Marine intelligent forecasting large model construction method

Abstract

The invention provides an ocean intelligent forecasting large model construction method, and relates to the field of ocean forecasting, and the method specifically comprises the following steps: obtaining multi-source ocean observation data, and constructing a high-resolution ocean analysis data set which is subjected to quality control, space-time registration, standardization and data set division processing through multi-source observation data fusion and numerical mode assimilation; constructing a basic prediction model, and performing multi-scale fusion on the frequency domain enhanced features and the spatial local features by using the basic prediction model; the trained basic prediction model is operated in a set area range, and an output result of the basic prediction model is recovered to an original physical quantity value through inverse standardization; and comparing the rolling output of the basic prediction model with observation data or a high-resolution mode result through a correction module, learning an error, outputting a correction quantity, and superposing the correction quantity with an original prediction result to obtain a prediction field. According to the technical scheme, the problem that the ocean forecasting model in the prior art cannot meet the requirement of a complex application scene is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of marine prediction, in particular to a marine intelligent prediction large model construction method. BACKGROUND

[0002] High-precision marine prediction has important application value in national defense security, shipping safety, fishery management, and offshore energy development. With the increasing frequency of climate variability and extreme events, the rapid changes in marine conditions have a direct impact on social security. Therefore, it is of great significance to develop a prediction method that can depict the evolution of the ocean at a higher spatial and temporal resolution.

[0003] Existing marine prediction mainly relies on numerical simulation methods. Numerical models based on ocean dynamics equations have played an important role in global circulation and climate research. However, such models have bottlenecks in improving spatial and temporal resolution. Achieving mesoscale eddy resolution simulation usually requires a fine grid of 2-5 kilometers, which brings high computational overhead and parameterization uncertainty. Even with advanced high-performance computing platforms, it is still difficult to achieve high-precision and high-timeliness operational prediction in a large range and long period.

[0004] In recent years, the rapid development of artificial intelligence has promoted the exploration and application of marine prediction large models. Existing research has shown that deep learning-based methods can improve the prediction of ocean temperature, salinity, and flow rate to some extent, and are superior to traditional numerical models in terms of computational efficiency and accuracy of some indicators, and have shown potential in high-resolution and rapid prediction. However, existing artificial intelligence marine large models still have obvious shortcomings. On the one hand, limited by the source of training data, most models rely on global public reanalysis data as the benchmark field, with a spatial resolution of about 1 / 12° and a time resolution of mostly daily, only a few works can reach the level of 6 hours. This level is still insufficient to analyze typical mesoscale and small-scale dynamic processes such as mesoscale eddies, fronts, and tidal jets, and it is difficult to meet the fine needs 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, and lack systematic solutions with stability and operational running capability. In summary, in the global range, the operational exploration of regionalized and ultra-high-resolution marine artificial intelligence large models is still in its infancy.

[0005] Therefore, there is a need for a marine intelligent prediction large model construction method that can be stably operated, fully utilize multi-source information, and meet the needs of complex application scenarios. SUMMARY

[0006] The main purpose of the present application is to provide a marine intelligent prediction large model construction method to solve the problem that the existing marine prediction model cannot meet the needs of complex application scenarios.

[0007] To achieve the above object, the application provides a marine intelligent prediction large model construction method, which specifically comprises the following steps: S1, obtaining multi-source marine observation data, constructing high-resolution marine analysis data set through multi-source observation data fusion and numerical model assimilation, quality control, space-time registration, standardization and data set division processing.

[0008] 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.

[0009] S3, running the trained basic prediction model in a set area range, and the output result of the basic prediction model is restored to the original physical quantity value through inverse standardization.

[0010] 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 the correction quantity, superimposing it with the original prediction result, and obtaining the prediction field.

[0011] 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.

[0012] S2.2, the composite loss function used in the training process of the basic prediction model is: Wherein, is a reference true value tensor, is a prediction tensor, represents norm, is an effective mask matrix, represents element-wise multiplication, is a total variation regularization term, , , , is the loss weight of each term.

[0013] S2.3, using an adaptive gradient type optimization method, combining a learning rate scheduling mechanism to optimize parameters; the learning rate is expressed as: Wherein, represents the learning rate of the tth iteration, is the initial learning rate, is a scheduling function that changes with the training step number.​​​

[0014] Further, step S1 specifically comprises the following steps: S1.1, obtaining multi-source ocean observation data including sea temperature, salinity, flow velocity and sea surface height through glider observation data, buoy observation data and satellite remote sensing products.

[0015] S1.2, removing missing points and outliers of observation data, and performing spatio-temporal registration processing on the observation data, matching the observation time to the model integration time window in the time dimension.

[0016] S1.3, performing simulation calculation at a kilometer-level grid scale using a regional high-resolution ocean numerical model to obtain numerical simulation data; and assimilating the observation data processed in step S1.2 into the numerical model, using a three-dimensional variation, four-dimensional variation or ensemble Kalman filter, mapping the model data to the observation space by constructing an observation operator, and combining the observation error covariance and the background error covariance to form an optimal analysis field, realizing the organic combination of observation data and numerical simulation data.

[0017] 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 the training process.

[0018] 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.

[0019] Further, step S2.1 specifically comprises the following steps: S2.1.1, inputting a sequence into the input layer : ; wherein, denotes the number of input variable channels, denotes the number of vertical layers, denotes the number of zonal grid numbers, denotes the number of meridional grid numbers.

[0020] S2.1.2, obtaining a feature representation through an encoder , and a frequency domain enhancement operation is represented as: ; wherein, denotes a three-dimensional Fourier transform, denotes an inverse Fourier transform, is a feature after frequency domain enhancement, is a frequency domain operator.

[0021] 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: ; 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.

[0022] Furthermore, step S3 specifically includes the following steps: 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.

[0023] S3.2, run the trained base prediction model within the defined area.

[0024] 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: ; 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.

[0025] Furthermore, step S4 specifically includes the following steps: 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.

[0026] 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.

[0027] 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: ; 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.

[0028] S4.4 performs inverse standardization on a variable-by-variable basis based on the statistics used during training, restoring the data to the physical quantity scale and outputting it as a business product.

[0029] 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.

[0030] The present invention has the following beneficial effects: 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

[0031] In order to more clearly illustrate the technical solutions in the specific embodiments or prior art of the present application, the drawings required to be used in the description of the specific embodiments or prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings. In the drawings: Figure 1 A flowchart of a marine intelligent prediction large model construction method of the present application is shown.

[0032] Figure 2 A low-resolution marine large model flow rate prediction effect diagram using prior art is shown.

[0033] Figure 3 A high-resolution regional marine large model flow rate prediction effect diagram constructed by using the marine intelligent prediction large model construction method provided by the present application is shown. DETAILED DESCRIPTION

[0034] The technical solutions of the present application will be described below in conjunction with the drawings, and obviously, the described embodiments are some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0035] As Figure 1 shown, a marine intelligent prediction large model construction method, specifically comprising the following steps: S1, obtaining multi-source marine observation data, constructing high-resolution marine analysis data set through multi-source observation data fusion and numerical model assimilation, quality control, time and space registration, standardization and data set division processing.

[0036] S2, constructing a basic prediction model, including: an encoder, a frequency domain enhancement unit and a decoder, using the basic prediction model to perform multi-scale fusion on the frequency domain enhanced features and spatial local features.

[0037] S3, running the trained basic prediction model in a set regional range, and the output result of the basic prediction model is restored to the original physical quantity value through inverse standardization.

[0038] S4, comparing the rolling output of the basic prediction model with the observation data or the high-resolution model result through a three-dimensional convolution residual correction module, learning the error and outputting the correction quantity, superimposing with the original prediction result to obtain the prediction field.

[0039] Specifically, step S1 specifically comprises the following steps: S1.1, obtain multi-source ocean observation data including sea temperature, salinity, flow velocity and sea surface height through submarine observation data, buoy observation data and satellite remote sensing products.

[0040] S1.2, remove missing points and outliers of observation data, and perform spatiotemporal registration processing on the observation data, and match the observation time to the model integration time window in the time dimension.

[0041] S1.3, use regional high-resolution ocean numerical model to perform simulation calculation at kilometer-level grid scale to obtain numerical simulation data; and the observation data processed by step S1.2 is assimilated into the numerical model, and the assimilation method adopts three-dimensional variation, four-dimensional variation or ensemble Kalman filter, the model data is mapped to the observation space through the construction of observation operator, and the optimal analysis field is formed by combining the observation error covariance and the background error covariance.

[0042] S1.4, standardize the data after assimilation to eliminate the dimensional difference between different physical quantities and ensure that each physical quantity is in a unified numerical range during the training process; a common way is to linearly scale each variable according to its historical minimum and maximum value, so that its distribution falls within the [-1, 1] interval: wherein, represents the original physical value, and are the minimum and maximum values of the variable, is the standardization result. The standardized data is stored in a unified format, and is divided into training set, validation set and test set according to time sequence, and the proportion can be 7:2:1 to ensure the effectiveness of model training and the generalization ability of prediction results. The standardization method can also use zero mean standardization, logarithmic transformation or other normalization methods suitable for physical variables.

[0043] S1.5, store the standardized data in a unified format, and divide it into training set, validation set and test set according to time sequence.

[0044] ​Specifically, the application constructs an encoding-decoding deep neural network on regional ocean multivariate three-dimensional spatiotemporal data, and introduces a frequency domain neural operator branch in a multi-scale feature channel to enhance the representation ability of small and medium scale dynamic processes. The basic prediction model is composed of a multi-level down-sampling encoder, a frequency domain enhancement unit, a multi-level up-sampling 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 realizes long-range dependence modeling by introducing Fourier or correlation transform at each scale; the decoder gradually recovers the spatial resolution by step-by-step up-sampling and splicing fusion; and the output layer completes the multi-channel reconstruction of variables and refines the results. Step S2 specifically includes the following steps:

[0045] 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 turn. The encoder extracts spatial and vertical features layer by layer, the frequency domain enhancement branch extracts long-range dependence features in the frequency domain, the decoder gradually recovers the spatial resolution and fuses with the encoder features, and outputs a high-resolution result of multiple variables.

[0046] S2.2, a composite loss function used in the training process of the basic prediction model is: wherein, is a reference true value tensor, is a predicted tensor, represents norm (such as represents L1 loss, represents L2 loss), is an effective mask matrix, represents element-wise multiplication, is a total variation regularization term, used to constrain the spatial smoothness of the predicted field, suppress local oscillation and enhance physical consistency; , , , are loss weights. The application uses a composite loss function to improve the prediction effect. In the training process, a composite loss strategy is adopted, taking into account the overall prediction accuracy, local physical consistency and spatial structure constraint. In the training, a dynamic learning rate scheduling, mixed precision calculation, gradient clipping and early stopping mechanism are combined to improve the convergence efficiency and prevent overfitting.

[0047] ​S2.3, the application adopts a multi-level optimization strategy in the training process of the basic prediction model to ensure efficiency, stability and reproducibility. First, in terms of hardware utilization, the basic prediction model supports distributed data parallel training, and different computing nodes realize gradient synchronization through the communication backend, so as to obtain approximate linear acceleration effect in a multi-GPU or multi-node environment. At the same time, mixed precision calculation is introduced to significantly reduce the memory occupation and improve the training speed under the premise of maintaining numerical accuracy.

[0048] In terms of parameter optimization, an adaptive gradient-based optimization method (such as AdamW or its equivalent variants) is adopted, combined with a learning rate scheduling mechanism to optimize parameters to improve convergence effect; the learning rate scheduling can adopt segmented attenuation, cosine annealing or cyclic restart, etc., which is generally expressed as: ; wherein, represents the learning rate of the t-th iteration, is the initial learning rate, is a scheduling function that changes with the training step number, which can be an exponential decay function, a segmented constant function or a cosine function, etc. This mechanism can maintain a high learning rate in the early stage of training to speed up convergence, and gradually reduce the learning rate in the later stage of training to stabilize the model.

[0049] In terms of training stability, gradient clipping, regularization terms and total variation constraints are introduced to avoid gradient explosion or overfitting in complex multi-scale data. At the same time, an early stopping mechanism is designed to automatically terminate training when the validation set performance does not improve for several rounds to reduce invalid calculations. The parameters of the basic prediction model are saved as checkpoint files during the training process to facilitate interruption and recovery and subsequent business deployment.

[0050] In terms of reproducibility, the application fixes the random number seed and controls the running mode of parallel operators to enable 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 in multiple experiments or across computing platforms, effectively reducing the risk of uncertainty.

[0051] Specifically, step S2.1 specifically includes the following steps: S2.1.1, input sequence of input layer is: ; wherein, represents the number of input variable channels, represents the number of vertical layers, represents the number of latitude grid numbers, represents the number of longitude grid numbers.

[0052] S2.1.2, the feature representation obtained by the encoder , the frequency domain enhancement operation is represented as: ; wherein, represents a three-dimensional Fourier transform, represents an inverse Fourier transform, is the feature after frequency domain enhancement, is a frequency domain operator, which is used to weight, suppress or sparsify different frequency components in the frequency domain, so as to highlight important features related to ocean dynamic processes. It can be a fixed threshold mask, which retains the main frequency components and filters out the secondary components by means of spectral sparsification method; It can also be a parameterized learnable operator, the weight of which is updated during the training process, so as to dynamically adjust the relative importance of frequency components.

[0053] S2.1.3, in order to avoid the size mismatch caused by the up-sampling and down-sampling processes, the feature alignment is realized by symmetric padding or interpolation method in the decoding stage, the aligned decoding feature and the frequency domain enhanced feature of the corresponding scale are spliced in the channel dimension, and then the effective fusion of spectral domain and spatial-temporal features is realized, which guarantees the synergistic enhancement of long-range dependence and local details in the prediction result, and can more accurately depict complex dynamic structures such as frontal surface and vortex. The fusion process is represented as: ; wherein, represents the feature of the decoding path at the i-th layer, derived from the convolutional layer output in the intermediate decoding stage, mainly containing spatial local features, represents the feature of the frequency domain enhancement unit at the i-th layer, represents the channel splicing operation, represents the combination mapping of convolution, normalization and nonlinear activation, is the fused feature. In the decoding stage, the decoding feature and the frequency domain enhanced feature are fused through alignment and splicing operation, so as to realize the synergistic expression of long-range dependence and local details.

[0054] In the decoding stage, the decoding feature and the frequency domain enhanced feature are fused through alignment and splicing operation, so as to realize the synergistic expression of long-range dependence and local details.

[0055] The encoder and the decoder can adopt different three-dimensional convolution or attention structures, the frequency domain enhancement unit can be replaced by Fourier neural operator, sparse frequency domain convolution or low-rank attention module, the up-sampling and down-sampling can use interpolation or transposed convolution, and the loss function can also adopt mean square error, Huber loss or other composite forms.

[0056] ​Specifically, after the training of the basic prediction model is completed, the target area is predicted using the basic prediction model. First, the latest multi-source observation data and numerical simulation data are inputted, and are formatted and standardized through the preprocessing process consistent with the training stage. Subsequently, the basic prediction model is run in the set regional range, and outputs multivariate spatiotemporal prediction results including temperature, salinity, flow rate, etc. The generated prediction results have high spatial resolution and temporal resolution, and can be refined to a kilometer-level spatial scale and an hour-level time scale, so as to meet the demand of regional refined prediction. In order to ensure the continuity of the prediction results in time, the rolling time step method is adopted to sequentially deduce the marine state in the future period.

[0057] Step S3 specifically includes the following steps: S3.1, the rolling time step method is adopted to sequentially deduce the marine state in the future period, and the latest multi-source observation data and numerical simulation data are inputted, and are formatted and standardized through the preprocessing process of steps S1.2-S1.4.

[0058] S3.2, the trained basic prediction model is run in the set regional range.

[0059] S3.3, the output result of the basic prediction model is restored to the original physical quantity value through inverse standardization, so as to ensure the direct usability of the business end. The process corresponds to the standardization method in the training stage one by one. The output multivariate spatiotemporal prediction results include temperature, salinity or flow rate. Taking linear interval scaling to [-1, 1] as an example, the inverse standardization formula is: ; Wherein, is the standardized value output by the basic prediction model, and are the minimum value and the maximum value of , respectively, is the restored physical quantity value.

[0060] Specifically, after the rolling prediction of the basic prediction model is completed, an independent three-dimensional convolution residual correction module is introduced to post-process and optimize the prediction results, so as to reduce the error accumulation caused by long-term rolling and improve the spatiotemporal consistency. The three-dimensional convolution residual correction module compares the output of the basic prediction model with the observation data or the high-resolution mode result, learns the error, and outputs a correction quantity, which is superimposed with the original prediction result to obtain a more stable prediction field.

[0061] 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 to assimilate the high-resolution regional simulation results of multi-source observation data as reference labels to form input-label pairs.

[0062] S4.2, the training samples are normalized or standardized variable by variable and depth by depth, and a spatial depth validity mask is introduced to shield the land grid, invalid layer or missing point, so that the loss is calculated only in the effective sea area or horizon.

[0063] S4.3, the basic prediction model and the three-dimensional convolution residual correction module are deployed in a decoupled manner in a multi-GPU environment to support distributed parallel training and batch inference, facilitate the embedding of business pipelines, and meet the needs of high-resolution and refined regional marine prediction. In a multi-GPU environment, distributed data parallel training is used, combined with mixed precision acceleration and memory optimization; the optimizer uses an adaptive method of weight decay, and cooperates with a segmented learning rate scheduling strategy; a random seed is set to ensure repeatability; the loss function is mainly based on mask-weighted root mean square error (RMSE) / mean absolute error (MAE) (only calculated in the valid mask), so as to avoid the interference of invalid values or abnormal values on gradient update. The loss function is set as: ; ; wherein, and represent the predicted value and the true value, represents the union of indexes marked as valid in the mask matrix, is the root mean square error, is the mean absolute error. The training and verification process is divided by time samples to ensure consistency with the business inference process.

[0064] S4.4, during business inference, the latest rolling prediction of the basic prediction model is taken as the input of the three-dimensional convolution residual correction module, and the error estimate or correction value is output by the three-dimensional convolution residual correction module, and the original field is synthesized to obtain the correction result; then the inverse standardization is performed according to the statistics during training to restore the physical quantity scale and output as business products.

[0065] The three-dimensional convolution residual correction module is deployed in a decoupled manner with the post-processing plug-in form of the basic prediction model, the interface is connected to the rolling prediction cache, and the inference calculation is completed in real time on a single or multi-GPU; the module supports batch / on-line mode and data parallel expansion, and is convenient to be incorporated into the existing business pipeline.

[0066] The above-mentioned revised process significantly reduces the systematic deviation and time accumulation error of the rolling prediction by unifying the modeling space-vertical correlation at the voxel level (longitude x latitude x depth) through a three-dimensional convolution residual structure, in combination with mask constraints and residual learning, and improves the amplitude and phase consistency of meso-and small-scale structures such as fronts and eddies; distributed mixed precision training and lightweight three-dimensional convolution design ensure low additional computing power overhead and business availability, which can be plugged in without changing the underlying model.

[0067] Specifically, 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 dimensioned up or down by convolution; a main branch convolution stack and a shortcut branch are arranged in each residual unit, and a 1x1x1 convolution is used to match the dimensions when the number of channels is inconsistent; the output layer is regressed to the target channel by three-dimensional convolution; the main branch includes a three-dimensional convolution layer for extracting spatial features, followed by a normalization operation and a nonlinear activation function to enhance the feature expression capability, 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 to ensure that the information is not lost. Finally, the output of the main branch and the shortcut branch are added at each point position to obtain the result of the residual connection.

[0068] The three-dimensional convolution residual correction module uses a three-dimensional convolution-residual unit as a main feature extractor, and the input channel is first dimensioned up or down by convolution; a main branch convolution stack and a shortcut branch are arranged in each residual unit, and a 1x1x1 convolution is used to match the dimensions when the number of channels is inconsistent; the output layer is regressed to the target channel by three-dimensional convolution. The network learns the "prediction error field" in a residual manner, and the final result is obtained by superimposing the correction amount and the original prediction during inference; optionally, layer normalization is introduced under the channel priority layout to stabilize the training.

[0069] Taking the flow rate as an example, compared with the existing low-resolution marine large model, the high-resolution regional marine artificial intelligence large model can more finely depict the spatial distribution characteristics of the regional marine flow rate, and can capture more complete marine dynamic processes. Figure 2 The prediction result of the low-resolution marine large model is shown, and the overall flow field is relatively smooth, and the dynamic process of meso-and small-scale is not well analyzed. In contrast, Figure 3 The prediction result of the high spatio-temporal resolution is shown, which can clearly present more detailed dynamic structures, and can more accurately reflect the meso-and small-scale changes such as tidal flow processes. This shows that the method breaks through the resolution bottleneck of the existing large model which relies on public reanalysis data, significantly improves the resolution of meso-and small-scale marine dynamic processes while maintaining fast prediction capability, and provides higher precision technical support for regional marine applications such as national defense safety, shipping safety, fishery management and offshore energy development.

[0070] Of course, the above description is not a limitation of the present application, and the present application is not limited to the above examples. Changes, modifications, additions or substitutions made by those skilled in the art within the spirit and scope of the present application should also be included in the protection scope of the present application.

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; 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 regional range, 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, which is superimposed with the original prediction result to obtain a prediction field; 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 type optimization method combined with a learning rate scheduling mechanism to optimize parameters; 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 specifically comprises the following steps: S1.1, obtaining multi-source ocean observation data through submarine observation data, buoy observation data and satellite remote sensing products, the multi-source ocean observation data including sea temperature, salinity, flow rate and sea surface height; 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 model integration time window in the time dimension; S1.3, using a regional high-resolution ocean numerical model to perform simulation calculation at a kilometer-level grid scale to obtain numerical simulation data; and assimilating the observation data processed in step S1.2 into the numerical model, using a three-dimensional variation, four-dimensional variation or ensemble Kalman filter, mapping the model data to the observation space by constructing an observation operator, and combining the observation error covariance and the background error covariance to form an optimal analysis field; 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, 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 the frequency domain operator; S2.1.3, using symmetric padding or interpolation to align the features in the decoding stage, and concatenating the aligned decoding features and the corresponding scale frequency domain enhanced features in the channel dimension to effectively fuse the spectral domain and spatial features, and the fusion process is expressed as: ； wherein, represents a decoding path at the layer of features, represents a frequency domain enhancement unit at the layer of features, 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, using a rolling time step to sequentially deduce the future period of the ocean state, using the latest obtained multi-source observation data and numerical simulation data as input, and performing format and standardization processing through the preprocessing procedures of steps S1.2-S1.4; S3.2, running the trained basic prediction model in a set regional range; S3.3, the output result of the basic prediction model is recovered to the original physical value through inverse standardization, and the multivariate spatiotemporal prediction result is outputted, including temperature, salinity or flow rate, and the inverse standardization formula is: ； wherein, is a standardized value output by the base prediction model, and is a minimum value and a maximum value, respectively, of is a restored physical quantity value.

5. The method according to claim 1, wherein, Step S4 specifically comprises 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 assimilated as a reference label to form an input-label pair; S4.2, the training sample is normalized or standardized variable by variable and depth by depth, and a spatial depth effectiveness mask is introduced to shield the land grid, invalid layer or missing point, so that the loss is calculated only on the effective sea area or horizon; S4.3, the optimizer adopts an adaptive method of weight decay, and cooperates with a segmented learning rate scheduling strategy; A random seed is set to ensure repeatability, and a loss function is set as: ； ； wherein, and pred and true represent the predicted and true values, respectively, pred and true represent the predicted and true values, respectively, is the root mean square error, is the mean absolute error; S4.4, the variables are inversely normalized according to the statistics during training, restored to the physical quantity scale and outputted as business products.

6. The method according to claim 5, wherein, The three-dimensional convolution residual correction module in step S4.1 comprises an input layer, a three-dimensional convolution layer, a residual unit and an output layer connected in sequence; the input channel is first subjected to convolution dimension increase or dimension reduction; a main branch and a shortcut branch are arranged 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