Typhoon multi-tide station water level prediction method based on dynamic attention map circulation network

Abstract

The invention provides a typhoon multi-tide-gauge-station water level prediction method based on a dynamic attention map circulation network. The method comprises the following steps: acquiring water level data and meteorological data of a plurality of tide gauges during a typhoon; performing time sequence characteristic processing on the acquired water level data and meteorological data, and creating a topological structure; defining a dynamic graph attention hierarchical aggregation network model structure, and constructing a dynamic attention graph loop network model; using the processed data and the graph topology structure to train and optimize a dynamic attention graph loop network model; and carrying out real-time prediction on the water levels of the coastal multi-tide stations in the future A hours in the typhoon process through the optimized dynamic attention map loop network model. According to the method, prediction precision and calculation efficiency can be effectively considered, and the problems of insufficient prediction precision and calculation power waste are solved.

Description

Technical Field

[0001] This invention relates to the field of marine disaster prevention and mitigation technology, specifically to a method for predicting water levels at multiple tide gauge stations during typhoons based on a dynamic attention graph cyclic network. Background Technology

[0002] Storm surges are among the most destructive marine disasters along my country's coast. Their sudden rise in water levels often causes dikes to overflow, seawater to flood inland, and urban flooding, resulting in significant casualties and economic losses. Therefore, achieving rapid and accurate forecasting of coastal water levels is of urgent and significant practical importance for improving the disaster prevention and mitigation capabilities of coastal cities and protecting the lives and property of the people.

[0003] Current water level prediction relies mainly on hydrodynamic numerical models, which require high-precision shoreline and topographic data and are extremely dependent on computing resources. They are also unable to meet the need for rapid forecast updates when typhoons are approaching. At the same time, these physical models are sensitive to real-time input errors in typhoon wind and pressure fields, and their parameterization schemes are highly empirical.

[0004] In recent years, although machine learning has been introduced into single-station water level prediction and achieved some success, existing methods such as LSTM and GRU treat each tide gauge station as an isolated node, ignoring the coordinated changes between adjacent stations during typhoons caused by topographic propagation and wave superposition. This leads to a rapid amplification of prediction errors when the path changes abruptly or the local wind field intensifies sharply. While graph neural networks using static adjacency matrices can consider the spatial relationships between stations, they cannot adjust the coupling strength between nodes in real time as the typhoon moves and the meteorological field evolves. This results in insufficient adaptability of the model during strongly non-stationary processes, making it difficult to balance prediction accuracy and computational efficiency. Summary of the Invention

[0005] In view of this, the problem to be solved by the present invention is to provide a method for predicting water levels at multiple tide gauge stations during typhoons based on dynamic attention graph cyclic networks. This method can effectively balance prediction accuracy and computational efficiency, and solve the problems of insufficient prediction accuracy and wasted computing power.

[0006] This invention solves the above-mentioned technical problems through the following technical means: This invention provides a method for predicting water levels at multiple tide gauge stations during typhoons based on dynamic attention graph cyclic networks, comprising the following steps:

[0007] Acquire water level data and meteorological data from multiple tide gauge stations during the typhoon;

[0008] The acquired water level and meteorological data are subjected to time-series feature processing, and a topology structure is created.

[0009] Define the structure of the dynamic graph attention hierarchy aggregation network model and construct the dynamic attention graph recurrent network model;

[0010] The dynamic attention graph recurrent network model is trained and optimized using the processed data and graph topology.

[0011] The optimized dynamic attention graph recurrent network model is used to predict the water levels of multiple tide gauge stations along the coast in real time within the next A hours during a typhoon.

[0012] Furthermore, the time-series feature processing of the acquired water level data and meteorological data includes selecting water level observation data and meteorological data from all tide gauge stations in the region, setting an a-hour sliding window, and using the sliding window method to create a time-series feature for each time point. This feature consists of the water level sequence of the past a hours and the corresponding meteorological elements, and the water level height at the next moment immediately following the window is used as the prediction label for the sample.

[0013] Furthermore, the meteorological elements include: east-west wind speed component at 10 meters altitude, north-south wind speed component at 10 meters altitude, air temperature at 2 meters altitude, dew point temperature at 2 meters altitude, surface air pressure, sea level air pressure, sea surface temperature, and precipitation.

[0014] Furthermore, the creation of the graph topology includes:

[0015] Each tide gauge station in the study area is used as a graph node to obtain the node feature matrix;

[0016] The dynamic coupling strength was measured by calculating the Pearson correlation coefficient of water level fluctuations between any two stations, using the water level sequence observed synchronously at various tide gauge stations during historical typhoons as a sample.

[0017] When the correlation coefficient exceeds a preset threshold, edge weights are established between the corresponding nodes.

[0018] Furthermore, the construction of the dynamic attention graph recurrent network model also includes:

[0019] The input layer receives the standardized 12-hour water level-meteorological joint features;

[0020] The spatial feature extraction layer uses a graph attention network to calculate the attention weights between nodes.

[0021] Hierarchical sampling and aggregation layer: Based on the top K selection strategy of graph sampling and aggregation network, weighted aggregation of first-order, second-order and higher-order neighborhood information is performed to extract multi-scale spatial features.

[0022] The temporal feature modeling layer uses gated recurrent units to capture the nonlinear temporal dependence between water level and meteorological field;

[0023] The output layer maps the fused spatiotemporal features to the water level height of multiple tide gauge stations in the next A hours through a linear layer.

[0024] Furthermore, the spatial feature extraction layer employs a dynamic attention mechanism to achieve adaptive coupling between tide gauge station nodes;

[0025] It passed Calculate the attention weights α between nodes ij ;

[0026] Where W∈R F′×F Let F' and F' be the trainable linear transformation matrix, respectively, and let a ∈ R. 2F′ h is a trainable attention vector. i and h j These are the feature vectors of nodes i and j in the current input layer, respectively, and N is the feature vector of node i. i h represents the neighborhood of node i. k N represents the neighborhood of node i i Let || denote the feature vector of any neighbor node k in the vector array, and || denote the vector concatenation operation. The LeakyReLU activation function ensures that the gradient of the negative input does not vanish.

[0027] Furthermore, the hierarchical sampling and aggregation layer employs a pattern sampling network with a top-K optimization strategy to extract multi-scale spatial features in stages;

[0028] Its core computational process is as follows: First, attention weights are calculated based on node features to filter key neighbors, and then node representations are generated through weighted aggregation and nonlinear transformation;

[0029] In the second-level aggregation, the model filters the set of key indirect neighbors using the attention weights from the first level:

[0030]

[0031] And calculate the normalized attention weights based on the output features of the upper layer:

[0032]

[0033] Finally, through cross-layer feature integration, a composite representation of the node is obtained:

[0034]

[0035] in, The set of indirect neighbors selected for node v. Here, N(u) represents the key direct neighbors selected by the first-order attention weight, N(u) is the set of direct neighbors of node u, and K is the number of neighbors selected by TopK. This represents the association strength between node u and its neighbor k in the first layer. This indicates the association strength between nodes v and w in the second layer; and These are the second-layer attention weight matrix and the feature transformation weight matrix, respectively, W. final The weight matrix for the final feature fusion; The output features of node v in the first layer. The output features of the indirect neighbor w in the first layer, Let represent the output feature vector of the second-order neighbor node m in the first layer, b be the trainable attention weight vector, σ be the non-linear activation function, LeakyReLU be the modified linear unit with leakage parameters, and CONCAT be the vector concatenation operation.

[0036] Furthermore, the time-series feature modeling layer uses a gated cyclic unit (GRU) as the core time-series processing module, and achieves nonlinear dynamic modeling of typhoon storm surge water level data through its dual-gating mechanism;

[0037] This layer receives the graph embedding vector X from the spatial feature extraction layer. t As a time-series input, using the update gate z t =σ(W z X t +U z h t-1 )+b z Controlling the retention ratio of historical information is achieved by resetting the gate. t =σ(W r X t +U r h t-1 +b r Adjusting candidate states The generation process ultimately results in A gradual update strategy integrates old and new memories;

[0038] Where, h t and h t-1 Represent the hidden states at the current time t and the previous time t-1, respectively; ⊙ is the Hadamard product, tanh is the hyperbolic tangent function, σ is the sigmoid function, and W... * U * and b * The trainable parameters are represented by the input weight matrix, recursive weight matrix, and bias vector of each type of gate, respectively. The updated gate parameters are W. z U z ,b z Reset the door corresponding to W r U r ,b r Candidate states correspond to W h U h b h .

[0039] The temporal feature modeling layer, spatial feature extraction layer, and hierarchical sampling and aggregation layer form a cascaded architecture, which combines the hidden state h from the previous time step. t-1 With the extracted spatial features GraphConv(X) t This is used to predict the current tide level.

[0040] Furthermore, the training and optimization of the dynamic attention graph recurrent network model using the processed data and graph topology also includes dividing the water level dataset into training, validation, and test sets according to independent typhoon events. During training, mean squared error (MSE) is used as the loss function, and its calculation formula is as follows: Where n is the number of test samples, y i These are actual observed values. The model predicts the values; the Adam optimizer is used to update the parameters, and the model parameters are iteratively optimized through backpropagation and gradient descent algorithms.

[0041] Furthermore, it also includes calculating the root mean square error (RMSE) metric using the test set during the model evaluation phase to assess the model's predictive performance. The calculation formula is as follows: Where n is the number of test samples, y i These are actual observed values. These are the model's predicted values.

[0042] As can be seen from the above technical solution, the beneficial effects of the present invention are as follows: The present invention provides a method for predicting water levels at multiple tide gauge stations during typhoons based on dynamic attention graph cyclic networks, comprising the following steps: acquiring water level data and meteorological data from multiple tide gauge stations during a typhoon; performing time-series feature processing on the acquired water level data and meteorological data, and creating a topology structure; defining a dynamic graph attention hierarchical aggregation network model structure, and constructing a dynamic attention graph cyclic network model; training and optimizing the dynamic attention graph cyclic network model using the processed data and graph topology structure; and using the optimized dynamic attention graph cyclic network model to predict the water levels at multiple tide gauge stations along the coast in real time within the next A hours during a typhoon. This method can effectively balance prediction accuracy and computational efficiency, solving the problems of insufficient prediction accuracy and wasted computing power. Attached Figure Description

[0043] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0044] Figure 1 This is a schematic diagram of the process of the present invention;

[0045] Figure 2 This is a schematic diagram illustrating the implementation of the dynamic graph attention network in this invention.

[0046] Figure 3 This is a schematic diagram of the graph aggregation network based on the top K-optimization strategy hierarchical sampling in this invention;

[0047] Figure 4 This is a schematic diagram of the spatiotemporal fusion structure of the gated loop unit and the graph convolution module in this invention. Detailed Implementation

[0048] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0049] Please see Figures 1-4 This invention provides a method for predicting typhoon water levels at multiple tide gauge stations based on a dynamic attention graph recurrent network, comprising the following steps:

[0050] Obtain water level data and meteorological data from multiple tide gauge stations during the typhoon.

[0051] The acquired water level and meteorological data are subjected to time-series feature processing, and a topology structure is created.

[0052] Specifically, water level observation data and meteorological data from all tide gauge stations within the region are selected. A 12-hour sliding window is set, and a time-series feature is created for each time point using the sliding window method. This feature consists of the water level sequence of the past 12 hours and the corresponding meteorological elements. The water level height at the next moment immediately following the window is used as the prediction label for this sample. The value of 'a' ranges from 6 to 18 hours. The meteorological elements include: east-west wind speed component at 10 meters height, north-south wind speed component at 10 meters height, air temperature at 2 meters height, dew point temperature at 2 meters height, surface air pressure, sea level air pressure, sea surface temperature, and precipitation. The water level height at the next moment immediately following the window is used as the prediction label for this sample.

[0053] The creation of the graph topology includes:

[0054] Each tide gauge station in the study area is used as a graph node to obtain the node feature matrix;

[0055] The dynamic coupling strength was measured by calculating the Pearson correlation coefficient of water level fluctuations between any two stations, using the water level sequence observed synchronously at various tide gauge stations during historical typhoons as a sample.

[0056] When the correlation coefficient exceeds a preset threshold, edge weights are established between the corresponding nodes.

[0057] Define the structure of the dynamic graph attention hierarchy aggregation network model and construct the dynamic attention graph recurrent network model;

[0058] Specifically, the construction of the dynamic attention graph recurrent network model further includes:

[0059] The input layer receives the standardized 12-hour water level-meteorological joint features;

[0060] The spatial feature extraction layer uses a graph attention network to calculate the attention weights between nodes; the spatial feature extraction layer uses a dynamic attention mechanism to achieve adaptive coupling between tide gauge station nodes.

[0061] It passed Calculate the attention weights α between nodes ij ;

[0062] Where W∈R F′×F Let F' and F' be the trainable linear transformation matrix, respectively, and let a ∈ R. 2F′ h is a trainable attention vector. i and h j These are the feature vectors of nodes i and j in the current input layer, respectively, and N is the feature vector of node i. i h represents the neighborhood of node i. k N represents the neighborhood of node i i The feature vector of any neighbor node k is used, where || represents the vector concatenation operation. The LeakyReLU activation function ensures that the negative input gradient does not vanish. This layer takes meteorological features and water level time series as inputs, and calculates multiple sets of attention weights in parallel through a multi-head attention mechanism. The node feature matrix is ​​formed by weighted aggregation using subsequent hierarchical sampling and aggregation layers, enabling the model to capture more key information.

[0063] The hierarchical sampling and aggregation layer, based on the top K-optimization strategy of graph sampling and aggregation network, performs weighted aggregation of first-order, second-order and higher-order neighborhood information to extract multi-scale spatial features; the hierarchical sampling and aggregation layer adopts a graph sampling network with a top K-optimization strategy to extract multi-scale spatial features in stages.

[0064] Its core computational process is as follows: First, attention weights are calculated based on node features to filter key neighbors, and then node representations are generated through weighted aggregation and nonlinear transformation;

[0065] In the second-level aggregation, the model filters the set of key indirect neighbors using the attention weights from the first level:

[0066]

[0067] And calculate the normalized attention weights based on the output features of the upper layer:

[0068]

[0069] Finally, through cross-layer feature integration, a composite representation of the node is obtained:

[0070]

[0071] in, The set of indirect neighbors selected for node v. Here, N(u) represents the key direct neighbors selected by the first-order attention weight, N(u) is the set of direct neighbors of node u, and K is the number of neighbors selected by TopK. This represents the association strength between node u and its neighbor k in the first layer. This indicates the association strength between nodes v and w in the second layer; and These are the second-layer attention weight matrix and the feature transformation weight matrix, respectively, W. final The weight matrix for the final feature fusion; The output features of node v in the first layer. The output features of the indirect neighbor w in the first layer, Let represent the output feature vector of the second-order neighbor node m in the first layer, b be the trainable attention weight vector, σ be the non-linear activation function, LeakyReLU be the modified linear unit with leakage parameters, and CONCAT be the vector concatenation operation.

[0072] The time-series feature modeling layer uses a gated cyclic unit (GRU) to capture the nonlinear time-series dependency between water level and meteorological field. The time-series feature modeling layer uses a gated cyclic unit (GRU) as the core time-series processing module, and realizes nonlinear dynamic modeling of typhoon storm surge water level data through its dual-gating mechanism.

[0073] This layer receives the graph embedding vector X from the spatial feature extraction layer. t As a time-series input, using the update gate z t =σ(W z X t +U z h t-1 )+b z Controlling the retention ratio of historical information is achieved by resetting the gate. t =σ(W r X t +U r h t-1 +b r Adjusting candidate states The generation process ultimately results in A gradual update strategy integrates old and new memories;

[0074] Where, ht and h t-1 Represent the hidden states at the current time t and the previous time t-1, respectively; ⊙ is the Hadamard product, tanh is the hyperbolic tangent function, σ is the sigmoid function, and W... * U * and b * The trainable parameters are represented by the input weight matrix, recursive weight matrix, and bias vector of each type of gate, respectively. The updated gate parameters are W. z U z ,b z Reset the door corresponding to W r U r ,b r Candidate states correspond to W h U h b h .

[0075] The temporal feature modeling layer, spatial feature extraction layer, and hierarchical sampling and aggregation layer form a cascaded architecture, which combines the hidden state h from the previous time step. t-1 With the extracted spatial features GraphConv(X) t To predict the current tide level

[0076] The output layer maps the fused spatiotemporal features to the water level height of multiple tide gauge stations for the next hour through a linear layer.

[0077] The processed data and graph topology are used to train and optimize the dynamic attention graph recurrent network model. Specifically, the water level dataset is divided into training, validation, and test sets according to independent typhoon events. Mean squared error (MSE) is used as the loss function during training, and its calculation formula is as follows: Where n is the number of test samples, y i These are actual observed values. The model predicts the values; the Adam optimizer is used to update the parameters, iteratively optimizing the model parameters through backpropagation and gradient descent algorithms. The training process includes forward propagation to calculate the predicted values, calculating the error based on the MSE loss function, and backpropagation to update the weights, and uses a validation set to monitor the training process to avoid overfitting. In the model evaluation phase, the root mean square error (RMSE) metric is calculated using the test set to evaluate the model's predictive performance; its calculation formula is: Where n is the number of test samples, y i These are actual observed values. These are the model's predicted values.

[0078] The optimized dynamic attention graph recurrent network model is used to predict the water levels of multiple tide gauge stations along the coast in real time within the next hour during the typhoon; where the value of A ranges from 0.5 to 3.

[0079] In this invention Figure 2 This demonstrates how to calculate the attention weights between nodes based on real-time meteorological data and dynamically adjust the spatial coupling strength between tide gauge stations. Figure 3 This is used to demonstrate how to achieve multi-scale spatial feature extraction by selecting high-order neighboring nodes; Figure 4 This demonstrates how to jointly model dynamic spatial features with temporal dependencies.

[0080] This invention deeply integrates real-time meteorological elements with the spatial network of coastal tide gauge stations through a dynamic attention graph recurrent network framework. Significant advantages were achieved in independent testing using historical data from eight tide gauge stations in the Tokai region of Japan: Regarding prediction error, the RMSE of this invention was reduced to 0.0474m, and the MAE to only 0.0356m, representing reductions of 55% and 53% respectively compared to the traditional single-station prediction model LSTM, and reductions of 38% and 39% respectively compared to the static graph model GCN-GRU. 2 The accuracy was improved to 0.9872, approaching the observed value. In terms of computational efficiency, the unified graph structure parallel training reduced the time to traverse 100 data points in a single round to 79 seconds, saving 57% compared to the single-station prediction model LSTM and 7% compared to the static graph model GCN-GRU. The system can perform synchronous, high-precision, and timely water level predictions for multiple tide gauge stations along the coast, providing a scientific basis for real-time storm surge early warning and emergency decision-making in coastal cities. This invention uses "real-time data-driven dynamic graph attention + hierarchical neighborhood sampling" as its core mechanism to jointly model the coastal tide gauge station network, achieving rapid synchronous prediction of water levels at multiple stations during typhoons. Specifically, in the water level prediction process, neighbor station information is adaptively aggregated through "hierarchical sampling based on the top K-optimization strategy of attention weights," and then the obtained multi-scale spatial features are fused with a gated temporal network to output the water levels of multiple stations at future times.

[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for predicting water levels at multiple tide gauge stations during typhoons based on a dynamic attention graph recurrent network, characterized in that, Includes the following steps: Acquire water level data and meteorological data from multiple tide gauge stations during the typhoon; The acquired water level and meteorological data are subjected to time-series feature processing, and a topology structure is created. Define the structure of the dynamic graph attention hierarchy aggregation network model and construct the dynamic attention graph recurrent network model; The dynamic attention graph recurrent network model was trained and optimized using processed water level data, meteorological data, and graph topology. The optimized dynamic attention graph recurrent network model is used to predict the water levels of multiple tide gauge stations along the coast in real time within the next A hours during a typhoon.

2. The method for predicting typhoon water levels at multiple tide gauge stations based on dynamic attention graph recurrent networks according to claim 1, characterized in that, The process of performing time-series feature processing on the acquired water level and meteorological data includes selecting water level observation data and meteorological data from all tide gauge stations in the region, setting an a-hour sliding window, and using the sliding window method to create time-series features for each time point. The time-series features consist of the water level sequence of the past a hours and the corresponding meteorological elements, and the water level height at the next moment immediately following the window is used as the prediction label for the water level observation data and meteorological data from all tide gauge stations in the selected region.

3. The method for predicting typhoon water levels at multiple tide gauge stations based on a dynamic attention graph cyclic network according to claim 2, characterized in that, The meteorological elements include: east-west wind speed component at 10 meters altitude, north-south wind speed component at 10 meters altitude, air temperature at 2 meters altitude, dew point temperature at 2 meters altitude, surface air pressure, sea level air pressure, sea surface temperature, and precipitation.

4. The method for predicting typhoon water levels at multiple tide gauge stations based on dynamic attention graph recurrent networks according to claim 1, characterized in that, Also includes: Each tide gauge station in the study area is used as a graph node to obtain the node feature matrix; The dynamic coupling strength was measured by calculating the Pearson correlation coefficient of water level fluctuations between any two stations, using the water level sequence observed synchronously at various tide gauge stations during historical typhoons as a sample. When the Pearson correlation coefficient exceeds a preset threshold, edge weights are established between the corresponding nodes.

5. The method for predicting typhoon water levels at multiple tide gauge stations based on a dynamic attention graph cyclic network according to claim 1, characterized in that, The construction of the dynamic attention graph recurrent network model also includes: The input layer receives the standardized 12-hour water level-meteorological joint features; The spatial feature extraction layer uses a graph attention network to calculate the attention weights between nodes. Hierarchical sampling and aggregation layer: Based on the top K selection strategy of graph sampling and aggregation network, weighted aggregation of first-order, second-order and higher-order neighborhood information is performed to extract multi-scale spatial features. The temporal feature modeling layer uses gated recurrent units to capture the nonlinear temporal dependence between water level and meteorological field; The output layer maps the fused spatiotemporal features to the water level height of multiple tide gauge stations in the next A hours through a linear layer.

6. The method for predicting typhoon water levels at multiple tide gauge stations based on a dynamic attention graph cyclic network according to claim 5, characterized in that, The spatial feature extraction layer employs a dynamic attention mechanism to achieve adaptive coupling between tide gauge station nodes. It passed Calculate the attention weights α between nodes ij ; Where W∈R F’×F Let F' and F' be the trainable linear transformation matrix, respectively, and let a ∈ R. 2F’ h is a trainable attention vector. i h is the feature vector of node i in the current input layer. j N is the feature vector of node j in the current input layer. i h represents the neighborhood of node i. k N represents the neighborhood of node i i Given the feature vector of any neighbor node k, ‖ represents the vector concatenation operation, and the LeakyReLU activation function ensures that the gradient of the negative input does not vanish.

7. The method for predicting typhoon water levels at multiple tide gauge stations based on a dynamic attention graph recurrent network according to claim 6, characterized in that, The hierarchical sampling and aggregation layers employ a top-K optimization strategy for pattern sampling networks to extract multi-scale spatial features in stages. Its core computational process is as follows: First, attention weights are calculated based on node features to filter key neighbors, and then node representations are generated through weighted aggregation and nonlinear transformation; In the second-level aggregation, the model filters the set of key indirect neighbors using the attention weights from the first level: And calculate the normalized attention weights based on the output features of the upper layer: Finally, through cross-layer feature integration, a composite representation of the node is obtained: in, The set of indirect neighbors selected for node v. Here, N(u) represents the key direct neighbors selected by the first-order attention weight, N(u) is the set of direct neighbors of node u, and K is the number of neighbors selected by TopK. This represents the association strength between node u and its neighbor k in the first layer. This indicates the association strength between nodes v and w in the second layer; and These are the second-layer attention weight matrix and the feature transformation weight matrix, respectively, W. final The weight matrix for the final feature fusion; The output features of node v in the first layer. The output features of the indirect neighbor w in the first layer, Let represent the output feature vector of the second-order neighbor node m in the first layer, b be the trainable attention weight vector, σ be the non-linear activation function, LeakyReLU be the modified linear unit with leakage parameters, and CONCAT be the vector concatenation operation.

8. The method for predicting typhoon water levels at multiple tide gauge stations based on a dynamic attention graph recurrent network according to claim 7, characterized in that, The temporal feature modeling layer uses a gated loop unit as the core temporal processing module; The temporal feature modeling layer receives a graph embedding vector X from the spatial feature extraction layer. t As a time-series input, using the update gate z t =σ(W z X t +U z h t-1 )+b z Controlling the retention ratio of historical information is achieved by resetting the gate. t =σ(W r X t +U r h t-1 +b r Adjusting candidate states The generation process ultimately results in A gradual update strategy integrates old and new memories; Where, h t and h t-1 Represent the hidden states at the current time t and the previous time t-1, respectively; ⊙ is the Hadamard product, tanh is the hyperbolic tangent function, σ is the sigmoid function, and W... * U * and b * The trainable parameters are represented by the input weight matrix, recursive weight matrix, and bias vector of each type of gate, respectively. The updated gate parameters are W. z U z ,b z Reset the door corresponding to W r U r ,b r Candidate states correspond to W h U h b h ; The temporal feature modeling layer, spatial feature extraction layer, and hierarchical sampling and aggregation layer form a cascaded architecture, which combines the hidden state h from the previous time step. t-1 With the extracted spatial features GraphConv(X) t This is used to predict the current tide level.

9. The method for predicting typhoon water levels at multiple tide gauge stations based on a dynamic attention graph cyclic network according to claim 1, characterized in that, The method of using the processed data and graph topology to train and optimize the dynamic attention graph recurrent network model also includes dividing the water level dataset into training, validation, and test sets according to independent typhoon events. During training, mean squared error is used as the loss function, and its calculation formula is as follows: Where n is the number of test samples, y i These are actual observed values. The model predicts the values; the Adam optimizer is used to update the parameters, and the model parameters are iteratively optimized through backpropagation and gradient descent algorithms.

10. The method for predicting typhoon water levels at multiple tide gauge stations based on a dynamic attention graph cyclic network according to claim 9, characterized in that, This also includes the calculation of the root mean square error (RMSE) metric using the test set during the model evaluation phase, which is used to assess the model's predictive performance. The calculation formula is as follows: Where n is the number of test samples, y i These are actual observed values. These are the model's predicted values.

Images