A method for correcting ocean surface current field prediction data based on a physically guided gating hybrid network

Abstract

The application discloses a marine surface layer flow field prediction data correction method based on a physical guided gating hybrid network, and belongs to the technical field of marine dynamic prediction data error correction and artificial intelligence marine modeling. The method acquires multi-source data and coordinate time characteristics; a basic solver network with a multilayer perception machine (MLP) as a main part is used to generate a basic correction flow field, and a two-dimensional surface layer momentum term is used to calculate a dynamic residual error; a space-time corrector Corrector network with a convolution long short-term memory network (ConvLSTM) as a main part is used to learn data-driven increments, missing external forcing terms and gating weights; a correction flow field is output by fusing physical injection and basic correction and data increments under the gating adjustment, and stable optimization is realized by adopting stage-by-stage training. The application can improve the stability and precision of marine surface layer prediction flow field correction under the whole and complex dynamic state.

Description

Technical Field

[0001] This invention relates to the field of ocean dynamic forecast data error correction and artificial intelligence ocean modeling technology, specifically to a method for correcting ocean surface flow field forecast data based on a physically guided gated hybrid network. Background Technology

[0002] Ocean surface currents are crucial variables describing air-sea interactions, material transport, maritime search and rescue, navigation safety, oil spill drift, and environmental protection in marine engineering. Current operational ocean current forecasts primarily rely on numerical circulation models and data assimilation systems. While these methods can provide forecasted current fields with continuous spatiotemporal coverage, they are limited by computational resolution, initial field errors, uncertainties in physical parameterization, and the lack of resolution in subgrid processes, often resulting in systematic biases in the output surface velocity fields.

[0003] With the development of deep learning technology, convolutional neural networks, recurrent neural networks, and sequence-to-sequence models have been used for post-processing of numerical weather prediction errors. These purely data-driven methods can learn the nonlinear mapping between the forecast field and the analysis field, and have certain effects in single-step or short-term correction tasks. However, in multi-step closed-loop rolling inference scenarios required by actual operations, the output of the current step of the model is recursively used as the input of the next step. Purely data-driven networks are prone to non-physical perturbations in the absence of dynamic constraints, which gradually accumulate during the autoregression process, leading to error propagation amplification and distribution drift.

[0004] To address these issues, existing methods have attempted to incorporate the Navier-Stokes equations or two-dimensional shallow-water momentum equations as physical constraints into neural network training. However, real-world operational ocean analysis typically integrates three-dimensional baroclinic dynamics, bottom friction, wind stress curl, complex topographic modulation, and assimilation increments. Directly requiring the residuals of the simplified two-dimensional surface equations to approach zero can easily lead to physical constraint mismatch, causing the network to learn compensation terms lacking clear physical meaning in order to reduce residuals, thus disrupting stability in closed-loop rolling.

[0005] Therefore, there is an urgent need for a method for ocean surface flow field correction that can both utilize multi-source ocean environmental data for error correction and introduce physical information in a controlled manner, so that the model can have correction accuracy, physical interpretability and closed-loop rolling stability under complex dynamic conditions. Summary of the Invention

[0006] (a) Technical problems to be solved

[0007] The purpose of this invention is to provide a method for correcting ocean surface flow field forecast data based on a physically guided gated hybrid network, in order to solve the problems of error accumulation and non-physical drift that easily occur in existing pure data-driven error correction models during multi-step closed-loop rolling, and the problem of constraint mismatch in simplified equations in traditional physical strong constraint methods in real operational analysis fields.

[0008] (II) Technical Solution

[0009] To achieve the above objectives, the present invention adopts the following technical solution.

[0010] Step 1: Acquisition of multi-source data.

[0011] The numerical forecast data for the sea area to be corrected includes surface current field, sea surface wind field, sea surface height field, spatial coordinates, and temporal information at continuous time intervals. Specifically, the surface current field includes an eastward velocity component (u) and a northward velocity component (v), and the sea surface wind field includes a 10m eastward wind component and a 10m northward wind component.

[0012] Step 2: Data preprocessing.

[0013] Based on the spatial resolution of the analyzed flow field and sea surface height, the high-resolution grid of the predicted flow field is downsampled to the analyzed flow field grid, while the reanalysis wind field is resampled to the analyzed flow field grid through bilinear interpolation. Specifically, the predicted flow field is downsampled, the sea surface wind field is resampled using bilinear interpolation, and the eastward and northward sea surface height gradients are calculated using the sea surface height field. Latitude and longitude are normalized and scaled, and continuous physical variables are Z-score normalized. The annual day sequence is mapped to sinusoidal and cosine time periodic features. Using a continuous historical time period of length L as the input window, longitude, latitude, time periodic features, predicted current velocity, sea surface wind, and sea surface height gradient are concatenated into a multi-channel spatiotemporal tensor.

[0014] Step 3: Generate the basic flow field.

[0015] The multi-channel spatiotemporal tensor input points are shared by a basic solver, Solver. Solver performs nonlinear mapping on the shared parameters of each grid point and outputs a basic corrected flow field. This basic corrected flow field serves as both the background state for the final prediction and the state input for subsequent physical residual calculations.

[0016] Step 4: Calculate the known dynamic terms and dynamic residuals.

[0017] Based on the baseline corrected flow field, sea surface wind field, and sea surface height gradient, known dynamic terms in the two-dimensional surface momentum equation that can be calculated relatively stably from the two-dimensional surface variables are calculated, including the advection term, Coriolis force term, barotropic pressure gradient term, wind stress-driven term, and horizontal diffusion term. Then, the actual temporal evolution tendency of the baseline corrected flow field is calculated using grid center difference, and the difference between the actual temporal evolution tendency and the known dynamic terms is defined as the dynamic residual.

[0018] Step 5: Spatiotemporal correction and forced discovery of missing data.

[0019] The historical spatiotemporal input window is fed into the spatiotemporal corrector. The corrector utilizes spatiotemporal recursive features to extract the evolution law of network learning error, and outputs the data-driven increment, missing external forcing term, and spatiotemporal gating matrix through decoupled output heads. Among them, the missing external forcing term is used to characterize the three-dimensional baroclinic effect not explicitly modeled by the two-dimensional surface momentum equation, the topographic modulation of varying water depth, the frictional dissipation of the bottom boundary layer, the subgrid turbulence parameterization error, and the assimilation increment.

[0020] Step 6: Gated physical injection.

[0021] The missing external forcing term output by the spatiotemporal corrector Corrector is transformed into a one-step equivalent velocity increment by using the macroscopic time step, velocity standard deviation, and scale conversion factor. The one-step equivalent velocity increment is then selectively injected using the spatiotemporal gating matrix and a learnable global injection scale activated by Softplus.

[0022] Step 7: Output fusion.

[0023] A physical-guided gating hybrid network, including a basic solver Solver, a spatiotemporal corrector Corrector, a spatiotemporal gating matrix, and a global injection scale, is trained so that the basic solver Solver outputs the basic corrected flow field, and the spatiotemporal corrector Corrector outputs the data-driven increment, the missing external forcing term, and the spatiotemporal gating matrix G.

[0024] Step 8: Model training.

[0025] The physical-guided gating hybrid network, comprising the basic solver Solver, the spatiotemporal corrector Corrector, the spatiotemporal gating matrix, and the global injection scale, is trained to output the basic corrected flow field from the basic solver Solver. This causes the output data of the spacetime corrector to drive the increment. Missing external compulsion items And the spacetime gating matrix G.

[0026] Step 9: Forecast data correction.

[0027] The surface current field, sea surface wind field, sea surface height field, spatial coordinates and time information of the numerical forecast data at the time to be corrected are input into the trained ocean surface current field forecast data correction model. The model then sequentially performs basic flow field generation, known dynamic terms and residual calculation, spatiotemporal correction and missing force detection, gated physical injection and output fusion to obtain the corrected ocean surface current field forecast.

[0028] (III) Beneficial Effects

[0029] Compared with the prior art, the present invention has at least the following beneficial effects:

[0030] First, this invention does not directly use the simplified two-dimensional shallow water equation as a strong constraint. Instead, it explicitly models the dynamic non-closure difference between the simplified equation and the real analysis field as a missing external forcing term, so that the discovery of physical terms and the correction of statistical errors work together in the same framework.

[0031] Second, this invention separates the modeling of stable background correction, high-frequency statistical error compensation, and dynamically meaningful missing object forced discovery through a decoupled structure of the basic solver, data-driven incremental branch, and physical injection branch, which helps to reduce target coupling in multi-objective joint training.

[0032] Third, this invention utilizes a spatiotemporal gating matrix to locally selectively adjust the physics injection, ensuring that missing external forcing terms are activated or suppressed only at the required locations and components, thus avoiding inappropriate corrections caused by full physics injection in weak dynamic regions or noisy regions.

[0033] Fourth, the present invention adopts a phased training strategy, first stabilizing the basic correction field, then learning the dynamic residual semantics, and finally performing joint optimization, which helps to reduce physical branch semantic drift and gating collapse, and improve stability under closed-loop rolling conditions. Attached Figure Description

[0034] Figure 1 This is a network architecture diagram of the Physically Guided Gated Hybrid Network (PGHN) described in this invention.

[0035] Figure 2 This is a network structure diagram of the basic solver Solver described in this invention.

[0036] Figure 3 This is a network structure diagram of the spacetime corrector Corrector described in this invention.

[0037] Figure 4 This is a diagram of the three-stage training strategy described in this invention.

[0038] Figure 5This is a line graph comparing the overall RMSE (Root Mean Squared Error) of this invention with other state-of-the-art models on the u-component.

[0039] Figure 6 This is a line graph comparing the overall RMSE of this invention with other state-of-the-art models on the v component.

[0040] Figure 7 This is a line graph comparing the overall MAE (Mean Absolute Error) of this invention with other state-of-the-art models on the u-component.

[0041] Figure 8 This is a line graph comparing the MAE of this invention with other state-of-the-art models on the v component.

[0042] Figure 9 This is a line graph comparing the RMSE improvement rate of the u component under different SSH gradient (ocean surface height gradient) intensity bins for this invention.

[0043] Figure 10 This is a line graph comparing the RMSE improvement rate of the v component under different SSH gradient (ocean surface height gradient) intensity bins for this invention.

[0044] Figure 11 This is a line graph comparing the RMSE improvement rate of the u component under different wind speed intensities according to the present invention.

[0045] Figure 12 This is a line graph comparing the RMSE improvement rate of the v component under different wind speed intensities according to the present invention. Detailed Implementation

[0046] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. The following description is used to explain the technical solutions of the present invention and is not intended to limit the scope of protection of the present invention.

[0047] Reference Figure 1 The ocean surface flow field forecast data correction method based on physical-guided gated hybrid network described in this invention includes the following overall process: multi-source data acquisition, data preprocessing, basic flow field generation, known dynamic terms and residual calculation, spatiotemporal correction and missing force discovery, gated physical injection, output fusion, model training, and forecast data correction.

[0048] Step 1: Acquisition of multi-source data.

[0049] The numerical weather prediction (NRT) data for continuous time intervals is used to obtain surface current field, sea surface wind field, sea surface height field, spatial coordinate information, and temporal information for the sea area to be corrected. Specifically, the surface current field includes an eastward velocity component (u) and a northward velocity component (v); the sea surface wind field includes a 10m eastward wind component and a 10m northward wind component; and the sea surface height field is used to calculate the spatial gradient of sea surface height in the east-west and north-south directions.

[0050] Step 2, data preprocessing.

[0051] Based on the spatial resolution of the analyzed flow field and sea surface height, the high-resolution grid of the predicted flow field is downsampled to the analyzed flow field grid, while the reanalysis wind field is resampled to the analyzed flow field grid using bilinear interpolation. Longitude and latitude are normalized and scaled. Continuous physical variables such as predicted current velocity, sea surface wind speed, and sea surface height gradient are Z-score normalized. The annual daily sequence is converted into sinusoidal and cosine time-period features. A historical spatiotemporal input window of length L is constructed according to chronological order. After the above processing, the input channels within each time window include longitude, latitude, time-period features, predicted current velocity, 10m sea surface wind field, and sea surface height gradient.

[0052] Step 3: Generate the basic correction flow field.

[0053] Reference Figure 2 The historical spatiotemporal input window of length L is input into the basic solver Solver. The Solver adopts a point-shared multilayer perceptron structure, shares network parameters for each effective ocean grid point, and maps the multi-source features of the grid point within the historical window to the basic eastward current velocity and basic northward current velocity.

[0054] In one alternative implementation, the Solver comprises four linear mapping layers, each with a hidden layer dimension of 256, and includes a GELU activation function, a LayerNorm normalization layer, and a Dropout random deactivation layer between the hidden layers. The base corrected flow field output by the Solver serves as the background state for the final prediction and also as the state input for subsequent physical residual calculations.

[0055] Step 4: Calculate the known dynamic term and residual.

[0056] To incorporate physical information into real-world marine operational analysis, this invention does not directly treat the residuals of the two-dimensional shallow water momentum equation as a strong constraint that must approach zero. Instead, it first constructs known dynamic terms that can be calculated relatively stably using two-dimensional surface variables. The complete two-dimensional shallow water momentum equation can be found in the following formula:

[0057]

[0058]

[0059] in Coriolis parameters, It is the acceleration due to gravity. , These are the wind stresses in the x and y directions, respectively. For reference density, Here, H is the horizontal eddy viscosity coefficient, and H is the reference ocean height. Let be the bottom friction dissipation coefficient. The left side of the equation represents the nonlinear advection acceleration term, and the terms on the right side represent, in order, the Coriolis force, surface wind stress driving force, horizontal diffusion (second derivative term in space), barotropic pressure gradient force, and bottom friction dissipation. Considering the presence of three-dimensional baroclinic processes, complex terrain modulation, bottom friction, subgrid turbulence parameterization errors, and assimilation increments in real operational analysis fields, this invention extracts the portion of the two-dimensional surface momentum framework that can be calculated from two-dimensional variables into known and computable external forcing terms. :

[0060]

[0061]

[0062] In the known dynamic terms mentioned above, the sea surface height gradient is used to characterize the effect of the barotropic pressure gradient, the 10m sea surface wind field is used to estimate the wind stress driving effect, the spatial first and second order differences of the velocity field are used to estimate the advection term and the horizontal diffusion term, and the Coriolis parameter is determined according to the latitude of the sea area under study.

[0063] For real-world flow field analysis, the simplified two-dimensional surface equations are often not completely closed. To characterize this non-closure difference, this invention explicitly introduces a missing external forcing term dynamically predicted by a neural network in the spatiotemporal context. The missing external forcing term is used to represent complex dynamic processes that are not explicitly characterized by known two-dimensional dynamic terms:

[0064]

[0065]

[0066] During training, the actual time evolution tendency of the basic correction flow field is calculated using the grid center difference operator, and the difference between this actual time evolution tendency and the known dynamic term is defined as the dynamic residual. The dynamic residual is not forced to be compressed to zero, but rather serves as a physical supervision target for learning in the absence of external coercion:

[0067]

[0068]

[0069] With the above settings, the missing external forcing terms learned by the network have residual semantics, which can characterize unresolved three-dimensional baroclinic effects, variable depth topographic modulation, bottom boundary layer frictional dissipation, subgrid turbulence parameterization errors, and data assimilation increments, rather than simply serving arbitrary empirical compensation for data fitting.

[0070] After determining the equation form, this paper explicitly decomposes the final flow field correction into three functionally decoupled components: "basic correction," "pure data-driven increment," and "dynamic missing term injection," in order to construct a final output that balances flexibility and stability. , :

[0071]

[0072]

[0073] in, Smooth background correction is provided by a front-mounted point-by-point multilayer perceptron (MLP). This is a purely data-driven increment from the output of the ConvLSTM convolutional long short-term memory network for sequence models, responsible for absorbing high-frequency errors. For adaptive spatiotemporal confidence gating, it is used to suppress inappropriate physics injection in weak dynamics or high-noise regions, thereby structurally limiting inappropriate physics injection, and , This injects scale scalars into global physics. It should be noted that in the equations... This is not the microscopic discrete integration step size used in numerical models to satisfy the CFL stability condition (usually on the order of seconds), but rather the macroscopic characteristic time scale corresponding to the interval between two forecast outputs (6 hours). This represents the equivalent velocity increment accumulated by the instantaneous missing acceleration within the macroscopic sampling interval.

[0074] Step 5: Spatiotemporal correction and missing-forced discovery.

[0075] Reference Figure 3 The historical spatiotemporal input window is fed into the spatiotemporal corrector, the Corrector. The Corrector uses a ConvLSTM as its spatiotemporal recursive backbone to extract the temporal dependence and local spatial propagation characteristics of flow field errors within the historical window. In one optional implementation, the Corrector comprises three stacked ConvLSTM units, each with a 3×3 kernel size and 16 hidden channels.

[0076] The corrector extracts the hidden state from the last time step as the spatiotemporal representation of the current moment, and reweights the hidden features through a local spatial attention module. This local spatial attention module can consist of pointwise convolution and a sigmoid activation function, used to enhance the model's attention to areas affected by strong ocean currents, complex terrain, and regions with significant error changes.

[0077] Features after reweighting The input decoupling output headers respectively obtain data-driven increments. Missing external compulsion items And the spatiotemporal gating matrix G. The final correction is functionally divided into three components: basic correction, data-driven increment, and dynamic missing term injection. To maintain consistency with the parallel computation logic of tensors in the channel dimension, the above combination relationship can also be abstracted into a vectorized form:

[0078]

[0079] Step six, gated physical injection.

[0080] Since the missing external forcing term in the Corrector output resides in the normalized implicit space and cannot be directly added to the normalized velocity field, this invention converts it into a one-step equivalent velocity increment using a macroscopic time step, velocity standard deviation, and scale conversion factor. :

[0081]

[0082] In the formula, The scale transformation factor for acceleration from microscopic to macroscopic features. The macroscopic step size for the time integral of the actual business forecast (e.g., the number of seconds corresponding to 6 hours). The standard deviation vector corresponding to the predicted flow field (used for normalization) (Maintaining homogeneous scale). In particular, this paper introduces a learnable global injection scale parameter. (Activated via Softplus to ensure it is always greater than 0). The introduction separates the modeling of global-scale regulation from that of local gating regulation: by It is responsible for the adaptive amplification of the overall numerical magnitude of the physical channel, and is controlled by the gating matrix. Specifically responsible for gating and adjusting local grid points.

[0083] Since the gating matrix can only activate or suppress the physically equivalent velocity increment without changing the direction of the missing external forcing term, it can reduce the risk of inappropriate physical corrections in weak dynamic regions or high-noise regions.

[0084] Step 7: Perform output fusion.

[0085] The corrected ocean surface velocity field is obtained by adding the basic corrected flow field, the data-driven increment, and the gated physical injection increment. The output fusion process can be expressed as: Corrected velocity field = Basic corrected velocity field + Data-driven increment + Gating matrix × Global injection scale × Equivalent velocity increment.

[0086]

[0087] Step 8: Train the model.

[0088] Reference Figure 4 The training process employs a three-stage progressive strategy. First, the effective ocean grid set is defined. Model prediction field and analysis field supervision target And data loss is calculated only at valid points in the ocean:

[0089]

[0090] The first stage is the Solver pre-training stage. In this stage, the Corrector is frozen, and only the Solver is trained. This minimizes the data loss between the base corrected flow field and the analysis field, thereby obtaining a numerically smooth, spatially stable background base field that can be used for differential residual calculations. The loss function for this stage is... for:

[0091]

[0092] In the formula, , The analysis and supervision targets are u and v.

[0093] The second stage is the physics discovery stage. In this stage, the Solver is frozen, only the Corrector is trained, and gated injection is disabled. This allows the missing external forcing term to approximate the dynamic residuals calculated from the basic corrected flow field, while weak amplitude canonicalization is used to suppress abnormally large acceleration outputs. The loss function for this stage is... for:

[0094]

[0095] in This indicates the stopping gradient operator, used to prevent the residual itself from being reshaped during backpropagation. for The weak amplitude regularization coefficient.

[0096] The third stage is the joint optimization stage. In this stage, the data-driven increment, spatiotemporal gating matrix, and global injection scale are activated, and data reconstruction error, residual consistency constraints, teacher anchoring constraints, gating budget terms, information entropy regularization, and increment magnitude regularization are jointly considered. To prevent the data fitting objective from directly dragging missing external forcing terms away from their residual semantics, gradient blocking can be applied to the physical branch.

[0097]

[0098] Simultaneously, the Corrector after the second stage can be used as a teacher model to continuously anchor the missing external forcing terms in the joint optimization stage. This teacher anchoring loss function... for:

[0099]

[0100] in, This indicates the missing acceleration term in the teacher model output.

[0101] To prevent the gating matrix from degenerating into either all-open or all-closed, a budget term for the average activation ratio of the gating can be set; to enhance the local selectivity of the gating, information entropy regularization can be set. Let the expected average activation ratios of the gating on the u and v components be respectively... , The structured penalty term is then defined as:

[0102]

[0103]

[0104] in, Indicates the entire study area (oceanic effective grid) Take the spatial average value, and calculate the budget item loss. This is used to constrain the average opening ratio of the gating space, preventing the model from degenerating into extreme states of overall over-injection or overall closure; information entropy term loss. This encourages the gating values ​​to differentiate towards either 0 or 1, thereby enhancing their local selectivity and binarization tendency. Furthermore, to avoid excessive amplitude growth of pure data increments during the joint phase, additional measures are taken... Introducing amplitude regularization loss In summary, the total loss function in the joint optimization phase... It can be represented as:

[0105]

[0106] in, Indicates data loss. Represents physical residual loss The weighting coefficients, Represents the teacher anchoring loss function The weighting coefficients, Indicates budget item loss The weighting coefficients, Represents the loss of the information entropy term. The weighting coefficients, Indicates amplitude regularization loss The weighting coefficients.

[0107] Step nine: Correct the forecast data.

[0108] During the operational inference phase, the numerical weather prediction data for the ocean surface current field, sea surface wind field, sea surface height field, spatial coordinates, and temporal information at the time to be corrected are input into the trained ocean surface current field forecast data correction model. The model first generates a basic correction current field based on historical spatiotemporal input windows using a Solver network. Then, it calculates known dynamic terms and dynamic residuals based on the basic correction current field, sea surface wind field, and sea surface height gradient. Subsequently, it outputs data-driven increments, missing external forcing terms, and a spatiotemporal gating matrix through a spatiotemporal corrector network, and performs dimensional transformation and gated physical injection on the missing external forcing terms. Finally, it fuses the basic correction current field, data-driven increments, and gated physical injection increments to obtain the corrected ocean surface current field. When multi-step correction is required, the currently obtained corrected ocean surface current field can be recursively backfilled as part of the input window for subsequent times to obtain multi-step forecast data correction results.

[0109] Example

[0110] This embodiment uses the correction of the ocean surface current field in the sea area between 112.2° and 119.2° east longitude and 12.4° and 20.4° north latitude as an example. Input data include the HYCOM numerical forecast surface current velocity field, the ERA5 reanalysis 10m sea surface wind field, the CMEMS analysis sea surface height field, and CMEMS analysis flow field monitoring data. Both the forecast and analyzed flow fields contain two components: eastward velocity u and northward velocity v. The sea surface wind field contains two components: 10m eastward wind and 10m northward wind. The sea surface height field is used to obtain the east-west and north-south sea surface height gradients. The temporal resolution for all data types is set to 6 hours.

[0111] In this embodiment, the training samples are continuous data from 00:00 on February 1, 2023 to 18:00 on November 30, 2023, and the test samples are continuous data from 00:00 on December 1, 2023 to 18:00 on December 31, 2023. The input time window length is L=4, and the output is the corrected eastward flow velocity and corrected northward flow velocity on the analysis grid.

[0112] Reference Figure 5, Figure 6 , Figure 7 , Figure 8 On the overall test set, the method of this invention was compared with the original numerical prediction results and comparative models such as TCN, LSTM, 3DCNN-Unet, and ATCNs. The results show that the average RMSE and MAE of the method of this invention on the u component are 0.1229 and 0.0955, respectively, which are better than the original numerical prediction results of 0.1531 and 0.1209; the average RMSE and MAE on the v component are 0.1251 and 0.0960, respectively, which are better than the original numerical prediction results of 0.1501 and 0.1184.

[0113] Compared with TCN, LSTM, 3DCNN-Unet and ATCNs, the method of the present invention achieves smaller average RMSE and MAE values ​​on both u and v components, indicating that the method of the present invention has better correction effect in terms of overall error control and flow field fitting capability.

[0114] Reference Figure 9 , Figure 10 , Figure 11 , Figure 12 Under different dynamic intensities, sea surface wind speed modulus and sea surface height gradient modulus were used as binning indicators, and the data were divided into five intervals according to the quantiles of the training set: 0-50%, 50-80%, 80-95%, 95-99%, and 99-100%. The results show that in the high-intensity interval of the sea surface height gradient (99-100%), the RMSE improvement rate of the method of this invention relative to the original forecast data reaches 29.51% on the u-component and 16.41% on the v-component; in the high-intensity interval of wind speed (99-100%), the RMSE improvement rate of the method of this invention relative to the original forecast data reaches 22.80% on the u-component and 24.34% on the v-component.

[0115] The innovation of this invention is reflected in the following aspects:

[0116] First, this invention explicitly models the dynamic non-closure difference between the two-dimensional surface momentum equation and the real operational analysis field as a learnable missing external forcing term, rather than directly requiring the simplified physical equation residuals to approach zero, thereby mitigating the constraint mismatch problem of traditional strong physical constraints in the real ocean operational analysis field.

[0117] Second, the present invention adopts a hybrid network structure in which the basic solver, the data-driven incremental branch and the physical discovery injection branch are decoupled from each other. The stable background correction, high-frequency statistical error compensation and the forced discovery of missing data with dynamic meaning are modeled separately, which helps to reduce the target coupling in multi-objective joint optimization.

[0118] Third, this invention introduces a spatiotemporal gating matrix and a global injection scale into the physical injection path, enabling the physical equivalent velocity increment to be selectively injected according to the local dynamic state, thus avoiding over-correction caused by full physical injection in weak dynamic regions or noisy regions.

[0119] Fourth, this invention adopts a phased training strategy, first obtaining a stable basic correction field, then learning missing external forcing terms with residual semantics, and finally performing joint optimization. Furthermore, it mitigates the semantic drift of physical branches and gating collapse through gradient blocking, teacher anchoring, gating budget terms, and information entropy regularization.

[0120] Fifth, this invention is applicable to multi-step closed-loop rolling inference scenarios, and can recursively fill the current correction result back into the subsequent input window, maintaining good error correction capability and relative stability under complex dynamic conditions such as strong winds and strong sea surface height gradients.

Claims

1. A method for ocean surface current field prediction data correction based on a physically guided gating hybrid network, characterized in that, include: (1) Acquisition of multi-source data: Acquire the surface current field, sea surface wind field, sea surface height field, spatial coordinates and time information of the sea area to be corrected; wherein, the surface current field of the numerical forecast includes an eastward current component and a northward current component, and the sea surface wind field includes a 10m eastward wind component and a 10m northward wind component. (2) Data preprocessing: Register the surface current field, sea surface wind field and sea surface height field of the numerical forecast to the same spatial grid, calculate the sea surface height gradient, and normalize the continuous physical variables, spatial coordinates and time period characteristics to construct a historical spatiotemporal input window of length L; (3) Basic flow field generation: The basic solver Solver, which shares the input points of the historical spatiotemporal input window, generates a normalized basic correction flow field. The basic correction flow field serves as the background state of the final correction output and the state field for physical residual calculation. (4) Calculation of known dynamic terms and residuals: Based on the basic corrected flow field, sea surface wind field and sea surface height gradient, calculate the known dynamic terms in the two-dimensional surface momentum equation, and use time difference and spatial difference to calculate the dynamic residuals of the basic corrected flow field relative to the known dynamic terms; (5) Spatiotemporal correction and missing coercion discovery: The historical spatiotemporal input window is input into the spatiotemporal corrector Corrector, which outputs data-driven increments, missing external coercion terms and spatiotemporal gating matrices; wherein, the missing external coercion terms are used to characterize the dynamic processes and assimilation increments not explicitly described by the two-dimensional surface momentum equations. (6) Gated physical injection: The missing external forcing term is transformed to obtain a one-step equivalent velocity increment, and the one-step equivalent velocity increment is selectively injected using the spatiotemporal gating matrix and the learnable global injection scale. (7) Output fusion: The basic corrected flow field, the data-driven increment, and the gated physical injection increment are fused to obtain the corrected ocean surface velocity field; (8) model training: training a physics-guided gating hybrid network including the base solver Solver, the space-time corrector Corrector, the space-time gating matrix, and the global injection scale, so that the base solver Solver outputs a base correction flow field , the space-time corrector Corrector outputs data-driven increments , missing external forcing terms , and the space-time gating matrix G; (9) Forecast data correction: Input the numerical forecast surface current field, sea surface wind field, sea surface height field, spatial coordinates and time information of the time to be corrected into the trained physical-guided gated hybrid network model to obtain the corrected ocean surface forecast current field.

2. The method according to claim 1, characterized in that, The data preprocessing includes: normalizing and scaling longitude and latitude; normalizing the forecast current velocity, sea surface wind, sea surface height gradient, and monitored current velocity using Z-score; downsampling the forecast flow field to the analysis flow field grid and resampling the sea surface wind field to the analysis flow field grid using bilinear interpolation; mapping the annual day sequence to sinusoidal and cosine time periodic features, and combining them with physical variables to form the input channel.

3. The method according to claim 1, characterized in that, The input channels of the historical spatiotemporal input window include at least: longitude, latitude, sine time period characteristics, cosine time period characteristics, predicted eastward current velocity, predicted northward current velocity, 10m eastward wind, 10m northward wind, sea surface height eastward gradient, and sea surface height northward gradient.

4. The method according to claim 1, characterized in that, The basic solver Solver is a point-shared multilayer perceptron. The multilayer perceptron receives the multi-source feature vectors of each grid point within a historical window of length L. After multilayer linear mapping, nonlinear activation, normalization and random deactivation processing, it outputs the basic eastward flow velocity and basic northward flow velocity of each grid point.

5. The method according to claim 1, characterized in that, The known dynamic terms include at least the advection term, the Coriolis force term, the barotropic pressure gradient term corresponding to the sea surface height gradient, the wind stress driving term, and the horizontal diffusion term; the dynamic residual is the difference between the actual time evolution tendency of the basic corrected flow field and the known dynamic terms.

6. The method according to claim 1, characterized in that, The spatiotemporal corrector includes a spatiotemporal feature extraction module with a Convolutional Long Short-Term Memory (ConvLSTM) network as its backbone, a local spatial attention module, and a decoupled output head; the decoupled output head outputs the data-driven increment, the missing external forcing term, and the spatiotemporal gating matrix, respectively. The local spatial attention module uses pointwise convolution and the Sigmoid activation function to generate spatial attention weights, and reweights the hidden features output by the ConvLSTM channel by channel to enhance the model's ability to perceive regions with strong dynamic changes.

7. The method according to claim 1, characterized in that, The gated physical injection satisfies the following form: the missing external forcing term g is multiplied by the macroscopic time step ΔT, the scale conversion factor and the velocity standard deviation to obtain the equivalent velocity increment in the normalized space; then the equivalent velocity increment is multiplied by the Sigmoid gate matrix G and the global injection scale activated by Softplus, so that the physical injection is adaptively activated or suppressed in terms of spatial location and velocity components.

8. The method according to claim 1, characterized in that, Model training consists of three stages: the first stage freezes the Corrector and trains the Solver, enabling the Solver to learn the numerically smoothed basic correction flow field; In the second stage, the Solver is frozen and gated physical injection is turned off, and the missing external forcing term of the Corrector output is trained to approximate the dynamic residual. The third stage freezes the Solver, activates the data-driven increment, gating matrix, and global injection scale, and performs joint optimization on the Corrector. In the third stage joint optimization, gradient blocking or teacher anchoring constraints are set for missing external coercives, and gating budget regularization and information entropy regularization are set for the spatiotemporal gating matrix to suppress semantic drift and gating collapse due to missing external coercives.

9. The method according to claim 1, characterized in that, The forecast data correction includes: inputting the numerical forecast surface current field, sea surface wind field, sea surface height field, spatial coordinates and time information of the time to be corrected into a trained physical-guided gated hybrid network model; the model then sequentially performs basic flow field generation, known dynamic terms and residual calculation, spatiotemporal correction and missing force detection, gated physical injection and output fusion to obtain the corrected ocean surface forecast current field.

Images