A marine environment parameter rapid refinement prediction method based on a graph neural network
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SOUTHWEST JIAOTONG UNIV
- Filing Date
- 2021-10-20
- Publication Date
- 2026-05-26
AI Technical Summary
Existing marine environmental parameter forecasting technologies rely on physical equations or empirical formulas, which leads to large errors. Furthermore, traditional neural networks exhibit significant biases in their predictions when environmental disturbances occur, making it difficult to meet the needs of marine disaster forecasting.
By combining graph neural networks with historical observations from Argo and data from the MODAS system, and correcting the forecasting system through online learning and numerical assimilation methods, the graph neural network is used to simulate ocean dynamic processes, and a ray projection algorithm is used for three-dimensional visualization mapping, thereby achieving rapid and refined forecasting of ocean environmental parameters.
It improves the accuracy and timeliness of marine environmental parameter prediction, reduces the error of physical model assumptions, and realizes rapid and refined forecasting and three-dimensional visualization of marine environmental parameters, supporting marine environmental protection.
Smart Images
Figure CN115879498B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for rapid and refined forecasting of marine environmental parameters, and more particularly to a method for rapid and refined forecasting of marine environmental parameters based on graph neural networks, belonging to the field of marine environmental numerical forecasting. Background Technology
[0002] my country is one of the countries in the world most severely affected by marine disasters. These disasters are diverse, widespread, frequent, and cause significant losses. Major marine disasters and sudden environmental events such as storm surges, seaweed blooms, and oil spills threaten the economic and social development of coastal areas and the safety of people's lives and property. Therefore, marine environmental protection has become a crucial issue that marine management and forecasting departments at all levels of the country urgently need to address.
[0003] Numerical prediction of the marine environment is a crucial support for marine environmental protection. Marine environmental numerical prediction typically relies on prediction models based on physical equations. The challenge of this approach is that it requires considering numerous natural processes to understand complex physical phenomena. Therefore, it often necessitates approximations to the physical model or the use of empirical formulas to analytically describe a particular physical process, leading to inherent errors. However, dynamic laws are inherent in time-series observations. Since machine learning does not require human assumptions about the physical model and is a data-driven statistical learning method, it spontaneously learns the mapping relationship from input to output based on the probability distribution of training data. Therefore, it is more suitable for simulating and predicting complex physical processes that cannot be accurately represented by analytical equations.
[0004] Currently, research on applying machine learning methods to numerical weather prediction is limited both domestically and internationally, and is mainly concentrated in the meteorological field. Due to limitations such as the complexity and variability of ocean dynamic processes, the scarcity of ocean observation data, and the limited means of network-based representation of the ocean environment, research on using machine learning techniques for ocean environmental forecasting is still in its early stages. Existing research primarily employs traditional convolutional neural networks (CNNs) or generative adversarial networks (GANs) based on CNNs. These traditional neural networks are typically based on the assumption of independent and identically distributed (i.i.d.) networks, achieving strong nonlinear fitting capabilities through nested operations such as multi-layer convolution, nonlinear activation functions, and pooling. When environmental disturbances occur, the actual sample distribution deviates from the training sample distribution, leading to unpredictable deviations in the prediction results, which hinders their practical application.
[0005] Graph neural networks (GNNs) are connection-based models that utilize information exchange between nodes to represent dependencies within a graph, enabling accurate simulation of fluid dynamics. GNNs establish a structure based on nodes, edges, and a graph. Compared to traditional neural network methods, they establish dependencies between nodes linked by edges, making them better suited for representing complex interactions, more closely resembling actual physical processes, and exhibiting better generalization capabilities. Therefore, research based on GNNs holds significant research importance and application value in addressing the urgent need for forecasting complex marine environmental parameters. Summary of the Invention
[0006] In view of the aforementioned shortcomings of existing marine environmental parameter forecasting technologies, this invention proposes a rapid and refined forecasting method for marine environmental parameters based on graph neural networks. Using historical observations from Argo and data from the MODAS system as datasets, a graph neural network is employed to predict future marine environmental parameters (underwater 3D temperature and salinity, 3D acoustic field). Through online learning and numerical assimilation, the forecasting system is corrected based on environmental information from a limited dataset without affecting the normal operation of the network model. A ray casting algorithm from volume rendering technology is used to perform 3D visualization mapping of the marine environmental parameters. This invention introduces graph neural networks for forecasting marine environmental parameters, eliminating the need for explicit mathematical descriptions of ocean dynamic processes. This improves the timeliness and accuracy of marine environmental parameter forecasts, enabling rapid and refined forecasting and supporting its application in marine environmental protection.
[0007] The technical solution adopted by this invention to solve the technical problem is as follows: a rapid and refined forecasting method for marine environmental parameters based on graph neural networks. This method uses historical observations from Argo and data from the MODAS system as datasets. It employs graph neural networks to predict marine environmental parameters for future times. Through online learning and numerical assimilation, the forecasting system is corrected based on environmental information from a small dataset without affecting the normal operation of the network model, thus obtaining the marine environmental parameters. The distribution of the three-dimensional sound field is obtained through inversion, and the ray casting algorithm in volume rendering technology is used to perform three-dimensional visualization mapping of the marine environmental parameters. The steps of the graph neural network algorithm are as follows:
[0008] Step S1: Divide the target sea area into grids according to the resolution of the collected data. The water body in each grid is called a fluid micro-cluster.
[0009] Step S2: Refine the modeling of microclusters through graph embedding, process and calculate the evolution of environmental parameters between microclusters to generate the final latent variables;
[0010] In the graph, each microcluster is described by nodes. Different microclusters have different physical characteristics, such as temperature, salinity, flow velocity, flow direction, and depth. These parameterized features are organized into one-dimensional vectors, serving as the feature vectors of their respective nodes. Finally, dependencies are established between nodes by creating "edges" that link them. The features of these edges can be the distance between nodes or the temperature conductivity coefficient between nodes. The design of these features requires analysis based on the physical characteristics of the data to determine the feature organization form most beneficial to network computation. The evolution of environmental parameters between microclusters is processed through M-step information transfer, forming a sequence of latent variables and generating the final latent variables.
[0011] Step S3: The decoder extracts dynamic information from the final latent variables and outputs the environmental parameter predictions for the target sea area.
[0012] Furthermore, the forecast system is calibrated using online learning and numerical assimilation methods. The calibration method steps are as follows:
[0013] Step S1: Modify the network parameters by adding a fine-tuning bias to the graph neural network;
[0014] θ * =θ+b (1)
[0015] In equation (1), θ * The values represent the changed network parameters, where θ represents the network parameters and b represents the fine-tuning bias.
[0016] Step S2: Constrain the network's updates to important key parameters by adding a penalty term related to importance to the loss function;
[0017] The importance of a parameter is measured by the accumulation of gradients. When a parameter is more relevant to the current task, that is, more important, the network will update the parameter more frequently in order to make the feature more significant. Conversely, if the relevance is small, the network will update the parameter less often, and the accumulated gradient will naturally be smaller than the former. The importance coefficient Ω is calculated as follows.
[0018]
[0019] The loss function is:
[0020]
[0021] Step S3: Regarding the correction of the prediction results, the error between the current prediction value and the acoustic tomography analysis value is used to correct the future prediction results of the network.
[0022]
[0023] Furthermore, a ray casting algorithm from volume rendering technology is used to perform three-dimensional visualization mapping of marine environmental parameters. The steps of the algorithm are as follows:
[0024] Step S1: Organize the obtained marine environmental parameters based on an octree model;
[0025] Step S2: Use the ray casting algorithm in volume rendering technology to perform three-dimensional visualization mapping of the marine environment;
[0026] The advantage of the ray projection algorithm is that it can penetrate the volume data at any angle. Through the orderliness and hierarchical nature of the octree model, it can quickly index and locate the incident point of the projected ray entering the volume data. Then, it can resample along the projected ray, which can comprehensively, efficiently and flexibly realize the three-dimensional visualization of marine environmental parameters. Combined with multi-threading technology and GPU acceleration technology, it greatly reduces the efficiency of original marine environmental feature mapping and three-dimensional rendering, and realizes the rapid reconstruction and visualization of marine environmental parameters.
[0027] After adopting the above technical solution, the beneficial effects of the present invention are as follows:
[0028] This invention proposes a rapid and refined forecasting method for marine environmental parameters based on graph neural networks (GNNs). GNNs are used to predict marine environmental parameters to improve forecasting speed. GNNs can accurately simulate fluid dynamic processes. GNNs establish a structure based on nodes, edges, and graphs. Compared to traditional neural network methods, they establish dependencies between nodes linked by edges, which is more conducive to representing complex interactions, more closely approximating actual physical processes, and exhibiting better generalization than traditional neural networks. Compared to existing model-based environmental forecasting methods, this invention eliminates the need for human assumptions about complex physical models, reducing errors caused by approximations of physical models or the use of empirical formulas to analytically describe a physical process. Compared to related research, this invention combines GNNs with rapid and refined forecasting of marine environmental parameters to perform high-resolution numerical forecasting of marine environmental parameters, overcoming the limitations of cognitive understanding of physical models of marine dynamic processes. More importantly, this invention uses methods such as neural networks to obtain marine environmental parameters, and obtains the distribution of three-dimensional sound fields through inversion. It uses three-dimensional visualization technology to map three-dimensional views of marine environmental parameters, improves the visualization of prediction results, and realizes rapid and refined forecasting of marine environmental parameters, providing support for their application in marine environmental protection. Attached Figure Description
[0029] Figure 1 This is a flowchart illustrating the method of the present invention.
[0030] Figure 2 This is a diagram illustrating the update process of the graph neural network in the method of this invention.
[0031] Figure 3 This is a diagram showing the encoder-processor-decoder structure of the graph neural network in the method of this invention.
[0032] Figure 4 This is a graphical representation method of the marine environment in the method of the present invention.
[0033] Figure 5 This is a flowchart of the numerical weather prediction system correction process in the method of the present invention.
[0034] Figure 6 This is a flowchart of the three-dimensional visualization processing of the marine environment in the method of the present invention. Detailed Implementation
[0035] The embodiments of the present invention will be further described below with reference to the accompanying drawings.
[0036] The present invention addresses the problems of existing model-based environmental forecasting methods that require approximations of physical models or the use of empirical formulas to analytically describe a physical process, resulting in errors. It utilizes graph neural networks to accurately simulate fluid dynamic processes, overcoming the limitations of understanding physical models of marine dynamic processes, thereby enabling rapid, high-resolution numerical forecasting of marine environmental parameters.
[0037] This invention discloses a rapid and refined forecasting method for marine environmental parameters based on graph neural networks. The method applies graph neural networks to the rapid and refined forecasting of marine environmental parameters, and uses graph neural networks to simulate physical models of marine dynamic processes. The method flow is as follows: Figure 1 As shown, using historical observations from Argo and data from the MODAS system as datasets, a graph neural network is employed to predict future marine environmental parameters. Through online learning and numerical assimilation, the prediction system is corrected based on environmental information from a limited dataset without affecting the normal operation of the network model, yielding marine environmental parameters. The distribution of the three-dimensional sound field is then obtained through inversion, and the ray casting algorithm from volume rendering technology is used to perform three-dimensional visualization mapping of the marine environmental parameters. This invention introduces a graph neural network for predicting marine environmental parameters, eliminating the need for explicit mathematical descriptions of ocean dynamic processes, improving the timeliness and accuracy of marine environmental parameter predictions, and enabling rapid and refined forecasting of marine environmental parameters, thus supporting their application in marine environmental protection.
[0038] Suppose X t ∈x represents the state of the ocean at time t. This represents the state trajectory of the same sea area over K consecutive time steps. This invention uses a simulator s: x→x to simulate the evolution of this physical dynamics, that is, causally mapping the current state to future states. The simulator calculates dynamic information reflecting how the current state changes and uses this information to update the current state to the future state. This simulator utilizes a parameterized function approximator d. θ The dynamic process is calculated by x→y, where θ can be trained and optimized, and Y∈y represents dynamic information, the semantics of which are determined by the update mechanism. In this invention, the update mechanism refers to the temperature conduction, mass transport, and flow of micro-clusters in nearshore fluids. The update mechanism can be viewed as using the current state... and d θ To predict the next state The function. The update process is as follows: Figure 2 As shown.
[0039] In this invention, to enable the model to simulate the evolution of nearshore environments under various conditions as comprehensively as possible, we use θ to control d. θ This allows it to cover temperature conduction, mass transport, and micro-group interactions under various conditions. The following d is defined: θ The model, its structure is as follows Figure 3 As shown.
[0040] The encoder transforms the input marine environmental state parameters into a graphical form G. 0 =Encoder(X), where G = (V, E, u). The node embedding V describes the state of the micro-cluster itself. The edge embedding E describes the pairwise relationships between micro-clusters, and the global information embedding u describes the global information. Since the input X is an environmental parameter represented in other forms, graph networks cannot directly process it. Therefore, abstracting the marine environment through graphs is a prerequisite for using graph networks to simulate the evolution of marine environmental parameters. The graph embedding method for marine environmental parameters is as follows: Figure 4 As shown.
[0041] The processor calculates the evolution of environmental parameters among micro-clusters, which, through M-step information transfer, forms a sequence of latent variables and generates the final latent variable G. M =Processor(G 0 ).
[0042] The decoder obtains the final hidden variable G. M The dynamic information is extracted and the final output diagram Y is formed, which is the parameter prediction of the marine environment.
[0043] In practical applications of networks, the influence of uncontrollable random factors gradually accumulates over time, leading to a gradual decrease in prediction accuracy and a deviation of forecast results from the actual process. To ensure the long-term stable operation of numerical weather prediction models, real-time model calibration is necessary. This invention employs online learning and numerical assimilation methods to calibrate the forecast system based on limited environmental information without affecting the normal operation of the network model. The calibration mainly consists of two parts: fine-tuning of the network model based on online learning and correction of forecast results based on numerical assimilation. The block diagram of the numerical weather prediction system calibration scheme is shown below. Figure 5 As shown. The steps of the correction method are as follows:
[0044] Step S1: Modify the network parameters by adding a fine-tuning bias to the graph neural network;
[0045] θ * =θ+b (1)
[0046] In equation (1), θ * The values represent the changed network parameters, where θ represents the network parameters and b represents the fine-tuning bias.
[0047] Step S2: Constrain the network's updates to important key parameters by adding a penalty term related to importance to the loss function;
[0048] The importance of a parameter is measured by the accumulation of gradients. When a parameter is more relevant to the current task, that is, more important, the network will update the parameter more frequently in order to make the feature more significant. Conversely, if the relevance is small, the network will update the parameter less and less, so the accumulated gradient will be smaller than the former. The importance coefficient Ω is calculated as follows.
[0049]
[0050] The loss function is:
[0051]
[0052] Step S3: Regarding the correction of the prediction results, the error between the current prediction value and the acoustic tomography analysis value is used to correct the future prediction results of the network.
[0053]
[0054] The flowchart of the three-dimensional visualization process for marine environment is as follows: Figure 6 As shown, the ray casting algorithm in volume rendering technology is used to perform three-dimensional visualization mapping of marine environmental parameters. The steps of the algorithm are as follows:
[0055] Step S1: Organize the obtained marine environmental parameters based on an octree model;
[0056] Step S2: Use the ray casting algorithm in volume rendering technology to perform three-dimensional visualization mapping of marine environmental parameters;
[0057] The advantage of the ray projection algorithm is that it can penetrate the volume data at any angle. Through the orderliness and hierarchical nature of the octree model, it can quickly index and locate the incident point of the projected ray entering the volume data. Then, it can resample along the projected ray, which can comprehensively, efficiently and flexibly realize the three-dimensional visualization of marine environmental parameters. Combined with multi-threading technology and GPU acceleration technology, it greatly reduces the efficiency of original marine environmental feature mapping and three-dimensional rendering, and realizes the rapid reconstruction and visualization of marine environmental parameters.
[0058] In summary, the rapid and refined forecasting method for marine environmental parameters based on graph neural networks proposed in this invention has the following characteristics:
[0059] 1) Compared with existing model-based environmental forecasting methods, this invention does not require human assumptions about complex physical models, reduces the errors caused by approximating physical models or using empirical formulas to analytically describe a physical process, and greatly improves the calculation speed.
[0060] 2) Compared with traditional machine learning algorithms, the method of this invention is closer to the actual physical process, and therefore has better timeliness and accuracy;
[0061] 3) Through three-dimensional visualization of marine environmental parameters, the marine environmental parameters can be displayed intuitively, comprehensively, multi-levelly and three-dimensionally, so as to grasp the real-time laws of the marine environment.
[0062] The above description is merely a preferred embodiment of the method of the present invention. It should be noted that, without departing from the essence of the present invention and its apparatus, several modifications may be made in actual implementation and should also be included within the scope of protection of the present invention.
Claims
1. A rapid and refined forecasting method for marine environmental parameters based on graph neural networks, which uses Argo historical observations and data from the MODAS marine environmental dynamic analysis system to train the graph neural network to predict marine environmental parameters at future times. The marine environmental parameters include underwater three-dimensional temperature, salinity and three-dimensional sound field. The forecasting system is corrected through online learning and numerical assimilation. The ray casting algorithm in volume rendering technology is used to perform three-dimensional visualization mapping of the marine environmental parameters. The steps of the graph neural network algorithm are as follows: Step S1: Divide the target sea area into grids according to the resolution of the collected data. The water body in each grid is called a fluid micro-cluster. Step S2: Refine the modeling of microclusters through graph embedding, process and calculate the evolution of environmental parameters between microclusters to generate the final latent variables; In the diagram, each microcluster is characterized by nodes. Different microclusters have different physical characteristics, including but not limited to temperature, salinity, flow velocity, flow direction, and depth. The parameterized features are organized into a one-dimensional vector format, which serves as the feature vector for each node. Finally, by establishing "edges" between different nodes, dependencies are established between the nodes. The features of the edges are the distance between nodes or the temperature conduction coefficient between nodes. The design of the edge features is based on the analysis of the physical characteristics of the data to determine the feature organization form most conducive to network computation. The evolution of environmental parameters between microclusters is processed and calculated. After M steps of information transmission, a sequence of latent variables is formed, and the final latent variables are generated. Step S3: The decoder extracts dynamic information from the final latent variables and outputs the environmental parameter predictions for the target sea area.
2. The method for rapid and refined forecasting of marine environmental parameters based on graph neural networks as described in claim 1, characterized in that: The forecast system is calibrated using online learning and numerical assimilation methods. The calibration method comprises the following steps: Step S1: Modify the network parameters by adding a fine-tuning bias to the graph neural network; （1） In equation (1), θ represents the corrected network parameters, b represents the fine-tuning bias; Step S2: Constrain the network's updates to important key parameters by adding a penalty term related to importance to the loss function; The importance of a parameter is measured by the accumulation of gradients. When a parameter is more relevant to the current task, that is, more important, the network will inevitably update the parameter more frequently in order to make the feature more significant. Conversely, if the relevance is small, the network will update the parameter less often, so the accumulated gradient will naturally be smaller than the former. The importance coefficient Ω is calculated as follows. (2) The loss function is: (3) Step S3: Regarding the correction of the prediction results, the error between the current prediction value and the acoustic tomography analysis value is used to correct the future prediction results of the network; (4)。 3. The method for rapid and refined forecasting of marine environmental parameters based on graph neural networks as described in claim 1, characterized in that: A ray casting algorithm from volume rendering technology is used to perform 3D visualization mapping of the marine environment. The steps of the algorithm are as follows: Step S1: Organize the obtained marine environmental parameters based on an octree model; Step S2: Use the ray casting algorithm in volume rendering technology to perform three-dimensional visualization mapping of marine environmental parameters; The advantage of the ray projection algorithm is that it can penetrate the volume data at any angle. Through the orderliness and hierarchical nature of the octree model, it can quickly index and locate the incident point of the projected ray entering the volume data. Then, it can resample along the projected ray, which can comprehensively, efficiently and flexibly realize the three-dimensional visualization of marine environmental parameters. Combined with multi-threading technology and GPU acceleration technology, it greatly reduces the efficiency of original marine environmental feature mapping and three-dimensional rendering, and realizes the rapid reconstruction and visualization of marine environmental parameters.