A scale-aware, time-coupled method and system for predicting sea surface temperature
By using time-by-time recursive temporal modeling and frequency domain attention, large-scale physical processes and small-scale stochastic processes are explicitly distinguished. Combined with the ocean mixing layer heat budget equation and cross-scale amplitude sensing fusion, the problem of insufficient stability and accuracy in sea surface temperature prediction in existing technologies is solved, and higher physical consistency and prediction accuracy are achieved.
Patent Information
- Application Number
- CN202611277042.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-21
- Publication Date
- 2026-11-17
AI Technical Summary
Existing sea surface temperature prediction methods suffer from insufficient stability and error accumulation in long-term predictions. They lack explicit modeling and state interaction of large-scale physical processes and small-scale stochastic processes within the same time step, which leads to the introduction of non-physical high-frequency fluctuations in prediction results in frontal or high-variability regions.
A time-series modeling approach is adopted, which explicitly divides and collaboratively models large-scale physical processes and small-scale stochastic processes through gated cyclic units (GRUs). Frequency domain attention and the heat budget equation of the ocean mixing layer are introduced for constraints and corrections. Combined with a cross-scale amplitude sensing fusion mechanism, cross-scale collaborative evolution is achieved.
It improves the physical consistency and long-term stability of sea surface temperature prediction, significantly enhances the spatiotemporal consistency and overall accuracy of the prediction results, and can accurately depict local dynamic changes while maintaining global trend stability.
Smart Images

Figure CN123115756A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of ocean surface temperature prediction technology, specifically relating to a time-series coupled ocean surface temperature prediction method and system based on scale perception. Background Technology
[0002] Sea surface temperature (SST) is a key physical parameter characterizing the thermal state of the ocean. It has important applications in climate change research, marine ecosystem monitoring, fisheries resource management, and extreme weather event prediction. Accurate prediction of the sea surface temperature field is of great significance for understanding ocean dynamic processes, ensuring maritime economic activities, and disaster prevention and mitigation.
[0003] Traditional sea surface temperature (SST) prediction methods, such as convolutional neural networks, recurrent neural networks, and their variants, typically rely directly on historical SST data for time-series modeling and recursive prediction. However, existing purely data-driven prediction methods still suffer from insufficient prediction stability and significant long-term prediction error accumulation. To address these issues, a new SST prediction method integrating physical and data-driven information is proposed. This method constructs a dual-branch coupled structure of physical process flow and complex process flow to jointly model deterministic physical processes and stochastic complex processes in the evolution of SST. Specifically, a bidirectional information compensation mechanism is used to perform hidden state difference between the two branches. This allows the physical process flow to focus on characterizing large-scale physical consistency evolution features, while the complex process flow is used to compensate for small-scale nonlinear disturbances that are difficult for the physical model to explicitly describe. Simultaneously, the ocean mixing layer heat budget equation is introduced as a physical constraint to regularize the prediction process, and an anomaly compensation fusion module corrects the physical prediction results, thereby improving the accuracy and physical plausibility of SST prediction to a certain extent.
[0004] However, the above methods still have shortcomings, which will be discussed in detail below:
[0005] While focusing on integrating physical information and data-driven results at the overall prediction level, this approach lacks a mechanism to simultaneously model large-scale physical processes and small-scale stochastic processes within a single time step. In most methods, large-scale physical predictions are typically generated as a whole first, and then uniformly corrected by a data-driven model. However, the two types of processes do not form an explicit parallel modeling and state interaction relationship within the same time step, making it impossible for the evolution of small-scale perturbations to directly reflect the changes in the large-scale physical state at the current moment.
[0006] The lack of effective constraints and guidance of small-scale changes by utilizing large-scale evolutionary trends in time-by-time predictions makes it difficult to characterize the synergistic effects of different scale processes on the evolution of sea surface temperature (SST). In medium- and long-term recursive predictions, the small-scale compensation branch often relies solely on historical observations or prediction results from the previous moment for updates, lacking time-by-time constraints from large-scale physical trends. This can easily introduce non-physical high-frequency fluctuations or error accumulation in frontal zones or high-variability regions, affecting the stability and physical consistency of the prediction results. Summary of the Invention
[0007] To address the aforementioned technical issues, this invention proposes a scale-aware, time-coupled sea surface temperature prediction method and system. It employs a time-by-time recursive time-series modeling approach, using a gated cyclic unit (GRU) as the core time-series framework. At each time step, it explicitly divides and collaboratively models large-scale physical processes and small-scale stochastic processes, thereby generating sea surface temperature prediction results that satisfy physical consistency while also characterizing local detailed changes.
[0008] To achieve the above objectives, the first aspect of this invention proposes a scale-aware, time-coupled sea surface temperature prediction method, comprising:
[0009] S1, Large-scale physical processes:
[0010] By obtaining the sea surface temperature field at the previous moment Ocean dynamic characteristics and static features The system constructs a frequency domain key, a frequency domain query, and a time domain value, and introduces spectral attention to obtain large-scale time domain features. After low-pass filtering and land-sea masking, these features are compared with the hidden state from the previous time step. The data is input into the gated recurrent unit (GRU) for memory updates, thus obtaining the large-scale physical state at the current moment. By introducing physical constraints, large-scale prediction results can be obtained.
[0011] S2, Small-scale stochastic processes:
[0012] By obtaining the sea surface temperature field at the previous moment Ocean dynamic characteristics Static characteristics And the large-scale hidden state given by S1 in the previous time step. Constructing frequency domain keys and time domain values to capture ocean dynamic characteristics Static characteristics And the large-scale hidden state of the previous moment The concatenation mapping is converted into a time-domain query, and spectral attention is introduced to obtain small-scale time-domain features. After high-pass filtering and sea-land masking, these features are compared with the small-scale hidden state of the previous time step. The data is input into the gated recurrent unit (GRU) for memory updates, resulting in the current small-scale random state. Then, the mapping yields small-scale prediction results;
[0013] S3, Cross-scale amplitude sensing fusion:
[0014] From the current sea surface temperature field Amplitude information is extracted to construct a dynamic fusion weight field, and cross-scale weighted fusion of large-scale and small-scale prediction results is performed to obtain the final sea surface temperature prediction field.
[0015] Furthermore, by obtaining the sea surface temperature field at the previous moment... Ocean dynamic characteristics and static features This involves constructing a frequency domain key, a frequency domain query, and a time domain value, and introducing spectral attention to obtain large-scale time domain features, including:
[0016] S101, Ocean Dynamics and static features After being mapped linearly and then transformed by Fourier transform, the data is concatenated and mapped to the frequency domain query, representing the sea surface temperature field at the previous time step. The frequency domain key, representing the sea surface temperature field at the previous moment, is obtained through linear mapping and Fourier transform. The time-domain value is obtained through linear mapping;
[0017] S102. Calculate the similarity between the frequency domain Query and the frequency domain Key through dot product operation, and obtain the large-scale frequency domain weight through normalization operation.
[0018] S103. After performing an inverse Fourier transform on the large-scale frequency domain weights, multiply them element-wise with the time domain value to obtain the large-scale time domain features.
[0019] Furthermore, after low-pass filtering and land-sea masking, the hidden state is compared with the previous time step. The data is input into the gated recurrent unit (GRU) for memory updates, thus obtaining the large-scale physical state at the current moment. By introducing physical constraints, large-scale prediction results are obtained, including:
[0020] S104. Large-scale temporal features, after Fourier transform, are compared with the low-pass mask used for low-pass filtering. Element-wise multiplication yields large-scale low-frequency domain features, and inverse Fourier transform is performed to obtain large-scale low-frequency components.
[0021] S105, Large-scale Low-Frequency Components and Land-Sea Masking Element-wise multiplication and residual fusion with large-scale temporal features yield large-scale fused temporal features. ;
[0022] S106, Large-scale fused temporal features Hidden state from the previous moment The data is fed into the GRU for memory updates, yielding the large-scale physical state at the current moment. ;
[0023] S107, Large-scale physical state Based on the sea surface temperature heat budget equation, the sea surface temperature change rate constrained by large-scale physical state is obtained by modulating the horizontal advection term, surface heat flux term and vertical mixing term.
[0024] S108. Using the forward Eulerian scheme, the sea surface temperature change rate is multiplied by the time step and then added to the sea surface temperature field at the previous moment to obtain the large-scale prediction result.
[0025] Furthermore, by obtaining the sea surface temperature field at the previous moment... Ocean dynamic characteristics Static characteristics And the large-scale hidden state given by S1 in the previous time step. Constructing frequency domain keys and time domain values to capture ocean dynamic characteristics Static characteristics And the large-scale hidden state of the previous moment The concatenation mapping is transformed into a time-domain query, and spectral attention is introduced to obtain small-scale time-domain features, including:
[0026] S201, Sea Surface Temperature Field The time-domain value, the sea surface temperature field, is obtained through linear mapping. Ocean dynamic characteristics Static characteristics With hidden state Frequency domain key, frequency domain dynamic features, frequency domain static features, and frequency domain hidden state features are obtained through linear mapping and Fourier transform, respectively; the frequency domain dynamic features, frequency domain static features, and frequency domain hidden state features are concatenated and mapped to the frequency domain query.
[0027] S202. For each spectral location, calculate the similarity between the frequency domain Query and the frequency domain Key through dot product operation and normalize it to obtain the small-scale frequency domain weight.
[0028] S203. After performing an inverse Fourier transform on the small-scale frequency domain weights, multiply them element-wise with the time domain Value to obtain the small-scale time domain features.
[0029] Furthermore, after high-pass filtering and land-sea masking, the small-scale hidden state from the previous time step is obtained. The data is input into the gated recurrent unit (GRU) for memory updates, resulting in the current small-scale random state. Then, the mapping yields small-scale prediction results, including:
[0030] S204, Fourier transform of small-scale temporal features and the high-pass mask used for high-pass filtering. Element-wise multiplication yields small-scale high-frequency domain features, and inverse Fourier transform is performed to obtain small-scale high-frequency components.
[0031] S205, Small-Scale High-Frequency Components and Land-Sea Masking Element-wise multiplication and residual fusion with small-scale temporal features yield small-scale fused temporal features. ;
[0032] S206, Small-scale fusion temporal features Compared to the previous small-scale hidden state The data is fed into the GRU for memory updates, resulting in a small-scale random state at the current time step. It also maps and outputs small-scale prediction results.
[0033] Furthermore, from the current sea surface temperature field Amplitude information is extracted to construct a dynamic fusion weight field, and cross-scale weighted fusion of large-scale and small-scale prediction results is performed to obtain the final sea surface temperature prediction field, including:
[0034] S301, Based on the sea surface temperature field at the current moment Perform a two-dimensional fast Fourier transform and calculate the frequency domain amplitude spectrum;
[0035] S302. The frequency domain amplitude spectrum is mapped back to the time domain through inverse Fourier transform to obtain the pixel-by-pixel fused weight field.
[0036] S303. Apply normalization constraints to the pixel-by-pixel fusion weight field to obtain cross-scale dynamic fusion coefficients;
[0037] S304. The large-scale prediction results and small-scale prediction results are weighted and fused by cross-scale dynamic fusion coefficients to obtain the final sea surface temperature prediction field.
[0038] To achieve the above objectives, according to a second aspect of the present invention, a scale-aware, time-coupled sea surface temperature prediction system is provided, comprising:
[0039] Large-scale physical process units are used to obtain the sea surface temperature field from the previous moment. Ocean dynamic characteristics and static features The system constructs a frequency domain key, a frequency domain query, and a time domain value, and introduces spectral attention to obtain large-scale time domain features. After low-pass filtering and land-sea masking, these features are compared with the hidden state from the previous time step. The data is input into the gated recurrent unit (GRU) for memory updates, thus obtaining the large-scale physical state at the current moment. By introducing physical constraints, large-scale prediction results can be obtained.
[0040] Small-scale stochastic process units are used to obtain the sea surface temperature field from the previous time step. Ocean dynamic characteristics Static characteristics And the large-scale hidden state given by S1 in the previous time step. Constructing frequency domain keys and time domain values to capture ocean dynamic characteristics Static characteristics And the large-scale hidden state of the previous moment The concatenation mapping is converted into a time-domain query, and spectral attention is introduced to obtain small-scale time-domain features. After high-pass filtering and sea-land masking, these features are compared with the small-scale hidden state of the previous time step. The data is input into the gated recurrent unit (GRU) for memory updates, resulting in the current small-scale random state. Then, the mapping yields small-scale prediction results;
[0041] A cross-scale amplitude sensing fusion unit is used to obtain data from the current sea surface temperature field. Amplitude information is extracted to construct a dynamic fusion weight field, and cross-scale weighted fusion of large-scale and small-scale prediction results is performed to obtain the final sea surface temperature prediction field.
[0042] Compared with the prior art, the advantages of the present invention are as follows:
[0043] This invention improves the physical consistency and long-term stability of sea surface temperature (SST) prediction by constructing a time-series modeling framework that explicitly decouples large-scale physical processes from small-scale stochastic processes. Within a unified time-step recursive framework, this invention constructs both large-scale physical processes and small-scale stochastic processes, modeling low-frequency physical evolution and high-frequency stochastic disturbances separately at each time step. Compared to existing methods that only fuse processes at the stage or result level, this invention explicitly decouples the two types of processes in the time dimension. This allows the large-scale branch to focus on characterizing the macroscopic low-frequency trend of SST, while the small-scale branch focuses on local high-frequency changes, thus avoiding interference between features of different scales in the same modeling space. This scale-decoupled time-series modeling approach effectively suppresses non-physical drift phenomena in medium- and long-term predictions, enabling the prediction results to maintain the stability of the global evolution trend while possessing stronger physical interpretability and long-term robustness.
[0044] This invention also achieves deep-level collaborative modeling between processes at different scales through a cross-scale dynamic fusion mechanism of frequency domain guidance and amplitude sensing, significantly enhancing the spatiotemporal consistency of prediction results. This invention introduces frequency domain analysis into the feature modeling process, explicitly distinguishing low-frequency and high-frequency components through Fast Fourier Transform, and using large-scale physical states to guide small-scale high-frequency disturbances spectrally in the frequency domain. Simultaneously, based on the frequency domain amplitude information of the current SST, it constructs amplitude-sensing dynamic fusion weights to adaptively balance the relative contributions of large-scale physical results and small-scale random results in different regions and at different times. Compared to existing methods that rely on static or fixed weight fusion strategies, this invention can enhance the modeling of small-scale details in high-variability regions with concentrated energy, and strengthen the constraint on large-scale trends in stable background regions. This achieves true cross-scale collaborative evolution, enabling sea surface temperature prediction to maintain a consistent global trend while accurately depicting local dynamic changes, significantly improving the overall accuracy and stability of the prediction results. Attached Figure Description
[0045] 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.
[0046] In the attached diagram:
[0047] Figure 1 This is a flowchart illustrating the scale-aware temporal coupling sea surface temperature prediction method provided in this embodiment.
[0048] Figure 2 This is a schematic diagram of data processing for a large-scale physical process unit provided in this embodiment;
[0049] Figure 3 This is a schematic diagram of data processing for a small-scale stochastic process unit provided in this embodiment;
[0050] Figure 4 This is a schematic diagram of the scale-aware temporally coupled sea surface temperature prediction system provided in this embodiment. Detailed Implementation
[0051] 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.
[0052] 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.
[0053] 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 indicated technical features. Therefore, 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. Example 1:
[0054] Please see Figures 1-3 This invention provides a time-series coupled method for predicting sea surface temperature based on scale awareness, comprising the following steps:
[0055] S1, Large-scale physical processes:
[0056] By obtaining the sea surface temperature field at the previous moment Ocean dynamic characteristics and static features The system constructs a frequency domain key, a frequency domain query, and a time domain value, and introduces spectral attention to obtain large-scale time domain features. After low-pass filtering and land-sea masking, these features are compared with the hidden state from the previous time step. The data is input into the gated recurrent unit (GRU) for memory updates, thus obtaining the large-scale physical state at the current moment. By introducing physical constraints, large-scale prediction results can be obtained.
[0057] S2, Small-scale stochastic processes:
[0058] By obtaining the sea surface temperature field at the previous moment Ocean dynamic characteristics Static characteristics And the large-scale hidden state given by S1 in the previous time step. Constructing frequency domain keys and time domain values to capture ocean dynamic characteristics Static characteristics And the large-scale hidden state of the previous moment The concatenation mapping is converted into a time-domain query, and spectral attention is introduced to obtain small-scale time-domain features. After high-pass filtering and sea-land masking, these features are compared with the small-scale hidden state of the previous time step. The data is input into the gated recurrent unit (GRU) for memory updates, resulting in the current small-scale random state. Then, the mapping yields small-scale prediction results;
[0059] S3, Cross-scale amplitude sensing fusion:
[0060] From the current sea surface temperature field Amplitude information is extracted to construct a dynamic fusion weight field, and cross-scale weighted fusion of large-scale and small-scale prediction results is performed to obtain the final sea surface temperature prediction field.
[0061] In some embodiments, the overall structure of the large-scale physical process uses an improved gated recurrent unit (GRU) as the temporal skeleton and adopts a time-by-time recursive modeling approach, with each large-scale physical process unit corresponding to a time interval. By explicitly introducing frequency domain interactions and physical constraints into time series modeling, we can enhance large-scale prediction capabilities while ensuring physical consistency and interpretability.
[0062] See Figure 2 In some embodiments, the sea surface temperature field at the previous moment is obtained. Ocean dynamic characteristics and static features This involves constructing a frequency domain key, a frequency domain query, and a time domain value, and introducing spectral attention to obtain large-scale time domain features, including:
[0063] S101, Ocean Dynamics and static features After being mapped linearly and then transformed by Fourier transform, the data is concatenated and mapped to the frequency domain query, representing the sea surface temperature field at the previous time step. The frequency domain key, representing the sea surface temperature field at the previous moment, is obtained through linear mapping and Fourier transform. The time-domain value is obtained through linear mapping;
[0064] S102. Calculate the similarity between the frequency domain Query and the frequency domain Key through dot product operation, and obtain the large-scale frequency domain weight through normalization operation.
[0065] S103. After performing an inverse Fourier transform on the large-scale frequency domain weights, multiply them element-wise with the time domain value to obtain the large-scale time domain features.
[0066] Specifically, ocean dynamic characteristics Depicting dynamic and thermodynamic processes that change over time, static characteristics Describe background constraints such as topography, water depth, and land-sea distribution.
[0067] Specifically, ocean dynamic characteristics and static features After being mapped linearly and then transformed by Fourier transform, the data is concatenated and mapped to the frequency domain query, representing the sea surface temperature field at the previous time step. The frequency domain key, representing the sea surface temperature field at the previous moment, is obtained through linear mapping and Fourier transform. The time-domain value is obtained through linear mapping; it is represented as:
[0068] ;
[0069] ;
[0070] ;
[0071] ;
[0072] ;
[0073] in, This represents the two-dimensional Fast Fourier Transform (FFT). This represents the dynamic characteristics of the ocean at the previous moment in a large-scale physical process. A learnable linear projection layer that maps linearly to large-scale frequency domain dynamic features. Represents the static characteristics of a large-scale physical process at the previous moment. A learnable linear projection layer that maps linearly to large-scale frequency domain static features. It represents large-scale frequency domain dynamic characteristics. Represents large-scale frequency domain static characteristics. This indicates a splicing operation. This represents a linear mapping between the spliced large-scale frequency domain dynamic features and large-scale frequency domain static features. This represents a large-scale frequency domain query. This represents the sea surface temperature field at the previous moment in a large-scale physical process. A learnable linear projection layer that maps linearly to a large-scale frequency domain key. Represents a large-scale frequency domain key. This represents the sea surface temperature field at the previous moment in a large-scale physical process. A learnable linear projection layer for large-scale time-domain values, through linear mapping. Represents a large-scale time-domain value.
[0074] Specifically, to characterize the coupling relationship between the aforementioned dynamic and static features of the ocean at different frequency components, this invention introduces a dot-product form of spectral attention in the frequency domain, applying it to each spectral location. The similarity between the frequency domain query and the frequency domain key is calculated using dot product, and the large-scale frequency domain weight is obtained through normalization, as follows:
[0075] ;
[0076] ;
[0077] in, Indicates the scaling factor. Indicates the position in the spectrum at the previous moment. Large-scale spectral attention scoring, i.e., frequency domain query and frequency domain key at spectral location. Similarity measure at locations Indicates the position in the spectrum Large-scale frequency domain weights at that location This indicates a normalization operation.
[0078] Specifically, the dot product operation can directly measure the consistency between the above-mentioned ocean dynamic and static characteristics at the same frequency, thereby forming a "spectrum-wise interactive weight", which has the advantages of simple calculation and numerical stability.
[0079] Specifically, the large-scale frequency domain weights are mapped back to the time domain using an inverse Fourier transform, enabling external information to guide the evolution of sea surface temperature (SST) in the time domain. The large-scale time domain features are then obtained by element-wise multiplication with the time domain value, as follows:
[0080] ;
[0081] ;
[0082] in, This represents the two-dimensional inverse fast Fourier transform (IFFT). Represents large-scale frequency domain weights. Represents a large-scale time-domain value. This represents the interaction weights in the large-scale frequency domain. The time-domain guiding term obtained by the two-dimensional inverse fast Fourier transform It represents large-scale time-domain features.
[0083] In some implementations, after low-pass filtering and land / sea masking, the hidden state is compared with the previous time step. The data is input into the gated recurrent unit (GRU) for memory updates, thus obtaining the large-scale physical state at the current moment. By introducing physical constraints, large-scale prediction results are obtained, including:
[0084] S104. Large-scale temporal features, after Fourier transform, are compared with the low-pass mask used for low-pass filtering. Element-wise multiplication yields large-scale low-frequency domain features, and inverse Fourier transform is performed to obtain large-scale low-frequency components.
[0085] S105, Large-scale Low-Frequency Components and Land-Sea Masking Element-wise multiplication and residual fusion with large-scale temporal features yield large-scale fused temporal features. ;
[0086] S106, Large-scale fused temporal features Hidden state from the previous moment The data is fed into the GRU for memory updates, yielding the large-scale physical state at the current moment. ;
[0087] S107, Large-scale physical state Based on the sea surface temperature heat budget equation, the sea surface temperature change rate constrained by large-scale physical state is obtained by modulating the horizontal advection term, surface heat flux term and vertical entrainment term.
[0088] S108. Using the forward Eulerian scheme, the sea surface temperature change rate is multiplied by the time step and then added to the sea surface temperature field at the previous moment to obtain the large-scale prediction result.
[0089] Specifically, considering that large-scale physical mechanisms (horizontal advection term, surface heat flux term, and vertical entrainment term) mainly exhibit smooth, low-frequency evolution characteristics, while high-frequency components are more likely to contain small-scale random perturbations, this invention applies a low-pass filter (LPF) to the large-scale time-domain features.
[0090] Specifically, the large-scale time-domain features, after undergoing Fourier transform, are compared with the low-pass mask used for low-pass filtering. Element-wise multiplication preserves low frequencies and truncates high frequencies to obtain large-scale low-frequency frequency domain features. Inverse Fourier transform of these features yields large-scale low-frequency components, represented as:
[0091] ;
[0092] ;
[0093] ;
[0094] in, This represents a two-dimensional Fast Fourier Transform. This represents the large-scale frequency domain representation obtained after performing a two-dimensional fast Fourier transform on the time-domain features; This indicates that a low-pass mask is applied to the frequency domain representation. Large-scale low-frequency frequency domain features are obtained by retaining low-frequency components. This indicates element-wise multiplication. This represents the two-dimensional inverse fast Fourier transform. It represents large-scale low-frequency components.
[0095] Specifically, unlike the approach of the sea surface temperature control equation which "progresses time across the entire domain", the land mask does not act on the final physical progression result, but rather on the large-scale low-frequency components, in order to retain large-scale low-frequency information only on the ocean grid.
[0096] Specifically, large-scale low-frequency components and land-sea masking Element-wise multiplication and residual fusion with large-scale temporal features yield large-scale fused temporal features. , represented as:
[0097] ;
[0098] ;
[0099] in, This represents the land and sea mask, where the ocean pixel is 1 and the land pixel is 0. This represents the large-scale low-frequency components after masking. This represents the temporal characteristics of large-scale fusion.
[0100] Specifically, the core constraints on large-scale physical processes originate from the classical theory of the heat budget of the ocean mixed layer, which is expressed as:
[0101] ;
[0102] in, Indicates a local storage item. Represents the horizontal translation term. Indicates vertically enclosed items. This represents the vertical temperature / velocity covariance term. This represents the combination of net atmospheric heating and vertical turbulent diffusion at the bottom of the mixing layer;
[0103] The temporal evolution of sea surface temperature (SST) obtained through the ocean mixing layer heat budget equation exhibits significant multi-scale coupling characteristics: on the one hand, long-term SST variations are mainly dominated by large-scale deterministic physical mechanisms such as air-sea coupling, advection transport, and mixing layer heat budget; on the other hand, unresolved processes such as diffusion, turbulent mixing, and stochastic atmospheric forcing introduce a large number of small-scale high-frequency random disturbances, causing SST to exhibit complex nonlinear fluctuations on local and short-term scales.
[0104] Therefore, based on the existing research [SSTODE: Ocean-Atmosphere Physics-Informed Neural ODEs for Sea Surface Temperature Prediction] and [A Space-Time Partial Differential Equation Based Physics-Guided Neural Network for Sea Surface Temperature Prediction], this invention further simplifies the formula to obtain a sea surface temperature control equation suitable for deep learning frameworks, used to explicitly characterize large-scale deterministic evolution processes, expressed as:
[0105] ;
[0106] in, The time derivatives represent the average sea surface temperature of the mixed layer. Over time The rate of change of SST characterizes the response of SST to the combined effects of various physical processes at the current moment and is the core quantity describing the temporal evolution of SST.
[0107] This is the horizontal advection term. and These represent the ocean surface in the meridional direction. and latitude The horizontal velocity component in the direction is an important dynamic source of large-scale, low-frequency changes in SST;
[0108] This is the surface heat flux term. It represents the net heat flux acting on the sea surface, including shortwave radiation, longwave radiation, latent heat, and sensible heat flux. The density of seawater, The specific heat capacity of seawater at constant pressure. The thickness of the mixing layer is represented by the surface heat flux term, which describes the direct heating or cooling effect of air-sea energy exchange on the mixing layer temperature. This is one of the main ways in which air-sea coupling influences the evolution of the SST.
[0109] This is a vertical entrainment. This indicates the vertical mixing rate at the bottom of the mixing layer. The temperature of the subsurface seawater at the bottom of the mixing layer is denoted as ; the vertical entrainment term reflects the vertical heat exchange process between the mixing layer and the lower water body. When mixing is enhanced or upwelling occurs, the subsurface water body is entrained into the mixing layer, thereby modulating the sea surface temperature.
[0110] The residual term (Unresolved Subgrid-Scale Processes) is used to characterize small-scale physical processes that are not explicitly described in the equations, including diffusion, turbulent mixing, subgrid-scale eddies, and stochastic atmospheric forcing. These processes are usually highly nonlinear and stochastic, and are difficult to accurately describe through deterministic parameterization. Therefore, they are uniformly included in the residual term to reflect the high-frequency stochastic disturbances of SST on local and short-term scales.
[0111] Specifically, in large-scale physical states Based on the sea surface temperature governing equation, the sea surface temperature change rate constrained by large-scale physical states is obtained by modulating the horizontal advection term, surface heat flux term, and vertical entrainment term, as follows:
[0112] [ ];
[0113] in, This represents the sea surface temperature field at the previous moment. Based on this, from large-scale physical states The constrained rate of change of sea surface temperature; and Representing large-scale physical states respectively The modulated longitudinal and latitudinal horizontal velocity components; Represents large-scale physical states The modulated net heat flux; Represents large-scale physical states The thickness of the modulated hybrid layer; Represents large-scale physical states The vertical mixing velocity at the bottom of the modulated mixed layer; Represents large-scale physical states Temperature of the subsurface seawater at the bottom of the modulated mixed layer; ρ represents seawater density; This indicates the specific heat capacity of seawater at constant pressure.
[0114] Specifically, large-scale fusion of temporal features Hidden state from the previous moment The data is fed into the GRU for memory updates, yielding the large-scale physical state at the current moment. , represented as:
[0115] ;
[0116] ;
[0117] ;
[0118] ;
[0119] in, This represents the reset gate of the large-scale GRU at time step t. This represents the Sigmoid activation function; The weighting parameters of the large-scale reset gate with respect to the temporal features after residual fusion are represented. This indicates that the large-scale reset gate relates to the large-scale hidden state at the previous moment. Weight parameters; This represents the offset parameters of the large-scale reset gate; This represents the update gate of the large-scale GRU at time step t. This indicates a large-scale update gate with respect to input features. Weight parameters; This indicates that the large-scale update gate relates to the large-scale hidden state at the previous time step. Weight parameters; This indicates the bias parameters of the large-scale update gate; This represents the candidate large-scale hidden state of the large-scale GRU at time step t; The weight parameters representing the candidate large-scale hidden states with respect to the temporal features after residual fusion; The weight parameters of the candidate large-scale hidden state with respect to the previous large-scale hidden state after being modulated by the reset gate are represented. This represents the large-scale hidden state of the previous moment; The bias parameter represents the candidate large-scale hidden state; This represents the current state at a large scale obtained after updating at time step t.
[0120] Specifically, using a forward Euler time ordinary differential equation solver, the sea surface temperature change rate is multiplied by the time step, and then added to the sea surface temperature field at the previous moment to obtain the large-scale prediction result, expressed as:
[0121] ;
[0122] in, Indicates the time step. This indicates the results of large-scale predictions.
[0123] Specifically, by explicitly characterizing the deterministic physical evolution of SST at low-frequency scales through the aforementioned large-scale physical processes, a physically consistent macroscopic state basis is provided for subsequent small-scale stochastic process modeling and cross-scale fusion.
[0124] Unlike the sea surface temperature governing equation, large-scale physical processes focus only on large-scale deterministic processes dominated by horizontal advection, surface heat flux, and vertical entrainment terms, while neglecting small-scale residual terms that are not explicitly analyzed, such as diffusion, turbulent mixing, and stochastic forcing. The modeling and fusion are uniformly accomplished by a small-scale stochastic process, specifically: high-pass filtering is used to extract small-scale high-frequency perturbation components, and then residual fusion is used to form small-scale fused temporal features. and the small-scale hidden state of the previous moment. The common-input gated recurrent unit (GRU) performs the recursive update, thereby enabling the processing of small-scale residual terms. Modeling.
[0125] Small-scale stochastic processes are used to characterize small-scale high-frequency disturbances caused by unexplained analytical mechanisms such as diffusion, turbulent mixing, subgrid-scale eddies, and stochastic atmospheric forcing. Consistent with large-scale physical processes, they employ a time-by-time recursive modeling approach, with each small-scale stochastic process unit corresponding to a time interval. Furthermore, a gated recurrent unit (GRU) is used as the temporal skeleton to update the memory of small-scale states. The difference lies in that the small-scale process is mainly reflected in high-frequency components, and the evolution of the small-scale process needs to be constrained by the macroscopic physical background. Therefore, this invention explicitly introduces the state of the large-scale physical process from the previous moment. This guides the modeling of small-scale features at the current moment, thereby achieving cross-scale collaboration between macro and micro scales.
[0126] See Figure 3 In some implementations, the sea surface temperature field at the previous moment is obtained. Ocean dynamic characteristics Static characteristics And the large-scale hidden state given by S1 in the previous time step. Constructing frequency domain keys and time domain values to capture ocean dynamic characteristics Static characteristics And the large-scale hidden state of the previous moment The concatenation mapping is transformed into a time-domain query, and spectral attention is introduced to obtain small-scale time-domain features, including:
[0127] S201, Sea Surface Temperature Field The time-domain value, the sea surface temperature field, is obtained through linear mapping. Ocean dynamic characteristics Static characteristics With hidden state Frequency domain key, frequency domain dynamic features, frequency domain static features, and frequency domain hidden state features are obtained through linear mapping and Fourier transform, respectively; the frequency domain dynamic features, frequency domain static features, and frequency domain hidden state features are concatenated and mapped to the frequency domain query.
[0128] S202. For each spectral location, calculate the similarity between the frequency domain Query and the frequency domain Key through dot product operation and normalize it to obtain the small-scale frequency domain weight.
[0129] S203. After performing an inverse Fourier transform on the small-scale frequency domain weights, multiply them element-wise with the time domain Value to obtain the small-scale time domain features.
[0130] Specifically, at time step t, the sea surface temperature field of the previous moment is input. Ocean dynamic characteristics Static characteristics And the large-scale hidden state of the previous moment Sea surface temperature field See linear mapping. Figure 3 The linear layer yields the time-domain value, representing the sea surface temperature field. Ocean dynamic characteristics Static characteristics With hidden state The frequency domain key, frequency domain dynamic features, frequency domain static features, and frequency domain hidden state features are obtained through linear mapping and Fourier transform, respectively, and are expressed as follows:
[0131] ;
[0132] ;
[0133] ;
[0134] ;
[0135] ;
[0136] in, This represents the dynamic characteristics of the ocean at the previous moment in a small-scale physical process. A learnable linear projection layer that maps linearly to small-scale time-domain values. This represents the two-dimensional Fast Fourier Transform (FFT). Represents the small-scale time-domain value. This represents the dynamic characteristics of the ocean at the previous moment in a small-scale physical process. A learnable linear projection layer that maps linearly to a small-scale frequency domain key. Represents the small-scale frequency domain key. Represents the dynamic characteristics of the ocean at the previous moment in a small-scale physical process. A learnable linear projection layer that maps linearly to small-scale frequency domain dynamic features. It represents the dynamic characteristics in the frequency domain at a small scale. Represents the static characteristics of a small-scale physical process at the previous moment. A learnable linear projection layer that maps linearly to small-scale frequency domain static features. Represents small-scale frequency domain static characteristics. Represents the hidden state at the previous time step in a small-scale physical process. A learnable linear projection layer that maps linearly to small-scale frequency domain hidden state features. Represents large-scale frequency domain hidden state features;
[0137] Frequency domain dynamic features, frequency domain static features, and frequency domain hidden state features are concatenated and mapped to a frequency domain query to achieve frequency domain interaction. This query is used to characterize the joint prior of "external forcing + geographical constraints + macroscopic physical state" on small-scale perturbations; represented as:
[0138] ;
[0139] in, This indicates a splicing operation. Represents a linear mapping. This represents a small-scale frequency domain query.
[0140] Specifically, for each spectral location The similarity between the frequency domain query and the frequency domain key is calculated by dot product operation and normalized to obtain the small-scale frequency domain weight, expressed as:
[0141] ;
[0142] ;
[0143] in, Indicates the scaling factor. Indicates the position in the spectrum at the previous moment. Small-scale spectral attention scoring, i.e., frequency domain query and frequency domain key at spectral positions. Similarity measure at locations Indicates the location Small-scale frequency domain weights at that location;
[0144] Specifically, the small-scale frequency domain weights are mapped back to the time domain using an inverse Fourier transform, enabling the macroscopic state to guide the small-scale evolution in the time domain. The small-scale time domain features are then obtained by element-wise multiplication with the time-domain value, as follows:
[0145] ;
[0146] ;
[0147] in, Two-dimensional Inverse Fast Fourier Transform (IFFT) Represents the small-scale frequency domain weights. Represents the small-scale time-domain value. This represents the time-domain guiding term obtained by performing a two-dimensional inverse fast Fourier transform on the frequency-domain interactive weights. It represents small-scale time-domain features.
[0148] In other implementations, after high-pass filtering and land-sea masking, the small-scale hidden state from the previous time step is compared with the hidden state at the previous time step. The data is input into the gated recurrent unit (GRU) for memory updates, resulting in the current small-scale random state. Then, the mapping yields small-scale prediction results, including:
[0149] S204, Fourier transform of small-scale temporal features and the high-pass mask used for high-pass filtering. Element-wise multiplication yields small-scale high-frequency domain features, and inverse Fourier transform is performed to obtain small-scale high-frequency components.
[0150] S205, Small-Scale High-Frequency Components and Land-Sea Masking Element-wise multiplication and residual fusion with small-scale temporal features yield small-scale fused temporal features. ;
[0151] S206, Small-scale fusion temporal features Compared to the previous small-scale hidden state The data is fed into the GRU for memory updates, yielding the current small-scale physical state. It also maps and outputs small-scale prediction results.
[0152] Specifically, since small-scale perturbations mainly correspond to high-frequency components, this invention uses the same frequency domain truncation parameters as large-scale physical processes. The small-scale time-domain features are then subjected to Fourier transform and compared with the high-pass mask used for high-pass filtering. Element-wise multiplication, retaining high-frequency components and truncating low-frequency components, yields small-scale high-frequency domain features. Inverse Fourier transform of these small-scale high-frequency domain features yields small-scale high-frequency components, represented as:
[0153] ;
[0154] ;
[0155] ;
[0156] in, This represents a two-dimensional Fast Fourier Transform. The small-scale frequency domain representation obtained by performing a two-dimensional fast Fourier transform on the small-scale time-domain features. This represents the small-scale high-frequency frequency domain features obtained by applying a high-pass mask to the small-scale frequency domain representation and retaining the high-frequency components. This represents the two-dimensional inverse fast Fourier transform. It represents high-frequency components at small scales.
[0157] Specifically, small-scale high-frequency components and land-sea masking Element-wise multiplication is performed, and residual fusion is performed with the small-scale temporal features to obtain the small-scale fused temporal features. , represented as:
[0158] ;
[0159] ;
[0160] in, This represents the high-frequency components after masking. This represents the temporal features of small-scale fusion.
[0161] Specifically, small-scale fusion of temporal features Compared to the previous small-scale hidden state The data is fed into the GRU for memory updates, resulting in a small-scale random state at the current time step. , represented as:
[0162] ;
[0163] ;
[0164] ;
[0165] ;
[0166] in, This represents the reset gate of the small-scale GRU at time step t. This represents the Sigmoid activation function. This represents the weight of the reset gate with respect to the small-scale temporal features of the input residual fusion. This indicates that the gate's small-scale hidden state from the previous moment has been reset. The weight, This indicates resetting the door bias term. This represents the update gate of the small-scale GRU at time step t. This represents the weight parameters of the update gate with respect to the small-scale temporal features of the input residual fusion. This indicates that the update gate is related to the small-scale hidden state of the previous time step. The weight parameters, This indicates that the gate bias term is being updated. This represents the candidate hidden state of the small-scale GRU at time step t. The weight parameters represent the small-scale temporal features of the candidate small-scale hidden states after residual fusion with respect to the input. This represents the candidate small-scale hidden state with respect to the previous small-scale hidden state after being modulated by the reset gate. This represents the small-scale hidden state of the previous moment. This represents the candidate state bias term. It represents the small-scale random state at the current moment.
[0167] Small-scale random state The mapping outputs the small-scale high-frequency perturbation information at the current moment, i.e., the small-scale prediction result. .
[0168] The aforementioned small-scale stochastic processes, while maintaining a feature processing paradigm consistent with large-scale physical processes, explicitly utilize the large-scale hidden state from the previous time step. The modeling of high-frequency disturbances is guided by high-pass filtering, which focuses on small-scale random processes, thereby providing high-frequency detailed representations for subsequent cross-scale dynamic fusion.
[0169] In some implementations, from the sea surface temperature field at the current moment Amplitude information is extracted to construct a dynamic fusion weight field, and cross-scale weighted fusion of large-scale and small-scale prediction results is performed to obtain the final sea surface temperature prediction field, including:
[0170] S301, Based on the sea surface temperature field at the current moment Perform a two-dimensional fast Fourier transform and calculate the frequency domain amplitude spectrum;
[0171] S302. The frequency domain amplitude spectrum is mapped back to the time domain through inverse Fourier transform to obtain the pixel-by-pixel fused weight field.
[0172] S303. Apply normalization constraints to the pixel-by-pixel fusion weight field to obtain cross-scale dynamic fusion coefficients;
[0173] S304. The large-scale prediction results and small-scale prediction results are weighted and fused by cross-scale dynamic fusion coefficients to obtain the final sea surface temperature prediction field.
[0174] Specifically, due to the significant differences in the spectral structure of ocean processes, the frequency domain amplitude of the SST can directly reflect the strength of the contribution of processes at different scales to the temperature field at the current moment. When the spectral amplitude at a certain location is larger, it often corresponds to stronger energy input or more significant scale coupling, and the model should rely more on the high-frequency details characterized by small-scale stochastic processes. Conversely, when the amplitude is smaller and the spectral energy is more stable, the evolution of the SST is closer to the smooth change dominated by large-scale physical mechanisms and should follow the deterministic trend of large-scale physical processes more. Therefore, this invention directly extracts amplitude information from the SST at the current moment, constructs dynamic fusion weights, and realizes cross-scale collaborative computing.
[0175] Specifically, in time step Based on the current sea surface temperature field Perform a two-dimensional fast Fourier transform and calculate the frequency domain amplitude spectrum, which is expressed as:
[0176] ;
[0177] ;
[0178] in, Represents a complex matrix. Table of two-dimensional Fast Fourier Transform, and Let them represent the real and imaginary parts of a complex number, respectively. Indicates time step The frequency domain amplitude spectrum.
[0179] Specifically, the frequency domain amplitude spectrum mapping is mapped back to the time domain through inverse Fourier transform to obtain the pixel-by-pixel fused weight field, expressed as:
[0180] ;
[0181] in, Two-dimensional inverse fast Fourier transform, Indicates time step Pixel-wise fusion weight field.
[0182] Specifically, a normalization constraint is applied to the pixel-by-pixel fusion weight field to obtain cross-scale dynamic fusion coefficients.
[0183] Represented as:
[0184] ;
[0185] in, For the Sigmoid function, For cross-scale dynamic fusion coefficients, The larger the value, the stronger the spectral amplitude at the current moment, corresponding to stronger small-scale perturbations and high-frequency detail contributions; The smaller the value, the more stable the spectral energy, corresponding to a more significant large-scale smooth evolution trend.
[0186] Specifically, The large-scale and small-scale prediction results are weighted and fused to obtain the final sea surface temperature prediction field, which is expressed as:
[0187] ;
[0188] in, This represents the final predicted sea surface temperature field.
[0189] Example 2:
[0190] As another embodiment of the present invention, a scale-aware temporal coupled sea surface temperature prediction system is proposed, which adopts a scale-aware temporal coupled model for sea surface temperature prediction (SST-STCM) architecture, including:
[0191] Large-scale physical process units are used to obtain the sea surface temperature field from the previous moment. Ocean dynamic characteristics and static features The system constructs a frequency domain key, query, and time domain value, and introduces spectral attention to obtain large-scale time domain features. After low-pass filtering and land-sea masking, these features are compared with the hidden state from the previous time step. The data is input into the gated recurrent unit (GRU) for memory updates, thus obtaining the large-scale physical state at the current moment. By introducing physical constraints, large-scale prediction results can be obtained.
[0192] Small-scale stochastic process units are used to obtain the sea surface temperature field from the previous time step. Ocean dynamic characteristics Static characteristics And the large-scale hidden state given by S1 in the previous time step. Constructing frequency domain keys and time domain values to capture ocean dynamic characteristics Static characteristics And the large-scale hidden state of the previous moment The concatenation mapping is converted into a time-domain query, and spectral attention is introduced to obtain small-scale time-domain features. After high-pass filtering and sea-land masking, these features are compared with the small-scale hidden state of the previous time step. The data is input into the gated recurrent unit (GRU) for memory updates, resulting in the current small-scale random state. Then, the mapping yields small-scale prediction results;
[0193] A cross-scale amplitude sensing fusion unit is used to obtain data from the current sea surface temperature field. Amplitude information is extracted to construct a dynamic fusion weight field, and cross-scale weighted fusion of large-scale and small-scale prediction results is performed to obtain the final sea surface temperature prediction field.
[0194] Example 3: Model Training
[0195] Dataset:
[0196] This invention uses the NOAA Optimum Interpolation Sea Surface Temperature version 2.1 (OISSTv2.1) dataset provided by the National Oceanic and Atmospheric Administration (NOAA) to acquire sea surface temperature data. The OISST v2.1 dataset integrates satellite remote sensing observation data and marine field observation data, providing a continuous sea surface temperature field with a spatial resolution of 0.25 degrees × 0.25 degrees and a temporal resolution of 1 day.
[0197] Training sample construction and dataset partitioning:
[0198] A sliding time window is used to construct training samples sequentially along the time axis, maintaining the temporal order between adjacent training samples. For multi-time-dependent sea surface temperature (SST) prediction tasks, after completing the prediction for the current time, the current prediction result is used as the input for the SST at the next time, and prediction results for multiple future time periods are obtained by recursively calculating each time period. This embodiment performs SST predictions for the next 1, 5, 7, 10, and 14 days, respectively.
[0199] To prevent data leakage in the future, the dataset is divided according to the chronological order of the samples. The first 85% of the samples in the chronological order are used as the training set, and the last 15% are used as the test set. The training set is used for model parameter learning, while the test set is only used for performance evaluation after model training and does not participate in model parameter updates.
[0200] Model parameters and training environment:
[0201] A scale-aware temporal coupling model was implemented using the PyTorch deep learning framework and trained on a workstation equipped with an NVIDIA GeForce RTX 3090 graphics processor with 24 GB of video memory, and accelerated by parallel computing through CUDA.
[0202] Both the large-scale physical processes and small-scale stochastic processes in the model use gated recurrent units (GRUs) as the temporal modeling skeleton. The large-scale physical processes employ low-pass filtering to retain low-frequency components, while the small-scale stochastic processes employ high-pass filtering to retain high-frequency components. Both low-pass and high-pass filtering use the same frequency cutoff parameter; in this embodiment, the frequency cutoff parameter is set to 20.
[0203] The batch size during model training is set to 32, meaning 32 training samples are read from the training set in each iteration. The Adam optimizer is used to update the learnable parameters in the model, with an initial learning rate of [missing value]. .
[0204] During model training, a cosine annealing learning rate adjustment strategy is used to dynamically reduce the learning rate, so that the model searches for parameters with a larger learning rate in the early stage of training and refines the parameters with a smaller learning rate in the later stage of training.
[0205] The maximum number of training epochs for the model is set to 300. To suppress overfitting, an early stopping strategy is employed during training; training is terminated prematurely if the model loss does not decrease further within a preset number of consecutive epochs. Therefore, the actual number of training epochs does not exceed 300.
[0206] Loss function:
[0207] This invention uses the mean squared error (MSE) as the objective function, takes the difference between the predicted SST and the true SST as the optimization signal, and updates the model parameters through a backpropagation algorithm. The training batch contains... The nth sample, the nth The true SST of each sample is The model prediction results are The loss function is expressed as:
[0208] ;
[0209] in, Describing the Frobenius norm, This represents the mean squared error loss function;
[0210] By minimizing the above loss function, the model can learn an effective representation and prediction capability for SST evolution in multi-scale coupled scenarios.
[0211] Experimental results:
[0212] This invention has been used in a systematic comparative experiment on sea surface temperature forecasting, covering multiple time spans from short-term (1 day) to medium- and long-term (14 days). The experiments selected Convolutional Long Short-Term Memory Network (ConvLSTM), Hierarchical VisionTransformer Using Shifted Windows (Swin-Transformer), Convolutional Networks for Biomedical Image Segmentation (U-NET), An Enhanced Fourier Neural Operator-Based Deep-Learning Model for Multiphase Flow (U-FNO) combined with U-Net, Fourier ForeCasting Neural Network (FourCastNet), Dense Dilated Convolutional LSTM (D2CLD2CL), and Space-Time PDE-guided Neural Network (STPDE-NET) as baseline models. The results were analyzed by measuring RMSE (root mean square error), MAE (mean absolute error), and R² (coefficient of determination). The overall root mean square error, Pearson correlation coefficient ρ, and segmented temperature RMSE (15RMSE, 20RMSE, 25RMSE) were used to comprehensively evaluate the model's performance under different forecast lead times. The experimental results are shown in Tables 3-1 to 3-5. Among them:
[0213] .
[0214] As shown in Table 3-1, in the 1-day short-term forecast task, the SST-STCM proposed in this invention exhibits significant advantages, achieving the lowest overall RMSE value of 0.1928, far superior to the RMSE values of Swin-Transformer (0.2193) and STPDE-NET (0.2131). Particularly in the 25°C high-temperature sea area with intense thermal activity, the proposed invention achieves an RMSE of 0.1604, significantly lower than the ConvLSTM (0.1863) and D2CL (0.1977). This advantage is mainly attributed to the multi-scale decoupling architecture and amplitude-sensing fusion mechanism employed in the model. By explicitly separating high-frequency components in the frequency domain, small-scale stochastic process units can focus on characterizing the microscopic perturbations induced by turbulent mixing and random forcing, avoiding the smoothing effect of local details in the global optimization of traditional models. Simultaneously, the dynamic fusion strategy based on frequency domain energy distribution allows the model to adaptively assign higher weights to small-scale details in high-temperature active regions with concentrated energy, thereby achieving accurate capture of complex thermal structures in instantaneous forecasts. Furthermore, the R² of the first day of this invention is as high as 0.9938, and the correlation coefficient ρ reaches 0.9967, indicating that the forecast results maintain a very high consistency with the actual values.
[0215] Table 3-1: Analysis of Sea Surface Temperature Forecast Indicators for the Next Day
[0216]
[0217] As the forecast lead time extends to the medium term of 5-7 days, the nonlinear dynamic characteristics of the sea surface temperature field become more complex, and the cumulative effect of errors begins to emerge. However, this invention still maintains a robust leading position. As shown in Table 3-2, in the forecast for the next 5 days, the RMSE of this invention is 0.5329, significantly better than the RMSE values of U-FNO (0.6000) and D2CL (0.6021). Although STPDE-NET shows strong competitiveness under physical constraints with an RMSE value of 0.5948, this invention still significantly outperforms the former in terms of MAE (0.4125) and MAE (0.4508).
[0218] Table 3-2: Analysis of Sea Surface Temperature Forecast Indicators for the Next 5 Days
[0219]
[0220] As shown in Table 3-3, this advantage is further amplified in the 7-day forecast. The RMSE of this invention is controlled at 0.6261, while the error of the classic ConvLSTM has diverged to 0.7340, and the Swin-Transformer has also risen to 0.7165. Even FourCastNet, designed for large-scale flow fields, has a significantly higher RMSE value of 0.8177 than this invention. This advantage at this stage verifies the effectiveness of the large-scale physical process unit and its built-in physical constraints. Unlike pure data-driven methods that rely solely on time-series memory, the large-scale branch of this invention explicitly performs advection transport and heat budget calculations in the low-frequency domain, ensuring that the macroscopic evolution trajectory strictly follows the physical conservation laws. This physical consistency constraint effectively suppresses feature ambiguity and non-physical drift in medium-term forecasts, enabling the model to more accurately extrapolate the spatial displacement of heat than simple memory units.
[0221] Table 3-3: Analysis of Sea Surface Temperature Forecast Indicators for the Next 7 Days
[0222]
[0223] In the more challenging long-term forecasting phase of 10-14 days, this invention demonstrates exceptional robustness and resistance to error accumulation. As shown in Tables 3-4, in the 10-day forecast, the RMSE of this invention is 0.7522, superior to FourCastNet and STPDE-NET. In the critical high-temperature range of 25°C, the RMSE of this invention is 0.6600, slightly better than the RMSE value of FourCastNet (0.663).
[0224] Table 3-4: Analysis of Sea Surface Temperature Forecast Indicators for the Next 10 Days
[0225]
[0226] As shown in Table 3-5, this invention once again established a comprehensive leading advantage in the 14-day forecast, with an overall RMSE of 0.8501, significantly lower than the TRMSE of U-NE (0.8665) and STPDE-NET (0.8750). Meanwhile, the RMSE of pure sequence models such as ConvLSTM and Swin-Transformer surged to over 1.22 and 1.20 respectively, exhibiting significant error amplification. This invention, with a high correlation coefficient ρ of 0.9552, demonstrates that even over long timeframes, the evolution trend of the predicted results remains highly consistent with the actual sea conditions. This long-term robustness is mainly attributed to the cross-scale interaction mechanism in the model. The model explicitly utilizes large-scale hidden states to guide the modeling of small-scale stochastic processes. This "macro-constrained micro" co-evolution strategy ensures that the generation of high-frequency details is always constrained by the large-scale physical background, effectively avoiding the accumulation and divergence of random errors in long-term time-series extrapolation, thus achieving a dual breakthrough in accuracy and stability in complex marine environment forecasting.
[0227] Table 3-5: Analysis of Sea Surface Temperature Forecast Indicators for the Next 14 Days
[0228]
[0229] In summary, although the forecast accuracy of all models decreased over time, the error growth of this invention remained the smallest, especially in the 25RMSE index, where it consistently maintained the lowest error level among similar methods from day 1 to day 14. This significant advantage fully validates the superiority of fusing large-scale deterministic processes, small-scale stochastic processes, and amplitude sensing. By decoupling large-scale deterministic processes and small-scale stochastic processes and dynamically balancing their weights using frequency domain energy features, the limitations of single-scale modeling were successfully overcome, achieving high-fidelity reconstruction of local details while ensuring long-term physical consistency. Whether in short-term precise capture (1 day) or long-term trend extrapolation (14 days), this invention demonstrated excellent forecast accuracy and stability, proving the advancement and applicability of the multi-scale decoupling and co-evolution modeling paradigm in forecasting complex marine environments.
[0230] The program product of this invention for implementing the above-described method may employ a portable compact disk read-only memory and include program code, and may run on a terminal device, such as a personal computer. However, the program product of this invention is not limited thereto. In this invention, the readable storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system, apparatus, or device.
[0231] It should be noted that a computer-readable storage medium may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium, capable of transmitting, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.
[0232] 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 time-series coupled sea surface temperature prediction method based on scale perception, characterized in that, Includes the following steps: S1, Large-scale physical processes: By obtaining the sea surface temperature field at the previous moment Ocean dynamic characteristics and static features The system constructs a frequency domain key, a frequency domain query, and a time domain value, and introduces spectral attention to obtain large-scale time domain features. After low-pass filtering and land-sea masking, these features are compared with the hidden state from the previous time step. The data is input into the gated recurrent unit (GRU) for memory updates, thus obtaining the large-scale physical state at the current moment. By introducing physical constraints, large-scale prediction results can be obtained. S2, Small-scale stochastic processes: By obtaining the sea surface temperature field at the previous moment Ocean dynamic characteristics Static characteristics And the large-scale hidden state given by S1 in the previous time step. Constructing frequency domain keys and time domain values to capture ocean dynamic characteristics Static characteristics And the large-scale hidden state of the previous moment The concatenation mapping is converted into a time-domain query, and spectral attention is introduced to obtain small-scale time-domain features. After high-pass filtering and sea-land masking, these features are compared with the small-scale hidden state of the previous time step. The data is input into the gated recurrent unit (GRU) for memory updates, resulting in the current small-scale random state. Then, the mapping yields small-scale prediction results; S3, Cross-scale amplitude sensing fusion: From the current sea surface temperature field Amplitude information is extracted to construct a dynamic fusion weight field, and cross-scale weighted fusion of large-scale and small-scale prediction results is performed to obtain the final sea surface temperature prediction field.
2. The method according to claim 1, characterized in that, By obtaining the sea surface temperature field at the previous moment Ocean dynamic characteristics and static features This involves constructing a frequency domain key, a frequency domain query, and a time domain value, and introducing spectral attention to obtain large-scale time domain features, including: S101, Ocean Dynamics and static features After being mapped linearly and then transformed by Fourier transform, the data is concatenated and mapped to the frequency domain query, representing the sea surface temperature field at the previous time step. The frequency domain key, representing the sea surface temperature field at the previous moment, is obtained through linear mapping and Fourier transform. The time-domain value is obtained through linear mapping; S102. Calculate the similarity between the frequency domain Query and the frequency domain Key through dot product operation, and obtain the large-scale frequency domain weight through normalization operation. S103. After performing an inverse Fourier transform on the large-scale frequency domain weights, multiply them element-wise with the time domain value to obtain the large-scale time domain features.
3. The method according to claim 2, characterized in that, After low-pass filtering and land-sea masking, compared with the hidden state of the previous time step... The data is input into the gated recurrent unit (GRU) for memory updates, thus obtaining the large-scale physical state at the current moment. By introducing physical constraints, large-scale prediction results are obtained, including: S104. Large-scale temporal features, after Fourier transform, are compared with the low-pass mask used for low-pass filtering. Element-wise multiplication yields large-scale low-frequency domain features, and inverse Fourier transform is performed to obtain large-scale low-frequency components. S105, Large-scale Low-Frequency Components and Land-Sea Masking Element-wise multiplication and residual fusion with large-scale temporal features yield large-scale fused temporal features. ; S106, Large-scale fused temporal features Hidden state from the previous moment The data is fed into the GRU for memory updates, yielding the large-scale physical state at the current moment. ; S107, Large-scale physical state Based on the sea surface temperature heat budget equation, the sea surface temperature change rate constrained by large-scale physical state is obtained by modulating the horizontal advection term, surface heat flux term and vertical entrainment term. S108. Using the forward Eulerian scheme, the sea surface temperature change rate is multiplied by the time step and then added to the sea surface temperature field at the previous moment to obtain the large-scale prediction result.
4. The method according to claim 2, characterized in that, By obtaining the sea surface temperature field at the previous moment Ocean dynamic characteristics Static characteristics And the large-scale hidden state given by S1 in the previous time step. Constructing frequency domain keys and time domain values to capture ocean dynamic characteristics Static characteristics And the large-scale hidden state of the previous moment The concatenation mapping is transformed into a time-domain query, and spectral attention is introduced to obtain small-scale time-domain features, including: S201, Sea Surface Temperature Field The time-domain value, the sea surface temperature field, is obtained through linear mapping. Ocean dynamic characteristics Static characteristics With hidden state Frequency domain key, frequency domain dynamic features, frequency domain static features, and frequency domain hidden state features are obtained through linear mapping and Fourier transform, respectively; the frequency domain dynamic features, frequency domain static features, and frequency domain hidden state features are concatenated and mapped to the frequency domain query. S202. For each spectral location, calculate the similarity between the frequency domain Query and the frequency domain Key through dot product operation and normalize it to obtain the small-scale frequency domain weight. S203. After performing an inverse Fourier transform on the small-scale frequency domain weights, multiply them element-wise with the time domain Value to obtain the small-scale time domain features.
5. The method according to claim 4, characterized in that, After high-pass filtering and land-sea masking, compared with the small-scale hidden state of the previous time step... The data is input into the gated recurrent unit (GRU) for memory updates, resulting in the current small-scale random state. Then, the mapping yields small-scale prediction results, including: S204, Fourier transform of small-scale temporal features and the high-pass mask used for high-pass filtering. Element-wise multiplication yields small-scale high-frequency domain features, and inverse Fourier transform is performed to obtain small-scale high-frequency components. S205, Small-Scale High-Frequency Components and Land-Sea Masking Element-wise multiplication and residual fusion with small-scale temporal features yield small-scale fused temporal features. ; S206, Small-scale fusion temporal features Compared to the previous small-scale hidden state The data is fed into the GRU for memory updates, resulting in a small-scale random state at the current time step. It also maps and outputs small-scale prediction results.
6. The method according to claim 5, characterized in that, From the current sea surface temperature field Amplitude information is extracted to construct a dynamic fusion weight field, and cross-scale weighted fusion of large-scale and small-scale prediction results is performed to obtain the final sea surface temperature prediction field, including: S301, Based on the sea surface temperature field at the current moment Perform a two-dimensional fast Fourier transform and calculate the frequency domain amplitude spectrum; S302. The frequency domain amplitude spectrum is mapped back to the time domain through inverse Fourier transform to obtain the pixel-by-pixel fused weight field. S303. Apply normalization constraints to the pixel-by-pixel fusion weight field to obtain cross-scale dynamic fusion coefficients; S304. The large-scale prediction results and small-scale prediction results are weighted and fused by cross-scale dynamic fusion coefficients to obtain the final sea surface temperature prediction field.
7. A scale-aware, time-coupled sea surface temperature prediction system, characterized in that, include: Large-scale physical process units are used to obtain the sea surface temperature field from the previous moment. Ocean dynamic characteristics and static features The system constructs a frequency domain key, a frequency domain query, and a time domain value, and introduces spectral attention to obtain large-scale time domain features. After low-pass filtering and land-sea masking, these features are compared with the hidden state from the previous time step. The data is input into the gated recurrent unit (GRU) for memory updates, thus obtaining the large-scale physical state at the current moment. By introducing physical constraints, large-scale prediction results can be obtained. Small-scale stochastic process units are used to obtain the sea surface temperature field from the previous time step. Ocean dynamic characteristics Static characteristics And the large-scale hidden state given by S1 in the previous time step. Constructing frequency domain keys and time domain values to capture ocean dynamic characteristics Static characteristics And the large-scale hidden state of the previous moment The concatenation mapping is converted into a time-domain query, and spectral attention is introduced to obtain small-scale time-domain features. After high-pass filtering and sea-land masking, these features are compared with the small-scale hidden state of the previous time step. The data is input into the gated recurrent unit (GRU) for memory updates, resulting in the current small-scale random state. Then, the mapping yields small-scale prediction results; A cross-scale amplitude sensing fusion unit is used to obtain data from the current sea surface temperature field. Amplitude information is extracted to construct a dynamic fusion weight field, and cross-scale weighted fusion of large-scale and small-scale prediction results is performed to obtain the final sea surface temperature prediction field.
