Sea level anomaly prediction method and system based on trend seasonal dynamic learning

Abstract

The present application relates to the technical field of marine data processing and spatio-temporal prediction modeling, and discloses a sea level anomaly prediction method and system based on trend seasonal dynamic learning, which comprises the following steps: inputting sea level anomaly data, ocean current data and wind field data, and performing multi-scale preprocessing; performing spatio-temporal trend and seasonal decomposition on the preprocessing result of the sea level anomaly data to obtain fused seasonal features and fused trend features; performing seasonal-space dynamic learning on the fused seasonal features and the preprocessing result of the wind field data to obtain enhanced seasonal-space features; performing trend-time dynamic learning on the fused trend features and the preprocessing result of the ocean current data to obtain enhanced trend-time features, then performing step-by-step diffusion propagation and step-by-step guided propagation respectively, and inputting the decoder to obtain the final prediction result. The accuracy, stability and physical interpretability of the sea level anomaly prediction are improved.

Description

Technical Field

[0001] This invention belongs to the field of marine data processing and spatiotemporal prediction modeling technology, and in particular relates to a method and system for predicting sea level height anomalies based on trend seasonal dynamic learning. Background Technology

[0002] Sea level anomaly (SLA) is an important physical quantity describing changes in ocean dynamic state. Its spatiotemporal evolution is simultaneously influenced by wind fields, ocean currents, thermal processes, and multi-scale dynamic processes within the ocean.

[0003] The goal of sea level height anomaly prediction is to forecast target ocean variables by utilizing the mapping relationship between historical observation data and future ocean conditions. With the continuous accumulation of ocean observation data and the development of deep learning technology, deep neural networks have been widely applied in the field of sea level height anomaly prediction. These methods are mainly categorized as follows: prediction methods based on recurrent neural networks, prediction methods based on convolutional neural networks and their hybrid structures, prediction methods based on graph neural networks and nonlinear dimensionality reduction, and prediction methods based on the Transformer architecture and end-to-end large models. While these data-driven methods perform well in fitting historical data, they still have limitations in handling the inherent multi-scale spatiotemporal characteristics of sea level changes. Existing methods often separate trend and seasonal terms and model them separately, or only focus on the temporal dependence of a single scale, failing to fully explore the intrinsic correlations and interactions between different time and spatial scales.

[0004] Meanwhile, due to the rise of deep learning, deep learning-based multi-scale spatiotemporal feature learning methods have been widely applied to ocean multi-scale modeling. Convolutional Long Short-Term Memory (ConvLSTM) networks combine the spatial feature extraction capabilities of convolutional networks with the temporal modeling capabilities of LSTM, but they are still limited by the parallel efficiency problem caused by the temporal dependence of RNNs. To address the problem that ordinary CNNs cannot effectively capture scale-changing sea surface eddies, the U-Net structure was introduced as the backbone network of the prediction model. Its encoder uses pooling operations to obtain feature maps at different scales, and the decoder integrates multi-scale features through upsampling and convolution operations, thereby capturing local eddy information at different scales in the observed sea surface height anomaly field. Furthermore, a spatiotemporal transformation attention module was designed, comprising two sub-modules: Spatial Transformation Attention (STA) and Temporal Transformation Attention (TTA). STA utilizes the two-dimensional attention mechanism of the visual Transformer to promote information interaction throughout the sea surface height anomaly space, effectively capturing long-distance spatial dependencies. TTA innovatively combines the time axis and spatial axis freely to form a spatiotemporal plane containing temporal and partial spatial information, utilizing the global information interaction capability of the visual Transformer to learn the spatiotemporal dynamic changes of the sea surface height anomaly field.

[0005] However, the above methods suffer from several drawbacks. First, the lack of effective separation during modeling can easily lead to the overlap of long-term background information and short-term fluctuation information during feature extraction, thus affecting the accuracy and stability of sea level height anomaly predictions. Second, the seasonal and trend components of sea level height anomalies correspond to different dominant driving factors. Unifying the modeling of these two types of components makes it difficult to characterize the dependencies between multi-source driving information and the target variable, limiting the ability to express complex ocean structural features and affecting the interpretability of the results. Third, the lack of clear physical basis for cross-scale information interaction makes it difficult for the model to simultaneously maintain large-scale background consistency and local detail representation capabilities, thereby affecting the stability of sea level height anomaly predictions.

[0006] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] To address or at least mitigate one or more of the above problems, a method and system for predicting sea level height anomalies based on trend seasonal dynamic learning is provided, which can further improve the accuracy, stability, and physical interpretability of sea level height anomaly prediction.

[0008] To achieve the above objectives, according to a first aspect of the present invention, a method for predicting sea level height anomalies based on trend seasonal dynamic learning is provided, comprising: S1. Acquire and preprocess data: Spatiotemporal sequences of sea level height anomalies, ocean currents, and wind fields were acquired, and multi-scale preprocessing operations were performed to obtain the preprocessing results of the corresponding data at each scale. S2, Spatiotemporal Trend Seasonal Decomposition: The preprocessing results of sea level height anomaly data at various scales were subjected to spatiotemporal trend seasonal decomposition to extract seasonal and trend terms in time and space. A cross-attention mechanism was introduced to obtain fused seasonal and trend features at various scales. S3, Spatiotemporal Dynamic Learning: The preprocessing results of the seasonal features and wind field data at each scale are converted into spatial node features according to the spatial grid position and then weighted and fused. The spatial node features and current time information are processed in a high-order interactive manner by combining the learnable spatial third-order parameter tensor, and the enhanced seasonal spatial features at each scale are obtained by graph convolution operation. The fusion trend features and ocean current data preprocessing results at each scale are converted into spatial node features according to the spatial grid position and then weighted and fused. The time node features and current time information are processed in a high-order interactive manner by combining the learnable time third-order parameter tensor, and the enhanced trend time features at each scale are obtained through graph convolution operation. S4, Fusion Prediction: The enhanced seasonal spatial features at each scale are propagated step by step, and the enhanced trend temporal features at each scale are guided propagated step by step to obtain seasonal cross-scale fusion features and trend cross-scale fusion features, respectively. After splicing and mapping, they are input into the bottleneck Transformer decoder for decoding and prediction to obtain the final prediction result.

[0009] Furthermore, spatiotemporal sequences of sea level height anomalies, ocean currents, and wind fields were obtained, and multi-scale preprocessing operations were performed on each to obtain the preprocessed results of the corresponding data at each scale, including: By constructing three branches—small-scale, medium-scale, and large-scale—downsampling operations were performed on the spatiotemporal sequences of sea level height anomalies, ocean currents, and wind fields, respectively, to obtain the preprocessing results of sea level height anomaly data, ocean current data, and wind field data at each scale.

[0010] Furthermore, the preprocessing results of sea level height anomaly data at each scale were subjected to spatiotemporal trend seasonal decomposition to extract temporal and spatial seasonal and trend terms. A cross-attention mechanism was then introduced to obtain fused seasonal and trend features at each scale, including: Spatial smoothing is performed on the preprocessing results of sea level height anomaly data at various scales in the spatial dimension to extract the spatial trend term. The spatial trend term is then subtracted from the preprocessing results of sea level height anomaly data to obtain the spatial seasonal term. Meanwhile, in the time dimension, time smoothing operation is performed on the preprocessing results of sea level height anomaly data at various scales to extract the time trend term. The time trend term is then subtracted from the preprocessing results of sea level height anomaly data to obtain the time seasonal term. Cross-attention operations on spatial trend items and temporal trend items yield fused trend features, while cross-attention operations on spatial seasonal items and temporal seasonal items yield fused seasonal features.

[0011] Furthermore, the preprocessing results of the fused seasonal features and corresponding scale wind field data at each scale are converted into spatial node features according to the spatial grid location, and then weighted and fused. A high-order interactive processing of the spatial node features and current time information is performed using a learnable spatial third-order parameter tensor, and the enhanced seasonal spatial features at each scale are obtained through graph convolution operations, including: The seasonal features at each scale are fused and the wind field data preprocessing results at the corresponding scale are converted into spatial node representations according to the spatial grid position to obtain wind field feature spatial nodes and seasonal feature spatial nodes. The wind field feature spatial nodes and seasonal feature spatial nodes are then fused by weighting coefficients to obtain a joint representation of wind field features and seasonal features. Each spatial node in the joint representation of wind field characteristics and seasonal characteristics is mapped to a vector space of the same dimension to obtain source spatial node embedding and target spatial node embedding. The current time t is used as the initial information to expand it into a time embedding of the same dimension. The source space node embedding, target space node embedding, and temporal embedding are trilinearly interacted through an introduced learnable spatial third-order parameter tensor to jointly model the dynamic dependencies between spatial nodes under different temporal conditions, obtain the original spatial association scores between nodes, and remove negative value interference through the ReLU activation function to obtain the spatial non-negative intermediate association representation. The spatial non-negative intermediate association representation is normalized to obtain the spatial dynamic association strength. The spatial dynamic association strengths between all source space nodes and target space nodes together constitute the spatial dynamic association matrix. By using graph convolution operations to propagate and aggregate the joint representation of wind field characteristics and seasonal characteristics, we obtain enhanced seasonal spatial characteristics at various scales.

[0012] Furthermore, the preprocessing results of the fused trend features and corresponding ocean current data at each scale are converted into spatial node features according to the spatial grid position, and then weighted and fused. A high-order interactive processing of the time node features and current time information is performed using a learnable third-order time parameter tensor, and the enhanced trend time features at each scale are obtained through graph convolution operations, including: The preprocessing results of the trend features and corresponding ocean current data at each scale are transformed into time node representations according to time steps to obtain ocean current feature time nodes and trend feature time nodes. The ocean current feature time nodes and trend feature time nodes are then fused by weighting coefficients to obtain a joint representation of ocean current features and trend features. The current spatial location n is used as the initial information to expand into a spatial embedding, and each time node in the spatial embedding is mapped to the same dimension as the spatial embedding to obtain the source time node embedding and the target time node embedding. The source time node embedding, target time node embedding, and spatial embedding are trilinearly interacted through an introduced learnable third-order time parameter tensor to jointly model the dynamic dependencies between time nodes under different spatial conditions, obtain the original temporal association scores between nodes, and remove negative interference through the ReLU activation function to obtain the non-negative intermediate temporal association representation. The non-negative intermediate temporal association representation is normalized to obtain the temporal dynamic association strength. The temporal dynamic association strengths between all source time nodes and target time nodes together constitute the temporal dynamic association matrix. By using graph convolution operations to propagate and aggregate the joint representation of ocean current features and trend features, enhanced trend temporal features at various scales are obtained.

[0013] Furthermore, the enhanced seasonal spatial features at each scale are propagated stepwise, and the enhanced trend temporal features at each scale are guided propagation stepwise, resulting in seasonal cross-scale fused features and trend cross-scale fused features. These are then concatenated and mapped before being input into the bottleneck Transformer decoder for decoding and prediction to obtain the final prediction results, including: The enhanced seasonal spatial features adopt a propagation mode that diffuses stepwise from small scale to mesoscale and then to large scale, resulting in seasonal cross-scale fusion features. The enhanced trend temporal features adopt a propagation mode that guides stepwise from large scale to mesoscale and then to small scale, resulting in trend cross-scale fusion features. Seasonal cross-scale fusion features and trend cross-scale fusion features are concatenated to obtain trend-seasonal joint fusion features. Trends, seasonality, and combined characteristics Mapped to query vector Key vector Sum value vector The data is then input into the bottleneck Transformer decoder for decoding and prediction to obtain the final prediction result.

[0014] To achieve the above objectives, according to a second aspect of the present invention, a sea level height anomaly prediction system based on trend seasonal dynamic learning is provided, comprising: The acquisition and preprocessing module is used to acquire and preprocess data: acquire spatiotemporal sequences of sea level height anomalies, ocean currents, and wind fields, and perform multi-scale preprocessing operations to obtain the preprocessed results of the corresponding data at each scale. The spatiotemporal trend seasonal decomposition module is used for spatiotemporal trend seasonal decomposition: the preprocessing results of sea level height anomaly data at each scale are subjected to spatiotemporal trend seasonal decomposition operations to extract the seasonal and trend terms in time and space, and a cross-attention mechanism is introduced to obtain the fused seasonal features and fused trend features at each scale. The spatiotemporal dynamic learning module is used for spatiotemporal dynamic learning: the fused seasonal features at each scale and the preprocessed wind field data at the corresponding scale are converted into spatial node features according to the spatial grid position, and weighted fusion is performed. The spatial node features and current time information are processed in a high-order interactive manner by combining the learnable spatial third-order parameter tensor, and the enhanced seasonal spatial features at each scale are obtained through graph convolution operation. The fusion trend features at each scale and the preprocessing results of ocean current data at the corresponding scale are converted into spatial node features according to the spatial grid position and then weighted and fused. The time node features and current time information are processed in a high-order interactive manner by combining the learnable time third-order parameter tensor, and the enhanced trend time features at each scale are obtained by graph convolution operation. The fusion prediction module propagates the enhanced seasonal spatial features at each scale in a step-by-step manner and guides the propagation of the enhanced trend temporal features at each scale in a step-by-step manner, thereby obtaining the seasonal cross-scale fusion features and the trend cross-scale fusion features. After splicing and mapping, these features are input into the bottleneck Transformer decoder for decoding and prediction to obtain the final prediction results.

[0015] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art: This invention can explicitly distinguish between trend and seasonal components in sea level height anomalies, thereby reducing the overlap of different physical attribute information in the unified modeling process. By designing a multi-scale spatiotemporal trend-seasonal decomposition mechanism, long-term slowly changing trend information and short-term rapidly fluctuating seasonal information are extracted at different scales. Furthermore, a cross-attention mechanism enhances the consistency between spatial and temporal decomposition results. Because trend and seasonal components are explicitly separated and specifically characterized, the model can more accurately depict the long-term background evolution and short-term disturbance changes in sea level height anomalies, thus improving the accuracy and stability of prediction results.

[0016] This invention combines multi-source driving information such as wind fields and ocean currents to establish differentiated dynamic relationship models for different physical components, thereby enhancing the ability to express complex ocean dynamic mechanisms. A seasonal-spatial dynamic learning module is designed for seasonal components, using wind field information to learn the spatial propagation relationships of seasonal components; a trend-temporal dynamic learning module is designed for trend components, using ocean current information to learn the temporal evolution relationships of trend components. Since seasonal and trend components correspond to different dominant driving factors and dynamic dependence forms, differentiated dynamic learning can more effectively reveal the intrinsic mechanisms in the formation and evolution of sea level anomalies, improving the stability and physical interpretability of prediction results.

[0017] This invention constructs a cross-scale propagation and fusion method consistent with physical mechanisms, enabling the model to simultaneously retain large-scale background information and local detail information. A small-scale-dominated, stepwise diffusion propagation method is used for seasonal components, while a large-scale-dominated, stepwise guided propagation method is used for trend components. This allows the two types of physical components to propagate and fuse across different scales according to their respective dominant laws. Compared with existing unified fusion methods, this approach better reflects the scale-dependent characteristics of ocean dynamic processes. Therefore, while ensuring large-scale background consistency, it enhances the ability to express local changes and detailed perturbations at small and medium scales, further improving the accuracy, stability, and physical interpretability of sea-level height anomaly predictions.

[0018] The specific embodiments of the present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0019] The accompanying drawings, as part of this invention, are provided to further illustrate the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention, but do not constitute an undue limitation thereof. Clearly, the drawings described below are merely some embodiments; those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0020] In the attached diagram: Figure 1 This is a flowchart illustrating the sea level height anomaly prediction method based on trend seasonal dynamic learning in this specific embodiment. Figure 2 This is a schematic diagram of the spatiotemporal trend seasonal decomposition module in this specific embodiment; Figure 3 This is a schematic diagram of the spatiotemporal dynamic learning module in this specific embodiment; Figure 4 This is a schematic diagram of the architecture of the sea level height anomaly prediction system based on trend seasonal dynamic learning in this specific embodiment. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0022] In the following description, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0023] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0024] Please see Figure 1 This invention provides a method for predicting sea level height anomalies based on trend seasonal dynamic learning, comprising the following steps: S1. Acquire and preprocess data: Spatiotemporal sequences of sea level height anomalies, ocean currents, and wind fields were acquired, and multi-scale preprocessing operations were performed to obtain the preprocessing results of the corresponding data at each scale. S2, Spatiotemporal Trend Seasonal Decomposition: The preprocessing results of sea level height anomaly data at each scale were subjected to spatiotemporal trend seasonal decomposition to extract seasonal and trend terms in time and space. A cross-attention mechanism was introduced to obtain fused seasonal and trend features at each scale. S3, Spatiotemporal Dynamic Learning: The preprocessing results of the seasonal features and wind field data at each scale are converted into spatial node features according to the spatial grid position and then weighted and fused. The spatial node features and current time information are processed in a high-order interactive manner by combining the learnable spatial third-order parameter tensor, and the enhanced seasonal spatial features at each scale are obtained by graph convolution operation. The fusion trend features and ocean current data preprocessing results at each scale are converted into spatial node features according to the spatial grid position and then weighted and fused. The time node features and current time information are processed in a high-order interactive manner by combining the learnable time third-order parameter tensor, and the enhanced trend time features at each scale are obtained through graph convolution operation. S4, Fusion Prediction: The enhanced seasonal spatial features at each scale are propagated step by step, and the enhanced trend temporal features at each scale are guided propagated step by step to obtain seasonal cross-scale fusion features and trend cross-scale fusion features, respectively. After splicing and mapping, they are input into the bottleneck Transformer decoder for decoding and prediction to obtain the final prediction result.

[0025] In some embodiments, spatiotemporal sequences of sea level height anomalies, ocean currents, and wind fields are acquired, and multi-scale preprocessing operations are performed on each to obtain the preprocessing results of the corresponding data at each scale, including: By constructing three branches—small-scale, medium-scale, and large-scale—downsampling operations were performed on the spatiotemporal sequences of sea level height anomalies, ocean currents, and wind fields, respectively, to obtain the preprocessing results of sea level height anomaly data, ocean current data, and wind field data at each scale.

[0026] Specifically, to extract data features at different scales, the input spatiotemporal sequences of sea level height anomalies, ocean currents, and wind fields are constructed into three branches: small-scale, meso-scale, and large-scale. Downsampling operations are then performed on these sequences at each scale to obtain the preprocessed results for sea level height anomalies, ocean currents, and wind fields at each scale. The downsampling operation is represented as follows: ; ; ; in, , , These represent the preprocessing results of sea level height anomaly data at large, medium, and small scales obtained through downsampling operations, respectively. , , This represents the preprocessing results of large-scale, medium-scale, and small-scale ocean current data obtained through downsampling operations. , , This represents the preprocessing results of large-, medium-, and small-scale wind field data obtained through downsampling operations. This indicates a downsampling operation. , , These represent the initial input spatiotemporal sequences of sea level height anomalies, ocean currents, and wind fields, respectively.

[0027] In practical applications, spatial smoothing is performed on the preprocessing results of sea level height anomaly data at various scales in the spatial dimension to extract the spatial trend term, i.e., the smoothed background information in space. The spatial seasonal term is obtained by subtracting the spatial trend term from the preprocessing results of sea level height anomaly data. Meanwhile, in the time dimension, time smoothing operation is performed on the preprocessing results of sea level height anomaly data at various scales to extract the time trend term, i.e. the slowly changing component over time. The time trend term is subtracted from the preprocessing results of sea level height anomaly data to obtain the time seasonal term. Cross-attention operations on spatial trend items and temporal trend items yield fused trend features, while cross-attention operations on spatial seasonal items and temporal seasonal items yield fused seasonal features.

[0028] For example, the preprocessing results of sea level height anomaly data at various scales are subjected to spatial smoothing, i.e., spatial moving average operation, to extract spatial trends, i.e., spatial smoothing background information, as follows: ; in, This indicates a space smoothing operation. Represents spatial trend terms at different scales. This represents the preprocessing results of sea level height anomaly data at different scales. Indicates different scales. , , These represent large-scale, medium-scale, and small-scale, respectively.

[0029] For example, subtract the spatial trend term from the preprocessed sea level height anomaly data. The spatial seasonal term is obtained and expressed as: ; in, This represents the spatial seasonality at different scales, used to describe local disturbances and rapid changes relative to a smooth spatial background.

[0030] For example, a time smoothing operation, i.e., a time moving average operation, is performed on the preprocessing results of sea level height anomaly data at various scales to extract the time trend term, i.e., the slowly changing component over time, which is represented as: ; in, This indicates a time smoothing operation. Represents the time trend at different scales. For example, subtracting the time trend term from the preprocessed sea level height anomaly data yields the time seasonal term, expressed as: ; These represent seasonal time terms at different scales, used to describe short-term fluctuations and periodic changes over time.

[0031] For example, since the spatial decomposition results (spatial seasonal and spatial trend terms) and the temporal decomposition results (temporal seasonal and temporal trend terms) characterize the changes in sea level height anomalies from the spatial and temporal dimensions, respectively, to further enhance the consistency and complementarity between the two, a cross-attention operation is performed on the spatial trend terms and temporal trend terms at different scales to obtain fused trend features at different scales, and a cross-attention operation is performed on the spatial seasonal terms and temporal seasonal terms at different scales to obtain fused seasonal features at different scales, expressed as follows: ; ; in, This indicates a trend-based cross-attention operation. Indicates the fusion trend characteristics at different scales; This indicates seasonal cross-attention operations. This indicates the seasonal characteristics of fusion at different scales.

[0032] Because the fusion of seasonal and trend-based features differ significantly in their formation mechanisms, dominant driving factors, and dynamic dependencies, using a uniform approach for subsequent modeling can easily lead to insufficient expression of the dynamic patterns and structural characteristics of the two types of features. Therefore, this invention constructs separate seasonal-spatial dynamic learning mechanisms and trend-temporal dynamic learning mechanisms. It models the spatial propagation relationship of fused seasonal features using wind field data and the temporal evolution relationship of fused trend-based features using ocean current data, thereby achieving differentiated dynamic learning of the two types of features.

[0033] The seasonal component mainly reflects the short-term rapid fluctuations, local disturbances and their spatial propagation process in sea level height anomalies. Such changes are more susceptible to external forcing factors such as wind fields and have a stronger spatial effect. Therefore, seasonal spatial dynamic learning was carried out. In this process, the seasonal spatial dynamic learning sequentially learns the spatial dynamic relationship between the trend wind field nodes at each day t, without considering the relationship between each day.

[0034] The trend component primarily reflects the long-term, slow changes in sea level anomalies, the large-scale background state adjustment process, and their continuous evolution over time. These changes are more susceptible to the sustained influence of ocean transport processes such as ocean currents, and their cumulative effect over time is stronger. Therefore, a trend-time dynamic learning module was designed. It is important to emphasize that this module sequentially learns the dynamic relationships between different time points at each spatial location n, while the relationships between different spatial locations are not considered in this module.

[0035] In some embodiments, the preprocessing results of the fused seasonal features and wind field data at each scale are converted into spatial node features according to the spatial grid position, and then weighted and fused. A high-order interactive processing of the spatial node features and current time information is performed using a learnable spatial third-order parameter tensor, and the enhanced seasonal spatial features at each scale are obtained through graph convolution operations; including: The seasonal features at each scale are fused and the wind field data preprocessing results at the corresponding scale are converted into spatial node representations according to the spatial grid position to obtain wind field feature spatial nodes and seasonal feature spatial nodes. The wind field feature spatial nodes and seasonal feature spatial nodes are then fused by weighting coefficients to obtain a joint representation of wind field features and seasonal features. Each spatial node in the joint representation of wind field characteristics and seasonal characteristics is mapped to a vector space of the same dimension to obtain source spatial node embedding and target spatial node embedding. The current time t is used as the initial information to expand it into a time embedding of the same dimension. The source space node embedding, target space node embedding, and temporal embedding are trilinearly interacted by an introduced learnable third-order spatial relation tensor to jointly model the dynamic dependencies between spatial nodes under different temporal conditions, obtain the original spatial association scores between nodes, and remove negative interference through the ReLU activation function to obtain the spatial non-negative intermediate association representation. The spatial non-negative intermediate association representation is normalized to obtain the spatial dynamic association strength. The spatial dynamic association strengths between all source space nodes and target space nodes together constitute the spatial dynamic association matrix. By using graph convolution operations to propagate and aggregate the joint representation of wind field characteristics and seasonal characteristics, we obtain enhanced seasonal spatial characteristics at various scales.

[0036] Specifically, the preprocessed ocean current data, which integrates trend characteristics and corresponding scales, will be subjected to seasonal-spatial dynamic learning: At any scale In the preprocessing results of wind field data that integrates seasonal characteristics and corresponding scales, each location in the spatial grid is regarded as a spatial node; let the first... The height of the spatial grid at each scale is Width is Then the first The number of spatial nodes at each scale is: ; Furthermore, the fused seasonal features at each scale are transformed into spatial node representations according to the spatial grid positions of the preprocessed wind field data at the corresponding scales, resulting in wind field feature spatial nodes and seasonal feature spatial nodes. These nodes are then weighted and fused using weighting coefficients to obtain a joint representation of the wind field features and seasonal features, expressed as: ; in, A combined representation of wind field characteristics and seasonal characteristics. These represent learnable weight parameters at different scales, used to dynamically adjust the intensity of the wind field's effect on seasonality at different scales. Represents spatial nodes that represent wind field characteristics. Spatial nodes representing seasonal characteristics.

[0037] To further model the dynamic dependencies between spatial nodes, this invention introduces source spatial node embedding, target spatial node embedding, and temporal embedding, all of which are D-dimensional, and also introduces a learnable spatial third-order parameter tensor. This approach allows for trilinear interactions between source space node embeddings, target space node embeddings, and temporal embeddings to jointly model the dynamic dependencies between spatial nodes under different temporal conditions. This yields the original spatial association scores between nodes. Negative interference is removed using the ReLU activation function to obtain a non-negative intermediate spatial association representation. Normalization of this representation yields the dynamic spatial association strength. The dynamic spatial association strengths between all source and target space nodes collectively constitute the dynamic spatial association matrix. The process of constructing the dynamic spatial association matrix is ​​expressed as follows: ; ; ; ; Where t represents the current time, p represents the target space node index, q represents the source space node index, D represents the feature dimension of the source space node embedding, target space node embedding and time embedding, and a, b and c all represent the dimension index of the embedding vector, with values ​​ranging from 1 to D. Let represent the parameters of the learnable spatial third-order parameter tensor corresponding to the i-th scale at the positions of the a, b, and c dimensions. This represents the element in the a-th feature dimension of the time embedding corresponding to the current time step t at the i-th scale. This represents the element embedded in the b-th feature dimension at the p-th target space node at the i-th scale. Let represent the element embedded in the c-th feature dimension at the q-th source space node at the i-th scale; Let represent the original spatial association score between the q-th source spatial node and the p-th target spatial node at time t at the i-th scale. max represents the ReLU operation, which retains the original association score when it is greater than zero, and sets it to zero when it is less than or equal to zero, in order to remove the interference of negative association values ​​on the propagation process of subsequent spatial nodes. This represents the spatial non-negative intermediate association between the q-th source space node and the p-th target space node after ReLU activation; This means that the non-negative intermediate association representation of the space is normalized along the dimension corresponding to the source space node index q, so that the association weights of different source space nodes received by the same target space node are comparable; This represents the spatial dynamic correlation strength between the q-th source space node and the p-th target space node at time t under the i-th scale after normalization. The element in the p-th row and q-th column of the spatial dynamic correlation matrix. This represents the spatial dynamic correlation matrix at time t at the i-th scale. This represents the number of spatial nodes at the i-th scale.

[0038] Specifically, graph convolution operations are used to propagate and aggregate the joint representation of wind field features and seasonal features to obtain enhanced seasonal spatial features at various scales, represented as follows: ; in, This represents the enhanced seasonal spatial features obtained after seasonal spatial dynamic learning. GCN stands for Graph Convolution, used to aggregate features. This represents a combined representation of wind field characteristics and seasonal characteristics.

[0039] Trend components primarily reflect the long-term, slow changes in sea level anomalies, large-scale background state adjustments, and their continuous evolution over time. They are more susceptible to the sustained influence of ocean transport processes such as ocean currents, and exhibit a stronger cumulative effect over time. Therefore, trend-time dynamic learning was employed. Specifically, trend-time dynamic learning sequentially learns the dynamic relationships between different time points at each spatial location n, while the relationships between different spatial locations are not considered in the trend-time dynamic learning process.

[0040] In other embodiments, the fused trend features at each scale and the preprocessed ocean current data at the corresponding scale are converted into spatial node features according to the spatial grid position, and then weighted and fused. A learnable third-order time parameter tensor is used to perform high-order interactive processing on the time node features and the current time information, and graph convolution operations are used to obtain the enhanced trend time features at each scale, including: The preprocessing results of the trend features and corresponding ocean current data at each scale are transformed into time node representations according to time steps to obtain ocean current feature time nodes and trend feature time nodes. The ocean current feature time nodes and trend feature time nodes are then fused by weighting coefficients to obtain a joint representation of ocean current features and trend features. The current spatial location n is used as the initial information to expand into a spatial embedding, and each time node in the spatial embedding is mapped to the same dimension as the spatial embedding to obtain the source time node embedding and the target time node embedding. The source time node embedding, target time node embedding, and spatial embedding are trilinearly interacted through the introduction of a learnable third-order temporal relation tensor to jointly model the dynamic dependencies between time nodes under different spatial conditions, obtain the original temporal association scores between nodes, and remove negative interference through the ReLU activation function to obtain the non-negative intermediate temporal association representation. The non-negative intermediate temporal association representation is normalized to obtain the dynamic temporal association strength. The dynamic temporal association strengths between all source time nodes and target time nodes together constitute the dynamic temporal association matrix. By using graph convolution operations to propagate and aggregate the joint representation of ocean current features and trend features, enhanced trend temporal features at various scales are obtained.

[0041] Specifically, trend-time dynamic learning is performed by integrating trend characteristics and ocean current data preprocessing results at corresponding scales: For the fused trend features at any scale i and the preprocessed ocean current data at the corresponding scale, different time steps at any spatial location n are considered as nodes. Let the length of the input time series at the i-th spatial scale be... The height of the spatial grid is Width is Then, at any spatial location, there are a total of Each time point; Furthermore, the trend features at each scale and the preprocessing results of ocean current data at the corresponding scale are transformed into time node representations according to time steps, resulting in ocean current feature time nodes and trend feature time nodes. These ocean current feature time nodes and trend feature time nodes are then weighted and fused using weighting coefficients to obtain a joint representation of the ocean current features and trend features, expressed as: ; in, A combined representation of ocean current characteristics and trends. This represents the learnable weight parameters at different scales. Indicates the time points that represent ocean current characteristics. Indicates time points that represent trend characteristics.

[0042] To further model the dynamic dependencies between time nodes, this invention introduces spatial embedding, source time node embedding, and target spatial node embedding, all of which are D-dimensional. A learnable third-order time parameter tensor is introduced. This approach allows for trilinear interaction of source time node embeddings, target time node embeddings, and spatial embeddings, jointly modeling the dynamic dependencies between time nodes under different spatial conditions. This yields the original temporal association scores between nodes, and negative interference is removed using the ReLU activation function to obtain a non-negative intermediate temporal association representation. Normalization of this representation yields the dynamic temporal association strength. The combined dynamic temporal association strengths between all source and target time nodes constitute the dynamic temporal association matrix. The construction process of the dynamic temporal association matrix is ​​as follows: ; ; ; ; in, Let represent the parameters of the learnable time-based third-order parameter tensor corresponding to the i-th scale at positions a, b, and c in the dimensions. This represents the element in the a-th feature dimension of the spatial embedding corresponding to the current spatial location at the i-th scale. This represents the element embedded in the b-th feature dimension at the p-th target time node at the i-th scale. Let represent the element embedded in the c-th feature dimension at the q-th source time node at the i-th scale; This represents the original temporal correlation score between the q-th source time node and the p-th target time node at spatial location n at the i-th scale. `max` represents the ReLU operation, which retains the original correlation score if it is greater than zero, and sets it to zero if it is less than or equal to zero, to remove the interference of negative correlation values ​​on the propagation process of subsequent spatial nodes. This represents the non-negative intermediate association between the q-th source time node and the p-th target time node after ReLU activation. This indicates that the non-negative intermediate association representation along the dimension corresponding to the source time node index q is normalized. This represents the temporal dynamic correlation strength between the q-th source time node and the p-th target time node at the ith scale, when the spatial location is n. The element in the p-th row and q-th column of the time-dependent dynamic correlation matrix. Represents the time-dynamic correlation matrix. This represents the number of nodes at the i-th scale.

[0043] Specifically, graph convolution operations are used to propagate and aggregate the joint representations of ocean current features and trend features to obtain enhanced temporal trend features at various scales, represented as follows: ; in, This represents the trend-enhanced features obtained after trend-based time-dynamic learning. GCN stands for Graph Convolution, used to aggregate features. Represents the time-dynamic correlation matrix. A combined representation of ocean current characteristics and trends.

[0044] In some embodiments, the enhanced seasonal spatial features at each scale are propagated stepwise, and the enhanced trend temporal features at each scale are guided stepwise, resulting in seasonal cross-scale fused features and trend cross-scale fused features, respectively. These are then concatenated and mapped before being input into the bottleneck Transformer decoder for decoding and prediction to obtain the final prediction result, including: The enhanced seasonal spatial features adopt a propagation mode that diffuses stepwise from small scale to mesoscale and then to large scale, resulting in seasonal cross-scale fusion features. The enhanced trend temporal features adopt a propagation mode that guides stepwise from large scale to mesoscale and then to small scale, resulting in trend cross-scale fusion features. Seasonal cross-scale fusion features and trend cross-scale fusion features are concatenated to obtain trend-seasonal joint fusion features. Trends, seasonality, and combined characteristics Mapped to query vector Key vector Sum value vector The data is then input into the bottleneck Transformer decoder for decoding and prediction to obtain the final prediction result.

[0045] Specifically, at any given time t, the enhanced seasonal spatial features at each scale propagate from small scale to medium scale and then to large scale in a stepwise manner, resulting in seasonal cross-scale fusion features. The process is represented as follows: ; in, This is represented as a small-scale enhanced seasonal spatial feature. This is represented as seasonal spatial characteristics enhanced at the mesoscale. This represents the enhanced seasonal spatial characteristics on a large scale. Indicates seasonal cross-scale fusion characteristics; For any spatial location n, the enhanced trend temporal features at each scale are propagated and fused stepwise from large scale to mesoscale and then to small scale to obtain the trend cross-scale fused features corresponding to spatial location n. The process is represented as follows: ; in, This indicates the trend-based temporal characteristics after large-scale enhancement. This indicates the trend-based temporal characteristics following mesoscale enhancement. This indicates the trend-based temporal characteristics after small-scale enhancement. This indicates a trend of cross-scale fusion characteristics.

[0046] Specifically, seasonal cross-scale fusion features and trend cross-scale fusion features are concatenated to obtain trend-seasonal joint fusion features. , is represented as: ; in, This indicates concatenation along the feature channel dimension, where C represents the feature channel dimension. Indicates seasonal cross-scale fusion characteristics. Indicates trend-based cross-scale fusion characteristics. This indicates the combined characteristics of seasonal trends.

[0047] Specifically, the trend and seasonal integration characteristics The mapping is done using query vector Q, key vector K, and value vector V, and then input into the bottleneck Transformer decoder for decoding and prediction to obtain the final prediction result, represented as: ; ; ; ; in, , , These are the learnable mapping matrices, This indicates the bottleneck Transformer decoder. The output represents the predicted sea level height anomaly at a future time, which is the final prediction result.

[0048] Example 2: See Figure 2-4Based on the same inventive concept, this invention also provides a sea level height anomaly prediction system based on trend seasonal dynamic learning, comprising: The acquisition and preprocessing module is used to acquire and preprocess data. Spatiotemporal sequences of sea level height anomalies, ocean currents, and wind fields were acquired, and multi-scale preprocessing operations were performed to obtain the preprocessing results of the corresponding data at each scale. The Spatiotemporal Trend Seasonal Decomposition module is used for spatiotemporal trend seasonal decomposition. The preprocessing results of sea level height anomaly data at various scales were subjected to spatiotemporal trend seasonal decomposition to extract seasonal and trend terms in time and space. A cross-attention mechanism was introduced to obtain fused seasonal and trend features at various scales. The spatiotemporal dynamic learning module is used for spatiotemporal dynamic learning: the fused seasonal features at each scale and the preprocessed wind field data at the corresponding scale are converted into spatial node features according to the spatial grid position, and then weighted and fused. The spatial node features and current time information are processed in a high-order interactive manner by combining the learnable spatial third-order parameter tensor, and the enhanced seasonal spatial features at each scale are obtained through graph convolution operation. The fusion trend features at each scale and the preprocessing results of ocean current data at the corresponding scale are converted into spatial node features according to the spatial grid position and then weighted and fused. The time node features and current time information are processed in a high-order interactive manner by combining the learnable time third-order parameter tensor, and the enhanced trend time features at each scale are obtained by graph convolution operation. Fusion prediction module: The enhanced seasonal spatial features at each scale are propagated step by step, and the enhanced trend temporal features at each scale are guided propagated step by step to obtain seasonal cross-scale fusion features and trend cross-scale fusion features, respectively. After splicing and mapping, they are input into the bottleneck Transformer decoder for decoding and prediction to obtain the final prediction result.

[0049] Example 3: This embodiment selects the Northwest Pacific Ocean as the study area, and the latitude and longitude range of the study area is as follows: The sea level height anomaly data comes from the global L4 sea level height reprocessing product provided by Copernicus Marine Service, while the wind field and ocean current data are from the global wind field and ocean current reprocessing product provided by Copernicus Marine Service.

[0050] At a spatial resolution of Under these conditions, the input data for the study area was divided into 64×64 regular latitude and longitude grids. The dataset from 2013 to 2022 was divided into training, validation, and test sets in chronological order. The training set comprised 80% of the dataset, the validation set 10%, and the remaining 10% was used as the test set.

[0051] The method is implemented in PyTorch. Comparative experiments were conducted to predict sea level height anomalies for the next 7 days using historical 14-day data and for the next 30 days using historical 30-day data. Training samples were constructed using a sliding window approach with a window step size of 4. For missing and anomaly-filled values, an effective region mask was constructed, and only grid points in the effective sea area were counted during loss calculation and evaluation. The AdamW optimizer was used during the training phase, with an initial learning rate of [missing information]. The weight decay coefficient is The batch size is 4, and the maximum number of training epochs is 30. All computational tasks are performed on an NVIDIA RTX 4090 GPU. Training uses the ReduceLROnPlateau learning rate scheduling strategy, where the learning rate is decayed to 0.5 when the validation set RMSE stops decreasing, and the minimum learning rate is set to... The loss function used is Smooth L1 Loss. To improve training stability, the validation set RMSE is used as the basis for model selection during training, and the model with the lowest validation set RMSE is saved as the final test model.

[0052] During model training, a batch of historical sea level height anomaly sequences, wind field sequences, and ocean current sequences are input into the prediction model. After multi-scale preprocessing, spatiotemporal trend-seasonal decomposition, spatiotemporal dynamic learning, and cross-scale fusion prediction, the future sea level height anomaly prediction results are obtained. A smoothed L1 loss with an effective region mask is calculated based on the prediction results and the corresponding real results. The gradients of each trainable parameter are calculated through backpropagation, and the model parameters are updated using the AdamW optimizer. After each training epoch, the RMSE is calculated on the validation set, and the model parameters with the lowest RMSE on the validation set are saved. After training, the saved model parameters are used to predict on the test set.

[0053] Example 4: To comprehensively evaluate the predictive performance of this method for the spatial field of sea level height anomalies, root mean square error (RMSE), mean absolute error (MAE), and mean square error (MSE) are used as quantitative evaluation indicators. Let the true sea level height anomaly value be... The model predicts the value. The total number of samples is N, and the three evaluation indicators are defined as follows: ; ; ; RMSE represents the square root of the mean squared prediction error, reflecting the average level of the model's overall prediction error. MAE represents the average absolute error between the predicted and actual values, providing a more intuitive reflection of the model's average deviation across all prediction points. MAE provides a more stable evaluation of the model's average prediction accuracy over the entire spatial range.

[0054] The comparative methods used include Convolutional Long Short-Term Memory Network (Convlstm), Earthformer, Fourier Neural Operator (FNO), Latent Spectral Models (LSM), Shifted Window Transformer (Swin Transformer), U-Net-enhanced Fourier Neural Operator (U-FNO), and U-Net Convolutional Neural Network (U-Net).

[0055] Table 1: Overall average prediction results of different models for short-term prediction tasks

[0056] Table 1 presents the overall average prediction results of different models for short-term prediction tasks. As can be seen from Table 1, the present invention achieves the best results in all three evaluation metrics: RMSE, MSE, and MAE. This indicates that the present invention has higher overall prediction accuracy in short-term SLA prediction tasks. Compared to U-Net, the present invention reduces RMSE from 0.1716 to 0.1675, a decrease of approximately 2.39%; MSE from 0.0294 to 0.0280, a decrease of approximately 4.76%; and compared to Swin Transformer, the present invention reduces MAE from 0.1176 to 0.1110, a decrease of approximately 5.61%. This demonstrates that in short-term prediction scenarios, the present invention not only reduces the overall squared error but also effectively reduces the mean absolute error, making the prediction results closer to the true values ​​at most spatial grid points.

[0057] Table 2: Average prediction results of different models for long-term prediction tasks

[0058] Table 2 presents the average prediction results of different models for long-term prediction tasks. As can be seen from Table 2, this invention also achieves optimal results. Compared with the suboptimal method U-Net, the RMSE of this invention decreases from 0.4935 to 0.4508, a reduction of approximately 8.65%; the MSE decreases from 0.2435 to 0.2032, a reduction of approximately 16.55%; and the MAE decreases from 0.3461 to 0.3140, a reduction of approximately 9.27%. This demonstrates that in more challenging medium- and long-term prediction tasks, this invention exhibits significant performance advantages, better stability, and stronger error suppression capabilities, making it suitable for medium- and long-term prediction.

[0059] Unless otherwise defined, the technical or scientific terms used in this invention shall have the general meaning understood by one of ordinary skill in the art to which this invention pertains. Words such as “a,” “an,” “an,” “the,” “the,” and “these” used in this invention do not indicate quantitative limitations and may be singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this invention are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or modules (units) is not limited to the listed steps or modules (units) but may include steps or modules (units) not listed, or may include other steps or modules (units) inherent to these processes, methods, products, or devices. Words such as “connected,” “linked,” and “coupled” used in this invention are not limited to physical or mechanical connections but may include electrical connections, whether direct or indirect. “A plurality” used in this invention refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. Normally, the character " / " indicates that the objects before and after it are in an "or" relationship. In this invention, the terms "first," "second," "third," etc., are used only to distinguish similar objects and do not represent a specific order of objects.

[0060] Obviously, the accompanying drawings are merely some examples or embodiments of the present invention. Those skilled in the art can apply the present invention to other similar situations based on these drawings without any creative effort. Furthermore, it is understood that although the work done in this development process may be complex and lengthy, certain design, manufacturing, or production modifications made according to the technical content disclosed in the present invention are merely conventional technical means and should not be considered as insufficient disclosure of the present invention.

[0061] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-described technical content to create equivalent embodiments without departing from the scope of the present invention. The implementation schemes in the above embodiments can also be further combined or replaced. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A sea level height anomaly prediction method based on trend seasonal dynamic learning, characterized in that, Includes the following steps: S1. Acquire and preprocess data: Spatiotemporal sequences of sea level height anomalies, ocean currents, and wind fields were acquired, and multi-scale preprocessing operations were performed to obtain the preprocessing results of the corresponding data at each scale. S2, Spatiotemporal Trend Seasonal Decomposition: The preprocessing results of sea level height anomaly data at various scales were subjected to spatiotemporal trend seasonal decomposition to extract seasonal and trend terms in time and space. A cross-attention mechanism was introduced to obtain fused seasonal and trend features at various scales. S3, Spatiotemporal Dynamic Learning: The preprocessing results of the seasonal features and wind field data at each scale are converted into spatial node features according to the spatial grid position and then weighted and fused. The spatial node features and current time information are processed in a high-order interactive manner by combining the learnable spatial third-order parameter tensor, and the enhanced seasonal spatial features at each scale are obtained by graph convolution operation. The fusion trend features and ocean current data preprocessing results at each scale are converted into spatial node features according to the spatial grid position and then weighted and fused. The time node features and current time information are processed in a high-order interactive manner by combining the learnable time third-order parameter tensor, and the enhanced trend time features at each scale are obtained through graph convolution operation. S4, Fusion Prediction: The enhanced seasonal spatial features at each scale are propagated step by step, and the enhanced trend temporal features at each scale are guided propagated step by step to obtain seasonal cross-scale fusion features and trend cross-scale fusion features, respectively. After splicing and mapping, they are input into the bottleneck Transformer decoder for decoding and prediction to obtain the final prediction result.

2. The method according to claim 1, characterized in that, Spatiotemporal sequences of sea level height anomalies, ocean currents, and wind fields were acquired, and multi-scale preprocessing operations were performed on each to obtain the preprocessed results of the corresponding data at each scale, including: By constructing three branches—small-scale, medium-scale, and large-scale—downsampling operations were performed on the spatiotemporal sequences of sea level height anomalies, ocean currents, and wind fields, respectively, to obtain the preprocessing results of sea level height anomaly data, ocean current data, and wind field data at each scale.

3. The method according to claim 1, characterized in that, The preprocessing results of sea level height anomaly data at various scales were subjected to spatiotemporal trend seasonal decomposition to extract temporal and spatial seasonal and trend components. A cross-attention mechanism was then introduced to obtain fused seasonal and trend features at each scale, including: Spatial smoothing is performed on the preprocessing results of sea level height anomaly data at various scales in the spatial dimension to extract the spatial trend term. The spatial trend term is then subtracted from the preprocessing results of sea level height anomaly data to obtain the spatial seasonal term. Meanwhile, in the time dimension, time smoothing operation is performed on the preprocessing results of sea level height anomaly data at various scales to extract the time trend term. The time trend term is then subtracted from the preprocessing results of sea level height anomaly data to obtain the time seasonal term. Cross-attention operations on spatial trend items and temporal trend items yield fused trend features, while cross-attention operations on spatial seasonal items and temporal seasonal items yield fused seasonal features.

4. The method according to claim 2, characterized in that, The preprocessing results of wind field data at each scale, which integrates seasonal features and corresponding scales, are converted into spatial node features according to spatial grid positions and then weighted and fused. A high-order interactive processing of spatial node features and current time information is performed using a learnable spatial third-order parameter tensor. Finally, enhanced seasonal spatial features at each scale are obtained through graph convolution operations, including: The seasonal features at each scale are fused and the wind field data preprocessing results at the corresponding scale are converted into spatial node representations according to the spatial grid position to obtain wind field feature spatial nodes and seasonal feature spatial nodes. The wind field feature spatial nodes and seasonal feature spatial nodes are then fused by weighting coefficients to obtain a joint representation of wind field features and seasonal features. Each spatial node in the joint representation of wind field characteristics and seasonal characteristics is mapped to a vector space of the same dimension to obtain source spatial node embedding and target spatial node embedding. The current time t is used as the initial information to expand it into a time embedding of the same dimension. The source space node embedding, target space node embedding, and temporal embedding are trilinearly interacted through an introduced learnable spatial third-order parameter tensor to jointly model the dynamic dependencies between spatial nodes under different temporal conditions, obtain the original spatial association scores between nodes, and remove negative value interference through the ReLU activation function to obtain the spatial non-negative intermediate association representation. The spatial non-negative intermediate association representation is normalized to obtain the spatial dynamic association strength. The spatial dynamic association strengths between all source space nodes and target space nodes together constitute the spatial dynamic association matrix. By using graph convolution operations to propagate and aggregate the joint representation of wind field characteristics and seasonal characteristics, we obtain enhanced seasonal spatial characteristics at various scales.

5. The method according to claim 2, characterized in that, The fused trend features at each scale and the preprocessed ocean current data at the corresponding scale are converted into spatial node features according to spatial grid positions, and then weighted and fused. A learnable third-order temporal parameter tensor is used to perform high-order interactive processing on the time node features and current time information. Finally, graph convolution operations are used to obtain the enhanced trend time features at each scale, including: The preprocessing results of the trend features and corresponding ocean current data at each scale are transformed into time node representations according to time steps to obtain ocean current feature time nodes and trend feature time nodes. The ocean current feature time nodes and trend feature time nodes are then fused by weighting coefficients to obtain a joint representation of ocean current features and trend features. The current spatial location n is used as the initial information to expand into a spatial embedding, and each time node in the spatial embedding is mapped to the same dimension as the spatial embedding to obtain the source time node embedding and the target time node embedding. The source time node embedding, target time node embedding, and spatial embedding are trilinearly interacted through an introduced learnable third-order time parameter tensor to jointly model the dynamic dependencies between time nodes under different spatial conditions, obtain the original temporal association scores between nodes, and remove negative interference through the ReLU activation function to obtain the non-negative intermediate temporal association representation. The non-negative intermediate temporal association representation is normalized to obtain the temporal dynamic association strength. The temporal dynamic association strengths between all source time nodes and target time nodes together constitute the temporal dynamic association matrix. By using graph convolution operations to propagate and aggregate the joint representation of ocean current features and trend features, enhanced trend temporal features at various scales are obtained.

6. The method according to claim 5, characterized in that, The enhanced seasonal spatial features at each scale are propagated stepwise, and the enhanced trend temporal features at each scale are guided propagation stepwise, resulting in seasonal cross-scale fused features and trend cross-scale fused features. These are then concatenated and mapped before being input into the bottleneck Transformer decoder for decoding and prediction to obtain the final prediction results, including: The enhanced seasonal spatial features adopt a propagation mode that diffuses stepwise from small scale to mesoscale and then to large scale, resulting in seasonal cross-scale fusion features. The enhanced trend temporal features adopt a propagation mode that guides stepwise from large scale to mesoscale and then to small scale, resulting in trend cross-scale fusion features. Seasonal cross-scale fusion features and trend cross-scale fusion features are concatenated to obtain trend-seasonal joint fusion features. Trends, seasonality, and combined characteristics Mapped to query vector Key vector Sum value vector The data is then input into the bottleneck Transformer decoder for decoding and prediction to obtain the final prediction result.

7. A sea level height anomaly prediction system based on trend seasonal dynamic learning, characterized in that, include: The acquisition and preprocessing module is used to acquire and preprocess data: acquire spatiotemporal sequences of sea level height anomalies, ocean currents, and wind fields, and perform multi-scale preprocessing operations to obtain the preprocessed results of the corresponding data at each scale. The spatiotemporal trend seasonal decomposition module is used for spatiotemporal trend seasonal decomposition: the preprocessing results of sea level height anomaly data at each scale are subjected to spatiotemporal trend seasonal decomposition operations to extract the seasonal and trend terms in time and space, and a cross-attention mechanism is introduced to obtain the fused seasonal features and fused trend features at each scale. The spatiotemporal dynamic learning module is used for spatiotemporal dynamic learning: the fused seasonal features at each scale and the preprocessed wind field data at the corresponding scale are converted into spatial node features according to the spatial grid position, and weighted fusion is performed. The spatial node features and current time information are processed in a high-order interactive manner by combining the learnable spatial third-order parameter tensor, and the enhanced seasonal spatial features at each scale are obtained through graph convolution operation. The fusion trend features at each scale and the preprocessing results of ocean current data at the corresponding scale are converted into spatial node features according to the spatial grid position and then weighted and fused. The time node features and current time information are processed in a high-order interactive manner by combining the learnable time third-order parameter tensor, and the enhanced trend time features at each scale are obtained by graph convolution operation. The fusion prediction module propagates the enhanced seasonal spatial features at each scale in a step-by-step manner and guides the propagation of the enhanced trend temporal features at each scale in a step-by-step manner, thereby obtaining the seasonal cross-scale fusion features and the trend cross-scale fusion features. After splicing and mapping, these features are input into the bottleneck Transformer decoder for decoding and prediction to obtain the final prediction results.

Images