A marine environment forecasting method based on the combination of machine learning and numerical forecasting
By combining generative adversarial networks, graph neural networks, and Bayesian recursive networks, the problems of data sparsity and extreme event forecasting in marine environmental forecasting were solved, achieving more accurate marine environmental forecasting.
Patent Information
- Application Number
- CN202511020533.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-24
- Publication Date
- 2025-09-16
- Estimated Expiration
- 2045-07-24
AI Technical Summary
The existing marine environment forecasting methods have covariance matrix estimation bias in data-sparse areas of marginal seas, which leads to failure of initial field optimization, and the lack of sample support of neural networks in extreme marine events leads to distorted forecast results.
A generative adversarial network is used to extract edge area features, combined with a graph neural network to analyze spatial autocorrelation, a forecast field is generated through a dual loss function, and extreme event features are extracted through spatiotemporal decoupling. The residual correction network and Bayesian recursive network are used to output the forecast results, forcing the vorticity conservation equation and physical constraints to be satisfied.
It alleviates the failure of initial field optimization when the marginal sea area data is sparse, reduces the dependence on the number of extreme event samples, corrects the forecast error in the data-sparse area, and improves the accuracy and robustness of the forecast.
Smart Images

Figure CN120524339B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of marine environment forecasting, and more specifically, to a marine environment forecasting method based on the combination of machine learning and numerical forecasting. Background Art
[0002] With the growing demand for marine resource development and navigation safety, marine environment forecasting technology is particularly important. Traditional marine environment forecasting methods mainly pre-process multi-source data such as satellite remote sensing and buoy observations and input them into numerical models to predict the spatiotemporal distribution of factors such as sea surface temperature, ocean current speed, and wave height. Actual measurements have shown that numerical models require exascale floating-point operations, resulting in excessively long calculation times. In addition, the parameterization of small-scale processes such as mesoscale eddies and internal waves is insufficient, making it difficult to meet the needs of refined forecasting.
[0003] To overcome the above-mentioned defects, the main improvement directions of existing technologies include: introducing data assimilation technology, by adding an assimilation module to the numerical model, minimizing the error covariance between the observed data and the model simulation values based on Bayesian estimation theory, and optimizing the initial field and boundary conditions; adopting statistical post-processing methods and using machine learning models to correct the errors of numerical forecast results; compared with traditional technologies, existing technologies have made breakthrough progress in computational efficiency and improved the timeliness of short-term forecasts of regional oceans.
[0004] However, it still has some shortcomings in actual use. For example, data assimilation technology has a strong dependence on the density of observation data. In areas with sparse data in marginal seas, insufficient effective observation samples in the assimilation window lead to covariance matrix estimation bias, which in turn causes the failure of initial field optimization, and ultimately causes regional forecasts to increase errors in wider sea areas; the strong dependence of neural networks on the distribution of training data causes them to produce distorted forecast results in extreme ocean events due to lack of sample support. Summary of the Invention
[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides a marine environment forecasting method based on the combination of machine learning and numerical forecasting, and solves the problems raised in the above-mentioned background technology through the following scheme.
[0006] To achieve the above object, the present invention provides the following technical solutions:
[0007] A marine environment forecasting method based on the combination of machine learning and numerical forecasting, including:
[0008] S1: Based on the first observation data corresponding to the target sea area obtained by real-time observation, a first observation feature is extracted from the edge area corresponding to the target sea area through a generative adversarial network;
[0009] S2: Based on the first observation feature, the spatial autocorrelation of the target sea area is analyzed through a graph neural network to generate a first prediction factor;
[0010] S3: generating a first forecast field by applying the first forecast factor to a double loss function;
[0011] S4: Based on the first forecast field, performing a feature forecasting operation, the feature forecasting operation including extracting a second observation feature corresponding to the extreme ocean event through spatiotemporal decoupling and generating a second forecast field;
[0012] S5: fusing the second observation feature with the second forecast field, and generating a second prediction factor through a residual correction network;
[0013] S6: Based on the second prediction factor, output the prediction result corresponding to the target sea area through the Bayesian recursive network.
[0014] Preferably, the S2 graph neural network adopts a physical constraint graph attention architecture, and the design of the physical constraint layer in the graph neural network specifically includes:
[0015] The target sea area is divided into grids, each grid unit is regarded as a node in the graph, and the node feature vector of each node is assigned as the first observation feature;
[0016] The residual calculation of the vorticity conservation equation is embedded in the hidden layer, and the nodes in the graph are , calculate its relative vorticity and stream function , and calculate the residual of the vorticity conservation equation , specifically expressed as:
[0017] ,
[0018] in, Expressed as the Jacobian term, Expressed as time;
[0019] When the node Calculated residuals When the residual value is greater than the preset threshold, the node is frozen. Gradient updates during backpropagation.
[0020] Preferably, the construction of the dual loss function in S3 must satisfy:
[0021] Minimizing the difference between the first observation data and a preset background field, wherein the preset background field is a historical forecast field consisting of sea surface temperature, salinity, and current velocity;
[0022] The vorticity conservation equation is forced to be satisfied to generate a first forecast field, which is the initial field at the end of the assimilation window.
[0023] Preferably, the S3 is to force the vorticity conservation equation to be satisfied and generate the first forecast field, which is specifically expressed as:
[0024]
[0025] in, Expressed as velocity field, is the velocity component of seawater in the meridional direction in the velocity field, Expressed as Rossby parameter; the velocity field includes the velocity components of seawater in the longitudinal and latitudinal directions.
[0026] Preferably, the step S4, extracting the second observational feature corresponding to the extreme ocean event through spatiotemporal decoupling, specifically includes:
[0027] Main pathway: Atrous convolutional layer extracts large-scale background fields;
[0028] Side branch pathways: Deformable convolutional layers adaptively focus on the mutation structures corresponding to extreme events.
[0029] Preferably, the step S4 of generating the second forecast field specifically includes:
[0030] The numerical model is run to generate a control equation for the second forecast field within a preset time. The control equation includes a momentum equation and a temperature equation. The momentum equation is specifically expressed as follows:
[0031] ,
[0032] in, Expressed as velocity field, Expressed as time, Expressed as the advection term of the velocity field, Expressed as the Coriolis parameter, Expressed as the vertical unit buoyancy vector, Expressed as the reference density of seawater, Expressed as seawater pressure, Expressed as the diffusion term of the velocity field, It is expressed as the influence factor of the friction force inside the fluid on the flow velocity; the temperature equation is specifically expressed as:
[0033] ,
[0034] in, Expressed as velocity field, Expressed as time, is the seawater temperature, Expressed as the diffusion rate of temperature in the fluid.
[0035] Preferably, the residual correction network in S5 includes a physical hard constraint layer and an LSTM layer, wherein the physical hard constraint layer is specifically represented as follows:
[0036]
[0037] in, It is represented as the output residual feature map after processing by the physical hard constraint layer, including the residual information corrected by the physical constraint. It is represented as the residual feature map input to the physical hard constraint layer, including the correction information extracted from the second forecast field and the second observation feature, is expressed as sea level height, Expressed as velocity field, Expressed as time.
[0038] Preferably, the step S6, outputting the forecast result corresponding to the target sea area through a Bayesian recursive network, specifically includes:
[0039] Calculate the confidence level corresponding to the prediction results of the target sea area;
[0040] When the confidence level is lower than the threshold, the extreme event corresponding to the second observation feature is migrated to the training data set of the generative adversarial network in S1, and the spatiotemporal decoupling of S4 and the residual network weights of S5 are updated.
[0041] Technical effects and advantages of the present invention:
[0042] 1. The present invention uses S2's physical constraint graph attention architecture and vorticity conservation equation residual calculation technology to embed the vorticity conservation equation residual calculation in the hidden layer, alleviating the initial field optimization failure problem caused by covariance matrix estimation bias when the marginal sea data is sparse, and enhancing the model's robustness to sparse data through physical constraints;
[0043] 2. Through the spatiotemporal decoupling of S4, the present invention achieves targeted extraction of characteristics of extreme ocean events and reduces the dependence on the number of extreme event samples;
[0044] 3. The present invention uses the physical hard constraint layer technology of the S5 residual correction network, combined with the sea surface height and velocity field constraint residual characteristic map, to correct the initial field error in data sparse areas and alleviate the problem of large forecast errors in marginal sea areas. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1 This is a flowchart of the steps of a method for forecasting the marine environment based on the combination of machine learning and numerical forecasting provided in accordance with an embodiment of the present application.
[0046] Figure 2This is a diagram of the execution architecture of a feature forecasting operation in a method for forecasting an ocean environment based on the combination of machine learning and numerical forecasting provided in an embodiment of the present application. DETAILED DESCRIPTION
[0047] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0048] The terms used in the following examples of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the specification of the present application, the singular expressions "a", "an", "said", "above", "the", and "this" are intended to include plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used in the present application refers to and includes any or all possible combinations of one or more of the listed items.
[0049] In the following, the terms "first" and "second" are used for descriptive purposes only and should not be understood to imply or suggest relative importance or implicitly indicate the number of the technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the embodiments of this application, unless otherwise specified, "plurality" means two or more.
[0050] As attached Figure 1 The ocean environment forecasting method shown here, which combines machine learning with numerical forecasting, first extracts edge area features based on real-time observation data, constructs a forecast field using a double loss function, extracts extreme event features through spatiotemporal decoupling, and then outputs forecast results for the target sea area. Specifically, the method includes the following steps:
[0051] S1: Based on the first observation data corresponding to the target sea area obtained by real-time observation, a first observation feature is extracted from the edge area corresponding to the target sea area through a generative adversarial network;
[0052] S2: Based on the first observation feature, the spatial autocorrelation of the target sea area is analyzed through a graph neural network to generate a first prediction factor;
[0053] S3: generating a first forecast field by applying the first forecast factor to a double loss function;
[0054] S4: Based on the first forecast field, performing a feature forecasting operation, the feature forecasting operation including extracting a second observation feature corresponding to the extreme ocean event through spatiotemporal decoupling and generating a second forecast field;
[0055] S5: fusing the second observation feature with the second forecast field, and generating a second prediction factor through a residual correction network;
[0056] S6: Based on the second prediction factor, output the prediction result corresponding to the target sea area through the Bayesian recursive network.
[0057] Specifically, in S1, the first observation data includes but is not limited to sea surface height, sea surface temperature, thermohaline profile, surface flow field, etc. After preprocessing the first observation data, the data features corresponding to the first observation data are extracted from the edge area corresponding to the target sea area through a generative adversarial network, that is, the first observation features. The first observation features include spatial structure features, time evolution features, conserved physical quantity features and extreme event identification features; the spatial structure features include but are not limited to the sea surface temperature front gradient intensity, abnormal vortex morphological parameters of sea surface height, surface flow field shear stress distribution, etc.; the time evolution features include but are not limited to the vertical oscillation phase of the thermohaline profile, the sea surface height change rate, the surface flow field rotation trend, etc.; the conserved physical quantity features include but are not limited to the spatial distribution of potential vorticity, vorticity advection transport flux, density stratification stability, etc.; the extreme event identification features include but are not limited to the range of typhoon-induced cold wake, the abnormal value of sea-air flux in strong wind area, the vortex-front interaction intensity index, etc.
[0058] In this embodiment, the sea surface height in the first observation data is obtained by L-band synthetic aperture radar, with a spatial resolution of no more than 1 km and a time resolution of 6 hours; the sea surface temperature is obtained by inversion of infrared sensors, with a spatial resolution of no more than 1 km and a time resolution of 6 hours; the temperature and salinity profile is collected by Argo profile buoys deployed at a water depth of 0 to 2000 m in the target sea area, with a time resolution of 24 hours; the surface flow field data is obtained by coastal radar stations and is set to cover a range of 200 km offshore; it should be noted that the marginal area corresponding to the target sea area is defined as an area that meets the following conditions at the same time: on the geographical boundary, the sea area 50 to 200 km away from the nearest coastline to exclude the interference of nearshore turbid water bodies; on the dynamic characteristics, based on the historical eddy kinetic energy gradient value being greater than or equal to 0.1m 2 / s 2 The ocean frontal zone is divided into two parts; and the data sparseness index is that the buoy observation density is less than 1 / 10 4 km 2 area.
[0059] It should be noted that the first observation data are preprocessed, and the preprocessing operations include time-space alignment, that is, the satellite data and the buoy data are uniformly interpolated to a regular grid of 0.1°×0.1°, and the timestamps are synchronized to UTC to ensure the consistency of the data in time and space dimensions; missing value processing, the missing buoy data in the marginal sea area are filled by the Kriging spatial interpolation method; Wiener filtering is performed on the SAR sea surface height data to eliminate the white hat noise interference that may occur when the wind speed is greater than 15m / s.
[0060] Furthermore, the generative adversarial network includes a generator and a discriminator, the generator performs 3D convolution and pooling operations on the first observation data, and at the same time, in the bottleneck layer of the generator, a differentiable vorticity equation is embedded as a physical conservation constraint.
[0061] In this embodiment, the training data set of the generative adversarial network is constructed based on the real-time observation data of the target sea area in the past 10 years, and the real-time observation data includes at least 3 strong typhoon events. Random coordinate transformation is applied to the open sea area to simulate the complex morphology of the marginal sea area for data enhancement. The total loss function of the generative adversarial network is designed in a weighted composite form based on the training data set. , specifically expressed as:
[0062] ,
[0063] in, Denoted as the adversarial loss term, Expressed as a physical conservation loss term, Denoted as the reconstruction loss term, 、 、 Respectively expressed as the weights of the adversarial loss term, the physical conservation loss term, and the reconstruction loss term; wherein the adversarial loss term The Wasserstein distance is used to measure the difference between the first observation feature distribution and the first observation data distribution to optimize the distribution of the first observation feature, which is specifically expressed as:
[0064] ,
[0065] in, Represented as the first observation data of the discriminator in the generative adversarial network The expected output is Represented as the first observation feature of the discriminator in the generative adversarial network The expected output of the physical conservation loss term The L2 norm of the residual of the vorticity equation is used to force the generated data to satisfy the physical laws, which can be expressed as:
[0066] ,
[0067] in, Expressed as the sea surface height field in the first observation feature, Expressed as time, and Represented as the ocean current in the first observation feature in the longitudinal direction and weft direction The velocity component of Expressed as the zonal vorticity parameter, Expressed as L2 norm; the reconstruction loss It is used to constrain the local consistency between generated data and sparse real observations, specifically expressed as:
[0068] ,
[0069] in, represents the number of first observations for computing the reconstruction loss, represents the index of the first observation for computing the reconstruction loss, Expressed as The first observed feature, Expressed as The first observation data.
[0070] Specifically, in S2, based on the first observation feature, a non-stationary covariance structure is modeled through a graph neural network to learn the error spatial autocorrelation of various sea areas, and the obtained dynamic covariance operator is used as the first predictor.
[0071] It should be noted that the data dimension of the first observation feature is [time step × number of longitude grids × number of latitude grids × number of channels], and the target sea area is divided into 0.1°×0.1° grids, each grid unit is regarded as a node in the graph, and the node feature vector of each node is assigned as the first observation feature.
[0072] In one possible implementation, the graph neural network adopts a physical constraint graph attention architecture, and the design of the physical constraint layer in the graph neural network includes: embedding the residual calculation of the vorticity conservation equation in the hidden layer, , calculate its relative vorticity and stream function , and calculate the residual of the vorticity conservation equation , specifically expressed as:
[0073] ,
[0074] in, Expressed as the Jacobian term, Expressed as time; the implementation of the physical constraint is: when the node Calculated residuals When the residual value is greater than the preset threshold, the node is frozen. Gradient updates during backpropagation.
[0075] In this embodiment, the graph neural network learns the spatial correlation characteristics of the error through spatial autocorrelation analysis, and realizes the spatial correlation characteristics of the error through the node through the graph convolution process. In the The feature update of the layer; and the goal of analyzing the spatial autocorrelation of the target sea area through the graph neural network is to minimize the covariance estimation error, and based on the minimization of the covariance estimation error, the covariance between nodes expressed in matrix form is generated, that is, the first prediction factor.
[0076] Specifically, in S3, the dynamic covariance operator is integrated into the variational assimilation framework to design a dual loss function.
[0077] In one possible implementation, the requirements that the dual loss function needs to meet are as follows: data fitting term: minimizing the difference between the first observation data and the preset background field, where the preset background field is a historical forecast field composed of sea surface temperature, salinity, and current velocity; physical conservation term: forcing the vorticity conservation equation to be satisfied to generate a first forecast field, which is the initial field at the end of the assimilation window.
[0078] It should be noted that the physical conservation term is implemented by embedding the vorticity conservation equation, which is specifically expressed as:
[0079]
[0080] in, Expressed as velocity field, is the velocity component of seawater in the meridional direction in the velocity field, Expressed as Rossby parameter; the velocity field includes the velocity components of seawater in the longitudinal and latitudinal directions.
[0081] Specifically, in S4, based on the first forecast field, a feature forecast operation is performed, and the feature forecast operation includes spatiotemporal decoupling, and the spatiotemporal decoupling is used to separate the normal characteristics and extreme event characteristics in the first forecast field, and then output the extreme event characteristics as the second observation feature; the forecast field is optimized, based on the first forecast field, by running the numerical model to generate an optimized second forecast field.
[0082] In this embodiment, the first forecast field is defined as a 0.1°×0.1° latitude and longitude grid resolution and a 10-minute time step, and the parameters include three-dimensional current velocity, temperature, salinity, and sea surface height.
[0083] In one possible implementation, the spatiotemporal decoupling is a dual-pathway structure, including: a main path: a dilated convolutional layer extracts large-scale background fields; and a side path: a deformable convolutional layer adaptively focuses on the mutation structure corresponding to extreme events.
[0084] It should be noted that the prior knowledge of extreme events is used to guide the decoupling of features. The prior knowledge includes the spatial scale, time scale and key identification features of various types of extreme events. A spatiotemporal decoupling convolutional network is used for spatiotemporal decoupling. The architecture of the spatiotemporal decoupling convolutional network includes a trunk path and a side network path. The side path incorporates a physical attention mechanism and multi-scale dilated convolution. The physical attention mechanism focuses on areas where extreme events may occur by calculating attention masks to automatically identify extreme event areas such as ocean fronts and vortices. The multi-scale dilated convolution uses convolution kernels with multiple expansion rates to capture the characteristics of extreme events at different scales.
[0085] In this embodiment, the main path adopts a 3×3×3 convolution kernel, the side path adopts a 5×5×1 convolution kernel, the step size of the time dimension is set to 2, and the step size of the spatial dimension is set to 1; the filling method adopts periodic boundary filling in the longitudinal direction and symmetrical filling in the latitudinal direction; the convolution kernels with multiple expansion rates include: the expansion rate of the first layer of void convolution is 1, the receptive field is 9km; the expansion rate of the second layer is 3, the receptive field is 27km; the expansion rate of the third layer is 6, and the receptive field is 54km.
[0086] In one possible implementation, generating the second forecast field includes: running a numerical model, wherein the core numerical model selects a regional ocean model (ROMS), integrating the first forecast field as a starting state, and generating a control equation for the second forecast field within a preset time, wherein the control equation includes a momentum equation and a temperature equation, wherein the momentum equation is specifically expressed as:
[0087] ,
[0088] in, Expressed as velocity field, Expressed as time, Expressed as the advection term of the velocity field, Expressed as the Coriolis parameter, Expressed as the vertical unit buoyancy vector, Expressed as the reference density of seawater, Expressed as seawater pressure, Expressed as the diffusion term of the velocity field, It is expressed as the influence factor of the friction force inside the fluid on the flow velocity; the temperature equation is specifically expressed as:
[0089] ,
[0090] in, Expressed as velocity field, Expressed as time, is the seawater temperature, Expressed as the diffusion rate of temperature in the fluid.
[0091] In this embodiment, the operation configuration of the numerical model is as follows: the horizontal resolution is set to 3 km, the vertical stratification is set to 32 layers, the time step is set to 90 seconds, the boundary conditions are nested using the HYCOM global field; the parameterization scheme adopts the KPP vertical mixing scheme and the GM mesoscale eddy parameterization scheme.
[0092] Specifically, in S5, based on the second observation feature and the second forecast field, a dynamically compatible correction value, ie, a second forecast factor, is generated through a residual correction network.
[0093] In one possible implementation, the residual correction network includes a physical hard constraint layer and an LSTM layer structure, and the hidden layer of the residual correction network includes a physical hard constraint layer, which is specifically expressed as follows:
[0094]
[0095] in, It is represented as the output residual feature map after the physical hard constraint layer processing, including the residual information corrected by the physical constraint, which is used to generate the dynamic compatibility correction amount. It is represented as the residual feature map input to the physical hard constraint layer, including the correction information extracted from the second forecast field and the second observation feature, is expressed as sea level height, Expressed as velocity field, Expressed as time.
[0096] It should be noted that the correction amount is obtained by forward propagation calculation of the residual correction network.
[0097] Specifically, in S6, based on the second prediction factor, a prediction result corresponding to the target sea area is output through a Bayesian recursive network.
[0098] It should be noted that based on the second prediction factor and the historical assimilated error data, spatiotemporal uncertainty propagation is performed through the Bayesian recursive network, and then the error propagation law is integrated. The prediction results of the target sea area are generated through multiple recursive calculations. The prediction results of the target sea area need to construct a confidence probability map to quantify the uncertainty, which is specifically expressed as:
[0099]
[0100] in, is expressed as the variance of the sea surface anomaly, is represented as the variance operator, In different sampling times Historical assimilation error data The predicted output.
[0101] In one possible implementation, by calculating the confidence corresponding to the prediction result of the target sea area, when the confidence is lower than a threshold, the cross-scenario knowledge distillation mechanism is initiated, including: when the confidence is lower than the threshold, the extreme event corresponding to the second observation feature is migrated to the training data set of the generative adversarial network in S1, and the spatiotemporal decoupling of S4 and the residual network weights of S5 are updated.
[0102] Secondly: The drawings of the embodiments disclosed in the present invention only involve structures related to the embodiments disclosed in the present invention. Other structures may refer to conventional designs. The same embodiment and different embodiments of the present invention may be combined with each other without conflict.
[0103] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A method for forecasting marine environment based on the combination of machine learning and numerical forecasting, characterized in that: include: S1: Based on the first observation data corresponding to the target sea area obtained by real-time observation, a first observation feature is extracted from the edge area corresponding to the target sea area through a generative adversarial network; S2: Based on the first observation feature, the spatial autocorrelation of the target sea area is analyzed through a graph neural network to generate a first prediction factor; S3: generating a first forecast field by applying the first forecast factor to a double loss function; S4: Based on the first forecast field, performing a feature forecasting operation, the feature forecasting operation including extracting a second observation feature corresponding to the extreme ocean event through spatiotemporal decoupling and generating a second forecast field; S5: fusing the second observation feature with the second forecast field, and generating a second prediction factor through a residual correction network; S6: Based on the second prediction factor, output the prediction result corresponding to the target sea area through the Bayesian recursive network.
2. The method for predicting the marine environment based on the combination of machine learning and numerical prediction according to claim 1, characterized in that: In S2, the graph neural network adopts a physical constraint graph attention architecture, and the design of the physical constraint layer in the graph neural network specifically includes: The target sea area is divided into grids, each grid unit is regarded as a node in the graph, and the node feature vector of each node is assigned as the first observation feature; The residual calculation of the vorticity conservation equation is embedded in the hidden layer, and the nodes in the graph are , calculate its relative vorticity and stream function , and calculate the residual of the vorticity conservation equation , specifically expressed as: , in, Expressed as the Jacobian term, Expressed as time; When the node Calculated residuals When the residual value is greater than the preset threshold, the node is frozen. Gradient updates during backpropagation.
3. The method for forecasting the marine environment based on the combination of machine learning and numerical forecasting according to claim 1, characterized in that: The construction of the S3 dual loss function must satisfy the following requirements: Minimizing the difference between the first observation data and a preset background field, wherein the preset background field is a historical forecast field consisting of sea surface temperature, salinity, and current velocity; The vorticity conservation equation is forced to be satisfied to generate a first forecast field, which is the initial field at the end of the assimilation window.
4. The method for forecasting the marine environment based on the combination of machine learning and numerical forecasting according to claim 3, characterized in that: The S3 is forced to satisfy the vorticity conservation equation and generate the first forecast field, which is specifically expressed as: , in, Expressed as velocity field, is the velocity component of seawater in the meridional direction in the velocity field, Expressed as Rossby parameter; The velocity field includes velocity components of seawater in the longitudinal and latitudinal directions.
5. The method for predicting the marine environment based on the combination of machine learning and numerical prediction according to claim 1, characterized in that: S4, extracting the second observational features corresponding to extreme ocean events through spatiotemporal decoupling, specifically includes: Main pathway: Atrous convolutional layer extracts large-scale background fields; Side branch pathways: Deformable convolutional layers adaptively focus on the mutation structures corresponding to extreme events.
6. The method for predicting the marine environment based on the combination of machine learning and numerical prediction according to claim 1, characterized in that: The step S4, generating the second forecast field, specifically includes: The numerical model is run to generate a control equation for the second forecast field within a preset time. The control equation includes a momentum equation and a temperature equation. The momentum equation is specifically expressed as follows: , in, Expressed as velocity field, Expressed as time, Expressed as the advection term of the velocity field, Expressed as the Coriolis parameter, Expressed as the vertical unit buoyancy vector, Expressed as the reference density of seawater, Expressed as seawater pressure, Expressed as the diffusion term of the velocity field, Expressed as the factor affecting the flow velocity due to the friction inside the fluid; The temperature equation is specifically expressed as: , in, Expressed as velocity field, Expressed as time, is the seawater temperature, Expressed as the diffusion rate of temperature in the fluid.
7. The method for forecasting the marine environment based on the combination of machine learning and numerical forecasting according to claim 1, characterized in that: The residual correction network in S5 includes a physical hard constraint layer and an LSTM layer, wherein the physical hard constraint layer is specifically represented as follows: , in, It is represented as the output residual feature map after processing by the physical hard constraint layer, including the residual information corrected by the physical constraint. It is represented as the residual feature map input to the physical hard constraint layer, including the correction information extracted from the second forecast field and the second observation feature, is expressed as sea level height, Expressed as velocity field, Expressed as time.
8. The method for forecasting the marine environment based on the combination of machine learning and numerical forecasting according to claim 1, characterized in that: The step S6, outputting the forecast result corresponding to the target sea area through the Bayesian recursive network, specifically includes: Calculate the confidence level corresponding to the prediction results of the target sea area; When the confidence level is lower than the threshold, the extreme event corresponding to the second observation feature is migrated to the training data set of the generative adversarial network in S1, and the spatiotemporal decoupling of S4 and the residual network weights of S5 are updated.
Citation Information
Patent Citations
Marine sparse observation data assimilation method based on deep learning
CN115758898A
Data and knowledge combined driven interpretable sea temperature prediction method and device
CN118536077A