Method for constructing marine intelligent prediction large model

Abstract

The application provides a marine intelligent prediction large model construction method, relates to the field of marine prediction, and specifically comprises the following steps: acquiring multi-source marine observation data, fusing the multi-source observation data, and assimilating numerical modes to construct a high-resolution marine analysis data set that is subjected to quality control, space-time registration, standardization and data set division processing; constructing a basic prediction model; using the basic prediction model to perform multi-scale fusion on frequency domain enhanced features and spatial local features; running the trained basic prediction model in a set regional range; the output result of the basic prediction model is restored to an original physical value through inverse standardization; the rolling output of the basic prediction model is compared with observation data or high-resolution mode results through a correction module, errors are learned, and a correction amount is output; the correction amount is superimposed on the original prediction result to obtain a prediction field. The technical scheme of the application overcomes the problem in the prior art that a marine prediction model cannot meet the needs of complex application scenarios.

Description

Technical Field

[0001] This invention relates to the field of marine forecasting, and more specifically to a method for constructing a large-scale intelligent marine forecasting model. Background Technology

[0002] High-precision ocean forecasting has significant applications in national defense, shipping safety, fisheries management, and offshore energy development. With increasing climate variability and the frequency of extreme events, rapid changes in ocean conditions directly impact social security. Therefore, developing forecasting methods capable of characterizing ocean evolution at higher spatial and temporal resolutions is of paramount importance.

[0003] Current ocean forecasting primarily relies on numerical simulation methods. Numerical models based on ocean dynamics equations have played a crucial role in global circulation and climate research after long-term development. However, these models face bottlenecks in improving spatiotemporal resolution. Achieving eddy-resolution simulations at small to medium scales typically requires fine grids on the order of 2–5 kilometers, which introduces extremely high computational costs and parameterization uncertainties. Even with advanced high-performance computing platforms, it remains difficult to achieve both high accuracy and timeliness in operational forecasts over large areas and long time periods.

[0004] In recent years, the rapid development of artificial intelligence has driven the exploration and application of large-scale ocean forecasting models. Existing research shows that deep learning-based methods can improve the prediction of ocean temperature, salinity, and current velocity to some extent, outperforming traditional numerical models in computational efficiency and accuracy of some indicators, and demonstrating potential in high-resolution and rapid forecasting. However, existing large-scale AI ocean models still have significant shortcomings. On the one hand, limited by the sources of training data, most models rely on globally available reanalysis data as a reference field, with spatial resolution typically around 1 / 12° and temporal resolution mostly at the daily level, with only a few achieving the 6-hour level. This level is insufficient to resolve typical small- and medium-scale dynamic processes such as eddies, fronts, and tidal jets, and cannot meet the detailed requirements of rapidly evolving environments. On the other hand, existing methods are insufficient in the fusion and utilization of multi-source observations and regional high-resolution simulation data, lacking a systematic solution with stability and operational capabilities. In summary, globally, the operational exploration of regionalized, ultra-high-resolution large-scale AI ocean models is still in its early stages.

[0005] Therefore, there is a need for a method for constructing a large-scale intelligent marine forecasting model that can operate stably, make full use of multi-source information, and meet the needs of complex application scenarios. Summary of the Invention

[0006] The main objective of this invention is to provide a method for constructing a large-scale intelligent marine forecasting model, so as to solve the problem that existing marine forecasting models cannot meet the needs of complex application scenarios.

[0007] To achieve the above objectives, this invention provides a method for constructing a large-scale intelligent ocean forecasting model, which specifically includes the following steps:

[0008] S1. Acquire multi-source ocean observation data. Through multi-source observation data fusion and numerical model assimilation, construct a high-resolution ocean analysis dataset that has undergone quality control, spatiotemporal registration, standardization, and dataset partitioning.

[0009] S2, construct the basic prediction model, including: encoder, frequency domain enhancement unit and decoder, and use the basic prediction model to perform multi-scale fusion of frequency domain enhanced features and spatial local features.

[0010] S3 runs the trained base prediction model within a set area, and the output of the base prediction model is restored to the original physical quantity value through inverse standardization.

[0011] S4 compares the rolling output of the basic prediction model with the observation data or high-resolution model results through the three-dimensional convolutional residual correction module, learns the error and outputs the correction amount, which is then superimposed with the original prediction result to obtain the prediction field.

[0012] Step S2 specifically includes the following steps:

[0013] S2.1, Construct the basic prediction model, which includes: an input layer, an encoder, a frequency domain enhancement unit, a decoder, and an output reconstruction layer connected in sequence.

[0014] S2.2, the composite loss function used in the training of the basic prediction model. for:

[0015] ;

[0016] in, For the reference truth tensor, To predict tensors, express Norm, For an effective mask matrix, This represents element-wise multiplication. For the total variational regularization term , , , The weights for each loss are determined.

[0017] S2.3 employs an adaptive gradient-based optimization method, combined with a learning rate scheduling mechanism, to optimize the parameters; the learning rate is expressed as:

[0018] ;

[0019] in, This represents the learning rate in the t-th iteration. The initial learning rate, This is a scheduling function that varies with the number of training steps.

[0020] Furthermore, step S1 specifically includes the following steps:

[0021] S1.1 Acquire multi-source ocean observation data through underwater mooring observation data, buoy observation data and satellite remote sensing products. Multi-source ocean observation data includes: sea temperature, salinity, current velocity and sea surface height.

[0022] S1.2, remove missing points and outliers from the observation data, perform spatiotemporal registration processing on the observation data, and match the observation time to the model integration time window in the time dimension.

[0023] S1.3, simulates the ocean using a regional high-resolution numerical model at a kilometer-scale grid to obtain numerical simulation data; the observation data processed in step S1.2 is then assimilated into the numerical model. The assimilation method employs three-dimensional variational, four-dimensional variational, or ensemble Kalman filtering. By constructing observation operators, the model data is mapped to the observation space, and the optimal analysis field is formed by combining the observation error covariance and the background error covariance, thus achieving an organic combination of observation data and numerical simulation data.

[0024] S1.4 Standardize the assimilated data to eliminate dimensional differences between different physical quantities and ensure that each physical quantity is within a uniform numerical range during training.

[0025] S1.5 stores the standardized data in a unified format and divides it into training, validation, and test sets according to the time series.

[0026] Furthermore, step S2.1 specifically includes the following steps:

[0027] S2.1.1, Input sequence of the input layer for:

[0028] ;

[0029] in, Indicates the number of input channels. Indicates the number of vertical floors. Indicates the number of latitudinal grids. This indicates the number of meridional grid cells.

[0030] S2.1.2, the feature representation is obtained through the encoder. The frequency domain enhancement operation is represented as:

[0031] ;

[0032] in, Represents a three-dimensional Fourier transform. Indicates the inverse Fourier transform. Features enhanced in the frequency domain It is a frequency domain operator.

[0033] S2.1.3, in the decoding stage, symmetrical padding or interpolation is used to achieve feature alignment. The aligned decoded features are then spliced ​​with the corresponding scale of frequency domain enhancement features in the channel dimension, thereby effectively fusing spectral and spatiotemporal features. The fusion process is represented as follows:

[0034] ;

[0035] in, Indicates the decoding path is at the . Features of the layer The frequency domain enhancement unit is represented in the first... Features of the layer This indicates a channel splicing operation. This represents a combined mapping of convolution, normalization, and nonlinear activation. These are the features after fusion.

[0036] Furthermore, step S3 specifically includes the following steps:

[0037] S3.1, the ocean state for future periods is successively extrapolated using a rolling time step method. The latest multi-source observation data and numerical simulation data are used as inputs and formatted and standardized through the preprocessing process of steps S1.2 to S1.4.

[0038] S3.2, run the trained base prediction model within the defined area.

[0039] S3.3, the output of the basic prediction model is restored to the original physical quantity values ​​through inverse standardization. The output multivariate spatiotemporal prediction results include: temperature, salinity, or flow velocity. The inverse standardization formula is:

[0040] ;

[0041] in, The standardized value output by the basic prediction model. and They are respectively The minimum and maximum values, The value of the physical quantity after restoration.

[0042] Furthermore, step S4 specifically includes the following steps:

[0043] S4.1 uses the rolling output of the basic prediction model as the input of the three-dimensional convolutional residual correction module, and uses the high-resolution regional simulation results of assimilated multi-source observation data as reference labels to form an input-label pair.

[0044] S4.2 performs variable-wise and depth-wise normalization or standardization on the training samples, and introduces a spatial depth validity mask to shield land grids, invalid layers or missing measurement points, so that the loss is calculated only on the effective sea area or layer.

[0045] S4.3, the optimizer employs an adaptive weight decay method, combined with a segmented learning rate scheduling strategy; a random seed is set to ensure repeatability; the loss function is set as follows:

[0046] ;

[0047] in, and These represent the predicted value and the actual value, respectively. This represents the set of valid indices marked in the mask matrix. This is the root mean square error.

[0048] S4.4 performs inverse standardization on a variable-by-variable basis using the statistics at the time of training, restores the scale to physical quantities, and outputs it as a business product.

[0049] Further, the 3D convolutional residual correction module in step S4.1 includes: an input layer, a 3D convolutional layer, residual units, and an output layer connected in sequence; the input channels are first subjected to convolutional dimensionality increase or decrease; a main branch and a shortcut branch are set in each residual unit, and a 1×1×1 convolution is used to match the dimension when the number of channels is inconsistent, and the output layer regresses to the target channel using 3D convolution; wherein, the main branch contains a layer of 3D convolution to extract spatial features, followed by normalization operation and nonlinear activation function, and then through a channel transformation layer to realize the recombination of information; the shortcut branch directly passes the input features to the output layer; finally, the output of the main branch and the shortcut branch are added at each point to obtain the result of the residual connection.

[0050] The present invention has the following beneficial effects:

[0051] This invention addresses the complexity and increasing demands for precision in regional ocean forecasting by proposing an intelligent forecasting method that integrates multi-source observations with high-resolution numerical model data. By assimilating multi-source observation data, including those from underwater moorings, buoys, shipboard surveys, and satellite remote sensing, into the regional ocean numerical model, a more refined and realistic three-dimensional field is obtained, significantly improving the data foundation quality for regional ocean prediction. In terms of model structure, a frequency domain enhancement unit is introduced into the traditional encoder-decoder framework, enabling multi-scale fusion of long-range frequency domain dependencies and spatial local features, thus better characterizing small- to medium-scale dynamic processes such as eddies and fronts in the regional ocean. During training, a composite loss function balancing overall accuracy and physical plausibility is designed, combined with mask constraints and total variational constraints, ensuring the continuity and stability of prediction results at the regional scale. In post-processing, an independent three-dimensional convolutional residual correction module is used to effectively correct systematic errors generated in regional ocean rolling predictions, reducing long-term cumulative bias. Furthermore, in terms of engineering deployment, the basic prediction model is decoupled from the three-dimensional convolutional residual correction module, achieving modular, distributed parallel, and mixed-precision computation, ensuring that the needs of refined operational regional ocean prediction can be met even under limited computing power. Attached Figure Description

[0052] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0053] Figure 1 A flowchart of a method for constructing a large-scale intelligent marine forecasting model according to the present invention is shown.

[0054] Figure 2 The image shows the effect of current velocity prediction using a low-resolution large ocean model in the existing technology.

[0055] Figure 3 The diagram shows the current velocity forecasting effect of a high-resolution regional ocean large model constructed using the ocean intelligent forecasting large model construction method provided by the present invention. Detailed Implementation

[0056] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0057] like Figure 1The method for constructing a large-scale intelligent ocean forecasting model, as shown, specifically includes the following steps:

[0058] S1. Acquire multi-source ocean observation data. Through multi-source observation data fusion and numerical model assimilation, construct a high-resolution ocean analysis dataset that has undergone quality control, spatiotemporal registration, standardization, and dataset partitioning.

[0059] S2, construct the basic prediction model, including: encoder, frequency domain enhancement unit and decoder, and use the basic prediction model to perform multi-scale fusion of frequency domain enhanced features and spatial local features.

[0060] S3 runs the trained base prediction model within a set area, and the output of the base prediction model is restored to the original physical quantity value through inverse standardization.

[0061] S4 compares the rolling output of the basic prediction model with the observation data or high-resolution model results through the three-dimensional convolutional residual correction module, learns the error and outputs the correction amount, which is then superimposed with the original prediction result to obtain the prediction field.

[0062] Specifically, step S1 includes the following steps:

[0063] S1.1 Acquire multi-source ocean observation data through underwater mooring observation data, buoy observation data and satellite remote sensing products. Multi-source ocean observation data includes: sea temperature, salinity, current velocity and sea surface height.

[0064] S1.2, remove missing points and outliers from the observation data, perform spatiotemporal registration processing on the observation data, and match the observation time to the model integration time window in the time dimension.

[0065] S1.3, use the regional high-resolution ocean numerical model to perform simulation calculations at the kilometer-scale grid to obtain numerical simulation data; and assimilate the observation data processed in step S1.2 into the numerical model. The assimilation method adopts three-dimensional variational, four-dimensional variational or ensemble Kalman filtering. By constructing observation operators, the model data is mapped to the observation space, and the optimal analysis field is formed by combining the observation error covariance and the background error covariance.

[0066] S1.4 Standardize the assimilated data to eliminate dimensional differences between different physical quantities and ensure that each physical quantity is within a uniform numerical range during training. A common approach is to linearly scale each variable according to its historical minimum and maximum values, so that its distribution falls within the interval [-1, 1].

[0067] ;

[0068] in, Represents the original physical value. and These are the minimum and maximum values ​​of the variable, respectively. To standardize the results, the standardized data is stored in a uniform format and divided into training, validation, and test sets according to the time series, with a ratio of 7:2:1, to ensure the effectiveness of model training and the generalization ability of prediction results. Standardization methods can also include zero-mean standardization, logarithmic transformation, or other normalization methods suitable for physical variables.

[0069] S1.5 stores the standardized data in a unified format and divides it into training, validation, and test sets according to the time series.

[0070] Specifically, this invention constructs an encoder-decoder deep neural network on multivariable three-dimensional spatiotemporal data of regional oceans, and introduces a frequency domain neural operator branch into the multi-scale feature pathway to enhance the representation ability of small- and medium-scale dynamic processes. The basic prediction model consists of a multi-level downsampling encoder, a frequency domain enhancement unit, a multi-level upsampling decoder, and an output reconstruction layer. The encoder extracts features from the input three-dimensional tensor step by step to obtain multi-scale features from local to global; the frequency domain enhancement unit achieves long-range dependency modeling by introducing Fourier or correlation transforms at each scale; the decoder gradually restores spatial resolution through step-by-step upsampling and stitching fusion; the output layer completes the multi-channel reconstruction of variables and refines the results. Step S2 specifically includes the following steps:

[0071] S2.1, Construct the basic prediction model, which includes: an input layer, an encoder, a frequency domain enhancement unit, a decoder, and an output reconstruction layer connected in sequence. The encoder extracts spatial and vertical features layer by layer, the frequency domain enhancement branch extracts long-range dependent features in the frequency domain, and the decoder restores the spatial resolution step by step and fuses it with the encoder features to output a high-resolution result with multiple variables.

[0072] S2.2, the composite loss function used in the training of the basic prediction model. for:

[0073] ;

[0074] in, For the reference truth tensor, To predict tensors, express norm (e.g.) Indicates L1 loss. (representing L2 loss) For an effective mask matrix, This represents element-wise multiplication. This is the total variational regularization term, used to constrain the spatial smoothness of the predicted field, suppress local oscillations, and enhance physical consistency. , , , The weights for each loss term are defined. This invention employs a composite loss function to improve prediction performance. During training, a composite loss strategy is used to balance overall prediction accuracy, local physical consistency, and spatial structure constraints. Dynamic learning rate scheduling, mixed precision calculation, gradient pruning, and early stopping mechanisms are combined during training to improve convergence efficiency and prevent overfitting.

[0075] S2.3, this invention employs a multi-level optimization strategy during the training of the basic prediction model to simultaneously ensure efficiency, stability, and reproducibility. Firstly, in terms of hardware utilization, the basic prediction model supports distributed data parallel training, with different computing nodes achieving gradient synchronization through a communication backend, thereby achieving near-linear acceleration in multi-GPU or multi-node environments. Simultaneously, mixed-precision computation is introduced, significantly reducing GPU memory usage and improving training speed while maintaining numerical accuracy.

[0076] For parameter optimization, adaptive gradient-based optimization methods (such as AdamW or its equivalent variants) are used, combined with a learning rate scheduling mechanism to optimize parameters and improve convergence performance. Learning rate scheduling can take various forms, such as piecewise decay, cosine annealing, or loop restart, and is generally expressed as follows:

[0077] ;

[0078] in, This represents the learning rate in the t-th iteration. The initial learning rate, The scheduling function varies with the number of training steps and can be an exponentially decaying function, a piecewise constant function, or a cosine function, etc. This mechanism can maintain a large learning rate in the early stages of training to accelerate convergence, and gradually reduce the learning rate in the later stages of training to stabilize the model.

[0079] To improve training stability, gradient clipping, regularization terms, and total variational constraints are introduced to prevent gradient explosion or overfitting under complex multi-scale data. An early stopping mechanism is also designed to automatically terminate training when validation set performance fails to improve within several rounds, reducing unnecessary computation. Basic prediction model parameters are periodically saved as checkpoint files during training for easy recovery from interruptions and subsequent business deployment.

[0080] Regarding reproducibility, this invention uses a fixed random number seed and controls the operation mode of parallel operators to repeatedly obtain consistent training results under the same input conditions. This design ensures that the training process and results of the basic prediction model are comparable and consistent across multiple experiments or computing platforms, effectively reducing the risks associated with uncertainty.

[0081] Specifically, step S2.1 includes the following steps:

[0082] S2.1.1, Input sequence of the input layer for:

[0083] ;

[0084] in, Indicates the number of input channels. Indicates the number of vertical floors. Indicates the number of latitudinal grids. This indicates the number of meridional grid cells.

[0085] S2.1.2, the feature representation is obtained through the encoder. The frequency domain enhancement operation is represented as:

[0086] ;

[0087] in, Represents a three-dimensional Fourier transform. Indicates the inverse Fourier transform. Features enhanced in the frequency domain It is a frequency domain operator used to weight, suppress, or sparsify different frequency components in the frequency domain to highlight important features related to ocean dynamic processes. A fixed threshold mask can be used to retain the main frequency components and filter out the secondary components by means of spectral sparsity methods; It can also be a parameterized learnable operator whose weights are updated during training, thereby dynamically adjusting the relative importance of frequency components.

[0088] S2.1.3, to avoid size mismatch during upsampling and downsampling, symmetrical padding or interpolation is used to align features during the decoding stage. The aligned decoded features are then concatenated with the corresponding frequency-domain enhanced features in the channel dimension, effectively fusing spectral and spatiotemporal features. This mechanism ensures the synergistic enhancement of long-range dependencies and local details in the prediction results, enabling more accurate characterization of complex dynamic structures such as fronts and vortices. The fusion process is represented as follows:

[0089] ;

[0090] in, Indicates the decoding path is at the . Features of the layer The output of the convolutional layers in the decoding stage mainly contains spatial local features. The frequency domain enhancement unit is represented in the first... Features of the layer This indicates a channel splicing operation. This represents a combined mapping of convolution, normalization, and nonlinear activation. These are the features after fusion.

[0091] During the decoding stage, the decoding features are fused with the frequency domain enhancement features through alignment and splicing operations, thereby achieving a coordinated expression of long-range dependencies and local details.

[0092] The encoder and decoder can employ different 3D convolution or attention structures. The frequency domain enhancement unit can be replaced with a Fourier neural operator, sparse frequency domain convolution, or low-rank attention module. Upsampling and downsampling can be performed using interpolation or transposed convolution, and the loss function can also be mean square error, Huber loss, or other composite forms.

[0093] Specifically, after the basic prediction model is trained, it is used to predict the target area. First, the latest multi-source observation data and numerical simulation data are used as input, and formatted and standardized through the same preprocessing procedure as in the training phase. Then, the basic prediction model is run within the defined area, outputting multivariate spatiotemporal prediction results including temperature, salinity, and current velocity. The generated prediction results have high spatial and temporal resolution, refined to kilometer-level spatial scales and hour-level temporal scales, meeting the needs of refined regional forecasts. To ensure the temporal continuity of the prediction results, a rolling time step approach is used to progressively extrapolate the ocean state for future periods.

[0094] Step S3 specifically includes the following steps:

[0095] S3.1, the ocean state for future periods is successively extrapolated using a rolling time step method. The latest multi-source observation data and numerical simulation data are used as inputs and formatted and standardized through the preprocessing process of steps S1.2 to S1.4.

[0096] S3.2, run the trained base prediction model within the defined area.

[0097] S3.3, the output of the basic prediction model is restored to its original physical quantity values ​​through inverse standardization to ensure direct usability for business applications. This process corresponds one-to-one with the standardization method used in the training phase. The output multivariate spatiotemporal prediction results include: temperature, salinity, or flow velocity. Taking a linear interval scaled to [-1, 1] as an example, the inverse standardization formula is:

[0098] ;

[0099] in, The standardized value output by the basic prediction model. and They are respectively The minimum and maximum values, The value of the physical quantity after restoration.

[0100] Specifically, after completing the rolling prediction of the basic prediction model, an independent 3D convolutional residual correction module is introduced to post-process and optimize the prediction results, thereby reducing the accumulation of errors caused by long-term rolling and improving spatiotemporal consistency. The 3D convolutional residual correction module compares the output of the basic prediction model with the observation data or high-resolution model results, learns the error and outputs the correction amount, which is then superimposed with the original prediction result to obtain a more stable prediction field.

[0101] Step S4 specifically includes the following steps:

[0102] S4.1 uses the rolling output of the basic prediction model as the input of the three-dimensional convolutional residual correction module, and uses the high-resolution regional simulation results of assimilated multi-source observation data as reference labels to form an input-label pair.

[0103] S4.2 performs variable-wise and depth-wise normalization or standardization on the training samples, and introduces a spatial depth validity mask to shield land grids, invalid layers or missing measurement points, so that the loss is calculated only on the effective sea area or layer.

[0104] S4.3 decouples the basic prediction model and the 3D convolutional residual correction module and deploys them in a multi-GPU environment, supporting distributed parallel training and batch inference. This facilitates embedding into operational pipelines and meets the requirements for high-resolution, refined regional marine forecasting. Distributed data parallelism is used for training in the multi-GPU environment, combining mixed-precision acceleration and memory optimization. The optimizer employs an adaptive weight decay method, coupled with a segmented learning rate scheduling strategy. A random seed is set to ensure repeatability. The loss function primarily uses mask-weighted root mean square error (RMSE) / mean absolute error (MAE) (calculated only within effective masks) to avoid interference from invalid or outlier values ​​on gradient updates. The loss function is set as follows:

[0105] ;

[0106] ;

[0107] in, and These represent the predicted value and the actual value, respectively. This represents the set of valid indices marked in the mask matrix. The root mean square error, The mean absolute error is used. The training and validation process is divided into rolling segments based on time samples to ensure consistency with the business inference process.

[0108] S4.4 During business inference, the latest rolling prediction of the basic prediction model is used as the input of the three-dimensional convolutional residual correction module. The error estimate or correction value output by the three-dimensional convolutional residual correction module is combined with the original field to obtain the correction result. Then, the variable-wise inverse standardization is performed according to the statistics at the time of training to restore it to the physical quantity scale and output as the business product.

[0109] The 3D convolutional residual correction module is deployed as a post-processing plugin, decoupled from the basic prediction model. It interfaces with the rolling prediction cache, and inference computation is completed in real time on single or multiple GPUs. The module supports batch / online modes and data parallel expansion, making it easy to integrate into existing business pipelines.

[0110] The above correction process uses a 3D convolutional residual structure to uniformly model spatial-vertical correlation at the voxel level (longitude × latitude × depth). Combined with mask constraints and residual learning, it significantly reduces the systematic bias and time accumulation error of rolling prediction, and improves the amplitude and phase consistency of small and medium-scale structures such as fronts and vortices. Distributed hybrid precision training and lightweight 3D convolutional design ensure low additional computing power and business availability, and can be used plug and play without modifying the basic model.

[0111] Specifically, the 3D convolutional residual correction module in step S4.1 includes: an input layer, a 3D convolutional layer, residual units, and an output layer connected in sequence. The input channels are first subjected to convolutional upscaling or downscaling. Within each residual unit, a main branch convolution stack and a shortcut branch are set up. When the number of channels is inconsistent, a 1×1×1 convolution is used to match the dimension. The output layer regresses to the target channel using 3D convolution. The main branch contains a 3D convolution layer for extracting spatial features, followed by normalization and a non-linear activation function to enhance feature representation. Then, a channel transformation layer is used to recombine the information. The shortcut branch directly passes the input features to the output layer to ensure no information loss. Finally, the main branch output and the shortcut branch output are added point-by-point to obtain the residual connection result.

[0112] The 3D convolutional residual correction module uses 3D convolutional residual units as trunk feature extractors. The input channels are first subjected to convolutional dimensionality increase / decrease. Within each residual unit, main branch convolutional stacking and shortcut branches are set. When the number of channels is inconsistent, 1×1×1 convolutions are used to match the dimension. The output layer regresses to the target channel using 3D convolution. The network learns the "prediction error field" in a residual manner. During inference, the correction amount is superimposed with the original prediction to obtain the final result. Optionally, layer normalization can be introduced under a channel-first layout to stabilize training.

[0113] Taking flow velocity as an example, the high-resolution regional ocean artificial intelligence model proposed in this invention can more finely characterize the spatial distribution characteristics of regional ocean flow velocity and capture a more complete ocean dynamic process compared with the existing low-resolution ocean model. Figure 2The results of a low-resolution large-scale ocean model are presented, showing a relatively smooth overall flow field but insufficient analysis of small- and medium-scale dynamic processes. In contrast, Figure 3 The image shows the prediction results of this invention at high spatiotemporal resolution, which can clearly present more detailed dynamic structures, especially reflecting small- and medium-scale changes such as tidal currents more accurately. This indicates that the method of this invention breaks through the resolution bottleneck of existing large models that rely on publicly available reanalysis data. While maintaining rapid prediction capabilities, it significantly improves the resolution of small- and medium-scale ocean dynamic processes, thus providing higher-precision technical support for applications such as regional ocean defense security, shipping safety, fisheries management, and offshore energy development.

[0114] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for constructing a marine intelligent prediction large model, characterized in that, Specifically comprising the following steps: S1, obtaining multi-source ocean observation data, constructing high-resolution ocean analysis data sets through multi-source observation data fusion and numerical model assimilation, which are quality controlled, time and space registered, standardized and data set divided; S1.1, obtaining multi-source ocean observation data through mooring observation data, buoy observation data and satellite remote sensing products, the multi-source ocean observation data including sea temperature, salinity, flow velocity and sea surface height; S2, constructing a basic prediction model, including an encoder, a frequency domain enhancement unit and a decoder, and using the basic prediction model to perform multi-scale fusion on the frequency domain enhanced features and spatial local features; S3, running the trained basic prediction model in a set region, and the output result of the basic prediction model is restored to the original physical quantity value through inverse standardization; S4, comparing the rolling output of the basic prediction model with the observation data or the high-resolution mode result through a three-dimensional convolution residual correction module, learning the error and outputting a correction quantity, superimposing the correction quantity with the original prediction result to obtain a prediction field; The three-dimensional convolution residual correction module is deployed in a decoupled manner with the basic prediction model as a post-processing plug-in, interfaces with the rolling prediction cache, and inference calculation is completed in real time on multiple GPUs; training is performed in a distributed data parallel manner under a multi-GPU environment, combined with mixed precision acceleration and memory optimization; Step S2 specifically comprises the following steps: S2.1, constructing a basic prediction model, the basic prediction model comprising an input layer, an encoder, a frequency domain enhancement unit, a decoder and an output reconstruction layer connected in sequence; S2.2, the composite loss function employed in the base prediction model training process is: ； wherein, is a reference true value tensor, is a predicted tensor, denotes a norm, is an effective mask matrix, denotes an element-wise multiplication, is a total variation regular term, , , , are loss weights; S2.3, using an adaptive gradient-based optimization method to optimize parameters combined with a learning rate scheduling mechanism; the learning rate is expressed as: ； wherein, denotes the learning rate for the t-th iteration, is an initial learning rate, is a schedule function that varies with the training step number.

2. The method according to claim 1, wherein, Step S1 further comprises the following steps: S1.2, removing missing points and outliers of the observation data, and performing time and space registration processing on the observation data, matching the observation time to the mode integration time window in the time dimension; S1.3, performing simulation calculation on a kilometer-level grid scale using a regional high-resolution ocean numerical model to obtain numerical simulation data; The observation data processed by step S1.2 is assimilated into the numerical model, and the assimilation method is three-dimensional variation, four-dimensional variation or ensemble Kalman filter; the model data is mapped to the observation space by constructing an observation operator, and the optimal analysis field is formed by combining the observation error covariance and the background error covariance; S1.4, standardizing the assimilated data to eliminate the dimensional differences between different physical quantities and ensure that each physical quantity is in a unified numerical range during training; S1.5, storing the standardized data in a unified format, and dividing them into a training set, a validation set and a test set according to the time sequence.

3. The method of claim 1, wherein the method comprises: Step S2.1 specifically comprises the following steps: S2.1.1, input layer input sequence is: ； wherein, represents the number of input variable channels, represents the number of vertical layers, represents the number of weft grid numbers, represents the number of warp grid numbers; S2.1.2, encoder-derived feature representation The frequency domain enhancement operation is represented as: ； wherein, denotes a three-dimensional Fourier transform, denotes an inverse Fourier transform, is the frequency domain enhanced feature, is a frequency domain operator; S2.1.3, aligning the features in the decoding stage using symmetric padding or interpolation, concatenating the aligned decoding features and the corresponding scale frequency domain enhanced features in the channel dimension, and then effectively fusing the spectral domain and spatial features, the fusion process being expressed as: ； wherein, represents a decoding path at the layer’s characteristics, represents a frequency domain enhancement unit at the layer’s characteristics, represents a channel concatenation operation, represents a combined mapping of convolution, normalization, and non-linear activation, is the fused feature.

4. The method according to claim 2, wherein, Step S3 specifically comprises the following steps: S3.1, sequentially deduce the ocean state of the future period in a rolling time step manner, taking the latest multi-source observation data and numerical simulation data as input, and formatting and standardizing the data through the preprocessing procedures of steps S1.2-S1.4; S3.2, run the trained basic prediction model in the set regional range; S3.3, the output result of the basic prediction model is restored to the original physical value through inverse standardization, and the multivariate spatiotemporal prediction result includes temperature, salinity or flow rate, and the inverse standardization formula is: ； wherein, the standardized value output by the base prediction model, and the minimum and maximum values of respectively, the restored physical quantity value.

5. The method according to claim 1, wherein, Step S4 specifically includes the following steps: S4.1, the rolling output of the basic prediction model is taken as the input of the three-dimensional convolution residual correction module, and the high-resolution regional simulation result of the multi-source observation data is taken as the reference label to form the input-label pair; S4.2, the training samples are normalized or standardized variable by variable and depth by depth, and spatial depth effectiveness masks are introduced to shield land grids, invalid layers or missing points, so that the loss is calculated only in the effective sea area or horizon; S4.3, the optimizer adopts an adaptive method with weight decay, and cooperates with a segmented learning rate scheduling strategy; Set a random seed to ensure repeatability; the loss function is set as: ； ； wherein, and pred and true represent the predicted and true values, respectively, represents the union of indices marked as valid in the mask matrix, is the root mean square error, is the mean absolute error; S4.4, normalize the variables according to the statistics during training, restore them to the physical quantity scale and output them as business products.

6. The method according to claim 5, wherein, The three-dimensional convolution residual correction module in step S4.1 includes an input layer, a three-dimensional convolution layer, a residual unit and an output layer connected in turn; the input channel is first subjected to convolution dimensionality increase or reduction; a main branch and a shortcut branch are set in each residual unit, and a 1x1x1 convolution is used to match the dimensions when the channel numbers are inconsistent; the output layer is regressed to the target channel by three-dimensional convolution; wherein the main branch contains a three-dimensional convolution layer for extracting spatial features, followed by a normalization operation and a nonlinear activation function, and then passes through a channel transformation layer to realize the recombination of information; the shortcut branch directly transmits the input features to the output layer; finally, the main branch output and the shortcut branch are added at each point position to obtain the result of residual connection.

Images