Search

CN-121765359-B - Regional sea wave forecasting method based on single-point buoy observation and graph neural network

CN121765359BCN 121765359 BCN121765359 BCN 121765359BCN-121765359-B

Abstract

The application discloses a regional sea wave forecasting method based on single-point buoy observation and a graphic neural network, and belongs to the field of ocean environment forecasting. The solution of MESHGRAPHNET model based on Graph Neural Network (GNN) and fused with single-point buoy observation is innovatively provided, so that the efficiency, precision and timeliness of effective wave height forecasting of sea waves are remarkably improved. The method comprises the steps of carrying out time sequence feature extraction and short time prediction on an observation sequence of a single buoy station through an LSTM network, then taking a prediction result output by the LSTM and ERA5 space wind field features together as node features of a space graph neural network, encoding the node wind field features and the side geometric features into a multidimensional potential space by an encoder of the graph neural network, carrying out side update, node aggregation and global state update through a message transmission layer to fuse local neighborhood information and global sea state, and finally outputting effective wave height at the time of prediction by a decoder.

Inventors

  • YU HAIQING
  • LV TING
  • CHEN SONGGUI
  • ZHU YINGTAO
  • LU LV

Assignees

  • 山东大学
  • 交通运输部天津水运工程科学研究所
  • 哈尔滨工程大学

Dates

Publication Date
20260505
Application Date
20260304

Claims (7)

  1. 1. A regional sea wave forecasting method based on single-point buoy observation and graph neural network is characterized in that an observation sequence of a single buoy station is subjected to time sequence feature extraction and short-time forecasting through an LSTM network, and then a forecasting result output by the LSTM and ERA5 spatial wind field features are used as node features of the spatial graph neural network together; The encoder of the graph neural network encodes the node wind field characteristics and the edge geometric characteristics into a multidimensional potential space, performs edge updating, node aggregation and global state updating through a message transmission layer to fuse local neighborhood information and global sea area states, and finally outputs effective wave heights at the pre-report time by a decoder; Comprising the steps of, Step (1), problem modeling and data preprocessing; Step (2), modeling by adopting LSTM; modeling a site effective wave height sequence by adopting a long-short-term memory network LSTM (least squares) to provide a time sequence forecasting result of a single site, and outputting a single-point forecasting value at a target moment; step (3), processing the graph neural network; step (3.1), inputting characteristics and single-point condition constraints; The effective wave height at the forecasting time is driven by a wind field sequence, and condition information of a single site is introduced as external constraint; Single point condition information is given by LSTM to site sequence The forecast value of the time is recorded as The method comprises The forecast value of the moment is used as a condition quantity input in the graph model, and is injected with the same value at each node, so that the consistency constraint is applied to the full-field prediction; For any node The node input characteristics consist of a historical wind field, a current wind field and single-point condition quantity, and the expression is as follows: ; Wherein, the A10 meter wind vector at the node; Step (3.2), constructing edge attributes; Constructing edge attribute based on two-dimensional position coordinates of grid nodes, and constructing edge attribute for any adjacent node pair Node is provided with Is the two-dimensional position coordinates of (a) Node Is the two-dimensional position coordinates of (a) Wherein And (3) with Respectively, generating a set of adjacent edges according to the connection relation of the vertexes of the triangular grid cells, and generating each directed edge for each directed edge Constructing a corresponding edge attribute vector; step (3.3), encoding-processing-decoding; step (3.3.1), encoding stage; FVCOM unstructured grid is represented as ; Each node With input feature vectors Each edge Having geometric edge properties The node coordinates are A section; point features and edge attributes are first mapped linearly to hidden space dimensions The initial hidden variable is obtained as shown in the following formula: Wherein, the Representing nodes The hidden representation at the initial node after encoding, Representing edges The coded initial edge hidden representation is marked as a layer index, and is used for distinguishing characteristic representations of different layers; Is a leachable linear mapping layer for projecting original nodes and edge features into dimensions The application takes the hidden space of ; Step (3.3.2), a treatment stage; The processor is composed of Stacking message passing layers, wherein each layer adopts a calculation paradigm of updating edge characterization firstly, and then aggregating and updating node characterization; The method comprises the steps of introducing a gating coefficient based on the consistency of wind direction and side direction in a side updating link, continuously modulating the strength of an edge message, naturally weakening the influence of a modulating item on the transmission when wind direction information is not representative, enabling the modulating item not to depend on the explicit judgment of sea condition types (storms or swells) but to serve as a controllable structure prior, enhancing the transmission of neighborhood information along the wind direction under the condition that the wind direction information has more interpretation power, and weakening the influence when the wind direction interpretation power is insufficient; step (3.3.3), decoding stage; Outputting final node hidden representation in processing stage Decoding to obtain a node effective wave height forecast result according to the following formula:
  2. 2. The method for regional ocean wave forecasting based on single-point buoy observation and graph neural networks according to claim 1, wherein in the message passing layer, single-point buoy data affects adjacent nodes.
  3. 3. The regional sea wave forecasting method based on the single-point buoy observation and graph neural network according to claim 1, wherein the step (1) comprises, Step (1.1), diagram representation; The FVCOM unstructured grid is represented as undirected graph G, expressed as: ; Wherein, the A set of grid nodes is represented, Representing a set of edges formed by spatially adjacent nodes; The node set is represented as , wherein, For the node index to be a function of the node index, The total number of the grid nodes; Edge set The method is composed of adjacent node pairs in grid topology, and is used for describing a space adjacent relation on a non-structural grid and providing a structural basis for subsequent space information propagation; step (1.2), task definition; The task is defined as a node level regression problem, and under the condition of given wind field information and single buoy station wave height, the effective wave height space distribution at the target moment is forecasted; For any node At the time of prediction Is effective wave height for prediction of (2) Respectively marked as the effective wave height and the reference effective wave height ; Given forecast time Selecting a time interval The wind field sequence within, constructs a node input, wherein, Representing a historical wind field window length; step (1.3), standardized treatment; And uniformly adopting Z-core standardization processing based on training set statistics for input characteristics and supervision output.
  4. 4. The regional sea wave forecasting method based on the single-point buoy observation and graph neural network according to claim 1, wherein the step (2) comprises, Step (2.1), sample construction; Constructing a supervision learning sample according to the effective wave height time sequence of the site in a sliding window mode; step (2.2), LSTM time sequence modeling; the LSTM network is used for modeling the time dependency relationship in the single site effective wave height sequence, and modeling the long-term trend information and the short-term fluctuation characteristic is realized by introducing a memory unit and a gating mechanism; for each input sample, the LSTM network reads the historical effective wave height sequence with the length of 64 step by step according to the time sequence, and at each time step, the network sequentially executes the following state updating process based on the effective wave height input value of the current time step and the hidden state of the last time step: step (2.3), single-point prediction output; the time sequence features are input into a linear output layer, mapped to generate a one-step advanced effective wave height predicted value of a target moment, and quantitative prediction of the single station significant wave height is realized; Step (2.4), model training and parameter updating; In the training stage, comparing the predicted effective wave height output by the LSTM network with the reference effective wave height at the corresponding target moment, adopting a regression error as a training target, and updating network parameters in an iterative optimization mode to enable the model to gradually learn the time evolution rule of the effective wave height of the single site; Step (2.5), reasoning stage and continuous forecast; In the reasoning stage, a sliding window input mode consistent with the training stage is adopted to predict the historical effective wave height sequence of the continuous time period window by window, and a single-point effective wave height value corresponding to the target moment is output, so that a single-station effective wave height forecasting sequence on a continuous time scale is formed.
  5. 5. The regional sea wave forecasting method based on single-point buoy observation and graph neural network according to claim 4, wherein the step (2.2) comprises, Firstly, modulating the state of a memory unit in the last time step through a forgetting door, and controlling the proportion of information to be reserved or attenuated in history memory, thereby avoiding the interference of irrelevant or outdated information on the current modeling process; secondly, screening the input information of the current time step through an input gate, and determining the degree of writing new information into a memory unit by combining candidate memory contents generated by the current input so as to introduce effective characteristics related to the current wave evolution; then, the history memory reserved by the forgetting gate is fused with the new memory introduced by the input gate, and the memory unit state of the current time step is updated to enable the network to continuously accumulate the history evolution information of the effective wave height in the sequence advancing process; Finally, modulating the current memory unit state through an output gate to generate a hidden state of the current time step, wherein the hidden state is used as an output representation of the current time step and also used as an input of state update of the next time step; and after the input of the complete history window sequence is completed, taking the hidden state corresponding to the last window time as the comprehensive time sequence characteristic representation of the history sequence, and representing the whole state of the effective wave height of the station before the predicted target time.
  6. 6. The regional sea wave forecasting method based on the single-point buoy observation and graph neural network according to claim 1, wherein the step (3.2) comprises, Step (3.2.1), distance calculation and standardization; Push-down type computing node And node Is defined by the geometric distance of: distance set for all sides Calculating the mean value and standard deviation, and obtaining a standardized distance according to a standardized formula ; Step (3.2.2), direction angle calculation and direction coding; The edge direction angle is calculated from the node coordinate difference as follows: Wherein, the Representing slave nodes Pointing node Is provided with a pair of side direction angles, Representing an arctangent function, typically ranging from ; The directional code is then constructed as follows: Wherein, the , Is the sine and cosine coding of the direction angle, used for avoiding the angle in the process of angle Discontinuous at the location; Step (3.2.3), splicing edge attributes; splicing the standardized distance and direction codes into edge attribute vectors according to a fixed sequence Constructing completed edge attributes The geometric descriptive quantity of the edge in the figure is input into an edge encoder and is mapped to the hidden space to participate in the process of message transmission and feature updating.
  7. 7. The regional sea wave forecasting method based on the single-point buoy observation and graph neural network according to claim 1, wherein the step (3.3.2) comprises, Determining the modulation factor of direction consistency, letting the node Is the forecast time of (2) Is of the wind vector of (1) Node Wind vector at the forecast time of (2) Firstly, the wind speed vectors at two ends are averaged according to the following steps to obtain the local wind speed vector on the edge: then proceed as follows Normalizing to obtain a unit wind direction vector: Wherein, the Representing a second norm (Euclidean norm), namely the Euclidean length; the edge direction vector is obtained by node geometric coordinate difference according to the following formula: normalized as unit direction vector of edge as follows: Wherein, the Determined by static grid geometry, independent of time; Finally, adopting the edge-level wind direction unit vector And unit vector of side direction The dot product of (2) measures the directional consistency and obtains the gating weight of the edge message through Sigmoid mapping according to the following formula: Wherein, the For a fixed scaling factor for adjusting the weight discrimination, Is a Sigmoid function of the code, Scalar weights obtained by the consistency of wind direction and edge direction are used for modulating the intensity of edge messages; Perform edge update, in the first In layer messaging, edge update takes hidden representation of two end nodes and hidden representation of current edge as input, and the function is updated Obtaining candidate edge representation and further introducing direction consistency gating weight And performing multiplicative modulation on the candidate edge representations to obtain updated edge hidden representations, wherein the calculation formula is as follows: Wherein, the Respectively represent the first Layer node And node Is defined by the node hidden vector of (a), Represent the first Layer node Pointing node Is the node Is informative for aggregation and edge feature updates, The resulting candidate edge hidden vectors are updated for the edges, For gating the weighted edge hidden vector, The method is realized by a multi-layer perceptron, and is formed by connecting a plurality of linear transformation layers and nonlinear activation functions in series according to a fixed sequence, and the general form is expressed as follows: Wherein, the The input vector is represented as such, In order for the parameters to be able to be learned, Through the nonlinear mapping, the original node characteristics and the edge attributes are projected to a unified hidden space representation, so that the characteristic expression capability is enhanced while the physical information integrity is maintained, and a unified representation basis is provided for the subsequent message transmission and characteristic update based on a graph structure; Performing node aggregation and node update, in the first place In layer messaging, for each node Converging all the incoming edges The carried side information; To distinguish between incoming edge aggregation, each undirected edge is typically implemented Is unfolded into two directional edges And (3) with Order-making machine Is a node Is a neighborhood of (1), then the aggregate vector is defined as Wherein, the Representing slave nodes Pointing node Is represented by a directed edge of (a); then splicing the node self representation with the aggregation message, and finishing node updating through a multi-layer perceptron: Wherein, the Representing a vector concatenation operation; For a pair of From 1 to Sequentially performing the following operations of calculating the gating weight of the consistency of the wind direction and the side direction Performing edge update to obtain candidate edge representation and pressing Multiplicative weighting to obtain edge representations Performing node edge message aggregation and completing node update to obtain 。

Description

Regional sea wave forecasting method based on single-point buoy observation and graph neural network Technical Field The application relates to the field of marine environment forecasting, in particular to a regional sea wave forecasting method based on single-point buoy observation and a graph neural network. Background The sea wave forecasting is a key technology in the fields of ocean engineering, shipping safety, disaster prevention, disaster reduction and the like, the existing sea wave numerical forecasting method is based on a model of a physical equation, such as WAVEWATCH III, SWAN and the like, is high in calculation complexity, depends on the accuracy of initial conditions and boundary conditions, and can finish forecasting of 7-10 days in the future only after time is spent for the refinement (the spatial resolution of meters to hundred meters) forecasting of a local sea area. Currently, data-driven machine learning methods provide new solution paths for sea wave forecasting, such as Convolutional Neural Networks (CNNs) or Recurrent Neural Networks (RNNs) based on regular grids, which are difficult to effectively process unstructured, irregularly distributed station/grid point data common in marine monitoring after application. The current more efficient techniques are based on spatio-temporal prediction models of the Graph Neural Network (GNN), such as the "MESHGRAPHNET" architecture proposed by team DeepMind (see references: p.w. Battaglia et al, "Relational inductive biases, DEEP LEARNING, and graph Networks", arXiv:1806.01261, 2018; and j.b. Ham et al, "Learning to Simulate Complex PHYSICS WITH GRAPH Networks", arXiv:2002.09405, 2020). DeepMind's MESHGRAPHNET uses GNN for fluid simulation, the core of which is to construct a graph on a triangular mesh, fusing neighborhood information by messaging. But the Graph Neural Network (GNN) has significant limitations in the field of sea wave forecasting. The method is characterized by comprising the steps of verifying ideal fluid/simple scenes, including physical simulation such as cylindrical detour and marble collision, not verifying real ocean wave data, failing to combine ocean observation conditions, driving a plurality of ocean wave forecasting models only by wind fields, and neglecting critical observation data in ocean forecasting, wherein no practical business design means such as sparse observation (such as single-point buoy) and other practical business constraints are not considered. Thus, the GNN of the prior art has potential, but has not been effectively combined with the sea wave forecast business requirements. In view of this, the present application has been made. Disclosure of Invention The application provides a regional sea wave forecasting method based on single-point buoy observation and a graph neural network, which aims to overcome the technical difficulties that the input is regularized and analyzed lattice point data and real-time observation data is not deeply coupled in the prior art, and innovatively bases on the graph neural network (Graph Neural Network, GNN) and fuses the solution of a MESHGRAPHNET model of single-point buoy observation so as to fill the blank of the prior art and remarkably improve the efficiency, precision and timeliness of sea wave effective wave height (SIGNIFICANT WAVE HEIGHT, hs) forecasting. In order to achieve the aim, the regional sea wave forecasting method based on single-point buoy observation and graph neural network provided by the application is that an observation sequence of a single buoy station is subjected to time sequence feature extraction and short-time forecasting through an LSTM network, and then a forecasting result output by the LSTM and ERA5 spatial wind field features are used as node features of the spatial graph neural network together; The encoder of the graph neural network encodes the node wind field characteristics and the edge geometric characteristics into a multidimensional potential space, performs edge updating, node aggregation and global state updating through a message transmission layer to fuse local neighborhood information and global sea area states, and finally outputs effective wave heights at the pre-report time by a decoder. Further, in the messaging layer, single point buoy data affects neighboring nodes. The regional sea wave forecasting method based on the single-point buoy observation and the graph neural network comprises the following steps of: Step (1), problem modeling and data preprocessing; Step (2), modeling by adopting LSTM; modeling a site effective wave height sequence by adopting a long-short-term memory network LSTM (least squares) to provide a time sequence forecasting result of a single site, and outputting a single-point forecasting value at a target moment; step (3), processing the graph neural network; step (3.1), inputting characteristics and single-point condition constraints; The effective wave height at the forecasting time is