Marine environment parameter prediction method based on multi-physical space-time adaptive memory network
By using a multi-physics spatiotemporal adaptive memory network, high-precision collaborative prediction of marine environmental parameters is achieved, solving the problems of high computational complexity and prediction instability in existing technologies, and improving the accuracy and stability of marine environmental prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-11
- Publication Date
- 2026-03-24
AI Technical Summary
Existing marine environment prediction models have high computational complexity in terms of high-resolution grids and multivariate coupling, making it difficult to meet the needs of rapid and high-precision prediction. Furthermore, initial condition errors and uncertainties in physical parameters lead to long-term prediction instability, failing to fully explore the nonlinear coupling relationship between temperature, salinity, dynamics, and wind stress in the marine system.
A multi-physics spatiotemporal adaptive memory network is adopted, which realizes the dependency modeling of multi-physical variables and multi-scale spatial perception through a multi-dimensional feature perception module. Combined with a multi-physics ocean module, dynamic constraints are explicitly integrated, and the spatiotemporal adaptive memory unit is used to enhance the modeling ability of long-term dynamic dependencies, thus constructing a hybrid paradigm of "physics-driven and data-driven parallel coupling".
While maintaining physical consistency, it improves the high-precision collaborative prediction capability of marine environmental parameters, enhances the prediction accuracy and stability of complex marine environments, and solves the problems of high computational complexity and prediction instability in existing technologies.
Smart Images
Figure CN121723908A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of marine environment prediction technology, and more specifically to a method for predicting marine environmental parameters based on multi-physics spatiotemporal adaptive memory networks. Background Technology
[0002] Currently, accurate prediction of multivariable ocean fields, including sea surface temperature, salinity, velocity, and wind stress, is crucial for understanding air-sea interactions, improving climate predictability, and ensuring the safety of marine operations. Current marine environmental predictions still primarily rely on physical mechanism-driven global ocean circulation models (OGCMs), such as ROMS and NEMO. These models, based on three-dimensional primitive equations and physical parameterization schemes, have long played a vital role in operational marine simulations.
[0003] However, due to its strong dependence on high-resolution grids, multivariate coupling, and data assimilation, the computational complexity of OGCM increases exponentially, making it difficult to meet the demand for fast and high-precision predictions even on ultra-large-scale parallel platforms. Furthermore, initial condition errors and physical parameter uncertainties are often amplified during multi-step integration, leading to long-term prediction instability and consequently affecting the coordinated control of multiple physical variables. Deep learning (DL) technology, as a general-purpose function approximator, possesses powerful representational capabilities and has shown great potential in spatiotemporal sequence prediction and complex physical field reconstruction in recent years. However, existing research has largely focused on prediction tasks for single physical variables (such as sea surface temperature (SST) or sea level height) or local regions, and has not yet fully explored the nonlinear coupling relationships between temperature, salinity, dynamics, and wind stress in the ocean system.
[0004] Therefore, how to achieve high-precision prediction of complex marine environments while maintaining physical consistency and multivariate collaborative prediction is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] In view of the above problems, this invention is proposed to provide a marine environmental parameter prediction method based on a multi-physics spatiotemporal adaptive memory network that overcomes or at least partially solves the above problems. While maintaining physical consistency and multi-variable collaborative prediction, it achieves high-precision prediction of complex marine environments.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a method for predicting marine environmental parameters based on a multi-physics spatiotemporal adaptive memory network, comprising: Obtain the environmental state parameters of the target ocean at the current moment; Based on the environmental state parameters input to the multi-dimensional feature perception module, the dependency modeling of multiple physical variables and multi-scale spatial perception are realized, and the depth features of the environmental state parameters are obtained. Based on the depth features input to the multiphysics ocean module, combined with historical state information, dynamic constraints are explicitly incorporated to obtain physically consistent prediction results. Based on the fusion of the physical consistency prediction result and the deep features, a physical prediction enhancement result is obtained; Based on the deep features, the spatiotemporal features are input to a spatiotemporal memory decoupling module that includes multiple spatiotemporal adaptive memory units to obtain spatiotemporal features; Based on the weighted fusion of the physical prediction enhancement results and the spatiotemporal features, the environmental state prediction parameters of the target ocean at the next moment are obtained.
[0007] In one embodiment, the multidimensional feature perception module includes: a channel correlation modeling unit, a multi-scale spatial perception unit, and a deep fusion gating mechanism unit; The environmental state parameters are input to the channel correlation modeling unit and the multi-scale spatial perception unit, respectively, to obtain channel attention features and spatial attention features. The channel attention features, spatial attention features, and environmental state parameters are input into the deep fusion gating mechanism unit to obtain the depth features.
[0008] In one embodiment, the channel correlation modeling unit includes: a first average pooling layer, a first convolutional layer, and a first activation layer; The environmental state parameters are sequentially input into the first average pooling layer, the first convolutional layer, and the first activation layer to obtain the channel attention features; The multi-scale spatial perception unit includes: a second average pooling layer, a max pooling layer, a second convolutional layer, a third convolutional layer, and a second activation layer. The environmental state parameters are respectively input to the second average pooling layer and the maximum pooling layer to obtain the average pooling feature and the maximum pooling feature. Spatial features are obtained by fusing the average pooling feature and the max pooling feature. Based on the spatial features, they are respectively input into the second convolutional layer and the third convolutional layer to obtain the first scale features and the second scale features respectively; The spatial attention features are obtained by linearly weighting the first and second scale features using learnable fusion weights and then inputting them into the second activation layer.
[0009] In one embodiment, the deep fusion gating mechanism unit includes: a fourth convolutional layer, a third activation layer, a fifth convolutional layer, and a fourth activation layer; The channel attention features, the spatial attention features, and the environmental state parameters are fused to obtain a first fused feature; Based on the fusion of the initial fusion features and the environmental state parameters, a second fusion feature is obtained; The third fusion feature is obtained by subtracting the second fusion feature from the environmental state parameters element by element. Based on the environmental state parameters, they are sequentially input into the fourth convolutional layer, the third activation layer, the fifth convolutional layer, and the fourth activation layer to obtain the gating function; The depth feature is obtained by fusing the gating function and the third fusion feature with the environmental state parameters.
[0010] In one embodiment, the multiphysics ocean module includes: a multiscale physics predictor and an input assimilation unit; Obtain historical environmental state parameters of the target ocean; The historical hidden state vector is obtained based on the historical environmental state parameters; The multi-scale fusion features are obtained by inputting the historical hidden state vector into the multi-scale physical predictor. Based on the fusion of the multi-scale fusion features and the historical hidden state vector, the historical physical prediction result is obtained; The historical physics prediction results and the deep features are input into the input assimilation unit to obtain the process physics prediction results; The physical consistency prediction result is obtained by fusing the process physical prediction result and the historical physical prediction result.
[0011] In one embodiment, the multi-scale physics predictor includes: a sixth convolutional layer, a seventh convolutional layer, and an eighth convolutional layer; The historical hidden state vectors are respectively input into the sixth and seventh convolutional layers to obtain fine-scale features and large-scale features. The multi-scale fused features are obtained by fusing the fine-scale features and the large-scale features and then inputting them into the eighth convolutional layer.
[0012] In one embodiment, the input assimilation unit includes: a channel correlation modeling unit, a ninth convolutional layer, and a fifth activation layer; The depth features are input into the channel correlation modeling unit to obtain the channel features; The first intermediate result is obtained by subtracting the channel characteristics from the historical physical prediction results element by element. Based on the channel features and the historical physics prediction results, they are sequentially input into the ninth convolutional layer and the fifth activation layer to obtain an approximate Kalman gain. The second intermediate result is obtained by fusing the approximate Kalman gain with the first intermediate result, and is used as the physical prediction result of the process.
[0013] In one embodiment, the multiple spatiotemporal adaptive memory units have the same structure, each including: a dual-channel memory subunit and a dynamically growing window subunit; Set the initial hidden state vector, initial time memory information, and initial spatiotemporal memory information at the initial moment; Based on the initial hidden state vector, the initial time memory information, the initial spatiotemporal memory information, and the deep features, the current hidden state vector, the updated time memory information, and the updated spatiotemporal memory information are obtained by inputting them into the dual-channel memory subunit. The first hidden state vector is obtained by inputting the current hidden state vector into the dynamically growing window sub-unit. Based on the deep features, the first hidden state vector, the updated time memory information, and the updated spatiotemporal memory information, the data is input to the next spatiotemporal adaptive memory unit for processing, and so on. The corresponding hidden state vector output by the last spatiotemporal adaptive memory unit is used as the spatiotemporal feature.
[0014] In one embodiment, the dual-channel memory subunit includes: a tenth convolutional layer, a sixth activation layer, an eleventh convolutional layer, a seventh activation layer, a twelfth convolutional layer, an eighth activation layer, a thirteenth convolutional layer, and a ninth activation layer; The depth features are sequentially input into the tenth convolutional layer and the sixth activation layer to obtain the first processed features; Based on the initial hidden state vector, it is sequentially input into the eleventh convolutional layer and the seventh activation layer to obtain the second processed feature; Based on the extraction and fusion of the first processing feature and the second processing feature, the incremental time memory information is obtained; The updated time memory information is obtained by fusing the incremental time memory information and the initial time memory information. Based on the initial spatiotemporal memory information, it is sequentially input into the twelfth convolutional layer and the eighth activation layer to obtain the third processing feature; Based on the extraction and fusion of the third processing feature and the first processing feature, the spatiotemporal memory information increment is obtained; The updated spatiotemporal memory information is obtained by fusing the spatiotemporal memory information increment and the initial spatiotemporal memory information. Based on the updated time memory information and the updated spatiotemporal memory information, the information is input into the thirteenth convolutional layer and the ninth activation layer to obtain the comprehensive memory features. Based on the comprehensive memory features, the first processing features, and the second processing features, the input gate features are obtained; The current hidden state vector is obtained by fusing the input gate features and the integrated memory features.
[0015] In one embodiment, the dynamically growing window subunit includes: a third average pooling layer, a buffer, a multi-head attention module, and a linear layer; The current hidden state vector is input to the third average pooling layer to obtain the query vector at the current time step; The window value is obtained based on the current buffer length and the preset growth step size of the buffer. The token with the most recent window value count is selected as the key and value based on the buffer. The query vector, the key, and the value are input into the multi-head attention module to obtain the multi-head attention weights. Based on the multi-head attention weights input to the linear layer, channel-level scaling factors are obtained; The first hidden state vector is obtained by fusing the current hidden state vector and the channel-level scaling factor.
[0016] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method for predicting marine environmental parameters based on a multi-physics spatiotemporal adaptive memory network, which has the following beneficial effects: 1. This invention proposes a Multidimensional Feature Perception Module (MDFPM): It designs a lightweight perception mechanism that integrates channel and spatial dimensions. Through the Channel Correlation Modeling Unit (CCMU), it efficiently models the dependencies between physical variables using local one-dimensional convolution without dimensionality reduction, avoiding parameter redundancy caused by fully connected components. Combined with the Multi-Scale Spatial Perception Unit (MSPU), it extracts significant region information from average pooling and max pooling features, and uses multi-scale convolution to capture local and global spatial structures. A gated residual fusion strategy is introduced to enhance the perception capability of key physical variables and significant spatial regions while maintaining computational efficiency.
[0017] 2. This invention proposes a Multiphysics Ocean Module (MPOC): This module constructs a deep neural network unit based on physical constraints, and through a multi-scale physical predictor combined with an input assimilation module, achieves a "prediction-correction" physical modeling paradigm. This unit explicitly incorporates ocean physical laws (such as salinity transport, temperature advection, and wind-stress coupling), providing interpretable prediction results while maintaining physical consistency.
[0018] 3. This invention proposes a Spatiotemporal Adaptive Memory Unit (STAMU): It proposes a memory decoupling dual-channel memory framework with multi-scale temporal semantic perception and evolution-driven adaptive attention reconstruction mechanism. Spatiotemporal memory is decoupled into two paths: a temporal memory stream that captures the long-term trend of sequence evolution and a spatial memory stream that retains multi-scale spatial structure information. These are two independently maintained but collaboratively interacting paths. By constructing an intelligent temporal query memory pool and dynamically adjusting the temporal receptive field based on the sequence evolution state, it achieves in-depth mining of historical spatiotemporal semantics and autonomous allocation of multi-head attention weights. This significantly improves the model's ability to capture long-term dependencies and complex dynamic features, effectively solving the "depth-time dilemma".
[0019] 4. Construct a hybrid paradigm of "physics-driven and data-driven parallel coupling". The physical modeling module and the data-driven module run in parallel: the former explicitly embeds multiple physical processes (such as salinity, sea surface temperature, velocity field, wind stress, etc.) to ensure that the prediction results conform to basic physical laws; the latter models and supplements complex, nonlinear, and uncertain dynamic features through deep neural networks. In the prediction stage, the outputs of the two branches are coupled through residual correction and weighted fusion, thereby balancing physical consistency and the flexibility of data-driven approaches.
[0020] 5. This invention achieves efficient dependency modeling and multi-scale spatial perception of multiple physical variables through the Multidimensional Feature Perception Module (MDFPM), explicitly incorporates dynamic constraints through the Multiphysics Ocean Module (MPOC) to ensure physical consistency, and enhances the modeling capability for long-term dynamic dependencies through the Spatiotemporal Adaptive Memory Unit (STAMU). Finally, it achieves high-precision collaborative prediction of multiple variables under the hybrid paradigm of "physical-driven and data-driven parallel coupling". Attached Figure Description
[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0022] Figure 1 This is a flowchart of the marine environmental parameter prediction method based on a multi-physics spatiotemporal adaptive memory network provided in an embodiment of the present invention.
[0023] Figure 2 This is a schematic diagram of the multi-physics spatiotemporal adaptive memory network structure provided in an embodiment of the present invention.
[0024] Figure 3 This is a schematic diagram of the multi-dimensional feature perception module structure provided in an embodiment of the present invention.
[0025] Figure 4 This is a schematic diagram of the multiphysics ocean module structure provided in an embodiment of the present invention.
[0026] Figure 5 This is a schematic diagram of the spatiotemporal adaptive memory unit structure provided in an embodiment of the present invention. Detailed Implementation
[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0028] Example 1 like Figure 1 As shown, this invention discloses a method for predicting marine environmental parameters based on a multi-physics spatiotemporal adaptive memory network, comprising the following steps. For ease of description, these steps are numbered S1 to S6, and these numbers are not intended to limit the sequential relationship between the various steps of this invention: S1 obtains the environmental state parameters of the target ocean at the current moment.
[0029] Furthermore, the environmental state parameters include: sea surface temperature, salinity, seawater current velocity vector, and sea surface wind stress vector; Sea surface temperature (SST), sea surface salinity (SSS), seawater velocity vector, and sea surface wind stress vector are key state variables characterizing air-sea interactions, energy and material transport processes, and the evolution of extreme ocean events such as ocean heat waves and storm surges. These parameters not only play a central role in climate system regulation, mesoscale eddy evolution (O(10 km)), and energy closure, but also directly affect the motion stability, path planning, and energy consumption assessment of surface vehicles.
[0030] S2 inputs environmental state parameters into a multi-dimensional feature perception module to achieve dependency modeling of multiple physical variables and multi-scale spatial perception, thereby obtaining deep features of environmental state parameters.
[0031] Furthermore, such as Figure 2As shown, the Multi-Physical Spatial-Temporal Adaptive Memory Network (PhySTNet) of this invention is composed of a Multi-Dimensional Feature Perception Module (MDFPM), a Multi-Physical Ocean Module (MPOC), and a spatiotemporal memory decoupling module including multiple Spatiotemporal Adaptive Memory Units (STAMU). It is used to input the environmental state parameters of the target ocean at the current moment and output the predicted environmental state parameters at the next moment.
[0032] Furthermore, such as Figure 3 As shown, the Multidimensional Feature Perception Module (MDFPM) includes: a Channel Correlation Modeling Unit (CCMU), a Multi-Scale Spatial Perception Unit (MSPU), and a Deep Fusion Gating Mechanism Unit. Environmental state parameters are input to the channel correlation modeling unit and the multi-scale spatial perception unit, respectively, to obtain channel attention features and spatial attention features. The memory channel attention features, spatial attention features, and environmental state parameters are input into the deep fusion gating mechanism unit to obtain deep features.
[0033] Furthermore, the channel correlation modeling unit includes: a first average pooling layer, a first convolutional layer, and a first activation layer; The environmental state parameters are sequentially input into the first average pooling layer, the first convolutional layer, and the first activation layer to obtain channel attention features.
[0034] Furthermore, the Multidimensional Feature Perception Module (MDFPM) integrates lightweight perception mechanisms of channel and spatial dimensions, balancing expressive power and computational efficiency, and enhancing the model's ability to model complex multi-physical variables.
[0035] Furthermore, to model the interdependencies between physical variable channels, this invention employs a Channel Correlation Modeling Unit (CCMU) to extract channel attention. Unlike traditional channel attention (such as SE-Net), CCMU avoids dimensionality reduction operations and uses local one-dimensional convolution to directly model neighborhood dependencies in the channel dimension, thereby improving feature representation while reducing computational overhead. The specific process is as follows: Environmental status parameters The input is fed into the first average pooling layer for global average pooling to obtain the channel response. z c The calculation formula is: ; Where B represents the batch size, C represents the number of channels, H represents the latitude coordinates, W represents the longitude coordinates, and c represents the channel index; The input is then fed into the first convolutional layer, Conv1D, to model local dependencies between channels in a local manner, avoiding parameter redundancy introduced by the traditional fully connected approach. Finally, channel attention features are obtained through the first activation layer (Sigmoid activation function). A c : ; This channel attention feature is used to enhance the response to channels with key physical variables.
[0036] Furthermore, the multi-scale spatial perception unit includes: a second average pooling layer, a max pooling layer, a second convolutional layer, a third convolutional layer, and a second activation layer; Environmental state parameters are input into the second average pooling layer and the maximum pooling layer, respectively, to obtain the average pooling characteristics and the maximum pooling characteristics. Spatial features are obtained by fusing average pooling features and max pooling features; Based on the spatial features, they are input into the second and third convolutional layers respectively to obtain the first scale features and the second scale features respectively; Spatial attention features are obtained by linearly weighting the first and second scale features using learnable fusion weights and then inputting them into the second activation layer.
[0037] Furthermore, the Multi-Dimensional Feature Perception Module (MDFPM) aims to characterize salient regions in feature maps, focusing on capturing spatial structures with physical significance such as abrupt changes and vortices. To this end, MDFPM introduces a Multi-scale Spatial Perception Unit (MSPU) to analyze features from average pooling. With max pooling characteristics Extracting spatially salient features: ; ; ; The spliced spatial features M are input into the second (3×3) and third (7×7) convolutional layers respectively to capture receptive fields of different scales, thus obtaining the first-scale features. Second-scale features : ; ; Ultimately, learnable fusion weights γ For first-scale features Second-scale features Linear weighting is applied, followed by activation through a second activation layer to obtain spatial attention features. A s : .
[0038] Furthermore, the deep fusion gating mechanism unit includes: a fourth convolutional layer, a third activation layer, a fifth convolutional layer, and a fourth activation layer; The channel attention features, spatial attention features, and environmental state parameters are fused to obtain the first fused feature; The second fusion feature is obtained by fusing the initial fusion features with environmental state parameters; The third fusion feature is obtained by subtracting the second fusion feature from the environmental state parameters element by element. The environmental state parameters are sequentially input into the fourth convolutional layer, the third activation layer, the fifth convolutional layer, and the fourth activation layer to obtain the gating function; Deep features are obtained by fusing gating functions and third fusion features with environmental state parameters.
[0039] Furthermore, to achieve coordinated regulation of channel and spatial attention while mitigating the risk of false inhibition caused by attention enhancement, the present invention designs a deep fusion gating mechanism unit in MDFPM: First, the channel attention features are... A c Spatial attention characteristics A s and environmental state parameters x t-1 After fusion, then with environmental state parameters x t-1 Adding elements one by one yields the second fusion feature. x enh : ; Based on environmental state parameters x t-1 The input is sequentially fed into the fourth convolutional layer and the third activation layer Re. lu Fifth convolutional layer and fourth activation layer Sigmoid , obtain the gate function g t-1 : ; By constructing a gating function, the degree of attention enhancement is adaptively controlled, effectively mitigating the problem of false enhancement while preserving sensitivity to real target features; Based on gating function g t-1 After fusion with the third fusion feature, and with environmental state parameters x t-1 Fusion yields deep features : .
[0040] Furthermore, Where S represents salinity and T represents sea surface temperature. u x and u y These represent the two horizontal components of the seawater velocity. τ x and τ y These represent the two components of wind stress, and H represents the latent space obtained by the encoder. S3 uses deep feature input to the multiphysics ocean module, combines historical state information, and explicitly incorporates dynamic constraints to obtain physically consistent prediction results.
[0041] Furthermore, this invention proposes a deep neural network for modeling ocean physical evolution, called the Multi-Physics Ocean Cell (MPOC). MPOC, with physical constraints at its core, combines multi-scale convolution and data assimilation concepts to achieve the organic fusion of physical models and observational data within a prediction-correction architecture. Its core idea is to construct a collaborative mechanism between physical prior paths and observational assimilation paths, enabling the model to possess stronger generalization and stability in practical applications such as long-term data, strong nonlinearity, and sparse observations. Its goal is to optimize the historical hidden state vectors... h t-1 The hidden state vector for the next time step is updated, resulting in the physical consistency prediction result. h t .
[0042] Furthermore, such as Figure 4 As shown, the multiphysics ocean module includes: a multiscale physics predictor and an input assimilation unit; Obtain historical environmental state parameters of the target ocean; Obtain the historical hidden state vector based on historical environment state parameters. h t-1 ; Based on historical hidden state vectors h t-1 Input to a multi-scale physics predictor to obtain multi-scale fused features ; Based on multi-scale fusion features With historical hidden state vector h t-1 By merging the results, we can obtain historical physics predictions. ; Historical Physics Prediction Results and depth features The results are fed into the input assimilation unit to obtain the process physical prediction results; Based on the fusion of process physics prediction results and historical physics prediction results, a physical consistency prediction result is obtained. h t .
[0043] Furthermore, the multi-scale physics predictor includes: a sixth convolutional layer, a seventh convolutional layer, and an eighth convolutional layer; The historical hidden state vectors are input into the sixth and seventh convolutional layers, respectively, to obtain fine-scale features and large-scale features. The multi-scale fused features are obtained by fusing fine-scale and large-scale features and then inputting them into the eighth convolutional layer. .
[0044] Furthermore, in this embodiment, the sixth convolutional layer uses 3×3 Conv2D convolution, the seventh convolutional layer uses 7×7 Conv2D convolution, and the eighth convolutional layer uses 1×1 Conv2D convolution.
[0045] Furthermore, the basic principle of multi-scale physics predictors is to use convolution to approximate partial derivatives. M P Spatial derivative and coefficient c i,j Combining to a certain differential order q These can be uniformly written as a combination of first-order / second-order spatial derivatives. This type of general linear partial differential equation encompasses a wide range of classical ocean physics models, such as momentum, continuity, and thermohaline transport equations.
[0046] ; On a discrete grid, the above equation is achieved using two sets of learnable convolutional kernels: 3×3 convolution --approximate Fine-scale derivatives; 7×7 convolution --approximate Isoscale derivative; In the process of modeling various ocean variables using physical neural networks, for sea surface temperature and salinity, the velocity field advection salinity or temperature is calculated. Corresponding formula The first-order term in the equation, ▽, represents the gradient operator in the two-dimensional case. ; Horizontal eddy diffusion Corresponding formula The second-order term in k Represents the horizontal diffusion (or viscosity) coefficient, ▽ 2 Represents the Laplace operator. ; For the speed of horizontal seawater ( u x , u y The same convolution yields nonlinear advection and viscous diffusion, automatically learning the wind stress coupling coefficient using the weights in the 1×1 convolution. ,approximate For wind stress It is then regarded as an externally forced quantity, and its original value is retained and passed into the subsequent gating mechanism.
[0047] The equivalent operation of a neural network is to concatenate the two sets of outputs along the channel dimension, and then perform a linear combination using a 1×1 convolution. The formula for this process is: .
[0048] Furthermore, the input assimilation unit includes: a channel correlation modeling unit (CCMU), a ninth convolutional layer, and a fifth activation layer; Deep features The input is fed into the channel correlation modeling unit (CCMU) to obtain channel features. ; Based on channel characteristics Compared with historical physics prediction results Subtracting element by element, we obtain the first intermediate result. ; Based on channel characteristics Compared with historical physics prediction results The inputs are sequentially fed into the ninth convolutional layer (Conv2D) and the fifth activation layer. Sigmoid The approximate Kalman gain is obtained. K t-1 : ; like Then it depends entirely on historical physics predictions; if Completely trust the current observation input; the approximate Kalman gain, a dynamic gating adjustment mechanism, can more fully integrate historical prediction information with input information, resulting in higher model accuracy; Based on approximate Kalman gain K t-1 Compared with the first intermediate result After fusion, the second intermediate result is obtained as the process physical prediction result. .
[0049] Furthermore, historical physics prediction results : ; Based on process physics prediction results and historical physics prediction results By merging the results, we can obtain physically consistent predictions. h t : .
[0050] S4 fused the physical consistency prediction results with deep features to obtain the physical prediction enhancement results.
[0051] Furthermore, the physical prediction enhancement results are as follows: .
[0052] S5 inputs deep features into a spatiotemporal memory decoupling module that includes multiple spatiotemporal adaptive memory units to obtain spatiotemporal features.
[0053] Furthermore, this invention constructs a spatiotemporal memory decoupling module composed of multiple STAMU units connected in series, decoupling spatiotemporal memory into two independent but interactive paths: a temporal memory flow and a spatial memory flow, significantly improving the model's expressiveness and training stability. Each time step contains multiple stacked STAMU layers (e.g., STAMU1 to STAMU4), with each layer independently maintaining its hidden state vector. Time memory information With spatiotemporal memory information It supports information dissemination across time and levels. Specifically, it supports time-based memory information. Propagated along the timeline, it is responsible for capturing long-term dependent trends in the sequence, while spatiotemporal memory information The hidden state flows along the network hierarchy (spatial direction) to preserve structural information and spatial evolution features. These two memory states are updated using independent gating mechanisms, maintaining information integrity while preventing mutual interference. Finally, in the output stage, these two memory states are fused to generate the hidden state at the current moment, achieving architectural optimization through decoupling in the update stage and fusion in the output stage. This mechanism not only enhances the model's ability to model dynamic scenes but also establishes clearer and more controllable channels for temporal and spatial dependencies along the gradient propagation path, effectively alleviating the "deep-in-time dilemma."
[0054] This "decoupling" design breaks the constraints of the hierarchical transmission of hidden states in traditional time memory streams, enabling the model to learn evolutionary patterns at different scales in parallel on multiple memory trajectories, thereby improving its responsiveness to complex spatiotemporal nonlinear phenomena (such as ocean mutations, slowly changing structures, and multi-scale interactions).
[0055] Furthermore, such as Figure 5As shown, multiple spatiotemporal adaptive memory units (STAMUs) have the same structure, all including: dual-channel memory sub-units and dynamically growing window sub-units; Set the initial hidden state vector, initial time memory information, and initial spatiotemporal memory information at the initial moment; Based on the initial hidden state vector, initial temporal memory information, initial spatiotemporal memory information and deep features, the current hidden state vector, updated temporal memory information and updated spatiotemporal memory information are obtained by inputting them into the dual-channel memory sub-unit. The first hidden state vector is obtained by inputting the current hidden state vector into the dynamically growing window sub-unit; The deep features, the first hidden state vector, the updated temporal memory information, and the updated spatiotemporal memory information are input into the next spatiotemporal adaptive memory unit for processing, and so on. The corresponding hidden state vector output by the last spatiotemporal adaptive memory unit is used as the spatiotemporal feature.
[0056] Furthermore, based on the deep features, the first hidden state vector, the updated time memory information, and the updated spatiotemporal memory information, the data is input into the next spatiotemporal adaptive memory unit STAMU2 for processing, correspondingly obtaining the second hidden state vector, the updated time memory information, and the updated spatiotemporal memory information. This data is then input into the next spatiotemporal adaptive memory unit STAMU3 for processing, and so on. This embodiment employs four spatiotemporal adaptive memory units. Therefore, the fourth hidden state vector output by the fourth spatiotemporal adaptive memory unit STAMU4 is used as the final output spatiotemporal feature. H t .
[0057] Furthermore, the dual-channel memory sub-unit includes: the tenth convolutional layer, the sixth activation layer, the eleventh convolutional layer, the seventh activation layer, the twelfth convolutional layer, the eighth activation layer, the thirteenth convolutional layer, and the ninth activation layer; The deep features are sequentially input into the tenth convolutional layer and the sixth activation layer to obtain the first processed features; The initial hidden state vector is sequentially input into the eleventh convolutional layer and the seventh activation layer to obtain the second processed feature. Based on the extraction and fusion of the first and second processing features, the incremental time memory information is obtained; Updated time memory information is obtained by fusing incremental time memory information with initial time memory information; Based on the initial spatiotemporal memory information, it is sequentially input into the twelfth convolutional layer and the eighth activation layer to obtain the third processed feature; Based on the extraction and fusion of the third processing feature and the first processing feature, the spatiotemporal memory information increment is obtained; Updated spatiotemporal memory information is obtained by fusing incremental spatiotemporal memory information with initial spatiotemporal memory information; Based on the updated temporal memory information and the updated spatiotemporal memory information, the input is given to the thirteenth convolutional layer and the ninth activation layer to obtain the comprehensive memory features; Based on the comprehensive memory features, the first processing features, and the second processing features, the input gate features are obtained; The current hidden state vector is obtained by fusing input gate features and comprehensive memory features.
[0058] Furthermore, for the temporal memory branch, the increment of temporal memory information Δ is first obtained through calculations using the input gate, forget gate, and Modulation Gate. C t : ; ; ; ; in, , and All of these represent the learnable parameters in the tenth convolutional layer. , and All of these represent the learnable parameters in the eleventh convolutional layer; Then, the forgetting mechanism is used to update time memory. Based on the fusion of the incremental time memory information and the initial time memory information, the updated time memory information is obtained: .
[0059] The spatiotemporal memory branch works similarly; the third processing feature and the first processing feature yield the spatiotemporal memory information increment: ; ; ; ; .
[0060] Furthermore, after the two memory updates are completed, they are concatenated to form a comprehensive memory vector. Based on the comprehensive memory features, the first processing features, and the second processing features, the input gate features are obtained. o t : ; Based on the updated temporal memory information and the updated spatiotemporal memory information, the input is given to the thirteenth convolutional layer and the ninth activation layer to obtain the comprehensive memory features. ;in, W 1×1 This indicates the thirteenth convolutional layer; The current hidden state vector is obtained by fusing input gate features and comprehensive memory features. : .
[0061] Furthermore, the dynamically growing window sub-unit includes: a third average pooling layer, a buffer, a multi-head attention module, and a linear layer; The current hidden state vector is input into the third average pooling layer to obtain the query vector at the current time step; The window value is obtained based on the current buffer length and the preset growth step size; The key and value are selected based on the number of most recent window values in the buffer. The multi-head attention weights are obtained by inputting the query vector, key, and value into the multi-head attention module. The channel-level scaling factor is obtained by inputting multi-head attention weights into a linear layer; The first hidden state vector is obtained by fusing the current hidden state vector with the channel-level scaling factor.
[0062] Furthermore, the prehidden state vector The input is fed into the third average pooling layer for global average pooling and flattened to obtain the query vector for the current time step. q t Maintain a buffer to store historical query tokens, based on the current buffer length. L With preset growth step size Step Dynamically calculate window value τ size: ; Where len represents the length of the buffer, indicating τ The value must be within the length range of the Buffer, and its value is a... L / Step Round down; Extract the most recent data from the buffer. τ Each token is used as a key k and a value v, and input into the Multi-head Attention module to calculate the multi-head attention weights. z t : ; Based on multi-head attention weights z t Input to a linear layer to obtain channel-level scaling factors. ε : ; in, W and b Represents the hyperparameters of a linear layer; Based on the current hidden state vector and channel-level scaling factor ε By fusion, the first hidden state vector is obtained. : ; in, λ This represents the scaling modulation hyperparameter, and finally... q t The gradient separation version is added to the buffer for attention computation in future time steps.
[0063] Furthermore, the gated convolution update maintains stable capture of local spatial patterns and short-term dynamics; the dual-memory mechanism can accumulate different spatiotemporal semantics over long periods; dynamic temporal attention allows units to adaptively adjust the historical length of the lookback at different stages, accommodating both early short dependencies and later long dependencies; incremental alignment regularization provides additional physical consistency constraints. In the implementation, attention only acts on the one-dimensional channel vector after global pooling, the window length is limited by dynamic computation, and the computational and memory costs are low; the forget gate with a fixed bias helps stabilize training; the buffer is explicitly maintained externally, which can retain stride information during long sequence inference and clear it during segmented processing to avoid leakage. This STAMU structure can significantly improve the capture and utilization of spatiotemporal features in tasks such as complex motion prediction and long-term dependency modeling.
[0064] S6 obtains the environmental state prediction parameters of the target ocean at the next moment by weighted fusion of physical prediction enhancement results and spatiotemporal features.
[0065] Furthermore, the physical prediction enhancement results are weighted by spatiotemporal features. α and β Weighted fusion yields the environmental state prediction parameters for the target ocean at the next moment. : .
[0066] This ensures the physical interpretability of the model while preserving the sensitivity of data-driven modeling to minor changes.
[0067] Example 2 Based on the same inventive concept, this invention also provides a marine environmental parameter prediction system based on a multi-physics spatiotemporal adaptive memory network, including: an input data acquisition module, a depth feature extraction module, a physical result prediction module, a prediction result enhancement module, a spatiotemporal feature acquisition module, and a prediction result output module; The input data acquisition module is used to acquire the environmental state parameters of the target ocean at the current moment; The deep feature extraction module is used to input environmental state parameters into the multi-dimensional feature perception module to realize the dependency modeling of multiple physical variables and multi-scale spatial perception, and obtain the deep features of environmental state parameters. The physical outcome prediction module is used to input the depth feature into the multi-physics ocean module, combine historical state information, and explicitly incorporate dynamic constraints to obtain physically consistent prediction results. The prediction result enhancement module is used to fuse the physical consistency prediction result with deep features to obtain the physical prediction enhancement result; The spatiotemporal feature acquisition module is used to obtain spatiotemporal features by inputting deep features into the spatiotemporal memory decoupling module, which includes multiple spatiotemporal adaptive memory units. The prediction result output module is used to obtain the environmental state prediction parameters of the target ocean at the next moment by weighted fusion of the physical prediction enhancement results and spatiotemporal features.
[0068] Furthermore, in this embodiment, the functional implementation methods of each functional module correspond one-to-one with the methods described above, and will not be repeated here.
[0069] Example 3 Based on the same inventive concept, the present invention also provides a computer device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When the processor executes the program stored in the memory, it can implement the marine environmental parameter prediction method based on a multi-physics spatiotemporal adaptive memory network as described in Example 1.
[0070] Based on the same inventive concept, the present invention also provides an electronic device, which includes a processor and a memory, wherein the memory stores instructions, characterized in that the instructions are loaded and executed by the processor to implement the marine environmental parameter prediction method based on a multi-physics spatiotemporal adaptive memory network as in Embodiment 1.
[0071] The electronic device may include a processor, a communications interface, a memory, and a communication bus, wherein the processor, communications interface, and memory communicate with each other via the communication bus. The processor can call logical instructions in the memory to execute the marine environmental parameter prediction method based on a multi-physics spatiotemporal adaptive memory network in Embodiment 1.
[0072] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0073] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.
[0074] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for predicting marine environmental parameters based on multi-physics spatiotemporal adaptive memory networks, characterized in that, include: Obtain the environmental state parameters of the target ocean at the current moment; Based on the environmental state parameters input to the multi-dimensional feature perception module, the dependency modeling of multiple physical variables and multi-scale spatial perception are realized, and the depth features of the environmental state parameters are obtained. Based on the depth features input to the multiphysics ocean module, combined with historical state information, dynamic constraints are explicitly incorporated to obtain physically consistent prediction results. Based on the fusion of the physical consistency prediction result and the deep features, a physical prediction enhancement result is obtained; Based on the deep features, the spatiotemporal features are input to a spatiotemporal memory decoupling module that includes multiple spatiotemporal adaptive memory units to obtain spatiotemporal features; Based on the weighted fusion of the physical prediction enhancement results and the spatiotemporal features, the environmental state prediction parameters of the target ocean at the next moment are obtained.
2. The method for predicting marine environmental parameters based on a multi-physics spatiotemporal adaptive memory network according to claim 1, characterized in that, The multi-dimensional feature perception module includes: a channel correlation modeling unit, a multi-scale spatial perception unit, and a deep fusion gating mechanism unit; The environmental state parameters are input to the channel correlation modeling unit and the multi-scale spatial perception unit, respectively, to obtain channel attention features and spatial attention features. The channel attention features, spatial attention features, and environmental state parameters are input into the deep fusion gating mechanism unit to obtain the depth features.
3. The method for predicting marine environmental parameters based on a multi-physics spatiotemporal adaptive memory network according to claim 2, characterized in that, The channel correlation modeling unit includes: a first average pooling layer, a first convolutional layer, and a first activation layer; The environmental state parameters are sequentially input into the first average pooling layer, the first convolutional layer, and the first activation layer to obtain the channel attention features; The multi-scale spatial perception unit includes: a second average pooling layer, a max pooling layer, a second convolutional layer, a third convolutional layer, and a second activation layer. The environmental state parameters are respectively input to the second average pooling layer and the maximum pooling layer to obtain the average pooling feature and the maximum pooling feature. Spatial features are obtained by fusing the average pooling feature and the max pooling feature. Based on the spatial features, they are respectively input into the second convolutional layer and the third convolutional layer to obtain the first scale features and the second scale features respectively; The spatial attention features are obtained by linearly weighting the first and second scale features using learnable fusion weights and then inputting them into the second activation layer.
4. The method for predicting marine environmental parameters based on a multi-physics spatiotemporal adaptive memory network according to claim 2, characterized in that, The deep fusion gating mechanism unit includes: a fourth convolutional layer, a third activation layer, a fifth convolutional layer, and a fourth activation layer; The channel attention features, the spatial attention features, and the environmental state parameters are fused to obtain a first fused feature; Based on the fusion of the initial fusion features and the environmental state parameters, a second fusion feature is obtained; The third fusion feature is obtained by subtracting the second fusion feature from the environmental state parameters element by element. Based on the environmental state parameters, they are sequentially input into the fourth convolutional layer, the third activation layer, the fifth convolutional layer, and the fourth activation layer to obtain the gating function; The depth feature is obtained by fusing the gating function and the third fusion feature with the environmental state parameters.
5. The method for predicting marine environmental parameters based on a multi-physics spatiotemporal adaptive memory network according to claim 2, characterized in that, The multiphysics ocean module includes: a multi-scale physics predictor and an input assimilation unit; Obtain historical environmental state parameters of the target ocean; The historical hidden state vector is obtained based on the historical environmental state parameters; The multi-scale fusion features are obtained by inputting the historical hidden state vector into the multi-scale physical predictor. Based on the fusion of the multi-scale fusion features and the historical hidden state vector, the historical physical prediction result is obtained; The historical physics prediction results and the deep features are input into the input assimilation unit to obtain the process physics prediction results; The physical consistency prediction result is obtained by fusing the process physical prediction result and the historical physical prediction result.
6. The method for predicting marine environmental parameters based on a multi-physics spatiotemporal adaptive memory network according to claim 5, characterized in that, The multi-scale physics predictor includes: a sixth convolutional layer, a seventh convolutional layer, and an eighth convolutional layer; The historical hidden state vectors are respectively input into the sixth and seventh convolutional layers to obtain fine-scale features and large-scale features. The multi-scale fused features are obtained by fusing the fine-scale features and the large-scale features and then inputting them into the eighth convolutional layer.
7. The method for predicting marine environmental parameters based on a multi-physics spatiotemporal adaptive memory network according to claim 5, characterized in that, The input assimilation unit includes: a channel correlation modeling unit, a ninth convolutional layer, and a fifth activation layer; The depth features are input into the channel correlation modeling unit to obtain the channel features; The first intermediate result is obtained by subtracting the channel characteristics from the historical physical prediction results element by element. Based on the channel features and the historical physics prediction results, they are sequentially input into the ninth convolutional layer and the fifth activation layer to obtain an approximate Kalman gain. The second intermediate result is obtained by fusing the approximate Kalman gain with the first intermediate result, and is used as the physical prediction result of the process.
8. The method for predicting marine environmental parameters based on a multi-physics spatiotemporal adaptive memory network according to claim 1, characterized in that, The spatiotemporal adaptive memory units described above have the same structure, each including: a dual-channel memory subunit and a dynamically growing window subunit; Set the initial hidden state vector, initial time memory information, and initial spatiotemporal memory information at the initial moment; Based on the initial hidden state vector, the initial time memory information, the initial spatiotemporal memory information, and the deep features, the current hidden state vector, the updated time memory information, and the updated spatiotemporal memory information are obtained by inputting them into the dual-channel memory subunit. The first hidden state vector is obtained by inputting the current hidden state vector into the dynamically growing window sub-unit. Based on the deep features, the first hidden state vector, the updated time memory information, and the updated spatiotemporal memory information, the data is input to the next spatiotemporal adaptive memory unit for processing, and so on. The corresponding hidden state vector output by the last spatiotemporal adaptive memory unit is used as the spatiotemporal feature.
9. The method for predicting marine environmental parameters based on a multi-physics spatiotemporal adaptive memory network according to claim 8, characterized in that, The dual-channel memory sub-unit includes: a tenth convolutional layer, a sixth activation layer, an eleventh convolutional layer, a seventh activation layer, a twelfth convolutional layer, an eighth activation layer, a thirteenth convolutional layer, and a ninth activation layer; The depth features are sequentially input into the tenth convolutional layer and the sixth activation layer to obtain the first processed features; Based on the initial hidden state vector, it is sequentially input into the eleventh convolutional layer and the seventh activation layer to obtain the second processed feature; Based on the extraction and fusion of the first processing feature and the second processing feature, the incremental time memory information is obtained; The updated time memory information is obtained by fusing the incremental time memory information and the initial time memory information. Based on the initial spatiotemporal memory information, it is sequentially input into the twelfth convolutional layer and the eighth activation layer to obtain the third processing feature; Based on the extraction and fusion of the third processing feature and the first processing feature, the spatiotemporal memory information increment is obtained; The updated spatiotemporal memory information is obtained by fusing the spatiotemporal memory information increment and the initial spatiotemporal memory information. Based on the updated time memory information and the updated spatiotemporal memory information, the information is input into the thirteenth convolutional layer and the ninth activation layer to obtain the comprehensive memory features. Based on the comprehensive memory features, the first processing features, and the second processing features, the input gate features are obtained; The current hidden state vector is obtained by fusing the input gate features and the integrated memory features.
10. The method for predicting marine environmental parameters based on a multi-physics spatiotemporal adaptive memory network according to claim 8, characterized in that, The dynamic growth window subunit includes: a third average pooling layer, a buffer, a multi-head attention module, and a linear layer; The current hidden state vector is input to the third average pooling layer to obtain the query vector at the current time step; The window value is obtained based on the current buffer length and the preset growth step size of the buffer. The token with the most recent window value count is selected as the key and value based on the buffer. The query vector, the key, and the value are input into the multi-head attention module to obtain the multi-head attention weights. Based on the multi-head attention weights input to the linear layer, channel-level scaling factors are obtained; The first hidden state vector is obtained by fusing the current hidden state vector and the channel-level scaling factor.