Search

CN-122022207-A - Storm surge intelligent forecasting method integrating causal structure and graph neural network

CN122022207ACN 122022207 ACN122022207 ACN 122022207ACN-122022207-A

Abstract

A storm surge intelligent forecasting method integrating a causal structure and a graph neural network relates to the technical field of artificial intelligence and ocean storm surge forecasting, and comprises the following steps of 1, multi-source data integration, 2, data preprocessing and standardization, 3, fusion modeling of storm surge weather space-time characteristics and water level time sequence characteristics by using a UNet-LSTM, 4, causal structure inference and graph structure learning, 5, obtaining a UNet-LSTM-DGCN combined model by using a graph convolution neural network to conduct final storm surge water level forecasting, 6, combined model training and verification, 7, forecasting result visualization, and 8, interpretation analysis. According to the method, storm tide forecasting is carried out by constructing an intelligent method combined with a dynamic graph neural network, dynamic causal relations in multi-station water level change under the influence of typhoons are effectively captured, forecasting accuracy is improved, and meanwhile decision basis inside a model is revealed through a visual graph structure.

Inventors

  • WANG ZHIFENG
  • ZHU ZHICHENG
  • WANG MENGKE

Assignees

  • 中国海洋大学

Dates

Publication Date
20260512
Application Date
20260415

Claims (8)

  1. 1. A storm surge intelligent forecasting method integrating a causal structure and a graph neural network is characterized by comprising the following steps: Step 1, multisource data fusion, namely collecting water level data of a plurality of hydrologic stations, weather stations and authoritative global weather forecast system data, wherein the water level data of the stations comprise historical and real-time water level changes; step 2, data preprocessing and standardization; the method comprises the steps of 3, realizing fusion modeling of storm tide weather space-time characteristics and water level time sequence characteristics by using a UNet-LSTM, completing water level preliminary prediction by using a UNet-LSTM combined model, extracting weather field space characteristics by using the UNet, modeling water level time sequence dependence by using the LSTM, and deeply fusing the two types of characteristics to output a preliminary prediction result; Step 4, causal structure inference and graph structure learning; step 5, obtaining a UNet-LSTM-DGCN combined model by using a graph convolution neural network to forecast the final storm surge level; Training and verifying a UNet-LSTM-DGCN combined model; Step 7, visualizing a prediction result; And 8, performing interpretation analysis.
  2. 2. The method for intelligently forecasting storm surge by fusing a causal structure and a graph neural network as claimed in claim 1, wherein the step 2 comprises the following specific steps: (a) Abnormal value detection and elimination, namely performing abnormal recognition and cleaning on time sequence data such as water level, wind speed and air pressure, performing abnormal value detection by adopting a method based on time sequence change rate and local neighborhood consistency test, and setting a water level time sequence as Corresponding time is Defining the water level change rate at adjacent moments: ; According to statistics of historical storm tide process of the research sea area, a reasonable water level change rate threshold interval is determined: [ , for the time when the change rate exceeds the interval, further calculating the average change rate in the local neighborhood If the point change rate deviates from the whole trend of the neighborhood remarkably and no physical continuous evolution support exists, judging an abnormal value caused by equipment failure or transmission error, and eliminating and marking the abnormal value as missing; (b) Interpolation of missing values, namely filling missing points in a one-dimensional water level time sequence by adopting linear time sequence interpolation, and setting time Level of water The missing has the effective observed values of front and back respectively , Corresponding time is , The interpolation formula is: ; For the two-dimensional meteorological lattice point field data of wind speed and air pressure, bilinear interpolation is adopted to realize space deletion complementation, so that multisource data is kept continuous and complete in space-time while the space distribution rationality of the meteorological lattice point field is maintained, and the coordinates of the points to be inserted are set as Around which 4 known lattice points have values of respectively First at Direction interpolation, again at The final value is obtained by directional interpolation, and the formula is as follows: ; ; ; (c) Data normalization, namely normalizing the preprocessed data by adopting a Z-score normalization method, mapping the data into standard normal distribution with the mean value of 0 and the standard deviation of 1, and setting a certain characteristic sequence as Its average value And standard deviation The calculation is as follows: ; ; ; Wherein, the Is a normalized characteristic value.
  3. 3. The method for intelligently forecasting storm surge by fusing a causal structure and a graph neural network as claimed in claim 2, wherein the step 3 comprises the following specific steps: (a) Aiming at the data of the meteorological lattice point fields of wind speed and air pressure, a UNet network is adopted to capture the spatial distribution and evolution rule of the meteorological lattice point fields, and the input meteorological lattice point fields are set as , wherein, Divided into a height and a width of the area matrix, The encoding stage extracts multi-scale space features through convolution layer downsampling, and the decoding stage restores the space dimension through upsampling and jump connection, and finally outputs the meteorological space features: , wherein, Is a feature dimension; (b) Using Is to weather space characteristics by convolution kernel of (a) Mapping to each site, and splicing with each site water level time sequence data in characteristic dimension to form site multi-characteristic matrix integrating weather space-time information , wherein, For the number of sites to be the number of sites, For the time step size of the time step, Is a feature dimension; (c) Modeling long-term and short-term time sequence dependence of the site multi-feature matrix by adopting an LSTM network, and finally outputting a preliminary forecast site storm tide level, wherein the specific formula is as follows: ; ; ; ; ; ; Wherein, forget the door Determining which information is to be removed from long-term memory; representing Sigmoid activation functions; is a weight matrix associated with the forget gate; Is the output result of the last time step; Is the input value of the current time step; Is an offset item of a forgetting door, and is input into the door Determines the current input The degree of influence on the memory of the updating unit; Is a weight matrix of the input gates, And Is a bias term; A weight matrix of candidate cell states, a memory cell For storing long-term dependencies in data; Representing a site multi-feature matrix.
  4. 4. The method for intelligently forecasting storm surge by fusing a causal structure and a graph neural network as claimed in claim 3, wherein the step 4 comprises the following specific steps: (a) Splicing the preliminary forecasting results of the preliminary time water levels of a plurality of stations and UNet-LSTM in the time dimension to form a continuous water level change matrix ; (B) Static global structure learning by two learnable node embedding layers , Mapping stable statistical characteristics of a site to a high-order characteristic space, and simultaneously introducing super parameters Controlling an activation function To avoid oversaturation of the characteristic values, the formula is: ; ; Wherein, the , Representing the output dimension of the embedded layer for core attribute association of the encoding site; Calculation of The difference of the outer products of the two embedded matrices forms an antisymmetric matrix The calculation is expressed as: ; ; finally, compressing the numerical value to the interval of [ -1, 1] through the tanh function, filtering invalid negative association by utilizing the ReLU function, and only preserving forward propagation association to obtain a static global adjacency matrix ; (C) Local dynamic structure learning, namely performing feature expansion on a site time sequence through time delay modeling and focus learning along with local association of dynamic change of a storm surge process, extracting global features of site time sequence by Conv 2D, performing dimension lifting by linear mapping to obtain a general time sequence feature matrix F, and then performing hidden feature number after dimension lifting According to time delay coefficient Resolution into factor fragments Fruit segment ; A characteristic representing the cause of the spread, Representing propagation response characteristics, learning responses of early propagation causes to later sites after dynamic time delay, calculating similarities between the cause segments and the fruit segments through batch matrix multiplication, quantifying time-varying time delay association strength in the whole period between the sites, dynamically adjusting element values of the similarity matrix along with storm surge development to capture association strength change between the sites when typhoon paths deviate, learning relations between the sites by using the similarities of the two segments to obtain a dynamic graph structure, and expressing a calculation formula as follows: ; ; ; Wherein, the Representing the linear layer parameters that can be learned, Representing a dynamic graph structure; (d) Graph fusion and sparsification the fusion formula is expressed as: ; Adopts Top-k sparsification strategy to pair Is kept with the maximum association strength The value is set to 0 and the rest is set to 0, the process is realized by a mask matrix, and a dynamic adjacent matrix is finally obtained ; ; ; Wherein, the ". If represents Hadamard product.
  5. 5. The method for intelligently forecasting storm surge by fusing a causal structure and a graph neural network as claimed in claim 4, wherein said step 5 comprises the following specific steps: First, to ensure that the own characteristics of each site are not lost in propagation, identity matrices are added to adjacent matrices Adding self-loop terms, summing according to rows to obtain node degree vector, and constructing degree matrix Then calculate the inverse square root of the degree matrix Symmetric normalization is achieved by matrix multiplication, and the process is expressed as: ; ; ; considering the complexity of storm tide time-space association, adopting a multi-layer stacked graph rolling structure, and setting the number of graph rolling layers as , And Respectively expressed as a prepositive time length and a forecast time length, and the characteristic dimension sequence is that Respectively expressing hidden layer dimensions of each layer, wherein each layer shares a normalized dynamic adjacency matrix Definition of the first The input features of the layers are The output is characterized by The hierarchical iteration formula and the final output are expressed as: ; ; Wherein, the Is the first The learnable weights of the layer linear layers, In order to obtain the final prediction result, 。
  6. 6. The method for intelligently forecasting storm surge by fusing a causal structure and a graph neural network as defined in claim 5, wherein the step 6 specifically comprises: The UNet-LSTM-DGCN combined model takes normalized time sequence data of storm tide water level and weather lattice point field data at the front moment as input, takes station water level at the corresponding moment as supervision label, adopts Mean Square Error (MSE) as a loss function, quantifies the deviation of a predicted value and a true value, and has the following calculation formula: ; Wherein, the As a total number of samples, Is the first The actual storm surge level at the moment, Predicted water level for model output; The UNet-LSTM-DGCN combined model is trained by updating network parameters by adopting an Adam optimizer, minimizing a loss function through a back propagation algorithm, dividing a data set into a training set and a verification set according to a ratio of 7:3, wherein the training set is used for parameter learning, the verification set is used for monitoring loss change in real time, and the overfitting phenomenon is restrained through a premature stopping strategy.
  7. 7. The method for intelligently forecasting the storm surge by fusing a causal structure and a graph neural network is characterized in that the step 7 specifically comprises the steps of drawing a comparison curve of a single-station actual measurement water level, a preliminary prediction sequence and a final prediction sequence, showing fitting effects on the storm surge water increasing, water reducing and peak processes, drawing a multi-station water level space thermodynamic diagram by combining regional geographic information, superposing causal related edges among stations, visually showing water level space distribution and propagation paths, showing a causal relation matrix obtained by learning in a network graph mode, showing causal influence intensity among stations in a weighted directed edge mode by using nodes, and clearly showing a core causal link of storm surge water level propagation.
  8. 8. The method for intelligently forecasting storm surge by fusing a causal structure and a graph neural network according to claim 7, wherein the step 8 specifically comprises quantifying the contribution degree of meteorological elements and historical water levels to a forecasting result through characteristic importance indexes, quantifying causal influence weights among different sites based on causal edges output by the graph structure, identifying key driving sites and leading effects of the key driving sites on water level changes of peripheral sites, explaining model decision basis from a physical mechanism level, combining forecasting residual errors with causal structure distribution, further positioning forecasting deviation caused by insufficient association capturing, and providing traceable and interpretable scientific support for model optimization and business application.

Description

Storm surge intelligent forecasting method integrating causal structure and graph neural network Technical Field The invention relates to the technical field of artificial intelligence and ocean storm surge prediction, in particular to a storm surge intelligent prediction method integrating a causal structure and a graph neural network. Background The traditional storm tide prediction is highly dependent on a complex numerical mode, and the coupling power process of ocean and atmosphere is restored through numerical calculation, so that the model needs to comprehensively consider the sea water power characteristics, complex shoreline and submarine topography, typhoon wind field, air pressure field and other physical elements, and describes the interaction between different time and space scales, so that the calculation amount is large, the time consumption is long, and the prediction result has certain complexity and uncertainty. From the technical development perspective, the storm surge forecasting method mainly comprises two generations of evolution, wherein the first generation is a numerical forecasting method based on a physical equation, the representative model comprises MIKE, ADCIRC, FVCOM and the like, and the second generation is an intelligent forecasting method based on deep learning. The Graph Neural Network (GNNs) has important significance in causal inference and interpretability analysis, the GNN can effectively represent and learn complex graph structure data, is suitable for modeling relations among variables, thus providing a good framework for causal inference, and can clearly identify which nodes or edges in the graph structure have the greatest contribution to a prediction result through the graph structure, thus providing interpretable feature significance and helping to understand the decision basis of a model. In summary, the prior art has the following drawbacks: 1. The numerical forecasting model represented by MIKE, ADCIRC, FVCOM is used for carrying out discrete solving on a shallow water equation set based on a finite difference or finite element method, a high-precision calculation grid is required to be constructed, the early preparation period of grid subdivision is as long as a plurality of weeks to a plurality of months, the equation set is required to be completely solved every time the numerical model is forecasted on a calculation level, the high-resolution simulation takes a plurality of hours to tens of hours, the forecasting requirements of high frequency and quick updating under emergency response are difficult to meet, and when the aggregate forecasting is required to be carried out to quantify uncertainty, the calculation resource requirement is multiplied, so that the application capability of the system under the scene with high timeliness requirement is further limited; Firstly, the models are purely data driving methods in nature, but not the causal relation among understanding variables, and an intelligent model generally lacks an explicit causal structure modeling module, so that the models are easy to learn false association and not real physical causal chains when facing extreme typhoon events with different training data distribution, and when storm surge propagation paths are dynamically changed due to different typhoon paths, the models cannot be adaptively adjusted, and are directly expressed as problems of underestimation of peak water level forecast, obvious time lag of a forecast sequence and an actual measurement sequence and the like, so that the robustness of the models under extreme disaster scenes is seriously restricted; 2. The existing deep learning model has the common problem of 'black box', the internal calculation process comprises multiple layers of nonlinear transformation, the characteristics gradually lose physical meaning in layer-by-layer transmission, the decision boundary is difficult to understand by human, and a predictor cannot know whether the model obtains a final prediction conclusion based on which sites, which moments and which physical factors, and the interpretable problem exists; 3. Although the graph neural network method starts to try to model the space dependency relationship among stations, the prior method is mostly based on a static graph structure, the method usually predefines an adjacent matrix according to the geographic distance among stations, and presumes that the relationship is fixed under different typhoon events and different time steps, however, the physical propagation process of storm surge has obvious dynamic characteristics, namely, when typhoons log in from different directions, the propagation path of storm surge, the sequence of affected stations and the strength of causality relationship among stations are all obviously changed, and the static graph structure is essentially a rigid constraint, so that the information transfer relationship among stations cannot be dynamically adjusted according