemporal encoder composed of 3D convolution and spatiotemporal attention mechanism; then, the quasi-geostrophic vorticity equation is embedded as a partial differential equation constraint, and the vorticity pseudo-energy residual of the characteristic field is calculated by automatic differentiation, so as to integrate the physical laws into the first ocean scale; at the same time, for the missing values ​​in the first ocean scale, the missing areas are identified by constructing a mask matrix. The missing areas in this embodiment include but are not limited to satellite orbit gaps, sparse buoy distribution areas, etc. Prior filling values ​​are generated based on the ConvLSTM network prediction trend, and mass conservation and heat conduction equation constraints are imposed on the missing areas during the generation process of the physical constraint adversarial network, and the generator loss function is integrated through the Lagrange multiplier method; then, the deconvolution layer is used as a decoder to reconstruct the spatiotemporal grid data that is aligned with the input data and standardized as the first key scale.

[0060] Furthermore, the discriminator of the physical constraint adversarial network adopts a dual-branch structure. The physical compliance branch calculates the matching degree between the reconstructed data and the vorticity pseudo-energy conservation equation, using the L2 norm as a constraint; the statistical truth branch uses spectral normalized convolution to discriminate the authenticity of the spatial power spectrum distribution of the input data.

[0061] Specifically, in S2, a cascaded spatiotemporal attention mechanism is used to extract first-scale features, which include multi-scale feature tensors of kilometer-level vortex features and hundred-kilometer-level circulation modal features.

[0062] In one possible implementation, the spatiotemporal attention mechanism adopts a cascaded architecture of a spatiotemporal graph neural network and a three-dimensional convolutional attention, and the construction steps include: using the standardized spatiotemporal grid corresponding to the first key scale as the node of the feature graph, the node includes four-dimensional features of sea surface temperature, sea surface height, latitudinal velocity and meridional velocity; based on the sea temperature gradient between adjacent nodes Shear with flow velocity, and use Gaussian kernel function to calculate dynamic adjacency matrix as edge weight ; In this embodiment, the structure of the feature map is reconstructed every 6 hours to capture the movement trajectory of the vortex; multiple layers of GATv2 are stacked, and multiple independent attention heads are used to calculate the feature correlation between nodes; the feature map output by the spatiotemporal graph neural network is maximum pooled along the time dimension to extract the key features of each spatial position in the time dimension; average pooling is performed along the spatial dimension to highlight important features and reduce the data dimension; a two-dimensional convolution kernel is used to generate a spatial weight map, and an attention matrix is ​​generated by cross-time slice weight fusion.

[0063] It should be noted that the spatiotemporal graph neural network and three-dimensional convolutional attention adopt a cascade optimization strategy. The output features of the spatiotemporal graph neural network are converted into grid data through spatial interpolation. The three-dimensional convolutional attention realizes multi-scale feature fusion, and then extracts the circulation mode of hundreds of kilometers. This embodiment prioritizes retaining the feature channels related to the geostrophic flow, the two-dimensional convolution kernel corresponds to the actual spatial range, and the cross-time slice weight fusion effectively suppresses the interference of weather-scale noise on the extraction of circulation mode. The attention matrix data represents the circulation mode.

[0064] In one possible implementation, obtaining the first scale feature includes: the spatiotemporal graph neural network uses a Gaussian kernel function to calculate a dynamic adjacency matrix to capture the local interaction feature map of the kilometer-scale vortex, which is specifically expressed as:

[0065] .

[0066] in, Represented as a node and nodes The edge weights between Represented as a node The sea temperature gradient at Represented as a node The sea temperature gradient at Expressed as the temperature gradient bandwidth coefficient, Expressed as the velocity correlation weight, Represented as a node The velocity vector at Represented as a node The velocity vector at .

[0067] It should be noted that the temperature gradient bandwidth coefficient is used to control the width of the Gaussian kernel function to affect the attenuation rate of the temperature gradient difference on the edge weight. Ensure the effective modeling of 50-100km scale interaction; the velocity correlation weight is used to adjust the importance of velocity correlation in the entire edge weight calculation, and the value range is between 0 and 1. The contribution ratio of dynamic characteristics and thermal characteristics is balanced; the velocity vector represents the speed and direction of the seawater flow at the target node, including the latitudinal velocity and the longitudinal velocity; the sea temperature gradient term between nodes Reflects the frontal strength difference at the vortex edge; velocity correlation term The coherent motion characteristics of the vortex core were quantified.

[0068] Specifically, in S3, based on the multi-scale feature tensor output by S2, a physical information-driven conditional variational autoencoder is designed, which combines thermodynamic constraints to generate feature data corresponding to extreme event samples that conform to physical laws as second-scale features.

[0069] In one possible implementation, generating the second-scale feature includes: physically encoding the thermodynamic features corresponding to the first-scale features, with the content of the physical encoding serving as a condition driven by physical information, the conditions including the depth of the tropical Pacific thermocline, the spatial gradient of the zonal wind stress anomaly, and the ocean mixed layer heat content anomaly; splicing the development stage of the extreme event with the features of the physical encoding; reconstructing the features through multi-layer 3D deconvolution of the decoder in the conditional variational autoencoder, introducing physical constraints before each layer to calculate the thermodynamic residual of the generated sample, and correcting the output through residual connection.

[0070] In this embodiment, the vortex characteristics obtained in S2 are spatially averaged and pooled to obtain the corresponding latitudinal wind stress gradient profile, and the circulation characteristics and the corresponding thermocline depth are screened; at the same time, the spatiotemporal evolution trend of the ocean mixed layer heat content anomaly is mined based on the attention matrix.

[0071] It should be noted that the physical constraints are reflected by the loss function, which is specifically expressed as:

[0072] .

[0073] in, Expressed as a physical constraint expressed as a loss function, Expressed as the variational lower bound loss, Expressed as physical constraint loss, Expressed as feature matching loss, 、 、 They are respectively represented by the preset weights of the variational lower bound loss, the physical constraint loss, and the feature matching loss. It should be noted that the variational lower bound loss is the KL divergence and the reconstruction loss based on the SSIM indicator. The feature matching loss is the SSTA correlation coefficient between the generated samples and historical extreme events in the Niño 3.4 region. The physical constraint loss is the degree to which the generated samples satisfy the physical laws. The calculation of the physical constraint loss in this embodiment is specifically expressed as follows:

[0074] .

[0075] in, Expressed as the number of generated samples, Represents the index of the generated sample, Expressed as The velocity field of the generated samples includes the velocity information of seawater at different locations and times. Expressed as The sea level height of the generated samples; it should be noted that, Expressed as the divergence of the velocity field, It is expressed as the abnormal increase or decrease of sea level height per unit time.

[0076] Specifically, during the S4 training process, a cross-modal feature interaction gate is deployed, and a gated recurrent unit is used to dynamically calculate the feature correlation matrix between the atmospheric and ocean modes contained in the first-scale features and the second-scale features. A tensor fusion operation is used to generate physically interpretable interaction features, and a hybrid optimization objective including a conservation law loss function and a data-driven loss function is further constructed to generate the third-scale features.

[0077] It should be noted that the cross-modal feature interaction gate splices the atmospheric features and ocean features contained in the first scale feature and the second scale feature, and performs spatiotemporal sequence reorganization, inputs the gated recurrent unit to obtain the feature correlation matrix between the atmospheric and ocean modes, and then obtains the interaction feature through tensor fusion operation projection. .

[0078] In this embodiment, the ocean features included in the first scale feature include but are not limited to vorticity field, velocity shear, sea surface height anomaly, latitudinal flow profile, meridional temperature gradient, mixed layer depth, etc.; the atmospheric features included in the first scale feature include but are not limited to latitudinal wind stress, meridional wind stress, sea surface sensible heat flux, latent heat flux, etc.; the ocean features included in the second scale feature include but are not limited to tropical Pacific sea surface temperature, thermocline depth, ocean heat content anomaly, equatorial undercurrent intensity, Kelvin wave and Rossby wave propagation phase, etc.; the atmospheric features included in the second scale feature include but are not limited to Walker circulation intensity index, convective activity, etc. The calculation formula of the tensor fusion operation is as follows:

[0079] .

[0080] in, It is represented as the generation-level normalization operation, Represented as the original atmospheric features that have not been processed by the cross-modal feature interaction gate, Represented as the original ocean features that have not been processed by the cross-modal feature interaction gate, Expressed as the characteristic correlation matrix between the atmospheric and oceanic modes, Expressed as a pair The tensor transpose operation, Expressed as matrix multiplication.

[0081] In one possible implementation, dynamically calculating the feature correlation matrix of the first scale feature and the second scale feature using a gated recurrent unit includes: quantifying the atmospheric feature and the ocean feature at each spatial position ( ) by calculating the atmospheric eigenvector With the ocean feature vector The cosine similarity between Measure, specifically expressed as:

[0082] .

[0083] It should be noted that the cosine similarity value The larger the value, the more ) The closer the directions of the two modal eigenvectors at the position are, the higher the correlation is. Furthermore, the global correlation between atmospheric and oceanic features is quantified. and ocean features Perform average pooling operation on the spatial dimension to obtain the vector of the overall features of the target area and , and calculate the mutual information based on the distribution of target area features , specifically expressed as:

[0084] .

[0085] in, Expressed as the joint probability distribution of atmospheric and oceanic characteristics, and Represented as marginal probability distributions of atmospheric and oceanic features, respectively.

[0086] It should be noted that mutual information is used to measure atmospheric characteristics and ocean features The degree of mutual dependence between the modes reflects the overall nonlinear correlation strength between the modes.

[0087] In one possible implementation, dynamically calculating the feature correlation matrix of the first scale feature and the second scale feature using a gated recurrent unit also includes: the gated recurrent unit dynamically adjusts the degree of fusion of the previous time step data and the current time step data through an internal update gate and a reset gate. When calculating local correlation, the update gate determines the proportion of the local correlation calculation result in the current time step that retains the previous time step data, while the reset gate controls whether to recalculate the local correlation, so that the local correlation can dynamically adapt to different spatiotemporal feature changes; when calculating global correlation, the gated recurrent unit uses a memory unit to store and update the historical information of the global correlation, so that the global correlation can be dynamically adjusted over time, thereby better capturing the long-term dependency between atmospheric and ocean features; further, the feature correlation matrix between the atmospheric and ocean modes is updated in the form of an exponential moving average. , specifically expressed as:

[0088] .

[0089] in, Expressed as the updated feature correlation matrix, Expressed as the feature correlation matrix of the previous time step, Represented as the feature correlation matrix of the current time step, It is expressed as a time decay factor; further, the time decay factor decays exponentially with the increase of the time step, which is specifically expressed as:

[0090] .

[0091] in, Expressed as the time since the last update, It is expressed as the decay constant, which is set to 6 hours in this embodiment.

[0092] It should be noted that the gated recurrent unit dynamically adjusts The value of enables the feature correlation matrix to be updated adaptively according to the current spatiotemporal feature changes; when the correlation between the atmospheric and ocean features changes significantly, the gated recurrent unit reduces The value of makes the current feature correlation matrix occupy a larger proportion in the update process; when the feature correlation is relatively stable, the gated recurrent unit will improve value to retain more historical information, thereby ensuring the stability and continuity of the feature correlation matrix.

[0093] In this embodiment, physical conservation constraints are introduced, and the vorticity pseudo-energy conservation equation is embedded in the conservation law loss function to constrain the characteristic evolution direction, so as to predict the relative vorticity field. and the true relative vorticity field The difference of eddy pseudo-energy integral value in the region , specifically expressed as:

[0094] .

[0095] in, Expressed as relative vorticity, Expressed as the number of dynamic feature correlation matrices within the target area, Represented as the index of the dynamic feature association matrix in the target area; the data-driven loss function uses the improved Wasserstein distance to evaluate the matching degree of the predicted field distribution to measure the model prediction feature distribution and the true feature distribution The difference between , specifically expressed as:

[0096] .

[0097] in, Denote as the discriminator function, Expressed as a linear interpolation of predicted features and true features, Expressed as a gradient penalty term, this embodiment , It represents the difference between the expected score of the discriminator for the predicted feature and the true feature.

[0098] It should be noted that the third scale feature is generated by the potential feature generated by physical conservation constraint, through projection and GELU activation function.

[0099] Specifically, in S5, the third-scale features are mapped to the physical quantity field through a trainable projection layer, that is, the third-scale features are converted into physical quantities directly related to ocean dynamics, including but not limited to components of relative vorticity and velocity fields.

[0100] In this embodiment, the relative vorticity The components of the velocity field are obtained by the third-scale feature through the tanh activation function and the preset projection matrix. It is generated by the Softsign activation function and the preset projection matrix; further, the discretized vorticity pseudo-energy conservation equation is used as the verification basis, which is specifically expressed as:

[0101] .

[0102] in, is expressed as the Rossby parameter.

[0103] It should be noted that the verification is based on substituting the physical quantity field obtained by converting the third-scale characteristics into the calculated conservation residual, thereby measuring the degree of deviation between the prediction and the physical law. The smaller the conservation residual, the more consistent the predicted physical quantity field is with the actual ocean dynamic process.

[0104] Furthermore, the residual field is processed by a preset coefficient generation network, which includes multiple convolutional layers and pooling layers to extract spatial and temporal feature patterns in the residual field. After multiple nonlinear transformations and feature compression operations, the coupling coefficient vector obtained is the cross-modal coupling coefficient. Each element in the coupling coefficient vector corresponds to the weight of each feature channel in the cross-modal feature interaction gate. By dynamically adjusting the coupling coefficient, dynamic adjustment of the weight distribution of the cross-modal feature interaction gate can be achieved.

[0105] Furthermore, when the conservation residual is large, the coupling coefficient will be adjusted to reduce the residual, so that the feature interaction gate can pay more attention to the feature channels that help meet the physical constraints; conversely, when the conservation residual is small, the coupling coefficient will maintain the current weight distribution and maintain the stability and continuity of the model.

[0106] Specifically, in S6, a two-layer feedback controller is designed. The upper layer dynamically adjusts the mixing ratio of the conservation law loss function and the data-driven loss function in S4 based on the real-time conservation residual output by the differentiable verification mechanism. The lower layer analyzes the available resources based on the cross-modal coupling coefficient to optimize the computing power allocation strategy of the spatiotemporal graph neural network and three-dimensional convolutional attention of the spatiotemporal attention mechanism in S2, so as to achieve the optimal resource ratio of physical constraints and data-driven.

[0107] Secondly: The drawings of the embodiments disclosed in the present invention only involve structures related to the embodiments disclosed in the present invention. Other structures may refer to conventional designs. The same embodiment and different embodiments of the present invention may be combined with each other without conflict.

[0108] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An artificial intelligence-based extended-scale ocean climate prediction method, characterized in that: include: S1: Obtain the first ocean scale including ocean satellite remote sensing and buoy observations, perform spatiotemporal alignment reconstruction through a physical constraint adversarial network, and output standardized grid data as the first key scale; S2: Deploy the spatiotemporal attention mechanism and extract first-scale features including vortex and circulation modes based on the first key scale; S3: Based on the first-scale features, the conditional variational autoencoder is driven to generate extreme event samples as the second-scale features; S4: Deploy a cross-modal feature interaction gate to dynamically calculate the feature correlation matrix of the first-scale features and the second-scale features, and construct a hybrid optimization objective that includes a conservation law loss function and a data-driven loss function to generate third-scale features; Calculating a feature correlation matrix of the first scale feature and the second scale feature includes: Quantify atmospheric and ocean characteristics at each spatial location ( ) by calculating the atmospheric eigenvector With the ocean feature vector The cosine similarity between Measure, specifically expressed as: ， Quantify the global correlation between atmospheric and oceanic characteristics by and ocean features Perform average pooling operation in the spatial dimension and calculate the mutual information based on the distribution of target area features , specifically expressed as: ， in, Expressed as the joint probability distribution of atmospheric and oceanic characteristics, and are expressed as marginal probability distributions of atmospheric and oceanic characteristics, respectively; The characteristic correlation matrix between the atmospheric and ocean modes is updated using the exponential moving average , specifically expressed as: ， in, Expressed as the updated feature correlation matrix, Expressed as the feature correlation matrix of the previous time step, Represented as the feature correlation matrix of the current time step, Expressed as a time decay factor; The time decay factor decays exponentially with the increase of the time step, which can be specifically expressed as: ， in, Expressed as the time since the last update, Expressed as the decay constant; S5: Based on the third scale feature, the cross-modal coupling coefficient is calculated using a differentiable verification mechanism embedded in the vorticity pseudo-energy conservation equation, and the weight distribution of the cross-modal feature interaction gate is dynamically adjusted; S6: Dynamically adjust the loss function proportional coefficient of the hybrid optimization objective and the computing power distribution of the spatiotemporal attention mechanism according to the prediction error distribution output by the differentiable verification mechanism.

2. The artificial intelligence-based extended-scale ocean climate prediction method according to claim 1, characterized in that: The S2, deployment of the spatiotemporal attention mechanism, specifically includes: A cascaded architecture using spatiotemporal graph neural networks and 3D convolutional attention; The spatiotemporal graph neural network uses the standardized spatiotemporal grid corresponding to the first key scale as nodes, and the nodes contain four-dimensional features of sea surface temperature, sea surface height, latitudinal velocity, and meridional velocity. Based on the sea temperature gradient and velocity shear between adjacent nodes, a Gaussian kernel function is used to calculate the dynamic adjacency matrix as the edge weight; The three-dimensional convolutional attention performs maximum pooling along the time dimension and average pooling along the spatial dimension, uses a two-dimensional convolution kernel to generate a spatial weight map and generates an attention matrix by fusing weights across time slices.

3. The artificial intelligence-based extended-scale ocean climate prediction method according to claim 2, characterized in that: The step S2, obtaining the first scale feature, specifically includes: The Gaussian kernel function is used to calculate the dynamic adjacency matrix to capture the local interaction characteristic map of the kilometer-scale vortex, which is specifically expressed as: ， in, Represented as a node and nodes The edge weights between Represented as a node The sea temperature gradient at Represented as a node The sea temperature gradient at Expressed as the temperature gradient bandwidth coefficient, Expressed as the velocity correlation weight, Represented as a node The velocity vector at Represented as a node The velocity vector at .

4. The method for predicting ocean climate at an extended scale based on artificial intelligence according to claim 1, characterized in that: The step S3, generating the second scale feature, specifically includes: Physically encoding the thermodynamic characteristics corresponding to the first scale characteristics; splicing the development stages of extreme events with the characteristics of the physical code; Feature reconstruction is performed through multi-layer 3D deconvolution of the decoder in the conditional variational autoencoder, and physical constraints are introduced before each layer to calculate the thermodynamic residual of the generated sample, and the output is corrected through residual connections.

5. The method for predicting ocean climate at an extended scale based on artificial intelligence according to claim 4, characterized in that: The S3, physical coding content as the physical information driven conditions include the tropical Pacific thermocline depth, the spatial gradient of the zonal wind stress anomaly, and the ocean mixed layer heat content anomaly extracted from the first scale features.

6. The artificial intelligence-based extended-scale ocean climate prediction method according to claim 1, characterized in that: The S4 hybrid optimization objective includes a conservation law loss function that constrains the feature evolution direction by embedding the vorticity pseudo-energy conservation equation, and a data-driven loss function that uses an improved Wasserstein distance to evaluate the matching degree of the predicted field distribution.

Images