A method and system for intelligent reconstruction and prediction of three-dimensional flow field of ocean

Abstract

The application discloses a method and system for intelligent reconstruction and prediction of a marine three-dimensional flow field, and relates to the technical field of data processing. The method comprises the following steps: obtaining multi-source observation data of the marine through a large language model and an edge monitoring device; preprocessing and fusing the multi-source observation data to obtain fused observation data; constructing a spatiotemporal correlation graph by taking a target observation parameter in the fused observation data as a node and taking the spatiotemporal correlation and the physical quantity coupling relationship between the target observation parameters as edges; determining reconstruction data corresponding to the multi-source observation data according to the spatiotemporal correlation graph; and predicting a three-dimensional flow field of the marine according to the reconstruction data and outputting a visualized three-dimensional flow field. The application can fuse and reconstruct the multi-source observation data, thereby improving the accuracy of the prediction of the three-dimensional flow field of the marine.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and system for intelligent reconstruction and prediction of three-dimensional ocean flow fields. Background Technology

[0002] Ocean three-dimensional flow field prediction is a core area of ​​ocean dynamics research, with significant value for climate change simulation, pollution diffusion prediction, resource development, and military strategy. However, traditional prediction methods face multiple technical bottlenecks: ① Scarcity of observational data: While satellite remote sensing technology can achieve global coverage, it can only acquire surface information. Underwater three-dimensional flow field reconstruction relies on costly and platform-stable profile measurements (such as ADCP and current meters), resulting in scarce data resources. ② High computational complexity of models: Traditional numerical models (such as POM and HYCOM) require solving physical equations, leading to extremely high computational costs and limited accuracy, making it difficult to meet real-time requirements. ③ Multivariate coupling challenges: Ocean flow fields involve spatiotemporal coupling of multiple variables such as temperature, salinity, and wind stress. Traditional statistical methods (such as ARIMA and SVR) struggle to capture nonlinear relationships and have short prediction timelines. In recent years, artificial intelligence technology has provided new ideas for ocean flow field prediction, but existing solutions still have limitations: ① Single-mode limitations: Some methods focus only on surface flow fields or vertical velocity reconstruction, lacking a holistic representation of the entire three-dimensional flow field. ② Strong data dependence: Deep learning models require large-scale historical data support, are prone to overfitting, and have poor interpretability. Summary of the Invention

[0003] In view of this, embodiments of this application provide a method and system for intelligent reconstruction and prediction of three-dimensional ocean flow fields, so as to improve the accuracy of predicting three-dimensional ocean flow fields.

[0004] One aspect of this application provides a method for intelligent reconstruction and prediction of a three-dimensional ocean flow field, the method comprising the following steps:

[0005] Acquire multi-source ocean observation data using large language models and edge monitoring equipment;

[0006] The multi-source observation data is preprocessed and fused to obtain fused observation data;

[0007] Using the target observation parameters in the fused observation data as nodes, and the spatiotemporal correlation and physical quantity coupling relationship between the target observation parameters as edges, a spatiotemporal correlation graph is constructed.

[0008] Based on the spatiotemporal correlation map, the reconstructed data corresponding to the multi-source observation data is determined;

[0009] The three-dimensional flow field of the ocean is predicted based on the reconstructed data, and the visualized three-dimensional flow field is output.

[0010] In some embodiments, acquiring multi-source observation data of the ocean through large language models and edge monitoring devices includes the following steps:

[0011] The user command is parsed using the large language model to obtain the profile observation data corresponding to the user command; the profile observation data is projected onto the spatiotemporal grid of the satellite altimeter to calculate the spatiotemporal matching degree of the profile observation data; the profile observation data with a spatiotemporal matching degree lower than a set threshold is deleted to obtain the first observation data;

[0012] The ADCP device in the edge monitoring device is used to acquire mobile ADCP data; Kalman filtering is used to remove noise from the mobile ADCP data; and then the flow velocity profile is reconstructed on the mobile ADCP data by the edge GPU to obtain the second observation data.

[0013] The first observation data and the second observation data are used as the multi-source observation data.

[0014] In some embodiments, the preprocessing and fusion of the multi-source observation data to obtain fused observation data includes the following steps:

[0015] The barotropic mode features and the first baroclinic mode features are extracted from the multi-source observation data;

[0016] Set the 3D mesh and boundary conditions for the FVCOM model;

[0017] The barotropic modal features and the first baroclinic modal features are projected onto the three-dimensional mesh of the FVCOM numerical model for spatiotemporal alignment.

[0018] The quality threshold is adjusted according to the task type, and the spatiotemporally aligned barotropic mode features and the first barotropic mode features are cleaned according to the quality threshold to obtain quality assessment features;

[0019] The edge monitoring device is used to extract vertical modal features from the acquired multi-source observation data;

[0020] Ekman flow features coupled with the barotropic mode features are extracted from the multi-source observation data obtained by the satellite in the edge monitoring device.

[0021] The quality assessment features, the vertical modal features, and the Ekman flow features are fused to obtain the fused observation data.

[0022] In some embodiments, the method further includes the following steps:

[0023] Adjust the feature weights of the target observation parameters according to the task requirements to adjust the corresponding edges in the spatiotemporal correlation graph.

[0024] In some embodiments, determining the reconstructed data corresponding to the multi-source observation data based on the spatiotemporal correlation map reconstruction includes the following steps:

[0025] The FVCOM model outputs a gridded hydrodynamic field based on the spatiotemporal correlation map;

[0026] The hydrodynamic field in the form of time-series data is input into the LSTM network, physical constraint terms are embedded into the LSTM network, and the sampling frequency is dynamically adjusted through an event-triggered mechanism.

[0027] Runoff contribution is calculated based on the SCS-CN model. Vertical velocity distribution is simulated through generalized unit hydrographs. A multi-scale attention mechanism is introduced to dynamically allocate computational resources to the mesoscale eddy region.

[0028] Optimize the parameters of the FVCOM model.

[0029] In some embodiments, optimizing the parameters of the FVCOM model includes the following steps:

[0030] Features of the three-dimensional flow field output by the FVCOM model and the fused observation data are extracted separately, and residuals are calculated.

[0031] The state space is constructed based on the residuals, and the FVCOM model is then subjected to reinforcement learning.

[0032] The weights of the FVCOM model are dynamically adjusted through a policy network and a value network; wherein the policy network is used as input to the residual feature map and outputs the action probability distribution; the value network is used as input to the state vector and outputs the expected reward.

[0033] The FVCOM model is coordinated and optimized based on the residual feature map transmitted by the edge monitoring device.

[0034] In some embodiments, the output visualization of the three-dimensional flow field includes the following steps:

[0035] The contour lines or vector maps of the three-dimensional flow field are output through a digital twin engine, AR / VR visualization terminal, BIM-GIS collaborative platform, or edge monitoring device.

[0036] Another aspect of this application embodiment provides a system for intelligent reconstruction and prediction of three-dimensional ocean current fields, the system comprising:

[0037] The data acquisition unit is used to acquire multi-source observation data of the ocean through large language models and edge monitoring equipment;

[0038] The data fusion unit is used to preprocess and fuse the multi-source observation data to obtain fused observation data;

[0039] The graph construction unit is used to construct a spatiotemporal correlation graph using the target observation parameters in the fused observation data as nodes and the spatiotemporal correlation and physical quantity coupling relationship between the target observation parameters as edges.

[0040] The data reconstruction unit is used to reconstruct and determine the reconstructed data corresponding to the multi-source observation data based on the spatiotemporal correlation map.

[0041] The prediction output unit is used to predict the three-dimensional flow field of the ocean based on the reconstructed data and output a visualization of the three-dimensional flow field.

[0042] Another aspect of this application embodiment provides an electronic device, including a processor and a memory;

[0043] The memory is used to store programs;

[0044] The processor executes the program to implement any of the methods described above.

[0045] Another aspect of this application provides a computer-readable storage medium storing a program that is executed by a processor to implement the method described in any of the above embodiments.

[0046] This application includes at least the following beneficial effects:

[0047] This application acquires multi-source ocean observation data using large language models and edge monitoring equipment; preprocesses and fuses the multi-source observation data to obtain fused observation data; constructs a spatiotemporal correlation graph using target observation parameters in the fused observation data as nodes and spatiotemporal correlations and physical quantity coupling relationships between target observation parameters as edges; reconstructs the corresponding reconstructed data based on the spatiotemporal correlation graph; and predicts and outputs a visualized three-dimensional ocean flow field based on the reconstructed data. This application's fusion and reconstruction of multi-source observation data improves the accuracy of predicting the three-dimensional ocean flow field. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 A flowchart illustrating a method for intelligent reconstruction and prediction of a three-dimensional ocean flow field provided in this application embodiment;

[0050] Figure 2 This is a schematic diagram of the structure of a system for intelligent reconstruction and prediction of a three-dimensional ocean flow field provided in an embodiment of this application. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0052] Before providing a detailed description of the embodiments of this application, some related technologies involved in the embodiments of this application will be described first, as follows:

[0053] Related technologies:

[0054] (1) Satellite remote sensing and Argo buoy joint reconstruction method:

[0055] A method is proposed that uses geostrophic and Ekman flows retrieved from satellite altimeters and scatterometers, combined with 1000-meter depth data from Argo buoys, to reconstruct a three-dimensional flow field by extracting the barotropic and first baroclinic modes. This method verifies the mapping relationship between satellite and Argo data, but it relies on multi-source data fusion and does not solve the challenge of direct remote sensing detection of sea surface vector flows.

[0056] (2) Traditional three-dimensional flow field numerical simulation:

[0057] Numerical models such as the Princeton Ocean Model (POM) are used to simulate the flow field by discretizing and solving the equations of motion. The advantage lies in the clear physical mechanisms, but it consumes significant computational resources and makes high-resolution real-time prediction difficult.

[0058] Other related technologies:

[0059] (1) EOF-LSTM coupled model for surface flow field prediction:

[0060] Existing methods extract spatial modes using empirical orthogonal functions (EOF) and combine them with long short-term memory networks (LSTM) to process time-series data for ocean surface current field prediction. However, these methods do not consider multivariate coupling and are only applicable to two-dimensional space.

[0061] (2) Deep learning model for vertical flow velocity reconstruction:

[0062] A deep convolutional neural network (CNN) based on sea surface height field is proposed to directly predict vertical current velocity field. Compared with traditional dynamic diagnostic methods (such as eSQG), it has higher accuracy and requires less data, but only focuses on vertical current velocity reconstruction.

[0063] (3) Transformer model for multivariate three-dimensional field prediction:

[0064] A 3D-Geoformer model is constructed, utilizing a spatiotemporal attention mechanism to simulate multivariate three-dimensional ocean-atmosphere fields (such as ENSO-related sea surface temperature and wind stress fields), with prediction lead times exceeding 18 months. This method overcomes the limitations of traditional RNNs, but requires multivariate datasets from specific regions.

[0065] Existing technologies have made progress in satellite data fusion, numerical simulation optimization, and deep learning applications, but problems remain, such as strong data dependence, insufficient multivariate coupling, and separation between vertical flow fields and full three-dimensional flow fields. This application aims to combine multi-source data with intelligent algorithms to solve the above-mentioned technical bottlenecks and achieve high-precision and timely intelligent prediction of ocean three-dimensional flow fields.

[0066] Disadvantages of existing technology:

[0067] (1) Strong data dependence and high acquisition cost: Traditional three-dimensional flow field reconstruction relies on high-cost, high-platform-stability profile measurement (such as ADCP, ocean current meter), and requires multi-source data (satellite + Argo) fusion, and data resources are scarce.

[0068] (2) High computational complexity, making real-time prediction difficult: Numerical models (such as POM) need to be solved based on physical equations, which consumes a lot of computational resources and makes it difficult to achieve high-resolution real-time prediction.

[0069] (3) Insufficient multivariate coupling and three-dimensional full flow field representation: Existing methods (such as EOF-LSTM and vertical velocity reconstruction models) only focus on a single variable (such as surface flow field or vertical velocity), and have not achieved overall three-dimensional full flow field representation and multivariate spatiotemporal coupling.

[0070] (4) Poor model interpretability and easy overfitting: Although deep learning models can handle complex relationships, they rely on large-scale historical data, are prone to overfitting, and lack physical mechanism explanation.

[0071] (5) Lack of direct remote sensing technology for sea surface vector flow: Traditional methods rely on indirect inversion (such as geostrophic flow + Ekman flow), which has not solved the problem of high-precision direct observation of sea surface vector flow.

[0072] The purpose of this invention, in view of the above-mentioned shortcomings, is to:

[0073] (1) Reduce data dependence and acquisition cost: By optimizing the mapping relationship between satellite remote sensing data and Argo buoy data, the dependence on high-cost profile measurement data is reduced, and efficient three-dimensional flow field reconstruction is achieved.

[0074] (2) Improve computational efficiency and real-time performance: Combine intelligent algorithms (such as deep learning) to replace some physical equations, reduce computational complexity, and meet the requirements of high-resolution real-time prediction.

[0075] (3) Realize three-dimensional full flow field and multivariate coupling representation: Through multimodal data fusion and intelligent algorithm design, the limitations of traditional models in coupling vertical flow field, horizontal flow field and multivariate (temperature, salinity, wind stress) are overcome, and a full three-dimensional flow field prediction system is constructed.

[0076] (4) Enhance model interpretability and stability: Introduce physical prior knowledge into the deep learning framework, balance data-driven and physical constraints, improve the model's generalization ability and reduce the risk of overfitting.

[0077] (5) Direct detection of sea surface vector flow: Develop vector flow inversion technology based on satellite altimeters and scatterometers to solve the problem of direct remote sensing observation of sea surface flow fields and improve the quality of data sources.

[0078] Reference Figure 1 This application provides a method for intelligent reconstruction and prediction of three-dimensional ocean current fields, specifically including the following steps S100~S140:

[0079] S100: Acquires multi-source ocean observation data through large language models and edge monitoring equipment;

[0080] S110: The multi-source observation data is preprocessed and fused to obtain fused observation data;

[0081] S120: Using the target observation parameters in the fused observation data as nodes and the spatiotemporal correlation and physical quantity coupling relationship between the target observation parameters as edges, construct a spatiotemporal correlation graph;

[0082] S130: Determine the reconstructed data corresponding to the multi-source observation data based on the spatiotemporal correlation map reconstruction;

[0083] S140: Based on the reconstructed data, predict the three-dimensional flow field of the ocean and output a visualized version of the three-dimensional flow field.

[0084] Optionally, acquiring multi-source ocean observation data through large language models and edge monitoring devices includes the following steps:

[0085] The user command is parsed using the large language model to obtain the profile observation data corresponding to the user command; the profile observation data is projected onto the spatiotemporal grid of the satellite altimeter to calculate the spatiotemporal matching degree of the profile observation data; the profile observation data with a spatiotemporal matching degree lower than a set threshold is deleted to obtain the first observation data;

[0086] The ADCP device in the edge monitoring device is used to acquire mobile ADCP data; Kalman filtering is used to remove noise from the mobile ADCP data; and then the flow velocity profile is reconstructed on the mobile ADCP data by the edge GPU to obtain the second observation data.

[0087] The first observation data and the second observation data are used as the multi-source observation data.

[0088] Optionally, the preprocessing and fusion of the multi-source observation data to obtain fused observation data includes the following steps:

[0089] The barotropic mode features and the first baroclinic mode features are extracted from the multi-source observation data;

[0090] Set the 3D mesh and boundary conditions for the FVCOM model;

[0091] The barotropic modal features and the first baroclinic modal features are projected onto the three-dimensional mesh of the FVCOM numerical model for spatiotemporal alignment.

[0092] The quality threshold is adjusted according to the task type, and the spatiotemporally aligned barotropic mode features and the first barotropic mode features are cleaned according to the quality threshold to obtain quality assessment features;

[0093] The edge monitoring device is used to extract vertical modal features from the acquired multi-source observation data;

[0094] Ekman flow features coupled with the barotropic mode features are extracted from the multi-source observation data obtained by the satellite in the edge monitoring device.

[0095] The quality assessment features, the vertical modal features, and the Ekman flow features are fused to obtain the fused observation data.

[0096] Optionally, the method further includes the following steps:

[0097] Adjust the feature weights of the target observation parameters according to the task requirements to adjust the corresponding edges in the spatiotemporal correlation graph.

[0098] Optionally, determining the reconstructed data corresponding to the multi-source observation data based on the spatiotemporal correlation map includes the following steps:

[0099] The FVCOM model outputs a gridded hydrodynamic field based on the spatiotemporal correlation map;

[0100] The hydrodynamic field in the form of time-series data is input into the LSTM network, physical constraint terms are embedded into the LSTM network, and the sampling frequency is dynamically adjusted through an event-triggered mechanism.

[0101] Runoff contribution is calculated based on the SCS-CN model. Vertical velocity distribution is simulated through generalized unit hydrographs. A multi-scale attention mechanism is introduced to dynamically allocate computational resources to the mesoscale eddy region.

[0102] Optimize the parameters of the FVCOM model.

[0103] Optionally, optimizing the parameters of the FVCOM model includes the following steps:

[0104] Features of the three-dimensional flow field output by the FVCOM model and the fused observation data are extracted separately, and residuals are calculated.

[0105] The state space is constructed based on the residuals, and the FVCOM model is then subjected to reinforcement learning.

[0106] The weights of the FVCOM model are dynamically adjusted through a policy network and a value network; wherein the policy network is used as input to the residual feature map and outputs the action probability distribution; the value network is used as input to the state vector and outputs the expected reward.

[0107] The FVCOM model is coordinated and optimized based on the residual feature map transmitted by the edge monitoring device.

[0108] Optionally, the output visualization of the three-dimensional flow field includes the following steps:

[0109] The contour lines or vector maps of the three-dimensional flow field are output through a digital twin engine, AR / VR visualization terminal, BIM-GIS collaborative platform, or edge monitoring device.

[0110] The following section will provide a detailed introduction and explanation of the solutions in the embodiments of this application, using specific application examples.

[0111] The solution in this embodiment includes 5 core modules, each of which is described in detail below:

[0112] Core Module 1: Multi-source observation data acquisition layer.

[0113] It integrates data sources such as satellite remote sensing (altimeter / scatterometer), shore-based high-frequency ground wave radar, mobile ADCP, weather stations, and Argo buoys. It proposes a three-in-one architecture of "command-driven - dynamic evaluation - edge collaboration," achieving for the first time intelligent, adaptive, and low-latency characteristics for multi-source marine data acquisition. The constructed modules consist of:

[0114] ① Command parsing engine: Based on a marine-specific model fine-tuned by GPT-4, it parses user commands to generate data requirement JSON (such as "Get ADCP current velocity data of East China Sea region at 14:00 on March 11, 2025, with an accuracy of ±0.1m / s").

[0115] ② Dynamic Data Quality Assessment Module:

[0116] Spatiotemporal consistency assessment: The spatiotemporal matching degree of the data source is calculated through the DUViT spatiotemporal attention mechanism (improvement of image registration technology).

[0117] Multi-dimensional quality indicators: including satellite data inversion error (radiative correction method), ADCP profile integrity (data integrity detection algorithm), and other indicators.

[0118] ③ Edge Collaborative Network: Lightweight processing nodes deployed on shore-based radar and mobile monitoring equipment to perform preprocessing tasks such as data compression and outlier filtering.

[0119] ④ Data Fusion Hub: Constructing a semantic association graph of multi-source data based on graph neural networks. The specific implementation steps for constructing the multi-source observation data acquisition layer are as follows:

[0120] Step 1: Instruction parsing and task decomposition. Input natural language instructions → Enhance semantic understanding through domain dictionaries (e.g., mapping "ADCP" to "mobile Doppler current profiler") to generate structured tasks.

[0121] Step 2: Dynamic data quality assessment and spatiotemporal consistency calculation: Project the ADCP profile data onto the spatiotemporal grid of the satellite altimeter using the DUViT model, and calculate the spatiotemporal matching degree.

[0122] .

[0123] Step 3: Edge collaborative data acquisition.

[0124] Intelligent task distribution: Prioritize sending high-priority tasks (matching degree > 0.85 and accuracy meet the standard) to ADCP devices, and adopt incremental update mode for weather stations (only transmit parameters that have changed by more than ±5%).

[0125] Edge preprocessing pipeline: Walk-through ADCP data: Kalman filtering is used to remove noise (data cleaning technique), and velocity profile reconstruction is performed through edge GPU (spatial analysis engine optimization).

[0126] Shore-based radar data: Automatic completion of missing sectors based on spatiotemporal correlation maps (improved image fusion method).

[0127] Step 4: Data fusion and quality verification.

[0128] Constructing a semantic association graph of multi-source data (semantic alignment technology extension): Nodes: ADCP flow velocity, satellite altimeter geostrophic flow, meteorological station wind stress; Edges: spatiotemporal correlation (Pearson correlation coefficient), physical quantity coupling relationship.

[0129] Quality verification: If the quality score of a data source is <0.6, a backup data source (such as Argo buoy) is automatically triggered to complete the data. The weight of the data source is dynamically adjusted through reinforcement learning (real-time data warehouse technology).

[0130] Core Module 2: Data Preprocessing and Fusion Layer.

[0131] Satellite data: Surface flow field inversion using geostrophic flow and Ekman flow;

[0132] Profile data: ADCP and Argo buoy data were used for vertical mode extraction;

[0133] Meteorological data: input of driving factors such as wind stress and air pressure.

[0134] We propose a three-in-one architecture of "dynamic modality selection, spatiotemporal alignment enhancement, and edge collaboration enhancement" to achieve intelligent, adaptive, and high-precision characteristics in the preprocessing of multi-source marine data.

[0135] Main modules for building a data preprocessing and fusion layer system:

[0136] ① Dynamic mode selection engine: Based on an improved algorithm of CEOF analysis, it automatically identifies the barotropic mode (90% kinetic energy) and the first barotropic mode.

[0137] ② Spatiotemporal alignment enhancement module: Satellite data: projected onto the FVCOM mesh through the DUViT model; Profile data: spatiotemporal interpolation using Kalman filtering (error <0.3m / s).

[0138] ③ Edge Collaborative Enhancement Network: A lightweight processing node deployed on ADCP and shore-based radar to perform flow profile reconstruction and feature extraction.

[0139] ④ Multimodal fusion hub: Construct a data semantic association graph based on graph neural network (nodes contain multi-dimensional features such as flow velocity, temperature, and salinity).

[0140] The specific implementation steps are as follows:

[0141] Step 1: Dynamic mode selection and initialization. Input the original dataset (satellite altimeter, ADCP, weather station, etc.), extract the barotropic mode (eigenvalue λ1>λ2) and the first baroclinic mode (eigenvalue λ2>λ3) through CEOF analysis, initialize the FVCOM 3D mesh, and set boundary conditions (such as tidal and wind stress).

[0142] Step 2: Spatiotemporal alignment enhancement, satellite data alignment: the geostrophic flow retrieved from the altimeter is projected onto the FVCOM grid, improving the spatial resolution to 2km×2km, and the Ekman flow retrieved from the scatterometer is spatiotemporally matched with Argo buoy data through the DUViT model (matching degree >0.85); profile data enhancement: ADCP data: noise is removed by Kalman filtering at the edge nodes (residual <0.1m / s), underway ADCP: velocity profile reconstruction is performed through the edge GPU.

[0143] Step 3: Dynamic quality assessment of multimodal data, constructing a multidimensional quality scoring system, and setting dynamic thresholds: automatically adjusting the quality thresholds according to the task type (such as oil spill diffusion prediction).

[0144] Step 4: Edge Collaborative Mode Decomposition, ADCP Data: Perform EOF decomposition at edge nodes to extract vertical modes (the first 3 principal modes explain 85% of the energy), and optimize mode weights through reinforcement learning (reward function based on FVCOM simulation error).

[0145] Satellite data: After complex transformation (us,t=us,t+ivs,t), scatterometer data is decomposed by CEOF to extract Ekman flow characteristics coupled with barotropic modes.

[0146] Step 5: Multimodal data semantic fusion to construct a semantic association graph with 18-dimensional features: Nodes: flow velocity, temperature, salinity, wind stress, air pressure, etc.; Edges: physical quantity coupling relationship, spatiotemporal correlation (Pearson coefficient > 0.7).

[0147] Dynamic graph attention mechanism: Automatically adjusts feature weights based on task requirements (such as pollution diffusion).

[0148] Step 6: Data assimilation and quality verification, residual analysis: Calculate the residual between the observed data and the model prediction (threshold ± 0.2 m / s). If the residual of a certain data source is greater than the threshold, automatically trigger the backup data source (such as Argo buoy) to complete the data. Reinforcement learning optimization: Dynamically adjust the coupling weights between FVCOM and the intelligent model through the PPO algorithm (improving convergence speed).

[0149] Core Module 3: Intelligent Refactoring Engine.

[0150] PI-LSTM time series model: Processes multivariate time series data (such as temperature, salinity, and flow velocity) and captures nonlinear dynamic characteristics;

[0151] Generalized unitline module: simulates the coupling of hydrological processes (such as runoff and evaporation) with three-dimensional flow fields;

[0152] FVCOM physical model: as the basic dynamic framework, it provides a gridded hydrodynamic field.

[0153] A three-in-one architecture of "command-driven, dynamic evaluation, and edge collaboration" is proposed to achieve intelligent, adaptive, and low-latency characteristics for multi-source marine data acquisition.

[0154] Intelligent Reconfiguration Engine System Architecture:

[0155] (1) Command parsing engine: Based on the marine domain-specific model fine-tuned by GPT-4, it parses user commands to generate data requirement JSON (such as "Get ADCP current velocity data of East China Sea region at 14:00 on March 11, 2025, with an accuracy of ±0.1m / s").

[0156] (2) Dynamic data quality assessment module: Spatiotemporal consistency assessment: calculates the spatiotemporal matching degree of the data source through the DUViT spatiotemporal attention mechanism (improvement of image registration technology); multi-dimensional quality indicators: including satellite data inversion error (radiocorrection method), ADCP profile integrity (data integrity detection algorithm) and other indicators.

[0157] (3) Edge Collaborative Network: Lightweight processing nodes deployed on shore-based radar and mobile equipment to perform preprocessing tasks such as data compression (adaptive compression algorithm) and outlier filtering.

[0158] (4) Data fusion hub: Construct a semantic association graph of multi-source data based on graph neural network (semantic alignment technology extension).

[0159] The specific implementation steps are as follows:

[0160] Step 1: Instruction parsing and task decomposition: Input natural language instructions → Enhance semantic understanding through domain dictionary (e.g., "ADCP" is mapped to "mobile Doppler current profiler").

[0161] Step 2: Dynamic data quality assessment and spatiotemporal consistency calculation: Project ADCP profile data onto the spatiotemporal grid of the satellite altimeter using the DUViT model, and calculate the spatiotemporal matching degree.

[0162] Step 3: Edge-coordinated data acquisition and intelligent task distribution: Prioritize sending high-priority tasks (matching degree > 0.85 and accuracy met) to ADCP devices, and adopt incremental update mode for weather stations (only transmit parameters that have changed by more than ±5%); Edge preprocessing pipeline: Mobile ADCP data: Use Kalman filtering to remove noise (data cleaning technology), and perform velocity profile reconstruction through edge GPU (spatial analysis engine optimization); Shore-based radar data: Automatically complete missing sectors based on spatiotemporal correlation maps (improved image fusion method).

[0163] Step 4: Data Fusion and Quality Verification. Construct a spatiotemporal correlation graph of multi-source data (with semantic alignment technology extension): Nodes: ADCP flow velocity, satellite altimeter geostrophic flow, meteorological station wind stress; Edges: spatiotemporal correlation (Pearson correlation coefficient), physical quantity coupling relationships. Quality Verification: If the quality score of a data source is <0.6, an alternative data source (such as Argo buoys) is automatically triggered for completion. Data source weights are dynamically adjusted through reinforcement learning (real-time data warehouse technology).

[0164] Step 5: Core computation of the intelligent refactoring engine:

[0165] ① FVCOM physical model initialization: Generate an unstructured triangular mesh based on SMS, set boundary conditions such as tidal and wind stress, and extract the barotropic mode (90% kinetic energy) and the first baroclinic mode through CEOF analysis.

[0166] ②PI-LSTM temporal modeling: Input temporal data such as temperature, salinity, and wind stress into the LSTM network, embed physical constraint terms (such as momentum equations), and dynamically adjust the sampling frequency through an event triggering mechanism (such as increasing it to 1Hz when the flow velocity changes abruptly).

[0167] ③ Generalized unitline coupling: Runoff contribution is calculated based on the SCS-CN model, and vertical velocity distribution is simulated through generalized unitlines (Guh distribution). A multi-scale attention mechanism is introduced to dynamically allocate computational resources to the mesoscale eddy region.

[0168] ④ Data assimilation and optimization: The PPO algorithm is used to optimize the coupling weight between FVCOM and the intelligent model. The residual threshold is set to ±0.2m / s. Adaptive mesh refinement technology: In the region where the velocity gradient is >0.5m / s, the mesh size is reduced from 2km to 500m.

[0169] Step 6: Prediction and visualization output, outputting a 3D flow field prediction (contour lines / vector) for the next 72 hours, integrating a pollution diffusion module, and simulating a schematic diagram of the oil spill transport path.

[0170] Core Module 4: Data Assimilation and Optimization Layer.

[0171] Reinforcement learning algorithms (such as PPO) optimize model parameters and combine them with residual inversion from observation data;

[0172] Multi-scale attention mechanism: balancing global modes (such as barotropic modes) with local details (such as mesoscale eddies).

[0173] The data assimilation and optimization layer module system architecture consists of:

[0174] (1) Residual feature extractor: a multi-scale residual analysis module based on spatiotemporal attention mechanism.

[0175] (2) Reinforcement learning optimizer: a dynamic weight adjustment engine using the PPO algorithm.

[0176] (3) Multi-scale fusion hub: A fusion interface that integrates FVCOM grid data and intelligent model prediction results.

[0177] (4) Edge-Center Communication Layer: Supports lightweight feature transmission.

[0178] Specific implementation process:

[0179] Step 1: Residual calculation and feature extraction. Input the three-dimensional flow field Qpred(x,y,z,t) predicted by the FVCOM model and the multi-source observation data O(x,y,z,t) (including satellite, ADCP, Argo, etc.). Output the residual field E(x,y,z,t) = O(x,y,z,t)−Qpred(x,y,z,t).

[0180] Step 2: Strengthen the construction of the learning environment.

[0181] State space:

[0182] The global residual variance is as follows:

[0183] ;

[0184] The local vortex intensity is as follows:

[0185] ;

[0186] Action space:

[0187] The data source weight adjustment coefficient wd∈[0.1,0.9];

[0188] The perturbations in the model parameters are as follows:

[0189] .

[0190] Step 3: Dynamic weight adjustment, policy network (Actor): input residual feature map, output action probability distribution.

[0191] The policy network (Actor) is a core component in reinforcement learning, its function being to generate a probability distribution of actions based on the current environmental state. In this embodiment, the policy network adopts a spatiotemporal attention-enhanced multilayer perceptron (MLP) architecture, with the specific structure as follows:

[0192] ① Input layer: Receives multi-scale residual feature maps (such as low-frequency, mid-frequency, and high-frequency features after spatial pyramid decomposition);

[0193] ② Spatiotemporal attention layer: Three-dimensional spatial pyramid decomposition (scale factors of 1×1×1, 2×2×2, 4×4×4) is used to perform multi-scale downsampling of residual features; a spatiotemporal attention module (STAttention) is introduced to focus on key flow field error regions (such as mesoscale vortex regions) through dynamic weight allocation.

[0194] ③ Policy network layer: It contains 3 fully connected layers (FC1: 256→128, FC2: 128→64, FC3: 64→32), with the activation function being Swish; the output layer uses the Softmax function to generate the action probability distribution, and the action space includes: the data source weight adjustment coefficient wd∈[0.1,0.9]; and the model parameter perturbation (continuous actions).

[0195] The dynamic weight adjustment of the policy network in the dynamic weight adjustment mechanism is the core innovation of this embodiment, and its specific implementation is as follows:

[0196] ① State-space modeling:

[0197] Global state: The global variance of the flow field residuals is:

[0198] .

[0199] Local state: Vortex intensity is:

[0200] ;

[0201] Extracted by eigenvalue decomposition of the Hessian matrix.

[0202] ②Motion space design:

[0203] Data source weight adjustment: KL divergence constraint (λ=0.01) of PPO algorithm is used to prevent weight degradation; dynamic weight allocation: satellite data weight is increased in frontal regions and ADCP data weight is increased in vortex regions; model parameter perturbation range is adaptively adjusted according to residual intensity.

[0204] ;

[0205] (α is 0.1 in the calm region and can reach 0.5 in the strong vortex region).

[0206] ③ Reinforcement learning training: The optimization algorithm adopts the PPO truncation strategy gradient (clip ratio=0.2), and updates 500 times in each iteration.

[0207] The Critic network takes a state vector as input and outputs the expected reward. The Critic system architecture consists of: ① Spatiotemporal attention layer: employing a three-dimensional spatial pyramid decomposition (scale factors 1×1×1, 2×2×2, 4×4×4) and a STAttention module; ② Multimodal fusion layer: integrating a semantic association graph of 18-dimensional features such as flow velocity, temperature, salinity, and wind stress; ③ Dynamic value evaluation layer: including physical constraint terms (such as the momentum equation) and an adaptive weight allocation module; ④ Edge collaboration interface: supporting feature quantization compression (SNORM encoding, bit rate ≤ 0.1bpp) and secure transmission protocols.

[0208] Specific implementation process of value network

[0209] Step 1: Input residual features, input a multi-scale residual field E(x,y,z,t) (including three resolutions: 1km, 500m, and 200m); Preprocessing: extract local spatial features through a three-dimensional convolutional layer, and normalize the residual field to [-1,1].

[0210] Step 2: Enhanced spatiotemporal attention, spatial pyramid decomposition and spatiotemporal attention calculation.

[0211] Spatial pyramid decomposition is the core module for multi-scale feature extraction. Its design goal is to capture the global and local features of flow field errors through spatial partitioning at different scales. Specifically, it is implemented as follows: ① Multi-scale spatial partitioning: A three-dimensional spatial pyramid decomposition strategy is adopted to divide the input residual field into three scales: fine scale (1×1×1): preserving high-resolution detailed features (such as velocity gradients in frontal regions); medium scale (2×2×2): balancing computational efficiency and feature representation capability (suitable for medium-scale vortex regions); coarse scale (4×4×4): capturing large-scale flow field structures (such as tidal-driven barotropic modes).

[0212] Features at each scale are generated through adaptive average pooling, avoiding the loss of edge information caused by traditional max pooling. ② Feature fusion mechanism: A weighted summation strategy is used to fuse multi-scale features.

[0213] ;

[0214] The weight coefficients are dynamically adjusted through reinforcement learning, and residual connections are introduced to alleviate the gradient vanishing problem.

[0215] The spatiotemporal attention mechanism focuses on key error regions through dynamic weight allocation, specifically implemented as follows:

[0216] ① 3D Spatiotemporal Attention Module (STAttention): Spatial attention: Employs 3D convolution to extract local spatial features (3×3×3 kernel size), and calculates intra-feature correlation through self-attention.

[0217] ;

[0218] Q, K, and V are generated from flow velocity, temperature and salinity, and wind stress characteristic maps, respectively. Temporal attention: The time series is divided into blocks of size b, and long-range dependencies (such as tidal periodic effects) are captured through attention calculations across time steps.

[0219] ② Dynamic weight allocation strategy and physical quantity coupling evaluation: calculate the cross-correlation coefficient between flow velocity and temperature and salinity, and dynamically adjust the weight of vortex intensity index; reinforcement learning optimization: use PPO algorithm to optimize attention weight, set the residual threshold to ±0.1m / s, and introduce KL divergence constraint (λ=0.02) to prevent uneven weight distribution.

[0220] Innovative technological breakthroughs include:

[0221] (1) The three-dimensional spatiotemporal pyramid attention mechanism combines three-dimensional spatial decomposition with self-attention mechanism to achieve global-local feature adaptive focusing of flow field error.

[0222] (2) Attention enhancement through multi-physics coupling: dynamic indicators such as vortex intensity and front position are introduced to dynamically adjust attention weights.

[0223] (3) Edge-center collaborative computing architecture, lightweight attention feature extraction (FLOPs<1G) is performed at the edge nodes, and parameter optimization is completed at the central server.

[0224] Step 3: Multimodal value fusion and physical quantity coupling evaluation: Calculate the velocity-temperature-salinity cross-correlation coefficient (increase the weight of temperature-salinity features by 10% when the threshold is >0.6); dynamically adjust the weight of the eddy intensity index; The value function is a core indicator used in reinforcement learning to evaluate the expected cumulative reward of a state or state-action, and its essence is to quantify the degree of achievement of the system's goal through mathematical modeling. In ocean current field reconstruction, the value function is used to dynamically evaluate the rationality of the data assimilation results, and guides the optimization of model parameters by minimizing the deviation between the prediction error and the physical constraints.

[0225] Step 4: Edge-Center Collaborative Optimization. Edge node tasks: Perform lightweight attention feature extraction (FLOPs<1G) and transmit feature maps via QUIC protocol (end-to-end latency<5ms). Center server tasks: Receive features and update value network parameters (using PPO algorithm, clip ratio=0.3), and distribute the optimized weights to edge nodes (quantization precision 8-bit fixed-point).

[0226] Step 5: Multi-scale fusion and parameter update:

[0227] Parameter perturbation injection: Apply perturbation δp to the FVCOM model parameters (such as turbulence coefficients). The perturbation range is controlled by an adaptive threshold: δp=σp⋅tanh(α⋅∣E∣).

[0228] Step 6: Edge-Center Collaborative Optimization. Edge node task: Perform lightweight residual feature extraction and transmit feature maps through quantization compression (SNORM). Center server task: Receive feature maps, update reinforcement learning parameters, and distribute the optimized weights to edge nodes.

[0229] Core Module 5: Prediction and Visualization Output Layer.

[0230] Real-time 3D flow field visualization (contours / vector maps); simulation of scenarios such as oil spill diffusion and pollution migration. The prediction and visualization output layer system consists of: ① Digital twin engine: integrating FVCOM models and deep learning prediction results, supporting dynamic parameter extrapolation; ② AR / VR visualization terminal: equipped with a 3D flow field rendering engine (OSGEarth) and an interactive SDK; ③ BIM-GIS collaborative platform: supporting real-time spatial analysis of IFC format building models and ocean flow field data; ④ Edge computing nodes: deploying lightweight models (FLOPs < 5G) to achieve real-time updates of flow field data.

[0231] The specific implementation process is as follows:

[0232] Step 1: Dynamic spatiotemporal alignment of multi-source data. Input data: satellite altimeter data (spatial resolution 25km×25km, temporal resolution 1 day), Argo buoy data (spatial resolution 1000m, temporal resolution 10 days), ADCP profile data (spatial resolution 500m, temporal resolution 1 hour), etc.

[0233] Dynamic weighted fusion algorithm:

[0234] Step 2: Deployment of the multimodal prediction model. Model architecture: ① Physical model layer: FVCOM 3D flow field model, coupled with dynamic parameters such as tides and wind stress; ② AI enhancement layer: a Transformer-based spatiotemporal sequence prediction network, with spatiotemporal sequences of satellite altimeter and Argo data as input; ③ Hybrid input layer is the core module for realizing dynamic fusion of multi-source ocean current field data. Its design goal is to overcome the limitations of traditional single data sources through multimodal data alignment and dynamic weight allocation. Specific components include a multi-source data interface module, a dynamic routing network, and a feature fusion module. The multi-source data interface module supports standardized access to various types of ocean observation data, such as satellite altimeter (25km resolution), Argo buoy (1km resolution), and ADCP profile data (500m resolution); it uses a spatiotemporal alignment algorithm (based on adaptive weighting of residual field intensity) to achieve dynamic spatiotemporal matching of multi-source data. The dynamic routing network (Router) adopts a sparse hybrid expert (MoE) architecture, containing 4 expert modules (each expert is an independent feedforward network) and a Top-K routing strategy; the router output probability distribution determines the data flow direction. The feature fusion module adopts a three-dimensional spatiotemporal pyramid fusion strategy. Spatial dimensions: 1×1×1 (detailed features), 2×2×2 (mesoscale vortices), 4×4×4 (large-scale flow fields); temporal dimension: long-range dependencies such as tidal cycles are captured through cross-time step attention.

[0235] Step 3: AR / VR interactive visualization, AR scene construction: Scan the coastline with the mobile phone camera and automatically load the BIM model of the corresponding area; overlay a three-dimensional flow field vector (color represents flow velocity, and arrow density represents vortex intensity) in the AR interface.

[0236] Users can view flow velocity profiles at specific locations via gestures (latency <200ms); VR scene design: users can enter a virtual ocean laboratory with a head-mounted display and control a virtual ship in real time to observe the flow field structure; flow field node data can be collected through data gloves to automatically generate flow velocity-temperature-salinity correlation heat maps.

[0237] Step 4: BIM-GIS collaborative application, data interface: spatially match the coordinates of the wharf pile foundation in the BIM model with the flow field prediction results; calculate the hydrodynamic parameters (such as impact force and mooring cable tension) when the ship is berthed; decision support: generate a conflict heat map between the typhoon path and the offshore wind farm layout; provide dynamic early warning of the risk of scour of the cross-sea bridge piers (based on CFD-DEM coupled simulation).

[0238] Step 5: Edge computing and real-time updates, lightweight deployment: Deploy edge servers (NVIDIA Jetson AGX Orin) at coastal observation stations to perform flow field data preprocessing; transmit compressed flow field data (bit rate ≤ 0.3bpp) to the central server via the QUIC protocol; dynamic update mechanism: trigger local model retraining (using a federated learning framework) when satellite or buoy data is updated; synchronize global digital twin parameters every 10 minutes.

[0239] In summary, this embodiment includes the following technical solutions:

[0240] (1) A multi-source data dynamic spatiotemporal alignment weighting algorithm is proposed, which proposes an adaptive weight allocation mechanism based on residual field intensity, breaking through the static spatiotemporal matching limitation of the traditional KNN registration method. The accuracy is improved by dynamically adjusting the weights of satellite, Argo, and ADCP data through a reinforcement learning policy network.

[0241] (2) Three-dimensional spatiotemporal pyramid attention enhancement mechanism: a spatiotemporal feature focusing technology combining three-dimensional spatial decomposition (1×1×1, 2×2×2, 4×4×4) and self-attention is proposed, which breaks through the limitations of traditional two-dimensional attention in capturing the global-local feature coupling of flow field error and improves the error matching degree.

[0242] (3) A multi-physics coupling value assessment model is introduced, which incorporates dynamic indicators such as vortex intensity and frontal position, and dynamically adjusts the weight of the value function. This achieves an increase in the weight of temperature and salinity characteristics when the velocity-temperature-salinity cross-correlation coefficient is >0.6.

[0243] (4) Edge-center collaborative training architecture: Lightweight edge nodes (FLOPs<1G) are designed to perform feature extraction, and the central server optimizes global parameters through a federated learning framework. Compared with traditional centralized computing, communication bandwidth utilization is improved and memory usage is reduced.

[0244] (5) Adaptive weight distribution penalty term: KL divergence constraint (λ=0.02) is introduced into the value network to prevent data source weight degradation and improve strategy stability.

[0245] (6) The AR / VR virtual-real collaborative analysis system dynamically displays flow field characteristics and engineering facilities (such as typhoon path-driven stress simulation of wharf structures), breaking through the static visualization limitations of the VolumeLIC algorithm. User operation response speed is improved, and the frame rate of VR scene flow field rendering is stable.

[0246] (7) A BIM-GIS deep integration platform is used to construct a BIM model containing 18 types of marine environmental parameters, enabling joint simulation of offshore wind farm site selection and flow field optimization. This breaks through the limitations of traditional GIS, which only supports spatial analysis, and supports real-time calculation of ship berthing hydrodynamic parameters (such as impact force and mooring cable tension).

[0247] (8) Dynamic noise injection routing strategy: Gaussian noise (σ=0.1) is added to the sparse hybrid expert (MoE) routing calculation to avoid over-centralization of the strategy. Compared with the noise-free method, the convergence speed is improved and the routing diversity is increased.

[0248] (9) Multimodal feature semantic alignment map. A semantic association map of 18-dimensional features such as flow velocity, temperature, and salinity is constructed by graph neural network to improve the accuracy of node similarity calculation.

[0249] (10) Edge inference under the federated learning framework executes a lightweight model at the edge node, protects data privacy through differential privacy, and optimizes global digital twin parameters.

[0250] (11) Multi-physics field coupling evaluation index: Introduce fluid dynamics-thermodynamics coupling index (such as the correlation between seawater stratification depth and velocity gradient) to optimize the design of offshore platform heat exchange system. Break through the traditional evaluation method that only focuses on a single physical quantity.

[0251] (12) Adaptive perturbation range dynamic adjustment: A method for dynamically adjusting the perturbation range of model parameters based on residual strength is proposed.

[0252] (13) Intelligent prediction by fusion of multiple methods: A method for intelligent prediction of ocean current field is proposed that integrates PI-LSTM machine learning model, FVCOM three-dimensional hydrodynamic model and generalized unit line method. It integrates the advantages of different methods and achieves the effect of complementary advantages.

[0253] (14) Adaptive Noise-Injected Sparse Hybrid Expert (MoE) Routing: Based on the Transformer architecture, this paper proposes adding Gaussian noise to the routing calculation to avoid over-centralization of policies. Compared with noise-free methods, it improves convergence speed and increases routing diversity.

[0254] (15) Multimodal feature semantic alignment map: a semantic association map of 18-dimensional features such as flow velocity, temperature and salinity is constructed through graph neural network, and the node similarity calculation accuracy is improved. It breaks through the feature fusion method that only relies on mode decomposition.

[0255] Reference Figure 2 This application provides a system for intelligent reconstruction and prediction of three-dimensional ocean flow fields, the system comprising:

[0256] The data acquisition unit is used to acquire multi-source observation data of the ocean through large language models and edge monitoring equipment;

[0257] The data fusion unit is used to preprocess and fuse the multi-source observation data to obtain fused observation data;

[0258] The graph construction unit is used to construct a spatiotemporal correlation graph using the target observation parameters in the fused observation data as nodes and the spatiotemporal correlation and physical quantity coupling relationship between the target observation parameters as edges.

[0259] The data reconstruction unit is used to reconstruct and determine the reconstructed data corresponding to the multi-source observation data based on the spatiotemporal correlation map.

[0260] The prediction output unit is used to predict the three-dimensional flow field of the ocean based on the reconstructed data and output a visualization of the three-dimensional flow field.

[0261] It is understood that the content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0262] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this application are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0263] Furthermore, although this application is described in the context of functional modules, it should be understood that, unless otherwise stated to the contrary, one or more of the described functions and / or features may be integrated into a single physical system and / or software module, or one or more functions and / or features may be implemented in a separate physical system or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding this application. Rather, given the properties, functions, and internal relationships of the various functional modules in the system disclosed herein, the actual implementation of the module will be understood within the scope of conventional technology for an engineer. Therefore, those skilled in the art can implement the application set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of this application, which is determined by the full scope of the appended claims and their equivalents.

[0264] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0265] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, system, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, system, or device). For the purposes of this specification, "computer-readable medium" can mean any system that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, system, or device.

[0266] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic systems) with one or more wires, portable computer disk drives (magnetic systems), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic systems, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0267] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0268] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0269] Although embodiments of this application have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the claims and their equivalents.

[0270] The above is a detailed description of the preferred embodiments of this application, but this application is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A method for intelligent reconstruction and prediction of ocean three-dimensional flow field, characterized in that, The method comprises the following steps: Obtaining multi-source observation data of the ocean through a large language model and an edge monitoring device; Preprocessing and fusing the multi-source observation data to obtain fused observation data; Constructing a spatiotemporal correlation graph with target observation parameters in the fused observation data as nodes and spatiotemporal correlation and physical quantity coupling between the target observation parameters as edges; Reconstructing the multi-source observation data according to the spatiotemporal correlation graph; Predicting a three-dimensional flow field of the ocean according to the reconstructed data and outputting a visualized three-dimensional flow field; The method of reconstructing the multi-source observation data according to the spatiotemporal correlation graph comprises the following steps: Outputting a gridded hydrodynamic field based on the spatiotemporal correlation graph through an FVCOM model; Inputting the hydrodynamic field in the form of time series data into an LSTM network, embedding a physical constraint term in the LSTM network, and dynamically adjusting the sampling frequency through an event triggering mechanism; Calculating runoff contribution based on an SCS-CN model, simulating vertical flow velocity distribution through a generalized unit line, introducing a multi-scale attention mechanism, and dynamically allocating computing resources to mesoscale vortex regions; Optimizing parameters of the FVCOM model; The method of optimizing parameters of the FVCOM model comprises the following steps: Extracting features of the three-dimensional flow field output by the FVCOM model and the fused observation data respectively and calculating residuals; Constructing a state space according to the residuals and enabling the FVCOM model to perform reinforcement learning; Dynamically adjusting weights of the FVCOM model through a policy network and a value network; wherein the policy network is used to input residual feature maps and output action probability distribution; the value network is used to input state vectors and output expected rewards; Coordinately optimizing the FVCOM model based on residual feature maps transmitted by the edge monitoring device.

2. The method for intelligent reconstruction and prediction of ocean three-dimensional flow field according to claim 1, characterized in that, The method of obtaining multi-source observation data of the ocean through a large language model and an edge monitoring device comprises the following steps: Analyzing user instructions using the large language model to obtain profile observation data corresponding to the user instructions; projecting the profile observation data into a spatiotemporal grid of a satellite altimeter to calculate a spatiotemporal matching degree of the profile observation data; deleting the profile observation data with a spatiotemporal matching degree lower than a set threshold to obtain first observation data; Obtaining underway ADCP data using an ADCP device in the edge monitoring device; removing noise of the underway ADCP data using Kalman filtering; and performing flow velocity profile reconstruction on the underway ADCP data through an edge GPU to obtain second observation data; The first observation data and the second observation data are used as the multi-source observation data.

3. The method of claim 1, wherein, The method of preprocessing and fusing the multi-source observation data to obtain fused observation data comprises the following steps: Extracting barotropic modal features and first baroclinic modal features from the multi-source observation data; Setting three-dimensional grids and boundary conditions of an FVCOM model; Projecting the barotropic modal features and the first baroclinic modal features onto the three-dimensional grids of the FVCOM numerical model for spatiotemporal alignment; adjusting a quality threshold according to a task type, and performing data cleaning on the barotropic modal feature and the baroclinic modal feature after spatiotemporal alignment according to the quality threshold, to obtain a quality evaluation feature; extracting a vertical modal feature from the obtained multi-source observation data by using the edge monitoring device; extracting an Ekman flow feature coupled with the barotropic modal feature from the multi-source observation data obtained by the satellite in the edge monitoring device; fusing the quality evaluation feature, the vertical modal feature, and the Ekman flow feature to obtain the fused observation data.

4. The method for intelligent reconstruction and prediction of ocean three-dimensional flow field of claim 1, wherein, The method further includes the following steps: adjusting the feature weight of the target observation parameter according to the task requirement to adjust the corresponding edge in the spatiotemporal correlation graph.

5. The method of intelligent reconstruction and prediction of ocean three-dimensional flow field according to any one of claims 1 to 4, characterized in that, The output visualization of the three-dimensional flow field includes the following steps: outputting the contour or vector diagram of the three-dimensional flow field by a digital twin engine, an AR / VR visualization terminal, a BIM-GIS collaborative platform, or the edge monitoring device.

6. A system for intelligent reconstruction and prediction of ocean three-dimensional flow field, characterized in that, The system includes: a data acquisition unit configured to acquire multi-source observation data of the ocean by a large language model and an edge monitoring device; a data fusion unit configured to preprocess and fuse the multi-source observation data to obtain fused observation data; a graph construction unit configured to construct a spatiotemporal correlation graph by taking a target observation parameter in the fused observation data as a node and taking a spatiotemporal correlation and a physical quantity coupling relationship between the target observation parameters as an edge; a data reconstruction unit configured to reconstruct and determine reconstruction data corresponding to the multi-source observation data according to the spatiotemporal correlation graph; a prediction output unit configured to predict a three-dimensional flow field of the ocean according to the reconstruction data and output a visualization of the three-dimensional flow field; wherein: the reconstruction and determination of the reconstruction data corresponding to the multi-source observation data according to the spatiotemporal correlation graph includes: outputting a gridded hydrodynamic field based on the spatiotemporal correlation graph by an FVCOM model; inputting the hydrodynamic field in the form of time series data into an LSTM network, embedding a physical constraint term in the LSTM network, and dynamically adjusting the sampling frequency by an event trigger mechanism; calculating runoff contribution based on an SCS-CN model, simulating vertical flow velocity distribution by a generalized unit line, introducing a multi-scale attention mechanism, and dynamically allocating computing resources to mesoscale vortex regions; optimizing parameters of the FVCOM model; the optimization of the parameters of the FVCOM model includes: extracting features of the three-dimensional flow field output by the FVCOM model and the fused observation data respectively and calculating residuals; constructing a state space according to the residuals and enabling the FVCOM model to perform reinforcement learning; dynamically adjusting the weight of the FVCOM model by a policy network and a value network; wherein the policy network is configured to input residual feature maps and output action probability distribution; and the value network is configured to input state vectors and output expected rewards; performing coordinated optimization on the FVCOM model based on residual feature maps transmitted by the edge monitoring device.

7. An electronic device, comprising: The electronic device includes a processor and a memory; the memory is configured to store a program; The processor executes the program to implement the method in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The storage medium stores a program, and the program is executed by a processor to implement the method in any one of claims 1-5.

Images