A nearshore sea wave prediction revision method based on physical factor coupling chebnet

Abstract

The present application relates to the technical field of ocean engineering, in particular to a nearshore sea wave prediction correction method based on physical factor coupling ChebNet, comprising the following steps: S1, constructing a dynamic graph model based on nearshore sea area sensing data, taking physical factors such as sea wave height, wind speed, wind direction and terrain elevation as graph nodes, using graph convolution to extract the space-time dependence between nodes, and generating a preliminary sea wave prediction; S2, introducing a space-time adaptive mechanism to dynamically adjust the coupling strength of physical factors in the graph and optimize the connection relationship between nodes; S3, using multi-order ChebNet to model the space-time variation of physical factors, iteratively correcting the prediction results, and outputting the final optimized sea wave prediction; the present application realizes accurate prediction, dynamic correction and strong robustness correction of nearshore sea waves by constructing a physical factor driven dynamic graph structure, introducing a space-time adaptive coupling mechanism and multi-order ChebNet high-order modeling, and improves the prediction accuracy, adaptability and space-time consistency.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of ocean engineering, and in particular to a nearshore sea wave prediction correction method based on physical factor coupling ChebNet. BACKGROUND

[0002] In recent years, with the rapid development of marine economy and coastal engineering, accurate prediction of nearshore sea waves is of great significance for port operation, offshore construction and marine disaster warning. The formation and propagation of sea waves are influenced by multiple physical factors, such as sea wave height, wind speed, wind direction and terrain elevation. These factors have strong spatiotemporal correlation. Traditional marine prediction methods gradually fail to meet the demand for high precision, real-time and complex coupling relationship modeling. Therefore, intelligent modeling methods that integrate multi-source physical data and graph neural networks have become an important research direction for improving the prediction ability of nearshore sea waves.

[0003] Existing sea wave prediction methods mostly use numerical models or deep learning models driven by a single physical factor, which fail to effectively consider the coupling relationship between multiple physical factors and have weak joint modeling capability for spatial structure and temporal evolution, resulting in poor generalization, insufficient adaptability and large prediction bias of the model in complex sea areas. In addition, some research based on graph neural networks does not introduce an adaptive mechanism, which cannot dynamically adjust the influence strength of each physical factor in the graph structure for different sea environments, thereby limiting the model's ability to capture the spatiotemporal variation characteristics of the actual ocean. SUMMARY

[0004] The application provides a nearshore sea wave prediction correction method based on physical factor coupling ChebNet. By constructing a dynamic graph structure model that integrates physical factors such as sea wave height, wind speed, wind direction and terrain elevation, combining graph convolution network to extract the spatiotemporal dependence relationship between nodes, introducing a spatiotemporal adaptive mechanism to dynamically optimize the coupling strength of nodes, and using multi-order ChebNet convolution to deeply model complex spatiotemporal variation processes, the method realizes high-precision correction of preliminary sea wave prediction results and significantly improves the accuracy and robustness of nearshore sea wave prediction.

[0005] A nearshore sea wave prediction correction method based on physical factor coupling ChebNet includes the following steps:

[0006] S1, based on the real-time sensing data of the nearshore sea area, a dynamic graph structure model is constructed, physical factors including sea wave height, wind speed, wind direction and terrain elevation are taken as graph nodes, and spatiotemporal dependence relationships between graph nodes are dynamically constructed through graph convolution operation to generate preliminary sea wave prediction results;

[0007] S2, the spatio-temporal adaptive mechanism is introduced, the coupling strength between the nodes of the graph is adjusted adaptively, the influence of each physical factor in the graph convolution operation is dynamically weighted according to the spatio-temporal characteristics of different sea areas, and the coupling relationship of the physical factors is optimized in real time;

[0008] S3, the spatio-temporal variation of the physical factors is modeled by using a multi-order ChebNet graph convolution layer, the preliminary sea wave prediction result is corrected based on the optimized coupling relationship of the physical factors through iterative updating, and the optimized sea wave prediction is output.

[0009] Optionally, the S1 comprises:

[0010] S11, a plurality of physical factors of a near-shore sea area are collected by a real-time sensor network, including sea wave height, wind speed, wind direction and terrain elevation, and the physical factors are respectively coded as nodes in a graph structure, the nodes are initialized by a graph convolution network (GCN), the physical characteristics and spatial position information of the nodes are converted into feature vectors in the graph structure, and preliminary node feature representation is formed;

[0011] S12, a spatio-temporal dependence relationship between the initialized nodes is dynamically constructed by a graph convolution operation, the connectivity and mutual relationship between the nodes are described by using an adjacency matrix, and a preliminary sea wave prediction result is generated based on the spatio-temporal dependence relationship.

[0012] Optionally, the S11 comprises:

[0013] S111, a plurality of physical factors of a near-shore sea area are collected by a real-time sensor network, including sea wave height , wind speed , wind direction and terrain elevation ;

[0014] S112, according to the types of the physical factors, each physical factor (sea wave height, wind speed, wind direction and terrain elevation) is taken as a node in a graph structure, and the physical characteristics (such as sea wave height, wind speed, etc.) and spatial position information of each node are converted into a feature vector;

[0015] S113, an adjacency matrix A of the graph is constructed to represent the spatio-temporal dependence relationship between different nodes, and an element of the adjacency matrix represents the connection strength between a node i and a node j;

[0016] S114, the node feature vector and the adjacency matrix A are taken as inputs and transmitted into a graph convolution network (GCN) for initialization, the features are propagated between the nodes and the representation of the nodes is updated through the connection relationship between the nodes;

[0017] S115, in the initialization of the graph convolution network, the feature representation of each node is updated by convolution operation, and finally a preliminary feature vector of each node is obtained as the preliminary node feature representation of the graph structure.

[0018] Optionally, the S12 comprises:

[0019] S121, by graph convolution operation, based on the adjacency matrix A and the preliminary feature vector update the node feature, capture the spatio-temporal dependence relationship, through multi-layer graph convolution operation, finally obtain the Kth layer node feature ;

[0020] S122, the node feature obtained by the graph convolution operation , which is passed to the output layer to generate a preliminary sea wave prediction result .

[0021] Optionally, the S2 comprises:

[0022] S21, a spatio-temporal adaptive weighting mechanism is introduced for each physical factor node (such as sea wave height, wind speed, etc.), by analyzing the spatio-temporal characteristics of different sea areas, the spatio-temporal weighting coefficient between each node and its neighbor node is calculated;

[0023] S22, in the graph convolution operation, the coupling strength between nodes is dynamically adjusted by using the adaptive mechanism, and according to the spatio-temporal weighting coefficient, the weight in the adjacency matrix is dynamically changed at each layer of graph convolution.

[0024] Optionally, the S21 comprises:

[0025] S211, by spatio-temporal variance analysis, the spatio-temporal characteristics of each physical factor node in different time and space environment are evaluated;

[0026] S212, by the spatial position difference and the physical feature difference between nodes, the spatio-temporal distance between each pair of nodes is calculated, which is used to quantify the spatio-temporal dependence relationship between nodes;

[0027] S213, based on the calculated spatio-temporal distance, the spatio-temporal weighting coefficient for each pair of nodes is calculated, which is used to represent the connection strength between nodes.

[0028] Optionally, the S22 comprises:

[0029] S221, in each layer of graph convolution operation, the element of the adjacency matrix is dynamically adjusted according to the spatio-temporal weighting coefficient , so as to change the coupling strength between nodes;

[0030] S222, in the graph convolution operation, the node feature propagation is performed using the dynamically adjusted adjacency matrix, and a dynamically adjusted node feature is obtained.

[0031] Optionally, the S3 comprises:

[0032] S31, a multi-order ChebNet graph convolution layer is used to model the spatio-temporal variation between physical factors, and the ChebNet approximates the graph convolution operation by using Chebyshev polynomials to capture the complex variation of the physical factors in space and time.

[0033] S32, in the multi-order ChebNet graph convolution layer, the spatio-temporal relationship of the physical factors is gradually optimized, and the preliminary sea wave prediction result is corrected by using the optimized physical factor coupling relationship through iterative updating, and finally the optimized sea wave prediction result is output.

[0034] Optionally, the S31 comprises:

[0035] S311, by Chebyshev polynomial expansion of the graph convolution kernel, a spectral domain approximation method is used to model the high-order dependence relationship between nodes.

[0036] S312, the node features optimized by the spatio-temporal weighting mechanism are used as input, and the multi-order convolution kernel is combined to extract the embedded expression in the complex spatial structure and time evolution process.

[0037] S313, in the ChebNet graph convolution of each order, the node features are fused with the neighborhood information layer by layer through multi-order adjacency propagation, and finally the high-dimensional spatio-temporal features used for downstream correction are output.

[0038] Optionally, the S32 comprises:

[0039] S321, by constructing the residual relationship between the preliminary sea wave prediction result and the multi-order ChebNet output feature, the error distribution is extracted.

[0040] S322, according to the optimization coupling strength of the nodes in the ChebNet convolution in the graph, different correction weights are assigned to the residual error.

[0041] S323, the preliminary prediction result and the residual correction amount are fused, and the final optimized sea wave prediction result is output.

[0042] The beneficial effects of the present application are:

[0043] The application can effectively capture the complex nonlinear coupling relationship between various physical variables in the nearshore sea area by constructing a dynamic graph structure model containing multiple physical factors such as wave height, wind speed, wind direction and terrain elevation, and introducing graph convolution operation to jointly model the spatial layout and time sequence between nodes. Compared with the traditional static modeling method, this method has the advantages of strong dynamic adaptability and rich structure expression capability, and can accurately depict the non-uniform and dynamic evolution environment in the nearshore area, thereby significantly improving the response ability and physical consistency of the preliminary sea wave prediction result to the original observation data.

[0044] The application introduces an adaptive weighting mechanism based on the change of the spatio-temporal characteristics of the nodes, dynamically adjusts the coupling strength between different physical factor nodes, and then adjusts the weight of the adjacency matrix in real time during the graph convolution process, so that it is more suitable for the local field change of the actual sea area. This mechanism fully considers the physical similarity, geographical distance and time sequence fluctuation characteristics between nodes, so that the convolution propagation path of the graph neural network in different sea areas and different time can have stronger local adaptability and global structure constraint, effectively enhancing the coupling modeling accuracy, spatial generalization ability and robustness of the model under heterogeneous physical fields and complex boundary conditions.

[0045] The application integrates a multi-order ChebNet graph convolution network to hierarchically model the high-dimensional spatio-temporal characteristics of physical factors, and based on the optimized graph structure relationship, the preliminary sea wave prediction result is modeled and distributed to estimate the residual error, and the dynamic correction of the prediction error and the directional correction of the structural deviation are realized under the framework of the multi-layer perceptron. Not only the potential physical constraint characteristics in the preliminary prediction are retained, but also the advantages of ChebNet in high-order dependence modeling are fully utilized, the spatio-temporal continuity and global consistency of the prediction result are improved through iterative propagation and embedding fusion, thereby greatly improving the accuracy, stability and adaptability of the overall prediction system in a variable environment. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the present application or prior art, the following will briefly introduce the drawings needed in the embodiments or prior art description. Obviously, the drawings in the following description only illustrate the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0047] Figure 1 The flowchart of the revision method of the embodiment of the present application is shown.

[0048] Figure 2 The schematic diagram of generating the preliminary sea wave prediction result of the embodiment of the present application is shown. DETAILED DESCRIPTION

[0049] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. Those skilled in the art may employ other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0050] like Figures 1-2 As shown, a nearshore wave forecast correction method based on physical factors coupled with ChebNet includes the following steps:

[0051] S1. Based on real-time sensing data of nearshore waters, a dynamic graph structure model is constructed. Physical factors including wave height, wind speed, wind direction, and terrain elevation are used as graph nodes. The spatiotemporal dependencies between graph nodes are dynamically constructed through graph convolution operations to generate preliminary wave forecast results.

[0052] S2 introduces a spatiotemporal adaptive mechanism, which dynamically adjusts the coupling strength between graph nodes to dynamically weight the influence of each physical factor in the graph convolution operation based on the spatiotemporal characteristics of different sea areas, and optimizes the coupling relationship of physical factors in real time.

[0053] S3 uses multi-level ChebNet graph convolutional layers to model the spatiotemporal changes of physical factors. Through iterative updates, it corrects the preliminary wave forecast results based on the optimized coupling relationship of physical factors and outputs the optimized wave forecast.

[0054] Through the above steps, a dynamic graph structure model integrating multiple physical factors is constructed. By introducing a spatiotemporal adaptive weighting mechanism and combining it with a multi-order ChebNet graph convolutional network, not only is accurate modeling of the complex physical field of nearshore waters achieved, but the coupling strength between nodes is also dynamically adjusted to adapt to the characteristics of different marine environments. Furthermore, through high-order feature extraction and residual correction mechanisms, the accuracy, stability, and adaptability to boundary conditions of wave forecast results are effectively improved.

[0055] S1 includes:

[0056] S11 collects various physical factors of nearshore waters through a real-time sensor network, including wave height, wind speed, wind direction, and topographic elevation. These physical factors are encoded as nodes in a graph structure. The nodes are initialized through a graph convolutional network (GCN), which transforms their physical characteristics and spatial location information into feature vectors in the graph structure, forming a preliminary node feature representation.

[0057] S12 dynamically constructs the spatiotemporal dependencies between initial nodes through graph convolution operations, uses the adjacency matrix to describe the connectivity and interrelationships between nodes, and generates preliminary wave forecast results based on the spatiotemporal dependencies.

[0058] Through the above steps, various physical factors in the coastal sea area are modeled in a graph structure, and the physical characteristics and spatial position information are fused by using a graph convolution network (GCN) to initialize the nodes, effectively build the spatial and temporal dependence relationship between the physical factors, and thus improve the structural perception and expression ability of the model to the complex physical environment.

[0059] S11 includes:

[0060] S111, collecting a plurality of physical factors from the coastal sea area through a real-time sensor network, including sea wave height , wind speed , wind direction , and terrain elevation ;

[0061] S112, according to the type of physical factors, each physical factor (sea wave height, wind speed, wind direction, terrain elevation) is taken as a node in the graph structure, and for each node, its physical characteristics (such as sea wave height, wind speed, etc.) and spatial position information are converted into a feature vector, represented as:

[0062] ;

[0063] wherein, is the feature vector of node i;

[0064] S113, constructing an adjacency matrix A of the graph to represent the spatial and temporal dependence relationship between different nodes, and the element of the adjacency matrix represents the connection strength between node i and node j, represented as:

[0065] ;

[0066] wherein, is the Euclidean distance between node i and node j, , are the spatial positions of node i and node j, respectively, is a smoothing term to prevent division by zero;

[0067] S114, taking the node feature vector and the adjacency matrix A as inputs and inputting them into a graph convolution network (GCN) for initialization, propagating the features through the connection relationship between the nodes and updating the representation of the nodes, represented as:

[0068] ;

[0069] wherein, is the feature representation of node i after k-layer graph convolution, is the neighbor node set of node i, , respectively the degree of node i and node j, is the weight matrix of the k-th layer, is an activation function;

[0070] S115, in the initialization of the graph convolution network, the feature representation of each node is updated by convolution operation, and finally the preliminary feature vector of each node is obtained , as the preliminary node feature representation of the graph structure, is represented as:

[0071] ;

[0072] wherein A is an adjacency matrix, is the initial feature vector of node i, is the weight matrix of the 0-th layer graph convolution.

[0073] Through the above steps, the sea wave height, wind speed, wind direction, terrain elevation and other time-varying physical factors and spatial position information are jointly coded into the node feature vector in the graph structure, and the spatio-temporal connection strength matrix is constructed combining the adjacency relationship between nodes, which not only effectively expresses the coupling relationship between different physical factors, but also significantly enhances the ability to describe the spatio-temporal dependence between nodes. Further, the graph convolution network is used to initialize and propagate the node features, so that the model can fully capture the local and global spatial correlation information, thereby realizing efficient modeling of the physical factor interaction structure in the complex marine environment in the initial stage.

[0074] S12 includes:

[0075] S121, through the graph convolution operation, based on the adjacency matrix A and the preliminary feature vector update the node feature, capture the spatio-temporal dependence relationship, through the multi-layer graph convolution operation, finally obtain the k-th layer node feature , represented as:

[0076] ;

[0077] wherein, the weight matrix of the 1st layer graph convolution, is the k-1-th layer node feature;

[0078] S122, the node feature obtained by the graph convolution operation is passed to the output layer to generate the preliminary sea wave prediction result , represented as:

[0079] ;

[0080] wherein, is the weight matrix of the output layer, This is the bias term for the output layer.

[0081] S2 includes:

[0082] S21 introduces a spatiotemporal adaptive weighting mechanism for each physical factor node (such as wave height, wind speed, etc.). By analyzing the spatiotemporal characteristics of different sea areas, the spatiotemporal weighting coefficient between each node and its neighboring nodes is calculated.

[0083] S22, in graph convolution operations, uses an adaptive mechanism to dynamically adjust the coupling strength between nodes. Based on the spatiotemporal weighting coefficients, the weights in the adjacency matrix are dynamically changed in each layer of graph convolution, so that the influence of each physical factor on the node feature update matches its importance in the spatiotemporal environment.

[0084] Through the above steps, a spatiotemporal adaptive weighting mechanism is introduced to dynamically calculate the coupling strength between nodes based on the spatiotemporal variation characteristics of different physical factors in a specific sea area, and adjust the adjacency matrix weights in real time during graph convolution operations. This enables the model to more accurately reflect the actual correlation between physical factors, thereby significantly improving the accuracy of feature propagation and the model's adaptability to complex environments.

[0085] S21 includes:

[0086] S211 assesses the spatiotemporal characteristics of each physical factor node by analyzing its changes under different time and spatial environments through spatiotemporal variance analysis, as follows:

[0087] ;

[0088] in, Physical factors related to node i variance Let be the physical factor value of node i at time t. Physical factors of node i within the time range T The mean of , where T is the number of times t;

[0089] S212 calculates the spatiotemporal distance between each pair of nodes based on the spatial location and physical characteristic differences between nodes. This is used to quantify the spatiotemporal dependencies between nodes and is represented as follows:

[0090] ;

[0091] in, Let i be the spatiotemporal distance between node i and node j. To adjust the hyperparameters for the influence of spatial distance and differences in physical characteristics on spatiotemporal distance, , These are the values ​​of the physical factors of nodes i and j at time t, respectively.

[0092] S213, based on the calculated spatiotemporal distance, calculating the spatiotemporal weighted coefficient for each pair of nodes , used to represent the connection strength between nodes, represented as:

[0093] ;

[0094] wherein, is the spatiotemporal weighted coefficient between node i and node j, is a hyperparameter for controlling the strength of the weighted coefficient.

[0095] S22 includes:

[0096] S221, in each layer of graph convolution operation, the element of the adjacency matrix is dynamically adjusted according to the spatiotemporal weighted coefficient , so as to change the coupling strength between nodes, represented as:

[0097] ;

[0098] wherein, is the adjusted adjacency matrix element, representing the spatiotemporal weighted connection strength between node i and node j;

[0099] S222, in the graph convolution operation, the dynamically adjusted adjacency matrix is used for node feature propagation, and the dynamically adjusted node feature is obtained, represented as:

[0100] ;

[0101] wherein, is the dynamically adjusted feature representation of node i at the kth layer.

[0102] S3 includes:

[0103] S31, a multi-order ChebNet graph convolution layer is used to model the spatiotemporal changes between physical factors, and the ChebNet approximates the graph convolution operation by using Chebyshev polynomials to capture the complex changes of physical factors in space and time;

[0104] S32, in the multi-order ChebNet graph convolution layer, the spatiotemporal relationship of the physical factors is gradually optimized, and the preliminary sea wave prediction result is corrected by using the optimized physical factor coupling relationship through iterative updating, and finally the optimized sea wave prediction result is output.

[0105] Through the above steps, the multi-order ChebNet graph convolution layer is introduced, the Chebyshev polynomial is used to realize efficient modeling of high-order graph structure, the complex spatial and temporal variation rules between physical factors can be deeply captured, and the accuracy of node feature expression is continuously improved through an iterative optimization process, finally the dynamic correction of the preliminary sea wave prediction result is realized, and the prediction accuracy, stability and spatiotemporal continuity are greatly improved.

[0106] S31 comprises:

[0107] S311, by Chebyshev polynomial expansion of the graph convolution kernel, the high-order dependence between nodes is modeled by using the spectral domain approximation method, and is expressed as:

[0108] ;

[0109] Wherein, is the Chebyshev polynomial convolution kernel of the lth order, is the Chebyshev polynomial convolution kernel of the l-1th order, is the Chebyshev polynomial convolution kernel of the l-2th order, is a symmetric normalized adjacency matrix representing the connection structure between nodes, is an identity matrix, is an initialized graph structure representation;

[0110] S312, the node feature optimized by the spatiotemporal weighting mechanism is taken as input, and the multi-order convolution kernel is combined to extract its embedded expression in the complex spatial structure and time evolution process, and is expressed as:

[0111] ;

[0112] Wherein, is the output feature of node i after L-order graph convolution, is the weight matrix of the lth order, and L is the maximum order of Chebyshev polynomial expansion;

[0113] S313, in the ChebNet graph convolution of each order, the node feature is fused with the neighborhood information through multi-order adjacency propagation layer by layer, and finally outputs the high-dimensional spatiotemporal feature used for downstream correction, and is expressed as:

[0114] ;

[0115] Wherein, is the final graph convolution output feature of node i, which is used for sea wave prediction correction, b is a bias term, and ReLU is an activation function.

[0116] Through the above steps, the Chebyshev polynomial expansion is used to realize efficient modeling of high-order dependence relationship of graph structure, the optimized node feature input multi-order convolution kernel is combined to extract embedded expression in complex spatial structure and time evolution, and further multi-order neighborhood propagation is used to enhance information fusion and upstream and downstream correlation expression between nodes, so that the description ability of the model to the dynamic change of the nearshore physical field is improved, and stable high-dimensional spatiotemporal feature support is provided for high-precision sea wave forecast revision.

[0117] S32 comprises:

[0118] S321, by constructing the residual relationship between the preliminary sea wave prediction result and the multi-order ChebNet output feature, the error distribution is extracted, which is represented as:

[0119] ;

[0120] wherein, is the prediction residual of node i, is the predicted feature of node i obtained by the multi-order ChebNet output;

[0121] ;

[0122] wherein, MLP is a multi-layer perceptron, is the first layer weight matrix of MLP, is the first layer bias vector of MLP, is a nonlinear activation function, is the second layer weight matrix of MLP, is the second layer bias vector of MLP;

[0123] S322, according to the optimized coupling strength of the node in the ChebNet convolution in the graph, different revision weights are assigned to the residual, which is represented as:

[0124] ;

[0125] wherein, is the revision value of node i;

[0126] S323, the preliminary prediction result and the residual revision quantity are fused, and the final optimized sea wave prediction result is output, which is represented as:

[0127] ;

[0128] wherein, is the optimized sea wave prediction result of node i.

[0129] Specific examples are as follows:

[0130] I. Scene setting:

[0131] Region selection:

[0132] A typical coastal zone is selected, which has complex seabed topography and significant wind-wave coupling disturbance.

[0133] Sensor layout:

[0134] The following sensor network is laid out in the target sea area:

[0135] Wave height sensor (4);

[0136] Wind speed and direction weather buoy (3);

[0137] Underwater topographic survey device (DEM data);

[0138] Tide gauge (1);

[0139] Data sampling frequency: 10 minutes / time, continuous collection of 24 hours of data.

[0140] Sensor data sample and graph structure initialization:

[0141] II. Sensor data sample and graph structure initialization:

[0142] The original data sample is shown in Table 1:

[0143] Table 1. Multi-source physical factor observation data of nearshore sea area

[0144] Sensing point Sea wave height (m) Wind speed (m / s) Wind direction (°) Topographic elevation (m) P1 1.2 6.8 140 -12.3 P2 1.6 7.4 155 -10.5 P3 0.9 5.6 130 -13.7 P4 1.1 6.1 135 -11.8

[0145] Graph structure modeling method:

[0146] Each sensing point is a graph node;

[0147] Adjacent physical points are connected, and the edge weight is initialized as the reciprocal of the Euclidean distance;

[0148] The node feature vector is , where is the wave height, is the wind speed, is the wind direction, is the terrain elevation.

[0149] III. Modeling details of the adaptive coupling mechanism:

[0150] Adaptive weighting coefficient calculation (take node P1 and neighbor P2 as an example):

[0151] Spatial distance weight: ;

[0152] Wind speed difference normalization: ;

[0153] Wind direction cosine similarity: ;

[0154] Final coupling strength:

[0155] ;

[0156] Pick , , The conclusion is:

[0157] ;

[0158] IV. ChebNet graph convolution modeling and correction results:

[0159] ChebNet build configuration:

[0160] Use a 3rd-order Chebyshev polynomial expansion;

[0161] Number of convolutional layers: 3;

[0162] Activation function: ReLU;

[0163] Training data: Wave height sequence measured in the past 7 days;

[0164] Loss function: MSE (predicted wave height vs. measured wave height)

[0165] Preliminary forecast results (at some times) are shown in Table 2:

[0166] Table 2 Comparison of wave forecast accuracy at different time points

[0167] Time Traditional physical model prediction (m) This method after the revised prediction (m) Measured wave height (m) 10:00 AM 1.32 1.25 1.23 10:10 AM 1.41 1.36 1.33 10:20 AM 1.35 1.30 1.29

[0168] Comparison of correction accuracy:

[0169] Original model RMSE: 0.14m;

[0170] This method RMSE: 0.06m;

[0171] The error decreased by 57.1%.

[0172] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0173] The above merely describes the preferred embodiments of the present application, and it should be pointed out that, for those skilled in the art, some improvements and refinements can be made without departing from the principles of the present application, and these improvements and refinements should also be considered as falling within the protection scope of the present application.

Claims

1. A method for nearshore sea wave forecast correction based on physical factor coupling ChebNet, characterized in that, The method comprises the following steps: S1, constructing a dynamic graph structure model based on real-time sensing data of the nearshore sea area, taking physical factors including sea wave height, wind speed, wind direction and terrain elevation as graph nodes, and dynamically constructing the spatio-temporal dependence relationship between the graph nodes through graph convolution operation to generate a preliminary sea wave prediction result; S2, introducing a spatio-temporal adaptive mechanism, dynamically weighting the influence of each physical factor in the graph convolution operation by adaptively adjusting the coupling strength between the graph nodes, and optimizing the coupling relationship of the physical factors in real time according to the spatio-temporal characteristics of different sea areas; S3, using a multi-order ChebNet graph convolution layer to model the spatio-temporal changes of the physical factors, and correcting the preliminary sea wave prediction result based on the optimized coupling relationship of the physical factors through iterative updating to output an optimized sea wave prediction result.

2. The method according to claim 1, wherein, The S1 comprises: S11, collecting various physical factors of the nearshore sea area through a real-time sensor network, including sea wave height, wind speed, wind direction and terrain elevation, and encoding the physical factors as nodes in a graph structure, initializing the nodes through a graph convolution network, converting the physical characteristics and spatial position information of the nodes into feature vectors in the graph structure, and forming a preliminary node feature representation; S12, dynamically constructing the spatio-temporal dependence relationship between the initialized nodes through graph convolution operation, describing the connectivity and mutual relationship between the nodes by using an adjacency matrix, and generating a preliminary sea wave prediction result based on the spatio-temporal dependence relationship. 3.The nearshore wave forecast correction method based on the physical factor coupled ChebNet according to claim 2, wherein, The S11 comprises: S111, collecting a plurality of physical factors from the near-shore sea area through a real-time sensor network, including sea wave height , wind speed , wind direction , terrain elevation ; S112, according to the types of physical factors, each physical factor is taken as a node in the graph structure, and the physical characteristics and spatial position information of each node are converted into a feature vector; S113, constructing an adjacency matrix A of the graph to represent the spatiotemporal dependency between different nodes, elements of the adjacency matrix A represent the connection strength between node i and node j; S114, the node feature vector and the adjacency matrix A as input, the incoming graph convolution network is initialized, the feature is propagated through the connection relationship between nodes, and the representation of the node is updated; S115, in the initialization of the graph convolution network, the feature representation of each node is updated through convolution operation, and finally the preliminary feature vector of each node is obtained as the preliminary node feature representation of the graph structure. , as the preliminary node feature representation of the graph structure.

4. The method according to claim 3, wherein, The S12 comprises: S121, through a graph convolution operation, based on the adjacency matrix A and the preliminary feature vector updating the node features, capturing the spatio-temporal dependencies, through a multi-layer graph convolution operation, to finally obtain the Kth layer node features ; S122, the node features obtained through the graph convolution operation passing it to the output layer to generate the preliminary sea wave prediction result .

5. The method of claim 4, wherein the method is characterized by, The S2 comprises: S21, introducing a spatio-temporal adaptive weighting mechanism for each physical factor node, calculating the spatio-temporal weighting coefficients between each node and its neighbor nodes by analyzing the spatio-temporal characteristics of different sea areas; S22, in the graph convolution operation, dynamically adjusting the coupling strength between the nodes by using the adaptive mechanism, and dynamically changing the weights in the adjacency matrix at each layer of graph convolution according to the spatio-temporal weighting coefficients.

6. The method of claim 5, wherein the method is characterized by, The S21 comprises: S211, evaluating the spatio-temporal characteristics of each physical factor node by analyzing the changes of each physical factor node in different time and space environments; S212, calculating the spatio-temporal distance between each pair of nodes by the spatial position difference and physical feature difference between the nodes, for quantifying the spatio-temporal dependence relationship between the nodes; S213, based on the calculated spatio-temporal distance, calculate a spatio-temporal weighting coefficient for each pair of nodes , for representing the connection strength between nodes.

7. The method according to claim 6, wherein, The S22 comprises: S221, in each layer of graph convolution operation, the element of the adjacency matrix According to the spatiotemporal weighting coefficient The coupling strength between nodes is dynamically adjusted, so that the coupling strength between nodes is changed; S222, in the graph convolution operation, using the dynamically adjusted adjacency matrix to propagate the node features, and obtaining the dynamically adjusted node features.

8. The method according to claim 7, wherein, The S3 comprises: S31, using a multi-order ChebNet graph convolution layer to model the spatio-temporal changes between the physical factors, and ChebNet approximates the graph convolution operation by using Chebyshev polynomials to capture the complex changes of the physical factors in space and time; S32, in the multi-order ChebNet graph convolution layer, the spatio-temporal relationship of the physical factors is gradually optimized, and the preliminary sea wave prediction result is corrected by using the optimized physical factor coupling relationship through iterative updating, and finally an optimized sea wave prediction result is output.

9. The method according to claim 8, wherein, The S31 comprises: S311, modeling the high-order dependence between nodes by Chebyshev polynomial expansion of the graph convolution kernel in a spectral domain approximation manner; S312, taking the node features optimized by the spatiotemporal weighting mechanism as input, combining the multi-order convolution kernel to extract the embedded expression of the node in the complex spatial structure and the time evolution process; S313, in the ChebNet graph convolution of each order, the node features are fused with the neighborhood information layer by layer through multi-order adjacency propagation, and finally output the high-dimensional spatiotemporal features for downstream correction.

10. The method of claim 9, wherein the method is characterized by, The S32 comprises: S321, by constructing the residual relationship between the preliminary sea wave prediction result and the multi-order ChebNet output feature, the error distribution is extracted; S322, according to the optimization coupling strength of the nodes in the graph in the ChebNet convolution, different correction weights are assigned to the residual; S323, the preliminary prediction result is fused with the residual correction amount, and the final optimized sea wave prediction result is output.

Images